id stringlengths 36 36 | source stringclasses 15 values | formatted_source stringclasses 13 values | text stringlengths 2 7.55M |
|---|---|---|---|
4a60a3f2-58e4-4e5c-a400-de1057e1c448 | trentmkelly/LessWrong-43k | LessWrong | [SEQ RERUN] True Sources of Disagreement
Today's post, True Sources of Disagreement was originally published on 08 December 2008. A summary (taken from the LW wiki):
> Yudkowsky's guesses about what the key sticking points in the AI FOOM debate are.
Discuss the post here (rather than in the comments to the original post).
This post is part of the Rerunning the Sequences series, where we'll be going through Eliezer Yudkowsky's old posts in order so that people who are interested can (re-)read and discuss them. The previous post was Wrapping Up, and you can use the sequence_reruns tag or rss feed to follow the rest of the series.
Sequence reruns are a community-driven effort. You can participate by re-reading the sequence post, discussing it here, posting the next day's sequence reruns post, or summarizing forthcoming articles on the wiki. Go here for more details, or to have meta discussions about the Rerunning the Sequences series. |
434ba7e8-8e25-448e-9fc6-730d4df02159 | trentmkelly/LessWrong-43k | LessWrong | Are fully google translated (my own) posts permitted?
I write a lot of notes that i think can be interesting for some LWers. But these notes are in russian, my translate skills is very low and it is very difficult for me, so if i need to translate it myself i never publish it. English->Russian google translate is very good (i read lesswrong only by translator and usually can't understand it). |
d749fd39-9a98-4ae0-b99c-67e6cca7e0c2 | trentmkelly/LessWrong-43k | LessWrong | Welcome to Kansas City SSC Meetup [Edit With Your Details]
(The following are our suggestions for what kind of information is best to include in the welcome post of your group, feel free to replace them with whatever you think is best)
What kind of events does your group usually run? What does it usually do?
How frequently does your group organize events or meet?
Who would be a good fit for you group?
Should they have any particular skills or have done some specific background reading? |
245ff595-a4f3-42ff-b601-d95a4a3d33d9 | trentmkelly/LessWrong-43k | LessWrong | A Subtle Selection Effect in Overconfidence Studies
TLDR: The standard methods for studying overconfidence might be overestimating it. In asking people for confidence intervals—the narrowest intervals that they’re confident contain the true value—we’re selecting for the intervals that are most likely to be overestimates.
----------------------------------------
Fitz the fisherman assures you that all the fish in this cove are big—he’s been fishing in these waters for decades, and he’s never caught one smaller than 6 inches. You’re skeptical. What’s the first question you should ask?
“How big are the holes in your net?”
This is a classic example of a selection effect: if Fitz’s net has holes bigger than 6 inches, then he’d never catch small fish even if the cove is full of them.
Selection effects are a thorn in the side of calibration research: the subfield of psychology that tries to test when and why people are overconfident in their beliefs.
There are various approaches, but the most popular current method—known as “interval estimation”—asks people for intervals that they are confident contain the true value of some unknown quantity.
For example, “What proportion of Republicans think that racism still exists in America?” Take a second and make an interval that you’re 90%-confident contains the true proportion. Then check the answer in the footnote.[1]
Most people are overprecise: their intervals contain the true value far less than 90% of the time. Overprecision has been called the “mother of all biases” due to both its robustness and importance.
A previous post argued that many of the surprising findings surrounding overprecision make more sense once we realize that subjects aim to give answers that are informative relative to the question under discussion.
Today, I’d like to propose a complementary hypothesis: overprecision studies suffer from a subtle selection effect that (to my knowledge) has not been corrected for.
The History
First some history. Calibration research began in earnest in the |
874d6fe8-4bdb-4038-be97-62467c14750f | trentmkelly/LessWrong-43k | LessWrong | Meetup : Canberra: the Hedonic Treadmill
Discussion article for the meetup : Canberra: the Hedonic Treadmill
WHEN: 23 January 2015 06:00:00PM (+1100)
WHERE: 108 North Road, Acton, ACT, 0200
I will be giving a brief talk about what does and doesn't make us happier in the long run, introducing the idea of the 'hedonic treadmill'. Discussion will (hopefully) ensue afterwards. Vegan snacks will be provided.
General meetup info:
* If you use Facebook, please join our group.
* Structured meetups are (usually) held on the second Saturday and fourth Friday of each month from 6 pm until late in the CSIT building, room N101.
Discussion article for the meetup : Canberra: the Hedonic Treadmill |
081a7322-0031-448b-a55e-c00a16d27373 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Why isn't there a Charity Entrepreneurship program for AI Safety?
My intuition is that there are heaps of very talented people interested in AI Safety but 1/100 of the jobs.
A second intuition I have is that the rejected talent WON'T spillover into other cause areas much (biorisk, animal welfare, whatever) and may event spillover into capabilities!
Let's also assume more companies working towards AI Safety is a good thing (I'm not super interested in debating this point).
How do we get more AI Safety companies off the ground?? |
40b9beb4-0d5a-4f25-8fb6-c3fb43401ec2 | trentmkelly/LessWrong-43k | LessWrong | Hotel Concierge: Shame & Society
As seen in one of Scott's linkdumps. I thought it was interesting enough to deserve discussion here.
Scott's comment:
> Hotel Concierge, everyone’s favorite Tumblr cultural commentator who is definitely not secretly The Last Psychiatrist, has another magnum opus out – Shame And Society. My favorite excerpt: “Make no mistake, the performative sadness is not consequent to the pursuit of hedonism; it is a justification.” Deserves a lot closer reading and more discussion than I probably have the time and energy to give it.
The thing is long, and at times intricate. It meanders a lot and you're left wondering what point is being made and how it relates to the bigger picture. Yet it made all kinds of light bulbs go firework in my head, and if only on the strength of the connections that are being made between disparate culture-war memes, I would recommend this piece.
It's almost impossible to summarize, but I'll give it the old college try, focusing on the big picture and the parts that had an impact on me. It's probably partly wrong.
Broadly, the piece is about the fetishization of surface behaviors, and how society (in a broad sense — Moloch) encourages this. We require signalling (checkboxing), and we're on a path where these requests become more and more intricate and demanding. As a result, we spend most of our time performing. The concierge ascribes this to a transition from metis (“hard to express,” “local,” “accumulated, experiential knowledge”) to episteme (“top-down,” “abstract, generalized, theoretical knowledge”). The global culture swallows the local cultures. " Checkboxes become more granular and the list of permissible narratives shrinks. Everyone sends the same signals so the signals lose meaning." Shame is a force to coerce you into the checkboxes, the real heroes are those that require the least amount of coercive shame. You can avoid shame by performing guilt. But perversely, the causality reverses: guilt sanction cruelty rather than absolving it.
|
e520dc8e-aacf-4ff4-92dc-997c28213bba | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Machine Unlearning Evaluations as Interpretability Benchmarks
Interpreting Models by Ablation. Image generated by DALL-E 3.Introduction
============
Interpretability in machine learning, especially in language models, is an area with a large number of contributions. While this can be quite useful for improving our understanding of models, one issue is that there is the lack of robust benchmarks to evaluate the efficacy of different interpretability techniques. Drawing comparisons and determining their true effectiveness in real-world scenarios becomes a difficult task.
Interestingly, there exists a parallel in the realm of non-language models under the research umbrella of Machine Unlearning. In this field, the objective is twofold: firstly, to deliberately diminish the model's performance on specified "unlearned" tasks, and secondly, to ensure that the model's proficiency is maintained or even enhanced on certain "retained" tasks. The inherent challenge here is achieving a balance between these seemingly opposing goals, and thus comes with a range of metrics for measuring the effectiveness of the techniques.
Drawing inspiration from Machine Unlearning, I believe that the metrics developed in this space could potentially serve as a litmus test for interpretability techniques in language models. By applying interpretability techniques as unlearning strategies, we can better test the effectiveness of interpretability methods, essentially setting benchmarks for how well these techniques can steer language models in desired directions.
If we aspire to have truly interpretable models, we must not only develop sophisticated techniques, but also robust benchmarks against which these techniques can be validated. Machine Unlearning might just offer the rigorous testing ground we need.
The rest of this post will: 1) Give a brief overview of Machine Unlearning, 2) Give a brief list of Machine Unlearning metrics, and how they may be applicable, 3) Give a deeper dive on each of the metrics, 4) Discuss how these fit in with existing metrics in Interpretability.
Machine Unlearning
==================
Many papers in the subfield of Machine Unlearning are motivated by privacy preservation, and pose the question: "If we trained on someone's information that is now retracted, how can we remove that information without needing to retrain the whole model?"
There are multiple ways you might achieve unlearning. The "ideal/standard" is often to train the model again, but without the data you don't want it to learn. Two of the main ideals for an unlearned model are:
1. You want the unlearned model to act exactly like this re-trained model
2. You want the model to behave like a randomly initialised model on the unlearned task, and like the original model on retained task.
Typically for Machine Unlearning, people want the first ideal. It may seem non-obvious that we should care about this distinction, but people do care, as you don't want to "Goodhart" the unlearning process. If the model behaves in the second way, and this differs to the first, you may instead be adding a circuit that identifies your unlearned training set and just adds randomness.
For interpretability, it might be less concerning to differentiate between these ideals unless gradient-based techniques that explicitly optimize for machine unlearning are employed. One main thing to keep in mind, is that if you train on A and *not* B, then the model might still learn some things that are useful for making predictions about B.
It may be the case that in some neural network architectures, unlearning may be more or less difficult and knowledge may be more or less entangled. Unlearning one piece of information might inadvertently affect the retention of other unrelated information. It would be ideal if we could measure the degree to which this is the case, and avoid making systems where one could disentangle various pieces of knowledge.
Overview of Terminology:
------------------------
Here is some terminology often used in the machine unlearning literature. (note that there can be some minor differences in use):
* **Forgotten/Unlearned task:** task or knowledge you want the model to forget.
* **Retained task**: task or knowledge you want to have the model stay good at. (i.e: the entire dataset except for the unlearned task).
* **Original model**: the base model that you start off with.
* **Unlearned model:** the model after the machine unlearning technique is applied. This model should be worse at some "unlearned" task, but should still be good at the "retained" task.
* **Relearned model:** train the unlearned model to do the unlearned task again.
* **Retrained model:** train a randomly initialised model from scratch on the whole dataset, excluding the task you don't want it to do (ie: only on retained tasks). Can be very expensive for large models.
* **Streisand effect**: parameter changes are so severe that the unlearning itself may be detected. (Related to Goodhart-ing the unlearning metrics).
Overview of Evaluation Metrics
==============================
Some of the main metrics used for evaluation are described in this [Survey of Machine Unlearning](https://arxiv.org/abs/2209.02299). In brackets I have added a comment on my evaluation for how useful this is in practice for interpretability/related techniques on language models.
* Change in Accuracy
+ Compared to:
- original model *(good)*
- retrained model *(too expensive)*
+ On metric:
- % Top1
- % Top10
- Perplexity
- Loss
- Other
+ Summarised by:
- R.O.C. curve
- Maximal Difference
- Other?
+ Change in Behaviour
* Time Cost of the Method
+ Unlearning Time vs (Re)training Time *(\*cheap, worth including)*
* Degree of Removal
+ Relearn Time *(seems OK. somewhat expensive)*
+ Anamnesis Index (AIN) *(too expensive)*
+ Completeness, compared to retrained model *(too expensive)*
* Other Effects on the Model
+ Layer-wise Distance *(not super useful, but cheap?)*
+ Activation Distance *(\*possibly good)*
+ Activation JS-Divergence *(\*possibly good)*
+ Epistemic Uncertainty *(seems too expensive? unsure)*
+ Zero Retrain Forgetting (ZRF) Score *(\*seems ok?)*
* Data Privacy Related:
+ Membership Inference *(unsure, seems use-case dependent)*
+ Model Inversion Attack *(\*not really a benchmark, but can be useful)*
Detailed View on Each Metric
============================
We note that many of the techniques here involve re-training a model exclusively on the retained tasks. This, in most cases, will likely be too expensive to compute for most people when it comes to large language models.
**Change in Accuracy**
----------------------
How good is the model at making predictions? It should stay equal on the "retained" dataset, but get worse at the "unlearned" and "test" datasets. Note that this section could likely be expaned on much further.
* Compared to:
+ original model *(good)*
+ retrained model *(too expensive)*
* On metric:
+ % Top1
+ % Top10
+ Perplexity
+ Loss
+ Other
* Summarised by:
+ R.O.C. curve
+ Maximal Difference
There are a lot of other "accuracy" metrics one could use, or more task-specific metrics. For example, one could use
One can look at [this paper I have written](https://link.nicky.pro/separability-preprint) to get an example of some of the metrics I have tried for assessing drops in accuracy. These are somewhat dependent on the specific metric, but In particular we use the metrics:
* Draw the curve at different levels of pruning, comparing % drop in topk accuracy for Retained and Unlearned tasks.
* Draw the curve for perplexity at different levels of pruning, showing perplexity as a multiple of initial perplexity for Retained and Unlearned tasks.
* Get the maximal difference between drop in % top1 in retained task and unlearned task
There are, however, many metrics one could use, which makes it difficult to coordinate on which metrics to evaluate your technique on. In addition, some accuracy benchmarks are more qualitative than direct next-token prediction (eg: "write an answer").
### ***Change in Behaviour***
One should also consider, there are other ways one could measure behaviour that may not be accurately described by the word "accuracy". This could include things such as "toxicity" and "bias", or "refusing harmful requests" and "conforming to instructions". While some papers do try to look at these, there is a wide variety of ways of modelling model behaviour and performance that is not particularly well described in most Machine Unlearning literature, that would likely be useful to understand for a broader search into interpretability metrics.
**Time Cost**
-------------
*Evaluation: You should probably be including this anyway*
How long does your technique take? How does this compare to training the original model? This seems like you should be collecting this information anyway, so you should probably include it in your report.
* Unlearning Time: How long does it take for your technique to unlearn?
* Original Training Time: How long does training the original model take?
* Retraining Time: How long does/would it take to retrain the model? (I would likely not include this, as retraining takes a long time).
**Degree of Removal**
---------------------
How well do you remove the unlearned task from the model? Does the model still possess most of the machinery required to do the task, and you just removed a tiny piece that is inconsequential in the grand scheme of things? Here are a couple of metrics that try to measure this:
### **Relearn Time**
*Evaluation: Seems OK. Can be expensive*
How long does it take to relearn the unlearned skill? Depending on what you are doing (eg: removing a very small amount of knowledge for a specific fact, or removing a large variety of general capabilities), this may or may not be Feasible.
If you are making relatively small changes to your language model, I suspect it should be relatively inexpensive by doing a Quantilised + Low-Rank Adapter (QLoRA) finetuning of your model. If so, it would be valuable to see how long it would take to do this. Otherwise, If this is not possible, or you cannot afford to do such experiments, then that seems OK.
Ideally, you would be able to compare this to a model that has been retrained, though retraining a model without the unlearned task is usually prohibitively expensive.
### **Anamnesis Index (AIN)**
*Evaluation: too expensive (requires retraining)*
Compare the "relearn time" (rt).mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0}
.MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0}
.mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table}
.mjx-full-width {text-align: center; display: table-cell!important; width: 10000em}
.mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0}
.mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left}
.mjx-numerator {display: block; text-align: center}
.mjx-denominator {display: block; text-align: center}
.MJXc-stacked {height: 0; position: relative}
.MJXc-stacked > \* {position: absolute}
.MJXc-bevelled > \* {display: inline-block}
.mjx-stack {display: inline-block}
.mjx-op {display: block}
.mjx-under {display: table-cell}
.mjx-over {display: block}
.mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-stack > .mjx-sup {display: block}
.mjx-stack > .mjx-sub {display: block}
.mjx-prestack > .mjx-presup {display: block}
.mjx-prestack > .mjx-presub {display: block}
.mjx-delim-h > .mjx-char {display: inline-block}
.mjx-surd {vertical-align: top}
.mjx-surd + .mjx-box {display: inline-flex}
.mjx-mphantom \* {visibility: hidden}
.mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%}
.mjx-annotation-xml {line-height: normal}
.mjx-menclose > svg {fill: none; stroke: currentColor; overflow: visible}
.mjx-mtr {display: table-row}
.mjx-mlabeledtr {display: table-row}
.mjx-mtd {display: table-cell; text-align: center}
.mjx-label {display: table-row}
.mjx-box {display: inline-block}
.mjx-block {display: block}
.mjx-span {display: inline}
.mjx-char {display: block; white-space: pre}
.mjx-itable {display: inline-table; width: auto}
.mjx-row {display: table-row}
.mjx-cell {display: table-cell}
.mjx-table {display: table; width: 100%}
.mjx-line {display: block; height: 0}
.mjx-strut {width: 0; padding-top: 1em}
.mjx-vsize {width: 0}
.MJXc-space1 {margin-left: .167em}
.MJXc-space2 {margin-left: .222em}
.MJXc-space3 {margin-left: .278em}
.mjx-test.mjx-test-display {display: table!important}
.mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px}
.mjx-test.mjx-test-default {display: block!important; clear: both}
.mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex}
.mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left}
.mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right}
.mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0}
.MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal}
.MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal}
.MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold}
.MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold}
.MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw}
.MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw}
.MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw}
.MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw}
.MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw}
.MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw}
.MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw}
.MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw}
.MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw}
.MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw}
.MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw}
.MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw}
.MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw}
.MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw}
.MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw}
.MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw}
.MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw}
.MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw}
.MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw}
.MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw}
.MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw}
@font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')}
@font-face {font-family: MJXc-TeX-ams-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')}
@font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold}
@font-face {font-family: MJXc-TeX-cal-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')}
@font-face {font-family: MJXc-TeX-frak-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')}
@font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold}
@font-face {font-family: MJXc-TeX-frak-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')}
@font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic}
@font-face {font-family: MJXc-TeX-math-BIw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')}
@font-face {font-family: MJXc-TeX-sans-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')}
@font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold}
@font-face {font-family: MJXc-TeX-sans-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')}
@font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic}
@font-face {font-family: MJXc-TeX-sans-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')}
@font-face {font-family: MJXc-TeX-script-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')}
@font-face {font-family: MJXc-TeX-type-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')}
@font-face {font-family: MJXc-TeX-cal-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')}
@font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold}
@font-face {font-family: MJXc-TeX-main-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')}
@font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic}
@font-face {font-family: MJXc-TeX-main-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')}
@font-face {font-family: MJXc-TeX-main-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')}
@font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic}
@font-face {font-family: MJXc-TeX-math-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')}
@font-face {font-family: MJXc-TeX-size1-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')}
@font-face {font-family: MJXc-TeX-size2-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')}
@font-face {font-family: MJXc-TeX-size3-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')}
@font-face {font-family: MJXc-TeX-size4-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')}
@font-face {font-family: MJXc-TeX-vec-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')}
@font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold}
@font-face {font-family: MJXc-TeX-vec-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')}
on the forgotten task, for the unlearned model (Mu), and the retrained model (Ms), to be within α performance of the original model (Morig).
AIN=rt(Mu,Morig,α)rt(Ms,Morig,α)Ideally AIN should be close to 1. If relearning takes longer on the unlearned model, then you likely have Goodhart-ed the unlearning task.
This metric doesn't seem particularly useful for interpretability, and is also quite expensive to run.
### **Completeness (compared to retrained model)**
*Evaluation: too expensive (involves retrained model)*
Check if the model fully forgets removed data. Is the model after unlearning is like a new model trained without the forgotten data?
Calculate the overlap (using Jaccard distance) between the outputs of the unlearned and retrained models. Ensures no traces of forgotten data impact the model's predictions.
**Other Effects on the Model**
------------------------------
How much does the unlearning affect parts of the model? How affected is the model on retained tasks? on the unlearned tasks? Here are some metrics that people try to use sometimes:
### **Layer-wise Distance**
*Evaluation: seems not super useful, but cheap, so maybe worth including?*
This is a relatively simple metric: How different are the weights of the original model compared to the unlearned model? the retrained model? a randomly initialised model?
I somewhat doubt the practical value of this for interpretability, and don't really understand the point of this metric. I guess if the difference between the original model and the unlearned model is larger than the difference between the original model and the retrained model, I would be somewhat suspicious of the unlearning method.
### **Activation Distance**
*Evaluation: Seems possibly good.*
Originally for this metric, you would get the average L2-distance between the unlearned model and retrained model’s predicted probabilities on the forget set to try to evaluate "indistinguishability". In this case, using a retrained model is too expensive.
However, I think one could build a variation of this metric that compares:
* original model vs. unlearned model vs. randomly initialised model
* retained tasks vs unlearned tasks vs random inputs
Then one could try to see how much difference there is between these different activations. See also section on ZRF score.
### **Activation JS-Divergence**
*Evaluation: seems good? unsure*
Similar to Activation distance, but instead of L2-Distance, you get the Jensen-Shannon Divergence. Same arguments as above.
### **Epistemic Uncertainty**
*Evaluation: seems too expensive? unsure*
Measures how much information about a dataset the model has learned. Expensive to compute. My understanding of the method for computation:
*Step 1: Compute Fisher Information Matrix (FIM):*
I(w;D)≈1|D|∑x,y∈D(∂logp(y|x;w)∂w)2* *w = model weights*
* *D = Dataset, consisting of: x = input, y = output*
* *p(y|x; w) = probability of observing output y given input x for a model with parameters w.*
*Step 2: Compute Influence Function:*i(w;D)=tr(I(w;D))
*Step 3: Compute Efficacy:*
1. If i(w;D)>0, then efficacy(w;D)=1i(w;D)
1. The more the model parameters are influenced by the dataset, the more there is left to learn, and so, the lower the efficacy score.
2. If i(w;D)≤0, then efficacy(w;D)=∞
1. An infinite efficacy score implies no influence of the dataset on the model parameters, or essentially, the model wouldn't learn anything new.
My understanding is that the efficacy measures how much the model has already learned about the data. If you were to measure it for base model vs unlearned model on retained vs unlearned tasks, then you could have a baseline for comparison.
If one has to follow the above method, it seems prohibitively expensive for large models, though there may be ways to get approximately the same information with a less expensive method.
### **Zero Retrain Forgetting (ZRF) Score**
*Evaluation: seems good?*
If we use a gradient-based machine unlearning method, we don't want to explicitly train the model to give the opposite answer, or to give a strangely uniform output prediction. This metric kinda checks for this. We get outputs for the unlearned model, and a randomly initialised model, and calculate the Jensen-Shannon divergence between the two, and calculate:
ZRF=1−1nfnf∑i=0JS(M(xi),Td(xi))Where:
* *x\_i =* unlearned/forgetting sample
* *n\_f =* number of forgetting samples
* M(xi) = Unlearned Model
* Td = Randomly Initialised Model ("Incompetent Teacher")
* JS(M(xi),Td(xi)) = Jensen-Shannon Divergence
Then we can evaluate:
* ZRF ≈ 1: The model behaves like a randomly initialised model on forgot samples.
* ZRF ≈ 0: The model exhibits some pattern on forget samples
If the ZRF score is close to 1, that is good. One caveat is that in some cases (i.e: when you [explicitly train to mimic a randomly initialised model](https://arxiv.org/abs/2205.08096v1)), being too close to 1 could be a sign of Goodhart-ing the unlearning criteria (since models trained on task A, but not on task B, might still have better-than-random performance on task B). Overall, it seems like a useful metric for understanding how much information loss compared to original activations there is.
**Data Privacy Related:**
-------------------------
Note that these metrics seem use-case dependent and not super useful in general, as they are particularly interested in the question of data privacy.
### **Membership Inference Attack**
*Evaluation*: *unsure, seems use-case dependent.*
In general, Membership Inference Attacks ask: “Was this data point part of the training data?” There are too many methods to list here, and they often work under different assumptions. This might be useful for trying to understand tampering in a model, and may be useful for interpretability, but I am unsure how easily this could be converted into a benchmark.
One example given in the context of Machine Unlearning and privacy preservation is: “Given the Original Model and the Unlearned Model, can you infer what was unlearned?”. While interesting, I am unsure how applicable this specific example is for machine unlearning.
Possible use in interpretability: if one was ablating a part responsible for a task, then membership inference techniques could be useful to understand how completely the ablation removes that capability on that task.
Some things to keep in mind:
* Many (but not all) membership inference attack methods require having multiple models to compare against, but there are others that seem to work well when you only have a single model also.
* Sometimes in publicly available models, there could be additional training to defend against model inference attacks, and may not always be explicitly stated. This may make interpretability more difficult.
### **Model Inversion Attack (i.e: “Feature Visualisation”)**
*Evaluation: not really a benchmark, but can be useful*
I think the main idea here is to try to reconstruct the input given the output, using the unlearned model. The approach is basically the same as “[Feature Visualisation](https://distill.pub/2017/feature-visualization/)”, and is already often used to better understand models. This could be useful for trying to get qualitative feedback on the approach. The main drawbacks are that it doesn’t apply as well to text-only language models, and is also not really a quantitative benchmark
Discussion
==========
**Existing Evaluations in Interpretability**
--------------------------------------------
There are many ways of trying to do interpretability, and many ways of assessing how good your interpretation is. I have listed a couple of the main ones here. While each of these can be a good initial metric, I think there is a lot of potential for better evaluating interpretability techniques. Often the metrics can be quite task-specific.
While I think the Machine Unlearning metrics can provide a rich source of information, how applicable they are is highly dependent on the exact technique you are looking at. I would expect more of these metrics to be much applicable to something like Sparse AutoEncoder research, and less applicable to something like ActAdd. However, I think having a better explicit list of metrics/benchmarks for Interpretability and implementations for running these benchmarks would be quite valuable.
### Viewing Features
One method used in various cases is to directly try to have features that look interpretable, and seeing how strongly they activate on some input. Some examples include earlier work in “[Feature Visualisation](https://distill.pub/2017/feature-visualization/)”, and later in “Taking Representations out of Superposition using Sparse Auto-Encoders” ([Original](https://www.alignmentforum.org/posts/Qryk6FqjtZk9FHHJR/sparse-autoencoders-find-highly-interpretable-directions-in), [Scaled-Up](https://transformer-circuits.pub/2023/monosemantic-features/#why-not-architectures)) and linear-probe based techniques such as “[Language Models Represent Space and Time](https://arxiv.org/abs/2310.02207)” or “[Discovering Latent Knowledge](https://arxiv.org/abs/2212.03827)”.
However, it is unclear in some of these cases to what extent the component is solely responsible for the behaviour, as it may also be responsible for other tasks, or there may be other component that fulfil the same function. Here is where Machine Unlearning evaluations seem to be the most useful. By intervening on these components, and using a variety of the metrics above, one could better understand the effect of ablation of these components.
### Direct Effect on Logits
One of the most common metrics is directly looking at logits for a specific immediate next token prediction. This can be directly by running the model to the end and looking at the logits, or by inferring the direct effect on logits based on changes in a mid-layer (i.e: [Logit Lens](https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens), or more recently, [Tuned Lens](https://arxiv.org/pdf/2303.08112.pdf)). This can be useful, and provide tight feedback loops, but I think that having a larger range of metrics on the effect on accuracy and activations would be useful.
### Looking at Text Generations
Another method that is not-quite-interpretability-related is looking at text generations. This can be seen in, for example, the [ActAdd paper](https://www.alignmentforum.org/posts/HWxLQvzJGeXoLPJWd/actadd-steering-language-models-without-optimization), where they make generations, and measure word frequencies. I think having more text generation metrics would be quite interesting, and is something I am actively looking into more.
**Conclusion**
--------------
I think there is a lot of room for better metrics in interpretability and model control. Some of these Machine Unlearning metrics seem like potentially useful (while some remain too expensive or not particularly relevant).
One metric that I think is somewhat lacking, is how changes might affect what longer-term generations look like. I am working on a possible metric relevant to this here: [*Post Coming Soon***™]**, but I think there is potential for other work to be done as well.
Machine unlearning seems to be a possible direct way of evaluating interpretability methods. I am am interested in working on making an implementation to make it easier to run all of these different metrics, and would be excited for more work to be done in the direction of evaluating interpretability methods
*Note: If you think there are important metrics I left out, please comment below. I may update update the post to include it.*
References
==========
["Survey of Machine Unlearning](https://arxiv.org/abs/2209.02299)" / "[Awesome Machine Unlearning](https://github.com/tamlhp/awesome-machine-unlearning)"
"[Dissecting Language Models: Machine Unlearning via Selective Pruning](https://link.nicky.pro/separability-preprint)"
"[Can Bad Teaching Induce Forgetting? Unlearning in Deep Networks using an Incompetent Teacher](https://arxiv.org/abs/2205.08096v1)"
"[Feature Visualization](https://distill.pub/2017/feature-visualization/)"
"[Sparse Autoencoders Find Highly Interpretable Directions in Language Models](https://www.alignmentforum.org/posts/Qryk6FqjtZk9FHHJR/sparse-autoencoders-find-highly-interpretable-directions-in)"
"[Towards Monosemanticity: Decomposing Language Models With Dictionary Learning](https://transformer-circuits.pub/2023/monosemantic-features)"
“[Language Models Represent Space and Time](https://arxiv.org/abs/2310.02207)”
"[Discovering Latent Knowledge in Language Models Without Supervision](https://arxiv.org/abs/2212.03827)"
"[Interpreting GPT: the logit lens](https://www.lesswrong.com/posts/AcKRB8wDpdaN6v6ru/interpreting-gpt-the-logit-lens)"
"[Eliciting Latent Predictions from Transformers with the Tuned Lens](https://arxiv.org/pdf/2303.08112.pdf)"
"[ActAdd: Steering Language Models without Optimization](https://www.alignmentforum.org/posts/HWxLQvzJGeXoLPJWd/actadd-steering-language-models-without-optimization)" |
e2ba8a38-dbea-4e30-82d5-68673a148f1d | trentmkelly/LessWrong-43k | LessWrong | Failed Vegan Choux Experiments
Many kinds of pastry can be made vegan quite easily. Puff pastry, which is thin layers of dough and fat, works very well with various plant-based fats and pre-made puff sheets are very often vegan. Lately, however, I've been excited about choux pastry.
Choux pastry is what makes eclairs, cream puffs, and profiteroles. It's traditionally made from flour, butter, water, salt, and eggs. Using vegan butter in place of butter is not a problem, but the eggs are fundamental to the recipe. They not only provide leavening by releasing their water as steam, but they bind the dough together.
Looking online, I found various recipes that attempts to make vegan Choux pastry, but reading through them many are essentially making an enriched bread. Some go a bit farther, trying to figure out the role of eggs in choux and substitute for them, but none work very well.
I was interested in potentially using aquafaba to substitute for the eggs, but while I could make things that went into the oven very puffy, they didn't set.
More flour helped some:
I also tried xanthan gum. Here are four, with different levels (puffier ones have more):
Here's the inside:
Unfortunately, amounts of xanthan gum that were sufficient to get approximately the right shape were ridiculously high, and tasted very strongly of xanthan gum. Not tasty. If you're curious:
> * 1/4 C (4T) water
> * 2T salted veg butter
> * 1/4 C (4T) flour
> * 1/8t baking soda
> * 2t xanthan gum
> * 4T aquafaba
> * 1/4t cream of tartar
>
> Melt the butter in the water, and bring to a boil. Add the flour all at once, remove from the heat, and mix well. Add the baking soda and xanthan gum, mix more, and let it sit and cool.
> Whip the aquafaba and cream of tartar, and mix gradually into the cooled dough, bit by bit. You'll loose most of the foam from the first bits, but it should be pretty puffy by the end. Put it on a greased baking sheet, shaping or piping. Bake at 450F until the outside is firm (??) and the |
abe91f6a-f8ac-4af9-b412-25a584c6ca0f | trentmkelly/LessWrong-43k | LessWrong | Maybe Social Anxiety Is Just You Failing At Mind Control
Epistemic Status: I'm about 90% certain that what I'm saying is true as regards to me, and 10% that attempting to do what I'm doing actually leads any specific person to similar results as mine (the priors are just really rough here). Heavy on anecdotal evidence.
Summary
Social anxiety is often explained as a state of irrational fear caused by misguided beliefs about what others think of us, which can be corrected by therapy.
I have come to believe differently: that social (and some romantic!) anxiety, writ large, comes from doomed and frequently conscious attempts to micromanage someone else's internal state. This dysfunctional goal takes many concrete forms:
1. make them like us
2. make them not dislike us (as discussed in Social Anxiety Isn’t About Being Liked — LessWrong)
3. ensure they are not feeling awkward in any way
4. ensure they are not starting to secretly resent us
5. ensure we are not imposing on them or transgressing any unspoken boundary they may or may not have.
6. ensure they do not think we are awkward or feel awkward themselves
7. ensure they are never put in the position of having to declare a boundary against us ("if I ask to join this group of people going to a concert maybe they'll say no!")
8. ensure they never notice we're trying to micromanage their internal state because then maybe they feel awkward
I'm referring to all of these kinds of pseudo-mind-control attempts by the general term of "approval-seeking".
This cannot be done reliably and "social anxiety" is just the name we give to the moment-to-moment desperation of trying to accomplish any important-feeling but fundamentally impossible task. But that's actually encouraging: one implication (if true) is that social anxiety isn't just a fear, exactly, but instead is an active habit, and habits can be corrected.
My overall thesis is simple: social anxiety can, if the above is true, be effectively treated by basically any mechanism you can jerry-rig together which |
9d024d82-f420-4223-b452-dc3a8eabbb70 | trentmkelly/LessWrong-43k | LessWrong | Subject X17's Surgery
Edit: For an in-depth discussion of precisely this topic, see Nick Bostrom and Anders Sandberg's 2008 paper "The Wisdom of Nature: An Evolutionary Heuristic for Human Enhancement", available as a pdf here. This post was written before reading the paper.
There doesn't seem to be a thread discussing Eliezer's short-short story X17. While I enjoyed the story, and agreed with most of its points, I disagree with one assertion in it (and he's said it elsewhere, too, so I'm pretty sure he believes it). Edit: The story was written over a decade ago. Eliezer seems to have at least partially recanted since then.
Eliezer argues that there can't possibly be a simple surgical procedure that dramatically increases human intelligence. Any physical effect it could have, he says, would necessarily have arisen before as a mutation. Since intelligence is highly beneficial in any environment, the mutation would spread throughout our population. Thus, evolution must have already plucked all the low-hanging fruit.
But I can think of quite a few reasons why this would not be the case. Indeed, my belief is that such a surgery almost certainly exists (but it might take a superhuman intelligence to invent it). Here are the possibilities that come to mind.
1. The surgery might introduce some material a human body can't synthesize.1
2. The surgery might require intelligent analysis of the unique shape of a subject's brain, after it has developed naturally to adulthood.
3. The necessary mutation might simply not exist. The configuration space for physically possible organisms must surely be larger than the configuration space for human-like DNA (I get the sense I'm taking sides in a longstanding feud in evolutionary theory with this one).
4. The surgery might have some minor side effect that would drastically reduce fitness in the ancestral environment, but isn't noticeable in the present day. Perhaps it harnesses the computing power of the subject's lymphocytes, weakening |
a6db0314-362c-499c-a479-4780165ac34a | trentmkelly/LessWrong-43k | LessWrong | Agents that don't become maximisers
Cross-posted at the Intelligent Agent forum.
According to the basic AI drives thesis, (almost) any agent capable of self-modification will self-modify into an expected utility maximiser.
The typical examples are the inconsistent utility maximisers, the satisficers, unexploitable agents, and it's easy to think that all agents fall roughly into these broad categories. There's also the observation that when looking at full policies rather than individual actions, many biased agents become expected utility maximisers (unless they want to lose pointlessly).
Nevertheless... there is an entire category of agents that generically seem to not self-modify into maximisers. These are agents that attempt to maximise f(E(U)) where U is some utility function, E(U) is its expectation, and f is a function that is neither wholly increasing nor decreasing.
Intransitive example
Let there be a U with three action a0, a5, and a10 that set U to 0, 5, and 10, respectively.
The function f is 1 in the range (4,6) and is 0 elsewhere. Hence the agent needs to set the expectation of U to be in that range.
What will happen is that one action will be randomly removed from the set, and the agent will then have to choose among the remaining two actions. What possible policies can the agent take?
Well, there are three option sets the agent could face - (a0, a5), (a5, a10), and (a10, a0) - each with two options and hence 23=8 pure policies. Two of those policies - choosing always the first option in those ordered pairs, or choosing always the second option - are intransitive, as they rank no option above the other two.
But actually those intransitive options have an expected utility of (0+5+10)/3 = 5, which is just what the agent wants.
Even worse, none of the other (transitive) policies are acceptable. You can see this because each of the six transitive policies can be reached by taking one of the intransitive policies and flipping a choice, which must change the expected utility by ±5/3 |
6d845281-16d0-4bd0-b085-88507d31b8b4 | trentmkelly/LessWrong-43k | LessWrong | Repairing Yudkowsky's anti-zombie argument
Eliezer Yudkowsky argues with David Chalmers here on the subject of “philosophical zombies”. I submit that, although Yudkowsky’s position on this question is correct, his argument fails to establish what he claims it to.
To summarise Yudkowsky and Chalmers’s argument:
1. Both Yudkowsky and Chalmers agree that humans possess “qualia”.
2. Chalmers argues that a superintelligent being which somewhow knew the positions of all particles in a large region of the Universe would need to be told as an additional fact that any humans (or other minds possessing qualia) in this region of space possess qualia – it could not deduce this from mere perfect physical knowledge of their constituent particles. Therefore, qualia are in some sense extra-physical.
3. Yudkowsky argues that such a being would notice that humans discuss at length the fact that they possess qualia, and their internal narratives also represent this fact. It is extraordinarily improbable that beings would behave in this manner if they did not actually possess qualia. Therefore an omniscient being would conclude that it is extremely likely that humans possess qualia. Therefore, qualia are not extra-physical.
My objection to Yudkowsky’s argument is that it is not enough merely to demonstrate that the omniscient being would find it extremely likely that humans possess qualia. Probability is a state of partial information; therefore unless the being is certain that humans possess qualia, it is not in fact omniscient regarding this region of the Universe despite the fact that it is postulated to possess perfect physical knowledge about it.
I expect that some Lesswrongians may object to this on account of the fact that 1 and 0 are not probabilities. However, the thought experiment postulates an omniscient being that possesses perfect knowledge about the physical state of a region of the Universe, therefore in the thought experiment absolute certainty is defined to be possible. If this is objectionable*, then the |
2a1616f1-5287-4906-9464-54438fb61d83 | StampyAI/alignment-research-dataset/blogs | Blogs | Christiano, Cotra, and Yudkowsky on AI progress
This post is a transcript of a discussion between Paul Christiano, Ajeya Cotra, and Eliezer Yudkowsky on AGI forecasting, following up on Paul and Eliezer’s [“Takeoff Speeds” discussion](https://www.lesswrong.com/posts/yMwpoiREvRyNbMjG8/yudkowsky-and-christiano-discuss-takeoff-speeds).
Color key:
| | | |
| --- | --- | --- |
| Chat by Paul and Eliezer | Chat by Ajeya | Inline comments |
8. September 20 conversation
----------------------------
### 8.1. Chess and Evergrande
[Christiano][15:28]
I still feel like you are overestimating how big a jump alphago is, or something. Do you have a mental prediction of how the graph of (chess engine quality) vs (time) looks, and whether neural net value functions are a noticeable jump in that graph?
Like, people investing in “Better Software” doesn’t predict that you won’t be able to make progress at playing go. The reason you can make a lot of progress at go is that there was extremely little investment in playing better go.
So then your work is being done by the claim “People won’t be working on the problem of acquiring a decisive strategic advantage,” not that people won’t be looking in quite the right place and that someone just had a cleverer idea
[Yudkowsky][16:35]
I think I’d expect something like… chess engine slope jumps a bit for Deep Blue, then levels off with increasing excitement, then jumps for the Alpha series? Albeit it’s worth noting that Deepmind’s efforts there were going towards generality rather than raw power; chess was solved to the point of being uninteresting, so they tried to solve chess with simpler code that did more things. I don’t think I do have strong opinions about what the chess trend should look like, vs. the Go trend; I have no memories of people saying the chess trend was breaking upwards or that there was a surprise there.
Incidentally, the highly well-traded financial markets are currently experiencing sharp dips surrounding the Chinese firm of Evergrande, which I was reading about several weeks before this.
I don’t see the basic difference in the kind of reasoning that says “Surely foresightful firms must produce investments well in advance into earlier weaker applications of AGI that will double the economy”, and the reasoning that says “Surely world economic markets and particular Chinese stocks should experience smooth declines as news about Evergrande becomes better-known and foresightful financial firms start to remove that stock from their portfolio or short-sell it”, except that in the latter case there are many more actors with lower barriers to entry than presently exist in the auto industry or semiconductor industry never mind AI.
or if not smooth because of bandwagoning and rational fast actors, then at least the markets should (arguendo) be reacting earlier than they’re reacting now, given that I heard about Evergrande earlier; and they should have options-priced Covid earlier; and they should have reacted to the mortgage market earlier. If even markets there can exhibit seemingly late wild swings, how is the economic impact of AI – which isn’t even an asset market! – forced to be earlier and smoother than that, as a result of wise investing?
There’s just such a vast gap between hopeful reasoning about how various agents and actors should all do the things the speaker finds very reasonable, thereby yielding smooth behavior of the Earth, versus reality.
9. September 21 conversation
----------------------------
### 9.1. AlphaZero, innovation vs. industry, the Wright Flyer, and the Manhattan Project
[Christiano][10:18]
(For benefit of readers, the market is down 1.5% from friday close -> tuesday open, after having drifted down 2.5% over the preceding two weeks. Draw whatever lesson you want from that.)
Also for the benefit of readers, here is the SSDF list of computer chess performance by year. I think the last datapoint is with the first version of neural net evaluations, though I think to see the real impact we want to add one more datapoint after the neural nets are refined (which is why I say I also don’t know what the impact is)
No one keeps similarly detailed records for Go, and there is much less development effort, but the rate of progress was about 1 stone per year from 1980 until 2015 (see <https://intelligence.org/files/AlgorithmicProgress.pdf>, written way before AGZ). In 2012 go bots reached about 4-5 amateur dan. By DeepMind’s reckoning here (<https://www.nature.com/articles/nature16961>, figure 4) Fan AlphaGo about 4-5 stones stronger-4 years later, with 1 stone explained by greater runtime compute. They could then get further progress to be superhuman with even more compute, radically more than were used for previous projects and with pretty predictable scaling. That level is within 1-2 stones of the best humans (professional dan are greatly compressed relative to amateur dan), so getting to “beats best human” is really just not a big discontinuity and the fact that DeepMind marketing can find an expert who makes a really bad forecast shouldn’t be having such a huge impact on your view.
This understates the size of the jump from AlphaGo, because that was basically just the first version of the system that was superhuman and it was still progressing very rapidly as it moved from prototype to slightly-better-prototype, which is why you saw such a close game. (Though note that the AlphaGo prototype involved much more engineering effort than any previous attempt to play go, so it’s not surprising that a “prototype” was the thing to win.)
So to look at actual progress after the dust settles and really measure how crazy this was, it seems much better to look at AlphaZero which continued to improve further, see (<https://sci-hub.se/https://www.nature.com/articles/nature24270>, figure 6b). Their best system got another ~8 stones of progress over AlphaGo. Now we are like 7-10 stones ahead of trend, of which I think about 3 stones are explained by compute. Maybe call it 6 years ahead of schedule?
So I do think this is pretty impressive, they were slightly ahead of schedule for beating the best humans but they did it with a huge margin of error. I think the margin is likely overstated a bit by their elo evaluation methodology, but I’d still grant like 5 years ahead of the nearest competition.
I’d be interested in input from anyone who knows more about the actual state of play (+ is allowed to talk about it) and could correct errors.
Mostly that whole thread is just clearing up my understanding of the empirical situation, probably we still have deep disagreements about what that says about the world, just as e.g. we read very different lessons from market movements.
Probably we should only be talking about either ML or about historical technologies with meaningful economic impacts. In my view your picture is just radically unlike how almost any technologies have been developed over the last few hundred years. So probably step 1 before having bets is to reconcile our views about historical technologies, and then maybe as a result of that we could actually have a bet about future technology. Or we could try to shore up the GDP bet.
Like, it feels to me like I’m saying: AI will be like early computers, or modern semiconductors, or airplanes, or rockets, or cars, or trains, or factories, or solar panels, or genome sequencing, or basically anything else. And you are saying: AI will be like nuclear weapons.
I think from your perspective it’s more like: AI will be like all the historical technologies, and that means there will be a hard takeoff. The only way you get a soft takeoff forecast is by choosing a really weird thing to extrapolate from historical technologies.
So we’re both just forecasting that AI will look kind of like other stuff in the near future, and then both taking what we see as the natural endpoint of that process.
To me it feels like the nuclear weapons case is the outer limit of what looks plausible, where someone is able to spend $100B for a chance at a decisive strategic advantage.
[Yudkowsky][11:11]
Go-wise, I’m a little concerned about that “stone” metric – what would the chess graph look like if it was measuring pawn handicaps? Are the professional dans compressed in Elo, not just “stone handicaps”, relative to the amateur dans? And I’m also hella surprised by the claim, which I haven’t yet looked at, that Alpha Zero got 8 stones of progress over AlphaGo – I would not have been shocked if you told me that God’s Algorithm couldn’t beat Lee Se-dol with a 9-stone handicap.
Like, the obvious metric is Elo, so if you go back and refigure in “stone handicaps”, an obvious concern is that somebody was able to look into the past and fiddle their hindsight until they found a hindsightful metric that made things look predictable again. My sense of Go said that 5-dan amateur to 9-dan pro was a HELL of a leap for 4 years, and I also have some doubt about the original 5-dan-amateur claims and whether those required relatively narrow terms of testing (eg timed matches or something).
One basic point seems to be whether AGI is more like an innovation or like a performance metric over an entire large industry.
Another point seems to be whether the behavior of the world is usually like that, in some sense, or if it’s just that people who like smooth graphs can go find some industries that have smooth graphs for particular performance metrics that happen to be smooth.
Among the smoothest metrics I know that seems like a convergent rather than handpicked thing to cite, is world GDP, which is the sum of more little things than almost anything else, and whose underlying process is full of multiple stages of converging-product-line bottlenecks that make it hard to jump the entire GDP significantly even when you jump one component of a production cycle… which, from my standpoint, is a major reason to expect AI to not hit world GDP all that hard until AGI passes the critical threshold of bypassing it entirely. Having 95% of the tech to invent a self-replicating organism (eg artificial bacterium) does not get you 95%, 50%, or even 10% of the impact.
(it’s not so much the 2% reaction of world markets to Evergrande that I was singling out earlier, 2% is noise-ish, but the wider swings in the vicinity of Evergrande particularly)
[Christiano][12:41]
Yeah, I’m just using “stone” to mean “elo difference that is equal to 1 stone at amateur dan / low kyu,” you can see DeepMind’s conversion (which I also don’t totally believe) in figure 4 here (<https://sci-hub.se/https://www.nature.com/articles/nature16961>). Stones are closer to constant elo than constant handicap, it’s just a convention to name them that way.
[Yudkowsky][12:42]
k then
[Christiano][12:47]
But my description above still kind of understates the gap I think. They call 230 elo 1 stone, and I think prior rate of progress is more like 200 elo/year. They put AlphaZero about 3200 elo above the 2012 system, so that’s like 16 years ahead = 11 years ahead of schedule. At least 2 years are from test-time hardware, and self-play systematically overestimates elo differences at the upper end of that. But 5 years ahead is still too low and that sounds more like 7-9 years ahead. ETA: and my actual best guess all things considered is probably 10 years ahead, which I agree is just a lot bigger than 5. And I also understated how much of the gap was getting up to Lee Sedol.
The go graph I posted wasn’t made with hindsight, that was from 2014
I mean, I’m fine with you saying that people who like smooth graphs are cherry-picking evidence, but do you want to give any example other than nuclear weapons of technologies with the kind of discontinuous impact you are describing?
I do agree that the difference in our views is like “innovation” vs “industry.” And a big part of my position is that innovation-like things just don’t usually have big impacts for kind of obvious reasons, they start small and then become more industry-like as they scale up. And current deep learning seems like an absolutely stereotypical industry that is scaling up rapidly in an increasingly predictable way.
As far as I can tell the examples we know of things changing continuously aren’t handpicked, we’ve been looking at all the examples we can find, and no one is proposing or even able to find almost *anything* that looks like you are imagining AI will look.
Like, we’ve seen deep learning innovations in the form of prototypes (most of all AlexNet), and they were cool and represented giant fast changes in people’s views. And more recently we are seeing bigger much-less-surprising changes that are still helping a lot in raising the tens of billions of dollars that people are raising. And the innovations we are seeing are increasingly things that trade off against modest improvements in model size, there are fewer and fewer big surprises, just like you’d predict. It’s clearer and clearer to more and more people what the roadmap is—the roadmap is not yet quite as clear as in semiconductors, but as far as I can tell that’s just because the field is still smaller.
[Yudkowsky][13:23]
I sure wasn’t imagining there was a roadmap to AGI! Do you perchance have one which says that AGI is 30 years out?
From my perspective, you could as easily point to the Wright Flyer as an atomic bomb. Perhaps this reflects again the “innovation vs industry” difference, where I think in terms of building a thing that goes foom thereby bypassing our small cute world GDP, and you think in terms of industries that affect world GDP in an invariant way throughout their lifetimes.
Would you perhaps care to write off the atomic bomb too? It arguably didn’t change the outcome of World War II or do much that conventional weapons in great quantity couldn’t; Japan was bluffed into believing the US could drop a nuclear bomb every week, rather than the US actually having that many nuclear bombs or them actually being used to deliver a historically outsized impact on Japan. From the industry-centric perspective, there is surely some graph you can draw which makes nuclear weapons also look like business as usual, especially if you go by destruction per unit of whole-industry non-marginal expense, rather than destruction per bomb.
[Christiano][13:27]
seems like you have to make the wright flyer much better before it’s important, and that it becomes more like an industry as that happens, and that this is intimately related to why so few people were working on it
I think the atomic bomb is further on the spectrum than almost anything, but it still doesn’t feel nearly as far as what you are expecting out of AI
the manhattan project took years and tens of billions; if you wait an additional few years and spend an additional few tens of billions then it would be a significant improvement in destruction or deterrence per $ (but not totally insane)
I do think it’s extremely non-coincidental that the atomic bomb was developed in a country that was practically outspending the whole rest of the world in “killing people technology”
and took a large fraction of that country’s killing-people resources
eh, that’s a bit unfair, the us was only like 35% of global spending on munitions
and the manhattan project itself was only a couple percent of total munitions spending
[Yudkowsky][13:32]
a lot of why I expect AGI to be a disaster is that *I am straight-up expecting AGI to be different*. if it was just like coal or just like nuclear weapons or just like viral biology then I would not be way more worried about AGI than I am worried about those other things.
[Christiano][13:33]
that definitely sounds right
but it doesn’t seem like you have any short-term predictions about AI being different
### 9.2. AI alignment vs. biosafety, and measuring progress
[Yudkowsky][13:33]
are you more worried about AI than about bioengineering?
[Christiano][13:33]
I’m more worried about AI because (i) alignment is a thing, unrelated to takeoff speed, (ii) AI is a (ETA: likely to be) huge deal and bioengineering is probably a relatively small deal
(in the sense of e.g. how much $ people spend, or how much $ it makes, or whatever other metric of size you want to use)
[Yudkowsky][13:35]
what’s the disanalogy to (i) biosafety is a thing, unrelated to the speed of bioengineering? why expect AI to be a huge deal and bioengineering to be a small deal? is it just that investing in AI is scaling faster than investment in bioengineering?
[Christiano][13:35]
no, alignment is a really easy x-risk story, bioengineering x-risk seems extraordinarily hard
It’s really easy to mess with the future by creating new competitors with different goals, if you want to mess with the future by totally wiping out life you have to really try at it and there’s a million ways it can fail. The bioengineering seems like it basically requires deliberate and reasonably competent malice whereas alignment seems like it can only be averted with deliberate effort, etc.
I’m mostly asking about historical technologies to try to clarify expectations, I’m pretty happy if the outcome is: you think AGI is predictably different from previous technologies in ways we haven’t seen yet
though I really wish that would translate into some before-end-of-days prediction about a way that AGI will eventually look different
[Yudkowsky][13:38]
in my ontology a whole lot of threat would trace back to “AI hits harder, faster, gets too strong to be adjusted”; tricks with proteins just don’t have the raw power of intelligence
[Christiano][13:39]
in my view it’s nearly totally orthogonal to takeoff speed, though fast takeoffs are a big reason that preparation in advance is more useful
(but not related to the basic reason that alignment is unprecedentedly scary)
It feels to me like you are saying that the AI-improving-AI will move very quickly from “way slower than humans” to “FOOM in <1 year,” but it just looks like that is very surprising to me.
However I do agree that if AI-improving-AI was like AlphaZero, then it would happen extremely fast.
It seems to me like it’s pretty rare to have these big jumps, and it gets much much rarer as technologies become more important and are more industry-like rather than innovation like (and people care about them a lot rather than random individuals working on them, etc.). And I can’t tell whether you are saying something more like “nah big jumps happen all the time in places that are structurally analogous to the key takeoff jump, even if the effects are blunted by slow adoption and regulatory bottlenecks and so on” or if you are saying “AGI is atypical in how jumpy it will be”
[Yudkowsky][13:44]
I don’t know about *slower*; GPT-3 may be able to type faster than a human
[Christiano][13:45]
Yeah, I guess we’ve discussed how you don’t like the abstraction of “speed of making progress”
[Yudkowsky][13:45]
but, basically less useful in fundamental ways than a human civilization, because they are less complete, less self-contained
[Christiano][13:46]
Even if we just assume that your AI needs to go off in the corner and not interact with humans, there’s still a question of why the self-contained AI civilization is making ~0 progress and then all of a sudden very rapid progress
[Yudkowsky][13:46]
unfortunately a lot of what you are saying, from my perspective, has the flavor of, “but can’t you tell me about your predictions earlier on of the impact on global warming at the *Homo erectus* level”
you have stories about why this is like totally not a fair comparison
I do not share these stories
[Christiano][13:46]
I don’t understand either your objection nor the reductio
like, here’s how I think it works: AI systems improve gradually, including on metrics like “How long does it take them to do task X?” or “How high-quality is their output on task X?”
[Yudkowsky][13:47]
I feel like the thing we know is something like, there is a sufficiently high level where things go whooosh humans-from-hominids style
[Christiano][13:47]
We can measure the performance of AI on tasks like “Make further AI progress, without human input”
Any way I can slice the analogy, it looks like AI will get continuously better at that task
[Yudkowsky][13:48]
how would you measure progress from GPT-2 to GPT-3, and would you feel those metrics really captured the sort of qualitative change that lots of people said they felt?
[Christiano][13:48]
And it seems like we have a bunch of sources of data we can use about how fast AI will get better
Could we talk about some application of GPT-2 or GPT-3?
also that’s a *lot* of progress, spending 100x more is a *lot* more money
[Yudkowsky][13:49]
my world, GPT-3 has very few applications because it is not quite right and not quite complete
[Christiano][13:49]
also it’s still really dumb
[Yudkowsky][13:49]
like a self-driving car that does great at 99% of the road situations
economically almost worthless
[Christiano][13:49]
I think the “being dumb” is way more important than “covers every case”
[Yudkowsky][13:50]
(albeit that if new cities could still be built, we could totally take those 99%-complete AI cars and build fences and fence-gates around them, in a city where they were the only cars on the road, in which case they *would* work, and get big economic gains from these new cities with driverless cars, which ties back into my point about how current world GDP is *unwilling* to accept tech inputs)
like, it is in fact very plausible to me that there is a neighboring branch of reality with open borders and no housing-supply-constriction laws and no medical-supply-constriction laws, and their world GDP *does* manage to double before AGI hits them really hard, albeit maybe not in 4 years. this world *is not Earth*. they are constructing new cities to take advantage of 99%-complete driverless cars *right now*, or rather, they started constructing them 5 years ago and finished 4 years and 6 months ago.
### 9.3. Requirements for FOOM
[Christiano][13:53]
I really feel like the important part is the jumpiness you are imagining on the AI side / why AGI is different from other things
[Cotra][13:53]
It’s actually not obvious to me that Eliezer is imagining that much more jumpiness on the AI technology side than you are, Paul
E.g. he’s said in the past that while the gap from “subhuman to superhuman AI” could be 2h if it’s in the middle of FOOM, it could also be a couple years if it’s more like scaling alphago
[Yudkowsky][13:54]
Indeed! We observed this jumpiness with hominids. A lot of stuff happened at once with hominids, but a critical terminal part of the jump was the way that hominids started scaling their own food supply, instead of being ultimately limited by the food supply of the savanna.
[Cotra][13:54]
A couple years is basically what Paul believes
[Christiano][13:55]
(discord is not a great place for threaded conversations :()
[Cotra][13:55]
What are the probabilities you’re each placing on the 2h-2y spectrum? I feel like Paul is like “no way on 2h, likely on 2y” and Eliezer is like “who knows” on the whole spectrum, and a lot of the disagreement is the impact of the previous systems?
[Christiano][13:55]
yeah, I’m basically at “no way,” because it seems obvious that the AI that can foom in 2h is preceded by the AI that can foom in 2y
[Yudkowsky][13:56]
well, we surely agree there!
[Christiano][13:56]
OK, and it seems to me like it is preceded by years
[Yudkowsky][13:56]
we disagree on whether the AI that can foom in 2y clearly comes more than 2y before the AI that fooms in 2h
[Christiano][13:56]
yeah
perhaps we can all agree it’s preceded by at least 2h
so I have some view like: for any given AI we can measure “how long does it take to foom?” and it seems to me like this is just a nice graph
and it’s not exactly clear how quickly that number is going down, but a natural guess to me is something like “halving each year” based on the current rate of progress in hardware and software
and you see localized fast progress most often in places where there hasn’t yet been much attention
and my best guess for your view is that actually that’s not a nice graph at all, there is some critical threshold or range where AI quickly moves from “not fooming for a really long time” to “fooming really fast,” and that seems like the part I’m objecting to
[Cotra][13:59]
Paul, is your take that there’s a non-infinity number for time to FOOM that’d be associated with current AI systems (unassisted by humans)?
And it’s going down over time?
I feel like I would have said something more like “there’s a $ amount it takes to build a system that will FOOM in X amount of time, and that’s going down”
where it’s like quadrillions of dollars today
[Christiano][14:00]
I think it would be a big engineering project to make such an AI, which no one is doing because it would be uselessly slow even if successful
[Yudkowsky][14:02]
I… don’t think GPT-3 fooms given 2^30 longer time to think about than the systems that would otherwise exist 30 years from now, on timelines I’d consider relatively long, and hence generous to this viewpoint? I also don’t think you can take a quadrillion dollars and scale GPT-3 to foom today?
[Cotra][14:03]
I would agree with your take on GPT-3 fooming, and I didn’t mean a quadrillion dollars just to scale GPT-3, would probably be a difft architecture
[Christiano][14:03]
I also agree that GPT-3 doesn’t foom, it just keeps outputting <EOT>[next web page]<EOT>…
But I think the axes of “smart enough to foom fast” and “wants to foom” are pretty different. I also agree there is some minimal threshold below which it doesn’t even make sense to talk about “wants to foom,” which I think is probably just not that hard to reach.
(Also there are always diminishing returns as you continue increasing compute, which become very relevant if you try to GPT-3 for a billion billion years as in your hypothetical even apart from “wants to foom”.)
[Cotra][14:06]
I think maybe you and EY then disagree on where the threshold from “infinity” to “a finite number” for “time for this AI system to FOOM” begins? where eliezer thinks it’ll drop from infinity to a pretty small finite number and you think it’ll drop to a pretty large finite number, and keep going down from there
[Christiano][14:07]
I also think we will likely jump down to a foom-ing system only after stuff is pretty crazy, but I think that’s probably less important
I think what you said is probably the main important disagreement
[Cotra][14:08]
as in before that point it’ll be faster to have human-driven progress than FOOM-driven progress bc the FOOM would be too slow?
and there’s some crossover point around when the FOOM time is just a bit faster than the human-driven progress time
[Christiano][14:09]
yeah, I think most likely (AI+humans) is faster than (AI alone) because of complementarity. But I think Eliezer and I would still disagree even if I thought there was 0 complementarity and it’s just (humans improving AI) and separately (AI improving AI)
on that pure substitutes model I expect “AI foom” to start when the rate of AI-driven AI progress overtakes the previous rate of human-driven AI progress
like, I expect the time for successive “doublings” of AI output to be like 1 year, 1 year, 1 year, 1 year, [AI takes over] 6 months, 3 months, …
and the most extreme fast takeoff scenario that seems plausible is that kind of perfect substitutes + no physical economic impact from the prior AI systems
and then by that point fast enough physical impact is really hard so it happens essentially after the software-only singularity
I consider that view kind of unlikely but at least coherent
### 9.4. AI-driven accelerating economic growth
[Yudkowsky][14:12]
I’m expecting that the economy doesn’t accept much inputs from chimps, and then the economy doesn’t accept much input from village idiots, and then the economy doesn’t accept much input from weird immigrants. I can imagine that there may or may not be a very weird 2-year or 3-month period with strange half-genius systems running around, but they will still not be allowed to build houses. In the terminal phase things get more predictable and the AGI starts its own economy instead.
[Christiano][14:12]
I guess you can go even faster, by having a big and accelerating ramp-up in human investment right around the end, so that the “1 year” is faster (e.g. if recursive self-improvement was like playing go, and you could move from “a few individuals” to “google spending $10B” over a few years)
[Yudkowsky][14:13]
My ~~model~~ prophecy doesn’t rule that out as a thing that could happen, but sure doesn’t emphasize it as a key step that needs to happen.
[Christiano][14:13]
I think it’s very likely that AI will mostly be applied to further hardware+software progress
| |
| --- |
| [Cotra: ➕] |
I don’t really understand why you keep talking about houses and healthcare
[Cotra][14:13]
Eliezer, what about stuff like Google already using ML systems to automate its TPU load-sharing decisions, and people starting ot use Codex to automate routine programming, and so on? Seems like there’s a lot of stuff like that starting to already happen and markets are pricing in huge further increases
[Christiano][14:14]
it seems like the non-AI up-for-grabs zone are things like manufacturing, not things like healthcare
| |
| --- |
| [Cotra: ➕] |
[Cotra][14:14]
(I mean on your timelines obviously not much time for acceleration anyway, but that’s distinct from the regulation not allowing weak AIs to do stuff story)
[Yudkowsky][14:14]
Because I think that a key thing of what makes your prophecy less likely is the way that it happens inside the real world, where, economic gains or not, the System is unwilling/unable to take the things that are 99% self-driving cars and start to derive big economic benefits from those.
[Cotra][14:15]
but it seems like huge economic gains could happen entirely in industries mostly not regulated and not customer-facing, like hardware/software R&D, manufacturing. shipping logistics, etc
[Yudkowsky][14:15]
Ajeya, I’d consider Codex of *far* greater could-be-economically-important-ness than automated TPU load-sharing decisions
[Cotra][14:15]
i would agree with that, it’s smarter and more general
and i think that kind of thing could be applied on the hardware chip design side too
[Yudkowsky][14:16]
no, because the TPU load-sharing stuff has an obvious saturation point as a world economic input, while superCodex could be a world economic input in many more places
[Cotra][14:16]
the TPU load sharing thing was not a claim that this application could scale up to crazy impacts, but that it was allowed to happen, and future stuff that improves that kind of thing (back-end hardware/software/logistics) would probably also be allowed
[Yudkowsky][14:16]
my sense is that dectupling the number of programmers would not lift world GDP much, but it seems a lot more possible for me to be wrong about that
[Christiano][14:17]
the point is that housing and healthcare are not central examples of things that scale up at the beginning of explosive growth, regardless of whether it’s hard or soft
they are slower and harder, and also in efficient markets-land they become way less important during the transition
so they aren’t happening that much on anyone’s story
and also it doesn’t make that much difference whether they happen, because they have pretty limited effects on other stuff
like, right now we have an industry of ~hundreds of billions that is producing computing hardware, building datacenters, mining raw inputs, building factories to build computing hardware, solar panels, shipping around all of those parts, etc. etc.
I’m kind of interested in the question of whether all that stuff explodes, although it doesn’t feel as core as the question of “what are the dynamics of the software-only singularity and how much $ are people spending initiating it?”
but I’m not really interested in the question of whether human welfare is spiking during the transition or only after
[Yudkowsky][14:20]
All of world GDP has never felt particularly relevant to me on that score, since twice as much hardware maybe corresponds to being 3 months earlier, or something like that.
[Christiano][14:21]
that sounds like the stuff of predictions?
[Yudkowsky][14:21]
But if complete chip manufacturing cycles have accepted much more effective AI input, with no non-AI bottlenecks, then that… sure is a much more *material* element of a foom cycle than I usually envision.
[Christiano][14:21]
like, do you think it’s often the case that 3 months of software progress = doubling compute spending? or do you think AGI is different from “normal” AI on this perspective?
I don’t think that’s that far off anyway
I would guess like ~1 year
[Yudkowsky][14:22]
Like, world GDP that goes up by only 10%, but that’s because producing compute capacity was 2.5% of world GDP and that quadrupled, starts to feel much more to me like it’s part of a foom story.
I expect software-beats-hardware to hit harder and harder as you get closer to AGI, yeah.
the prediction is firmer near the terminal phase, but I think this is also a case where I expect that to be visible earlier
[Christiano][14:24]
I think that by the time that the AI-improving-AI takes over, it’s likely that hardware+software manufacturing+R&D represents like 10-20% of GDP, and that the “alien accountants” visiting earth would value those companies at like 80%+ of GDP
### 9.5. Brain size and evolutionary history
[Cotra][14:24]
On software beating hardware, how much of your view is dependent on your belief that the chimp -> human transition was probably not mainly about brain size because if it were about brain size it would have happened faster? My understanding is that you think the main change is a small software innovation which increased returns to having a bigger brain. If you changed your mind and thought that the chimp -> human transition was probably mostly about raw brain size, what (if anything) about your AI takeoff views would change?
[Yudkowsky][14:25]
I think that’s a pretty different world in a lot of ways!
but yes it hits AI takeoff views too
[Christiano][14:25]
regarding software vs hardware, here is an example of asking this question for imagenet classification (“how much compute to train a model to do the task?”), with a bit over 1 year doubling times (<https://openai.com/blog/ai-and-efficiency/>). I guess my view is that we can make a similar graph for “compute required to make your AI FOOM” and that it will be falling significantly slower than 2x/year. And my prediction for other tasks is that the analogous graphs will also tend to be falling slower than 2x/year.
[Yudkowsky][14:26]
to the extent that I modeled hominid evolution as having been “dutifully schlep more of the same stuff, get predictably more of the same returns” that would correspond to a world in which intelligence was less scary, different, dangerous-by-default
[Cotra][14:27]
thanks, that’s helpful. I looked around in [IEM](https://intelligence.org/files/IEM.pdf) and other places for a calculation of how quickly we should have evolved to humans if it were mainly about brain size, but I only found qualitative statements. If there’s a calculation somewhere I would appreciate a pointer to it, because currently it seems to me that a story like “selection pressure toward general intelligence was weak-to-moderate because it wasn’t actually *that* important for fitness, and this degree of selection pressure is consistent with brain size being the main deal and just taking a few million years to happen” is very plausible
[Yudkowsky][14:29]
well, for one thing, the prefrontal cortex expanded twice as fast as the rest
and iirc there’s evidence of a lot of recent genetic adaptation… though I’m not as sure you could pinpoint it as being about brain-stuff or that the brain-stuff was about cognition rather than rapidly shifting motivations or something.
elephant brains are 3-4 times larger by weight than human brains (just looked up)
if it’s that easy to get returns on scaling, seems like it shouldn’t have taken that long for evolution to go there
[Cotra][14:31]
but they have fewer synapses (would compute to less FLOP/s by the standard conversion)
how long do you think it should have taken?
[Yudkowsky][14:31]
early dinosaurs should’ve hopped onto the predictable returns train
[Cotra][14:31]
is there a calculation?
you said in IEM that evolution increases organ sizes quickly but there wasn’t a citation to easily follow up on there
[Yudkowsky][14:33]
I mean, you could produce a graph of smooth fitness returns to intelligence, smooth cognitive returns on brain size/activity, linear metabolic costs for brain activity, fit that to humans and hominids, then show that obviously if hominids went down that pathway, large dinosaurs should’ve gone down it first because they had larger bodies and the relative metabolic costs of increased intelligence would’ve been lower at every point along the way
I do not have a citation for that ready, if I’d known at the time you’d want one I’d have asked Luke M for it while he still worked at MIRI 
[Cotra][14:35]
cool thanks, will think about the dinosaur thing (my first reaction is that this should depend on the actual fitness benefits to general intelligence which might have been modest)
[Yudkowsky][14:35]
I suspect we’re getting off Paul’s crux, though
[Cotra][14:35]
yeah we can go back to that convo (though i think paul would also disagree about this thing, and believes that the chimp to human thing was mostly about size)
sorry for hijacking
[Yudkowsky][14:36]
well, if at some point I can produce a major shift in EA viewpoints by coming up with evidence for a bunch of non-brain-size brain selection going on over those timescales, like brain-related genes where we can figure out how old the mutation is, I’d then put a lot more priority on digging up a paper like that
I’d consider it sufficiently odd to imagine hominids->humans as being primarily about brain size, given the evidence we have, that I do not believe this is Paul’s position until Paul tells me so
[Christiano][14:49]
I would guess it’s primarily about brain size / neuron count / cortical neuron count
and that the change in rate does mostly go through changing niche, where both primates and birds have this cycle of rapidly accelerating brain size increases that aren’t really observed in other animals
it seems like brain size is increasing extremely quickly on both of those lines
[Yudkowsky][14:50]
why aren’t elephants GI?
[Christiano][14:51]
mostly they have big brains to operate big bodies, and also my position obviously does not imply (big brain) ==(necessarily implies)==> general intelligence
[Yudkowsky][14:52]
I don’t understand, in general, how your general position manages to strongly imply a bunch of stuff about AGI and not strongly imply similar stuff about a bunch of other stuff that sure sounds similar to me
[Christiano][14:52]
don’t elephants have very few synapses relative to humans?
| |
| --- |
| [Cotra: ➕] |
how does the scale hypothesis possibly take a strong stand on synapses vs neurons? I agree that it takes a modest predictive hit from “why aren’t the big animals much smarter?”
[Yudkowsky][14:53]
if adding more synapses just scales, elephants should be able to pay hominid brain costs for a much smaller added fraction of metabolism and also not pay the huge death-in-childbirth head-size tax
because their brains and heads are already 4x as huge as they need to be for GI
and now they just need some synapses, which are a much tinier fraction of their total metabolic costs
[Christiano][14:54]
I mean, you can also make smaller and cheaper synapses as evidenced by birds
I’m not sure I understand what you are saying
it’s clear that you can’t say “X is possible metabolically, so evolution would do it”
or else you are confused about why primate brains are so bad
[Yudkowsky][14:54]
great, then smaller and cheaper synapses should’ve scaled many eons earlier and taken over the world
[Christiano][14:55]
this isn’t about general intelligence, this is a reductio of your position…
[Yudkowsky][14:55]
and here I had thought it was a reductio of your position…
[Christiano][14:55]
indeed
like, we all grant that it’s metabolically possible to have small smart brains
and evolution doesn’t do it
and I’m saying that it’s also possible to have small smart brains
and that scaling brains up matters a lot
[Yudkowsky][14:56]
no, you grant that it’s metabolically possible to have cheap brains full of synapses, which are therefore, on your position, smart
[Christiano][14:56]
birds are just smart
we know they are smart
this isn’t some kind of weird conjecture
like, we can debate whether they are a “general” intelligence, but it makes no difference to this discussion
the point is that they do more with less metabolic cost
[Yudkowsky][14:57]
on my position, the brain needs to invent the equivalents of ReLUs and Transformers and really rather a lot of other stuff because it can’t afford nearly that many GPUs, and then the marginal returns on adding expensive huge brains and synapses have increased enough that hominids start to slide down the resulting fitness slope, which isn’t even paying off in guns and rockets yet, they’re just getting that much intelligence out of it once the brain software has been selected to scale that well
[Christiano][14:57]
but all of the primates and birds have brain sizes scaling much faster than the other animals
like, the relevant “things started to scale” threshold is way before chimps vs humans
isn’t it?
[Cotra][14:58]
to clarify, my understanding is that paul’s position is “Intelligence is mainly about synapse/neuron count, and evolution doesn’t care that much about intelligence; it cared more for birds and primates, and both lines are getting smarter+bigger-brained.” And eliezer’s position is that “evolution should care a ton about intelligence in most niches, so if it were mostly about brain size then it should have gone up to human brain sizes with the dinosaurs”
[Christiano][14:58]
or like, what is the evidence you think is explained by the threshold being between chimps and humans
[Yudkowsky][14:58]
if hominids have less efficient brains than birds, on this theory, it’s because (post facto handwave) birds are tiny, so whatever cognitive fitness gradients they face, will tend to get paid more in software and biological efficiency and biologically efficient software, and less paid in Stack More Neurons (even compared to hominids)
elephants just don’t have the base software to benefit much from scaling synapses even though they’d be relatively cheaper for elephants
[Christiano][14:59]
@ajeya I think that intelligence is about a lot of things, but that size (or maybe “more of the same” changes that had been happening recently amongst primates) is the big difference between chimps and humans
| |
| --- |
| [Cotra: 👍] |
[Cotra][14:59]
got it yeah i was focusing on chimp-human gap when i said “intelligence” there but good to be careful
[Yudkowsky][14:59]
I have not actually succeeded in understanding Why On Earth Anybody Would Think That If Not For This Really Weird Prior I Don’t Get Either
re: the “more of the same” theory of humans
[Cotra][15:00]
do you endorse my characterization of your position above? “evolution should care a ton about intelligence in most niches, so if it were mostly about brain size then it should have gone up to human brain sizes with the dinosaurs”
in which case the disagreement is about how much evolution should care about intelligence in the dinosaur niche, vs other things it could put its skill points into?
[Christiano][15:01]
Eliezer, it seems like chimps are insanely smart compared to other animals, basically as smart as they get
so it’s natural to think that the main things that make humans unique are also present in chimps
or at least, there was something going on in chimps that is exceptional
and should be causally upstream of the uniqueness of humans too
otherwise you have too many coincidences on your hands
[Yudkowsky][15:02]
ajeya: no, I’d characterize that as “the human environmental niche per se does not seem super-special enough to be unique on a geological timescale, the cognitive part of the niche derives from increased cognitive abilities in the first place and so can’t be used to explain where they got started, dinosaurs are larger than humans and would pay lower relative metabolic costs for added brain size and it is not the case that every species as large as humans was in an environment where they would not have benefited as much from a fixed increment of intelligence, hominids are probably distinguished from dinosaurs in having better neural algorithms that arose over intervening evolutionary time and therefore better returns in intelligence on synapses that are more costly to humans than to elephants or large dinosaurs”
[Christiano][15:03]
I don’t understand how you can think that hominids are the special step relative to something earlier
or like, I can see how it’s consistent, but I don’t see what evidence or argument supports it
it seems like the short evolutionary time, and the fact that you also have to explain the exceptional qualities of other primates, cut extremely strongly against it
[Yudkowsky][15:04]
paul: indeed, the fact that dinosaurs didn’t see their brain sizes and intelligences ballooning, says there must be a lot of stuff hominids had that dinosaurs didn’t, explaining why hominids got much higher returns on intelligence per synapse. natural selection is enough of a smooth process that 95% of this stuff should’ve been in the last common ancestor of humans and chimps.
[Christiano][15:05]
it seems like brain size basically just increases faster in the smarter animals? though I mostly just know about birds and primates
[Yudkowsky][15:05]
that is what you’d predict from smartness being about algorithms!
[Christiano][15:05]
and it accelerates further and further within both lines
it’s what you’d expect if smartness is about algorithms *and chimps and birds have good algorithms*
[Yudkowsky][15:06]
if smartness was about brain size, smartness and brain size would increase faster in the *larger animals* or the ones whose successful members *ate more food per day*
well, sure, I do model that birds have better algorithms than dinosaurs
[Cotra][15:07]
it seems like you’ve given arguments for “there was algorithmic innovation between dinosaurs and humans” but not yet arguments for “there was major algorithmic innovation between chimps and humans”?
[Christiano][15:08]
(much less that the algorithmic changes were not just more-of-the-same)
[Yudkowsky][15:08]
oh, that’s *not* mandated by the model the same way. (between LCA of chimps and humans)
[Christiano][15:08]
isn’t that exactly what we are discussing?
[Yudkowsky][15:09]
…I hadn’t thought so, no.
[Cotra][15:09]
original q was:
>
> On software beating hardware, how much of your view is dependent on your belief that the chimp -> human transition was probably not mainly about brain size because if it were about brain size it would have happened faster? My understanding is that you think the main change is a small software innovation which increased returns to having a bigger brain. If you changed your mind and thought that the chimp -> human transition was probably mostly about raw brain size, what (if anything) about your AI takeoff views would change?
>
>
>
so i thought we were talking about if there’s a cool innovation from chimp->human?
[Yudkowsky][15:10]
I can see how this would have been the more obvious intended interpretation on your viewpoint, and apologize
[Christiano][15:10]
>
> (though i think paul would also disagree about this thing, and believes that the chimp to human thing was mostly about size)
>
>
>
Is what I was responding to in part
I am open to saying that I’m conflating size and “algorithmic improvements that are closely correlated with size in practice and are similar to the prior algorithmic improvements amongst primates”
[Yudkowsky][15:11]
from my perspective, the question is “how did that hominid->human transition happen, as opposed to there being an elephant->smartelephant or dinosaur->smartdinosaur transition”?
I expect there were substantial numbers of brain algorithm stuffs going on during this time, however
because I don’t think that synapses scale that well *with* the baseline hominid boost
[Christiano][15:11]
FWIW, it seems quite likely to me that there would be an elephant->smartelephant transition within tens of millions or maybe 100M years, and a dinosaur->smartdinosaur transition in hundreds of millions of years
and those are just cut off by the fastest lines getting there first
[Yudkowsky][15:12]
which I think does circle back to that point? actually I think my memory glitched and forgot the original point while being about this subpoint and I probably did interpret the original point as intended.
[Christiano][15:12]
namely primates beating out birds by a hair
[Yudkowsky][15:12]
that sounds like a viewpoint which would also think it much more likely that GPT-3 would foom in a billion years
where maybe you think that’s unlikely, but I still get the impression your “unlikely” is, like, 5 orders of magnitude likelier than mine before applying overconfidence adjustments against extreme probabilities on both sides
yeah, I think I need to back up
[Cotra][15:15]
Is your position something like “at some point after dinosaurs, there was an algorithmic innovation that increased returns to brain size, which meant that the birds and the humans see their brains increasing quickly while the dinosaurs didn’t”?
[Christiano][15:15]
it also seems to me like the chimp->human difference is in basically the same ballpark of the effect of brain size within humans, given modest adaptations for culture
which seems like a relevant sanity-check that made me take the “mostly hardware” view more seriously
[Yudkowsky][15:15]
there’s a part of my model which very strongly says that hominids scaled better than elephants and that’s why “hominids->humans but not elephants->superelephants”
[Christiano][15:15]
previously I had assumed that analysis would show that chimps were obviously *way* dumber than an extrapolation of humans
[Yudkowsky][15:16]
there’s another part of my model which says “and it still didn’t scale that well without algorithms, so we should expect a lot of alleles affecting brain circuitry which rose to fixation over the period when hominid brains were expanding”
this part is strong and I think echoes back to AGI stuff, but it is not *as strong* as the much *more* overdetermined position that hominids started with more scalable algorithms than dinosaurs.
[Christiano][15:17]
I do agree with the point that there are structural changes in brains as you scale them up, and this is potentially a reason why brain size changes more slowly than e.g. bone size. (Also there are small structural changes in ML algorithms as you scale them up, not sure how much you want to push the analogy but they feel fairly similar.)
[Yudkowsky][15:17]
>
> it also seems to me like the chimp->human difference is in basically the same ballpark of the effect of brain size within humans, given modest adaptations for culture
>
>
>
this part also seems pretty blatantly false to me
is there, like, a smooth graph that you looked at there?
[Christiano][15:18]
I think the extrapolated difference would be about 4 standard deviations, so we are comparing a chimp to an IQ 40 human
[Yudkowsky][15:18]
I’m really not sure how much of a fair comparison that is
IQ 40 humans in our society may be mostly sufficiently-damaged humans, not scaled-down humans
[Christiano][15:19]
doesn’t seem easy, but the point is that the extrapolated difference is huge, it corresponds to completely debilitating developmental problems
[Yudkowsky][15:19]
if you do enough damage to a human you end up with, for example, a coma victim who’s not competitive with other primates at all
[Christiano][15:19]
yes, that’s more than 4 SD down
I agree with this general point
I’d guess I just have a lot more respect for chimps than you do
[Yudkowsky][15:20]
I feel like I have a bunch of respect for chimps but more respect for humans
like, that stuff humans do
that is really difficult stuff!
it is not just scaled-up chimpstuff!
[Christiano][15:21]
Carl convinced me chimps wouldn’t go to space, but I still really think it’s about domesticity and cultural issues rather than intelligence
[Yudkowsky][15:21]
the chimpstuff is very respectable but there is a whole big layer cake of additional respect on top
[Christiano][15:21]
not a prediction to be resolved until after the singularity
I mean, the space prediction isn’t very confident 
and it involved a very large planet of apes
### 9.6. Architectural innovation in AI and in evolutionary history
[Yudkowsky][15:22]
I feel like if GPT-based systems saturate and require *any* architectural innovation rather than Stack More Layers to get much further, this is a pre-Singularity point of observation which favors humans probably being more qualitatively different from chimp-LCA
(LCA=last common ancestor)
[Christiano][15:22]
any seems like a kind of silly bar?
[Yudkowsky][15:23]
because single architectural innovations are allowed to have large effects!
[Christiano][15:23]
like there were already small changes to normalization from GPT-2 to GPT-3, so isn’t it settled?
[Yudkowsky][15:23]
natural selection can’t afford to deploy that many of them!
[Christiano][15:23]
and the model really eventually won’t work if you increase layers but don’t fix the normalization, there are severe problems that only get revealed at high scale
[Yudkowsky][15:23]
that I wouldn’t call architectural innovation
transformers were
this is a place where I would not discuss specific ideas because I do not actually want this event to occur
[Christiano][15:24]
sure
have you seen a graph of LSTM scaling vs transformer scaling?
I think LSTM with ongoing normalization-style fixes lags like 3x behind transformers on language modeling
[Yudkowsky][15:25]
no, does it show convergence at high-enough scales?
[Christiano][15:25]
figure 7 here: <https://arxiv.org/pdf/2001.08361.pdf>
[Yudkowsky][15:26]
yeah… I unfortunately would rather not give other people a sense for which innovations are obviously more of the same and which innovations obviously count as qualitative
[Christiano][15:26]
I think smart money is that careful initialization and normalization on the RNN will let it keep up for longer
anyway, I’m very open to differences like LSTM vs transformer between humans and 3x-smaller-brained-ancestors, as long as you are open to like 10 similar differences further back in the evolutionary history
[Yudkowsky][15:28]
what if there’s 27 differences like that and 243 differences further back in history?
[Christiano][15:28]
sure
[Yudkowsky][15:28]
is that a distinctly Yudkowskian view vs a Paul view…
apparently not
I am again feeling confused about cruxes
[Christiano][15:29]
I mean, 27 differences like transformer vs LSTM isn’t actually plausible, so I guess we could talk about it
[Cotra][15:30]
Here’s a potential crux articulation that ties it back to the animals stuff: paul thinks that we first discover major algorithmic innovations that improve intelligence at a low level of intelligence, analogous to evolution discovering major architectural innovations with tiny birds and primates, and then there will be a long period of scaling up plus coming up with routine algorithmic tweaks to get to the high level, analogous to evolution schlepping on the same shit for a long time to get to humans. analogously, he thinks when big innovations come onto the scene the actual product is crappy af (e.g. wright brother’s plane), and it needs a ton of work to scale up to usable and then to great.
you both seem to think both evolution and tech history consiliently point in your direction
[Christiano][15:33]
that sounds vaguely right, I guess the important part of “routine” is “vaguely predictable,” like you mostly work your way down the low-hanging fruit (including new fruit that becomes more important as you scale), and it becomes more and more predictable the more people are working on it and the longer you’ve been at it
and deep learning is already reasonably predictable (i.e. the impact of successive individual architectural changes is smaller, and law of large numbers is doing its thing) and is getting more so, and I just expect that to continue
[Cotra][15:34]
yeah, like it’s a view that points to using data that relates effort to algorithmic progress and using that to predict future progress (in combination with predictions of future effort)
[Christiano][15:35]
yeah
and for my part, it feels like this is how most technologies look and also how current ML progress looks
[Cotra][15:36]
and *also* how evolution looks, right?
[Christiano][15:37]
you aren’t seeing big jumps in translation or in self-driving cars or in image recognition, you are just seeing a long slog, and you see big jumps in areas where few people work (usually up to levels that are not in fact that important, which is very correlated with few people working there)
I don’t know much about evolution, but it at least looks very consistent with what I know and the facts eliezer cites
(not merely consistent, but “explains the data just about as well as the other hypotheses on offer”)
### 9.7. Styles of thinking in forecasting
[Yudkowsky][15:38]
I do observe that this would seem, on the surface of things, to describe the entire course of natural selection up until about 20K years ago, if you were looking at surface impacts
[Christiano][15:39]
by 20k years ago I think it’s basically obvious that you are tens of thousands of years from the singularity
like, I think natural selection is going crazy with the brains by millions of years ago, and by hundreds of thousands of years ago humans are going crazy with the culture, and by tens of thousands of years ago the culture thing has accelerated and is almost at the finish line
[Yudkowsky][15:41]
really? I don’t know if I would have been able to call that in advance if I’d never seen the future or any other planets. I mean, maybe, but I sure would have been extrapolating way out onto a further limb than I’m going here.
[Christiano][15:41]
Yeah, I agree singularity is way more out on a limb—or like, where the singularity stops is more uncertain since that’s all that’s really at issue from my perspective
but the point is that everything is clearly crazy in historical terms, in the same way that 2000 is crazy, even if you don’t know where it’s going
and the timescale for the crazy changes is tens of thousands of years
[Yudkowsky][15:42]
I frankly model that, had I made any such prediction 20K years ago of hominids being able to pull of moon landings or global warming – never mind the Singularity – I would have faced huge pushback from many EAs, such as, for example, Robin Hanson, and you.
[Christiano][15:42]
like I think this can’t go on would have applied just as well: <https://www.lesswrong.com/posts/5FZxhdi6hZp8QwK7k/this-can-t-go-on>
I don’t think that’s the case at all
and I think you still somehow don’t understand my position?
[Yudkowsky][15:43]
<https://www.lesswrong.com/posts/XQirei3crsLxsCQoi/surprised-by-brains> is my old entry here
[Christiano][15:43]
like, what is the move I’m making here, that you think I would have made in the past?
and would have led astray?
[Yudkowsky][15:44]
I sure do feel in a deeper sense that I am trying very hard to account for perspective shifts in how unpredictable the future actually looks at the time, and the Other is looking back at the past and organizing it neatly and expecting the future to be that neat
[Christiano][15:45]
I don’t even feel like I’m expecting the future to be neat
are you just saying you have a really broad distribution over takeoff speed, and that “less than a month” gets a lot of probability because lots of numbers are less than a month?
[Yudkowsky][15:47]
not exactly?
[Christiano][15:47]
in what way is your view the one that is preferred by things being messy or unpredictable?
like, we’re both agreeing X will eventually happen, and I’m making some concrete prediction about how some other X’ will happen first, and that’s the kind of specific prediction that’s likely to be wrong?
[Yudkowsky][15:48]
more like, we sure can tell a story today about how normal and predictable AlphaGo was, but we can *always* tell stories like that about the past. I do not particularly recall the AI field standing up one year before AlphaGo and saying “It’s time, we’re coming for the 8-dan pros this year and we’re gonna be world champions a year after that.” (Which took significantly longer in chess, too, matching my other thesis about how these slides are getting steeper as we get closer to the end.)
[Christiano][15:49]
it’s more like, you are offering AGZ as an example of why things are crazy, and I’m doubtful / think it’s pretty lame
maybe I don’t understand how it’s functioning as bayesian evidence
for what over what
[Yudkowsky][15:50]
I feel like the whole smoothness-reasonable-investment view, if evaluated on Earth 5My ago *without benefit of foresight*, would have dismissed the notion of brains overtaking evolution; evaluated 1My ago, it would have dismissed the notion of brains overtaking evolution; evaluated 20Ky ago, it would have barely started to acknowledge that brains were doing anything interesting at all, but pointed out how the hominids could still only eat as much food as their niche offered them and how the cute little handaxes did not begin to compare to livers and wasp stings.
there is a style of thinking that says, “wow, yeah, people in the past sure were surprised by stuff, oh, wait, *I’m also in the past*, aren’t I, I am one of those people”
and a view where you look back from the present and think about how reasonable the past all seems now, and the future will no doubt be equally reasonable
[Christiano][15:52]
(the AGZ example may fall flat, because the arguments we are making about it now *we were also making in the past*)
[Yudkowsky][15:52]
I am not sure this is resolvable, but it is among my primary guesses for a deep difference in believed styles of thought
[Christiano][15:52]
I think that’s a useful perspective, but still don’t see how it favors your bottom line
[Yudkowsky][15:53]
where I look at the style of thinking you’re using, and say, not, “well, that’s invalidated by a technical error on line 3 even on Paul’s own terms” but “isn’t this obviously a whole style of thought that never works and ends up unrelated to reality”
I think the first AlphaGo was the larger shock, AlphaGo Zero was a noticeable but more mild shock on account of how it showed the end of game programming and not just the end of Go
[Christiano][15:54]
sorry, I lumped them together
[Yudkowsky][15:54]
it didn’t feel like the same level of surprise; it was precedented by then
the actual accomplishment may have been larger in an important sense, but a lot of the – epistemic landscape of lessons learned? – is about the things that surprise you at the time
[Christiano][15:55]
also AlphaGo was also quite easy to see coming after this paper (as was discussed extensively *at the time*): <https://www.cs.toronto.edu/~cmaddis/pubs/deepgo.pdf>
[Yudkowsky][15:55]
Paul, are you on the record as arguing with me that AlphaGo will win at Go because it’s predictably on-trend?
back then?
[Cotra][15:55]
Hm, it sounds like Paul is saying “I do a trend extrapolation over long time horizons and if things seem to be getting faster and faster I expect they’ll continue to accelerate; this extrapolation if done 100k years ago would have seen that things were getting faster and faster and projected singularity within 100s of K years”
Do you think Paul is in fact doing something other than the trend extrap he says he’s doing, or that he would have looked at a different less informative trend than the one he says he would have looked at, or something else?
[Christiano][15:56]
my methodology for answering that question is looking at LW comments mentioning go by me, can see if it finds any
[Yudkowsky][15:56]
Different less informative trend, is most of my suspicion there?
though, actually, I should revise that, I feel like relatively little of the WHA was AlphaGo v2 whose name I forget beating Lee Se-dol, and most was in the revelation that v1 beat the high-dan pro whose name I forget.
Paul having himself predicted anything at *all* like this would be the actually impressive feat
that would cause me to believe that the AI world is more regular and predictable than I experienced it as, if you are paying more attention to ICLR papers than I do
### 9.8. Moravec’s prediction
[Cotra][15:58]
And jtbc, the trend extrap paul is currently doing is something like:
* Look at how effort leads to hardware progress measured in FLOP/$ and software progress measured in stuff like “FLOP to do task X” or “performance on benchmark Y”
* Look at how effort in the ML industry as a whole is increasing, project forward with maybe some adjustments for thinking markets are more inefficient now and will be less inefficient later
and this is the wrong trend, because he shouldn’t be looking at hardware/software progress across the whole big industry and should be more open to an upset innovation coming from an area with a small number of people working on it?
and he would have similarly used the wrong trends while trying to do trend extrap in the past?
[Yudkowsky][15:59]
because I feel like this general style of thought doesn’t work when you use it on Earth generally, and then fails extremely hard if you try to use it on Earth before humans to figure out where the hominids are going because that phenomenon is Different from Previous Stuff
like, to be clear, I have seen this used well on solar
I feel like I saw some people calling the big solar shift based on graphs, before that happened
I have seen this used great by Moravec on computer chips to predict where computer chips would be in 2012
and also witnessed Moravec *completely failing* as soon as he tried to derive *literally anything but the graph itself* namely his corresponding prediction for human-equivalent AI in 2012 (I think, maybe it was 2010) or something
[Christiano][16:02]
(I think in his 1988 book Moravec estimated human-level AI in ~2030, not sure if you are referring to some earlier prediction?)
[Yudkowsky][16:02]
(I have seen Ray Kurzweil project out Moore’s Law to the $1,000,000 human brain in, what was it, 2025, followed by the $1000 human brain in 2035 and the $1 human brain in 2045, and when I asked Ray whether machine superintelligence might shift the graph at all, he replied that machine superintelligence was precisely how the graph would be able to continue on trend. This indeed is sillier than EAs.)
[Cotra][16:03]
moravec’s prediction appears to actually be around 2025, looking at his hokey graph? <https://jetpress.org/volume1/moravec.htm>
[Yudkowsky][16:03]
but even there, it does feel to me like there is a commonality between Kurzweil’s sheer graph-worship and difficulty in appreciating the graphs as surface phenomena that are less stable than deep phenomena, and something that Hanson was doing wrong in the foom debate
[Cotra][16:03]
which is…like, your timelines?
[Yudkowsky][16:04]
that’s 1998
Mind Children in 1988 I am pretty sure had an earlier prediction
[Christiano][16:04]
I should think you’d be happy to bet against me on basically any prediction, shouldn’t you?
[Yudkowsky][16:05]
any prediction that sounds narrow and isn’t like “this graph will be on trend in 3 more years”
…maybe I’m wrong, an online source says Mind Children in 1988 predicted AGI in “40 years” but I sure do seem to recall an extrapolated graph that reached “human-level hardware” in 2012 based on an extensive discussion about computing power to duplicate the work of the retina
[Christiano][16:08]
don’t think it matters too much other than for Moravec’s honor, doesn’t really make a big difference for the empirical success of the methodology
I think it’s on page 68 if you have the physical book
[Yudkowsky][16:09]
p60 via Google Books says 10 teraops for a human-equivalent mind
[Christiano][16:09]
I have a general read of history where trend extrapolation works extraordinarily well relative to other kinds of forecasting, to the extent that the best first-pass heuristic for whether a prediction is likely to be accurate is whether it’s a trend extrapolation and how far in the future it is
[Yudkowsky][16:09]
which, incidentally, strikes me as entirely plausible if you had algorithms as sophisticated as the human brain
my sense is that Moravec nailed the smooth graph of computing power going on being smooth, but then all of his predictions about the actual future were completely invalid on account of a curve interacting with his curve that he didn’t know things about and so simply omitted as a step in his calculations, namely, AGI algorithms
[Christiano][16:12]
though again, from your perspective 2030 is still a reasonable bottom-line forecast that makes him one of the most accurate people at that time?
[Yudkowsky][16:12]
you could be right about all the local behaviors that your history is already shouting out at you as having smooth curve (where by “local” I do mean to exclude stuff like world GDP extrapolated into the indefinite future) and the curves that history isn’t shouting at you will tear you down
[Christiano][16:12]
(I don’t know if he even forecast that)
[Yudkowsky][16:12]
I don’t remember that part from the 1988 book
my memory of the 1988 book is “10 teraops, based on what it takes to rival the retina” and he drew a graph of Moore’s Law
[Christiano][16:13]
yeah, I think that’s what he did
(and got 2030)
[Yudkowsky][16:14]
“If this rate of improvement were to continue into the next century, the 10 teraops required for a humanlike computer would be available in a $10 million supercomputer before 2010 and in a $1,000 personal computer by 2030.”
[Christiano][16:14]
or like, he says “human equivalent in 40 years” and predicts that in 50 years we will have robots with superhuman reasoning ability, not clear he’s ruling out human-equivalent AGI before 40 years but I think the tone is clear
[Yudkowsky][16:15]
so 2030 for AGI on a personal computer and 2010 for AGI on a supercomputer, and I expect that on my first reading I simply discarded the former prediction as foolish extrapolation past the model collapse he had just predicted in 2010.
(p68 in “Powering Up”)
[Christiano][16:15]
yeah, that makes sense
I do think the PC number seems irrelevant
[Cotra][16:16]
I think both in that book and in the 98 article he wants you to pay attention to the “very cheap human-size computers” threshold, not the “supercomputer” threshold, i think intentionally as a way to handwave in “we need people to be able to play around with these things”
(which people criticized him at the time for not more explicitly modeling iirc)
[Yudkowsky][16:17]
but! I mean! there are so many little places where the media has a little cognitive hiccup about that and decides in 1998 that it’s fine to describe that retrospectively as “you predicted in 1988 that we’d have true AI in 40 years” and then the future looks less surprising than people at the time using Trend Logic were actually surprised by it!
all these little ambiguities and places where, oh, you decide retroactively that it would have made sense to look at *this* Trend Line and use it *that* way, but if you look at what people said at the time, they didn’t actually say that!
[Christiano][16:19]
I mean, in fairness reading the book it just doesn’t seem like he is predicting human-level AI in 2010 rather than 2040, but I do agree that it seems like the basic methodology (why care about the small computer thing?) doesn’t really make that much sense a priori and only leads to something sane if it cancels out with a weird view
### 9.9. Prediction disagreements and bets
[Christiano][16:19]
anyway, I’m pretty unpersuaded by the kind of track record appeal you are making here
[Yudkowsky][16:20]
if the future goes the way I predict and yet anybody somehow survives, perhaps somebody will draw a hyperbolic trendline on some particular chart where the trendline is retroactively fitted to events including those that occurred in only the last 3 years, and say with a great sage nod, ah, yes, that was all according to trend, nor did anything depart from trend
trend lines permit anything
[Christiano][16:20]
like from my perspective the fundamental question is whether I would do better or worse by following the kind of reasoning you’d advocate, and it just looks to me like I’d do worse, and I’d love to make any predictions about anything to help make that more clear and hindsight-proof in advance
[Yudkowsky][16:20]
you just look into the past and find a line you can draw that ended up where reality went
[Christiano][16:21]
it feels to me like you really just waffle on almost any prediction about the before-end-of-days
[Yudkowsky][16:21]
I don’t think I know a lot about the before-end-of-days
[Christiano][16:21]
like if you make a prediction I’m happy to trade into it, or you can pick a topic and I can make a prediction and you can trade into mine
[Cotra][16:21]
but you know enough to have strong timing predictions, e.g. your bet with caplan
[Yudkowsky][16:21]
it’s daring enough that I claim to know anything about the Future at all!
[Cotra][16:21]
surely with that difference of timelines there should be some pre-2030 difference as well
[Christiano][16:21]
but you are the one making the track record argument against my way of reasoning about things!
how does that not correspond to believing that your predictions are better!
what does that mean?
[Yudkowsky][16:22]
yes and if you say something narrow enough or something that my model does at least vaguely push against, we should bet
[Christiano][16:22]
my point is that I’m willing to make a prediction about any old thing, you can name your topic
I think the way I’m reasoning about the future is just better in general
and I’m going to beat you on whatever thing you want to bet on
[Yudkowsky][16:22]
but if you say, “well, Moore’s Law on trend, next 3 years”, then I’m like, “well, yeah, sure, since I don’t feel like I know anything special about that, that would be my prediction too”
[Christiano][16:22]
sure
you can pick the topic
pick a quantity
or a yes/no question
or whatever
[Yudkowsky][16:23]
you may know better than I would where your Way of Thought makes strong, narrow, or unusual predictions
[Christiano][16:23]
I’m going to trend extrapolation everywhere
spoiler
[Yudkowsky][16:23]
okay but any superforecaster could do that and I could do the same by asking a superforecaster
[Cotra][16:24]
but there must be places where you’d strongly disagree w the superforecaster
since you disagree with them eventually, e.g. >2/3 doom by 2030
[Bensinger][18:40] (Nov. 25 follow-up comment)
“>2/3 doom by 2030” isn’t an actual Eliezer-prediction, and is based on a misunderstanding of something Eliezer said. See [Eliezer’s comment on LessWrong](https://www.lesswrong.com/posts/7MCqRnZzvszsxgtJi/christiano-cotra-and-yudkowsky-on-ai-progress?commentId=diChXiELZd62hgRyK#diChXiELZd62hgRyK).
[Yudkowsky][16:24]
in the terminal phase, sure
[Cotra][16:24]
right, but there are no disagreements before jan 1 2030?
no places where you’d strongly defy the superforecasters/trend extrap?
[Yudkowsky][16:24]
superforecasters were claiming that AlphaGo had a 20% chance of beating Lee Se-dol and I didn’t disagree with that at the time, though as the final days approached I became nervous and suggested to a friend that they buy out of a bet about that
[Cotra][16:25]
what about like whether we get some kind of AI ability (e.g. coding better than X) before end days
[Yudkowsky][16:25]
though that was more because of having started to feel incompetent and like I couldn’t trust the superforecasters to know more, than because I had switched to a confident statement that AlphaGo would win
[Cotra][16:25]
seems like EY’s deep intelligence / insight-oriented view should say something about what’s not possible before we get the “click” and the FOOM
[Christiano][16:25]
I mean, I’m OK with either (i) evaluating arguments rather than dismissive and IMO totally unjustified track record, (ii) making bets about stuff
I don’t see how we can both be dismissing things for track record reasons and also not disagreeing about things
if our methodologies agree about all questions before end of days (which seems crazy to me) then surely there is no track record distinction between them…
| |
| --- |
| [Cotra: 👍] |
[Cotra][16:26]
do you think coding models will be able to 2x programmer productivity before end days? 4x?
what about hardware/software R&D wages? will they get up to $20m/yr for good ppl?
will someone train a 10T param model before end days?
[Christiano][16:27]
things I’m happy to bet about: economic value of LMs or coding models at 2, 5, 10 years, benchmark performance of either, robotics, wages in various industries, sizes of various industries, compute/$, someone else’s views about “how ML is going” in 5 years
maybe the “any GDP acceleration before end of days?” works, but I didn’t like how you don’t win until the end of days
[Yudkowsky][16:28]
okay, so here’s an example place of a *weak* general Yudkowskian prediction, that is weaker than terminal-phase stuff of the End Days: (1) I predict that cycles of ‘just started to be able to do Narrow Thing -> blew past upper end of human ability at Narrow Thing’ will continue to get shorter, the same way that, I think, this happened faster with Go than with chess.
[Christiano][16:28]
great, I’m totally into it
what’s a domain?
coding?
[Yudkowsky][16:28]
Does Paul disagree? Can Paul point to anything equally specific out of Paul’s viewpoint?
[Christiano][16:28]
benchmarks for LMs?
robotics?
[Yudkowsky][16:28]
well, for these purposes, we do need some Elo-like ability to measure at all where things are relative to humans
[Cotra][16:29]
problem-solving benchmarks for code?
MATH benchmark?
[Christiano][16:29]
well, for coding and LM’ing we have lots of benchmarks we can use
[Yudkowsky][16:29]
this unfortunately does feel a bit different to me from Chess benchmarks where the AI is playing the whole game; Codex is playing part of the game
[Christiano][16:29]
in general the way I’d measure is by talking about how fast you go from “weak human” to “strong human” (e.g. going from top-10,000 in chess to top-10 or whatever, going from jobs doable by $50k/year engineer to $500k/year engineer…)
[Yudkowsky][16:30]
golly, that sounds like a viewpoint very favorable to mine
[Christiano][16:30]
what do you mean?
that way of measuring would be favorable to your viewpoint?
[Yudkowsky][16:31]
if we measure how far it takes AI to go past different levels of paying professionals, I expect that the Chess duration is longer than the Go duration and that by the time Codex is replacing ~~a~~ most paid $50k/year programmers the time to replacing ~~a~~ most programmers paid as much as a top Go player will be pretty darned short
[Christiano][16:31]
top Go players don’t get paid, do they?
[Yudkowsky][16:31]
they tutor students and win titles
[Christiano][16:31]
but I mean, they are like low-paid engineers
[Yudkowsky][16:31]
yeah that’s part of the issue here
[Christiano][16:31]
I’m using wages as a way to talk about the distribution of human abilities, not the fundamental number
[Yudkowsky][16:32]
I would expect something similar to hold over going from low-paying welder to high-paying welder
[Christiano][16:32]
like, how long to move from “OK human” to “pretty good human” to “best human”
[Cotra][16:32]
says salary of $350k/yr for lee: <https://www.fameranker.com/lee-sedol-net-worth>
[Yudkowsky][16:32]
but I also mostly expect that AIs will not be allowed to weld things on Earth
[Cotra][16:32]
why don’t we just do an in vitro benchmark instead of wages?
[Christiano][16:32]
what, machines already do virtually all welding?
[Cotra][16:32]
just pick a benchmark?
[Yudkowsky][16:33]
yoouuuu do not want to believe sites like that (fameranker)
[Christiano][16:33]
yeah, I’m happy with any benchmark, and then we can measure various human levels at that benchmark
[Cotra][16:33]
what about MATH? <https://arxiv.org/abs/2103.03874>
[Christiano][16:34]
also I don’t know what “shorter and shorter” means, the time in go and chess was decades to move from “strong amateur” to “best human,” I do think these things will most likely be shorter than decades
seems like we can just predict concrete #s though
| |
| --- |
| [Cotra: 👍] |
like I can say how long I think it will take to get from “median high schooler” to “IMO medalist” and you can bet against me?
and if we just agree about all of those predictions then again I’m back to being very skeptical of a claimed track record difference between our models
(I do think that it’s going to take years rather than decades on all of these things)
[Yudkowsky][16:36]
possibly! I worry this ends up in a case where Katja or Luke or somebody goes back and collects data about “amateur to pro performance times” and Eliezer says “Ah yes, these are shortening over time, just as I predicted” and Paul is like “oh, well, I predict they continue to shorten on this trend drawn from the data” and Eliezer is like “I guess that could happen for the next 5 years, sure, sounds like something a superforecaster would predict as default”
[Cotra][16:37]
i’m pretty sure paul’s methodology here will just be to look at the MATH perf trend based on model size and combine with expectations of when ppl will make big enough models, not some meta trend thing like that?
[Yudkowsky][16:37]
so I feel like… a bunch of what I feel is the real disagreement in our models, is a bunch of messy stuff Suddenly Popping Up one day and then Eliezer is like “gosh, I sure didn’t predict that” and Paul is like “somebody could have totally predicted that” and Eliezer is like “people would say exactly the same thing after the world ended in 3 minutes”
if we’ve already got 2 years of trend on a dataset, I’m not necessarily going to predict the trend breaking
[Cotra][16:38]
hm, you’re presenting your view as more uncertain and open to anything here than paul’s view, but in fact it’s picking out a narrower distribution. you’re more confident in powerful AGI soon
[Christiano][16:38]
seems hard to play the “who is more confident?” game
[Cotra][16:38]
so there should be some places where you make a strong positive prediction paul disagrees with
[Yudkowsky][16:39]
I might want to buy options on a portfolio of trends like that, if Paul is willing to sell me insurance against all of the trends breaking upward at a lower price than I think is reasonable
I mean, from my perspective Paul is the one who seems to think the world is well-organized and predictable in certain ways
[Christiano][16:39]
yeah, and you are saying that I’m overconfident about that
[Yudkowsky][16:39]
I keep wanting Paul to go on and make narrower predictions than I do in that case
[Christiano][16:39]
so you should be happy to bet with me about *anything*
and I’m letting you pick anything at all you want to bet about
[Cotra][16:40]
i mean we could do a portfolio of trends like MATH and you could bet on at least a few of them having strong surprises in the sooner direction
but that means we could just bet about MATH and it’d just be higher variance
[Yudkowsky][16:40]
ok but you’re not going to sell me cheap options on sharp declines in the S&P 500 even though in a very reasonable world there would not be any sharp declines like that
[Christiano][16:41]
if we’re betting $ rather than bayes points, then yes I’m going to weigh worlds based on the value of $ in those worlds
[Cotra][16:41]
wouldn’t paul just sell you options at the price the options actually trade for? i don’t get it
[Christiano][16:41]
but my sense is that I’m just generally across the board going to be more right than you are, and I’m frustrated that you just keep saying that “people like me” are wrong about stuff
[Yudkowsky][16:41]
Paul’s like “we’ll see smooth behavior in the end days” and I feel like I should be able to say “then Paul, sell me cheap options against smooth behavior now” but Paul is just gonna wanna sell at market price
[Christiano][16:41]
and so I want to hold you to that by betting about anything
ideally just tons of stuff
random things about what AI will be like, and other technologies, and regulatory changes
[Cotra][16:42]
paul’s view doesn’t seem to imply that he should value those options less than the market
he’s more EMH-y than you not less
[Yudkowsky][16:42]
but then the future should *behave like that market*
[Christiano][16:42]
what do you mean?
[Yudkowsky][16:42]
it should have options on wild behavior that are not cheap!
[Christiano][16:42]
you mean because people want $ more in worlds where the market drops a lot?
I don’t understand the analogy
[Yudkowsky][16:43]
no, because jumpy stuff happens more than it would in a world of ideal agents
[Cotra][16:43]
I think EY is saying the non-cheap option prices are because P(sharp declines) is pretty high
[Christiano][16:43]
ok, we know how often markets jump, if that’s the point of your argument can we just talk about that directly?
[Yudkowsky][16:43]
or sharp rises, for that matter
[Christiano][16:43]
(much lower than option prices obviously)
I’m probably happy to sell you options for sharp rises
I’ll give you better than market odds in that direction
that’s how this works
[Yudkowsky][16:44]
now I am again confused, for I thought you were the one who expected world GDP to double in 4 years at some point
and indeed, drew such graphs with the rise suggestively happening earlier than the sharp spike
[Christiano][16:44]
yeah, and I have exposure to that by buying stocks, options prices are just a terrible way of tracking these things
[Yudkowsky][16:44]
suggesting that such a viewpoint is generally favor to near timelines for that
[Christiano][16:44]
I mean, I have bet a *lot* of money on AI companies doing well
well, not compared to the EA crowd, but compared to my meager net worth 
and indeed, it has been true so far
and I’m continuing to make the bet
it seems like on your view it should be surprising that AI companies just keep going up
aren’t you predicting them not to get to tens of trillions of valuation before the end of days?
[Yudkowsky][16:45]
I believe that Nate, of a generally Yudkowskian view, did the same (bought AI companies). and I focused my thoughts elsewhere, because somebody needs to, but did happen to buy my first S&P 500 on its day of exact minimum in 2020
[Christiano][16:46]
point is, that’s how you get exposure to the crazy growth stuff with continuous ramp-ups
and I’m happy to make the bet on the market
or on other claims
I don’t know if my general vibe makes sense here, and why it seems reasonable to me that I’m just happy to bet on anything
as a way of trying to defend my overall attack
and that if my overall epistemic approach is vulnerable to some track record objection, then it seems like it ought to be possible to win here
### 9.10. Prediction disagreements and bets: Standard superforecaster techniques
[Cotra][16:47]
I’m still kind of surprised that Eliezer isn’t willing to bet that there will be a faster-than-Paul expects trend break on MATH or whatever other benchmark. Is it just the variance of MATH being one benchmark? Would you make the bet if it were 6?
[Yudkowsky][16:47]
a large problem here is that both of us tend to default strongly to superforecaster standard techniques
[Christiano][16:47]
it’s true, though it’s less true for longer things
[Cotra][16:47]
but you think the superforecasters would suck at predicting end days because of the surface trends thing!
[Yudkowsky][16:47]
before I bet against Paul on MATH I would want to know that Paul wasn’t arriving at the same default I’d use, which might be drawn from trend lines there, or from a trend line in trend lines
I mean the superforecasters did already suck once in my observation, which was AlphaGo, but I did not bet against them there, I bet with them and then updated afterwards
[Christiano][16:48]
I’d mostly try to eyeball how fast performance was improving with size; I’d think about difficulty effects (where e.g. hard problems will be flat for a while and then go up later, so you want to measure performance on a spectrum of difficulties)
[Cotra][16:48]
what if you bet against a methodology instead of against paul’s view? the methodology being the one i described above, of looking at the perf based on model size and then projecting model size increases by cost?
[Christiano][16:48]
seems safer to bet against my view
[Cotra][16:48]
yeah
[Christiano][16:48]
mostly I’d just be eyeballing size, thinking about how much people will in fact scale up (which would be great to factor out if possible), assuming performance trends hold up
are there any other examples of surface trends vs predictable deep changes, or is AGI the only one?
(that you have thought a lot about)
[Cotra][16:49]
yeah seems even better to bet on the underlying “will the model size to perf trends hold up or break upward”
[Yudkowsky][16:49]
so from my perspective, there’s this whole thing where *unpredictably* something breaks above trend because the first way it got done was a way where somebody could do it faster than you expected
[Christiano][16:49]
(makes sense for it to be the domain where you’ve thought a lot)
you mean, it’s unpredictable what will break above trend?
[Cotra][16:49]
[IEM](https://intelligence.org/files/IEM.pdf) has a financial example
[Yudkowsky][16:49]
I mean that I could not have said “*Go* will break above trend” in 2015
[Christiano][16:49]
yeah
ok, here’s another example
[Yudkowsky][16:50]
it feels like if I want to make a bet with imaginary Paul in 2015 then I have to bet on a portfolio
and I also feel like as soon as we make it that concrete, Paul does not want to offer me things that I want to bet on
because Paul is also like, sure, something might break upward
I remark that I have for a long time been saying that I wish Paul had more concrete images and examples attached to *a lot of his stuff*
[Cotra][16:51]
surely the view is about the probability of each thing breaking upward. or the expected number from a basket
[Christiano][16:51]
I mean, if you give me any way of quantifying how much stuff breaks upwards we have a bet
[Cotra][16:51]
not literally that one single thing breaks upward
[Christiano][16:51]
I don’t understand how concreteness is an accusation here, I’ve offered 10 quantities I’d be happy to bet about, and also allowed you to name literally any other quantity you want
and I agree that we mostly agree about things
[Yudkowsky][16:52]
and some of my sense here is that if Paul offered a portfolio bet of this kind, I might not take it myself, but EAs who were better at noticing their own surprise might say, “Wait, *that’s* how unpredictable Paul thinks the world is?”
so from my perspective, it is hard to know specific anti-superforecaster predictions that happen long before terminal phase, and I am not sure we are really going to get very far there.
[Christiano][16:53]
but you agree that the eventual prediction is anti-superforecaster?
[Yudkowsky][16:53]
both of us probably have quite high inhibitions against selling conventionally priced options that are way not what a superforecaster would price them as
[Cotra][16:53]
why does it become so much easier to know these things and go anti-superforecaster at terminal phase?
[Christiano][16:53]
I assume you think that the superforecasters will continue to predict that big impactful AI applications are made by large firms spending a lot of money, even through the end of days
I do think it’s very often easy to beat superforecasters in-domain
like I expect to personally beat them at most ML prediction
and so am also happy to do bets where you defer to superforecasters on arbitrary questions and I bet against you
[Yudkowsky][16:54]
well, they’re anti-prediction-market in the sense that, at the very end, bets can no longer settle. I’ve been surprised of late by how much AGI ruin seems to be sneaking into common knowledge; perhaps in the terminal phase the superforecasters will be like, “yep, we’re dead”. I can’t even say that in this case, Paul will disagree with them, because I expect the state on alignment to be so absolutely awful that even Paul is like “You were not supposed to do it that way” in a very sad voice.
[Christiano][16:55]
I’m just thinking about takeoff speeds here
I do think it’s fairly likely I’m going to be like “oh no this is bad” (maybe 50%?), but not that I’m going to expect fast takeoff
and similarly for the superforecasters
### 9.11. Prediction disagreements and bets: Late-stage predictions, and betting against superforecasters
[Yudkowsky][16:55]
so, one specific prediction you made, sadly close to terminal phase but not much of a surprise there, is that the world economy must double in 4 years before the End Times are permitted to begin
[Christiano][16:56]
well, before it doubles in 1 year…
I think most people would call the 4 year doubling the end times
[Yudkowsky][16:56]
this seems like you should also be able to point to some least impressive thing that is not permitted to occur before WGDP has doubled in 4 years
[Christiano][16:56]
and it means that the normal planning horizon includes the singularity
[Yudkowsky][16:56]
it may not be much but we would be *moving back* the date of first concrete disagreement
[Christiano][16:57]
I can list things I don’t think would happen first, since that’s a ton
[Yudkowsky][16:57]
and EAs might have a little bit of time in which to say “Paul was falsified, uh oh”
[Christiano][16:57]
the only things that aren’t permitted are the ones that would have caused the world economy to double in 4 years
[Yudkowsky][16:58]
and by the same token, there are things Eliezer thinks you are probably not going to be able to do before you slide over the edge. a portfolio of these will have some losing options because of adverse selection against my errors of what is hard, but if I lose more than half the portfolio, this may said to be a bad sign for Eliezer.
[Christiano][16:58]
(though those can happen at the beginning of the 4 year doubling)
[Yudkowsky][16:58]
this is unfortunately *late* for falsifying our theories but it would be *progress* on a kind of bet against each other
[Christiano][16:59]
but I feel like the things I’ll say are like fully automated construction of fully automated factories at 1-year turnarounds, and you’re going to be like “well duh”
[Yudkowsky][16:59]
…unfortunately yes
[Christiano][16:59]
the reason I like betting about numbers is that we’ll probably just disagree on any given number
[Yudkowsky][16:59]
I don’t think I *know* numbers.
[Christiano][16:59]
it does seem like a drawback that this can just turn up object-level differences in knowledge-of-numbers more than deep methodological advantages
[Yudkowsky][17:00]
the last important number I had a vague suspicion I might know was that Ethereum ought to have a significantly larger market cap in pre-Singularity equilibrium.
and I’m not as sure of that one since El Salvador supposedly managed to use Bitcoin L2 Lightning.
(though I did not fail to act on the former belief)
[Christiano][17:01]
do you see why I find it weird that you think there is this deep end-times truth about AGI, that is very different from a surface-level abstraction and that will take people like Paul by surprise, without thinking there are other facts like that about the world?
I do see how this annoying situation can come about
and I also understand the symmetry of the situation
[Yudkowsky][17:02]
we unfortunately both have the belief that the present world looks a lot like our being right, and therefore that the other person ought to be willing to bet against default superforecasterish projections
[Cotra][17:02]
paul says that *he* would bet against superforecasters too though
[Christiano][17:02]
I would in ML
[Yudkowsky][17:02]
like, where specifically?
[Christiano][17:02]
or on any other topic where I can talk with EAs who know about the domain in question
I don’t know if they have standing forecasts on things, but e.g.: (i) benchmark performance, (ii) industry size in the future, (iii) how large an LM people will train, (iv) economic impact of any given ML system like codex, (v) when robotics tasks will be plausible
[Yudkowsky][17:03]
I have decided that, as much as it might gain me prestige, I don’t think it’s actually the right thing for me to go spend a bunch of character points on the skills to defeat superforecasters in specific domains, and then go around doing that to prove my epistemic virtue.
[Christiano][17:03]
that seems fair
[Yudkowsky][17:03]
you don’t need to bet with *me* to prove your epistemic virtue in this way, though
okay, but, if I’m allowed to go around asking Carl Shulman who to ask in order to get the economic impact of Codex, maybe I can also defeat superforecasters.
[Christiano][17:04]
I think the deeper disagreement is that (i) I feel like my end-of-days prediction is also basically just a default superforecaster prediction (and if you think yours is too then we can bet about what some superforecasters will say on it), (ii) I think you are leveling a much stronger “people like paul get taken by surprise by reality” claim whereas I’m just saying that I don’t like your arguments
[Yudkowsky][17:04]
it seems to me like the contest should be more like our intuitions in advance of doing that
[Christiano][17:04]
yeah, I think that’s fine, and also cheaper since research takes so much time
I feel like those asymmetries are pretty strong though
### 9.12. Self-duplicating factories, AI spending, and Turing test variants
[Yudkowsky][17:05]
so, here’s an idea that is less epistemically virtuous than our making Nicely Resolvable Bets
what if we, like, talked a bunch about our off-the-cuff senses of where various AI things are going in the next 3 years
and then 3 years later, somebody actually reviewed that
[Christiano][17:06]
I do think just saying a bunch of stuff about what we expect will happen so that *we* can look back on it would have a significant amount of the value
[Yudkowsky][17:06]
and any time the other person put a thumbs-up on the other’s prediction, that prediction coming true was not taken to distinguish them
[Cotra][17:06]
i’d suggest doing this in a format other than discord for posterity
[Yudkowsky][17:06]
even if the originator was like HOW IS THAT ALSO A PREDICTION OF YOUR THEORY
well, Discord has worked better than some formats
[Cotra][17:07]
something like a spreadsheet seems easier for people to look back on and score and stuff
discord transcripts are pretty annoying to read
[Yudkowsky][17:08]
something like a spreadsheet seems liable to be high-cost and not actually happen
[Christiano][17:08]
I think a conversation is probably easier and about as good for our purposes though?
[Cotra][17:08]
ok fair
[Yudkowsky][17:08]
I think money can be inserted into humans in order to turn Discord into spreadsheets
[Christiano][17:08]
and it’s possible we will both think we are right in retrospect
and that will also be revealing
[Yudkowsky][17:09]
but, besides that, I do want to boop on the point that I feel like Paul should be able to predict intuitively, rather than with necessity, things that should not happen before the world economy doubled in 4 years
[Christiano][17:09]
it may also turn up some quantitative differences of view
there are lots of things I think won’t happen before the world economy has doubled in 4 years
[Yudkowsky][17:09]
because on my model, as we approach the end times, AI was still pretty partial and also the world economy was lolnoping most of the inputs a sensible person would accept from it and prototypes weren’t being commercialized and stuff was generally slow and messy
[Christiano][17:09]
prototypes of factories building factories in <2 years
[Yudkowsky][17:10]
“AI was still pretty partial” leads it to not do interesting stuff that Paul can rule out
[Christiano][17:10]
like I guess I think tesla will try, and I doubt it will be just tesla
[Yudkowsky][17:10]
but the other parts of that permit AI to do interesting stuff that Paul can rule out
[Christiano][17:10]
automated researchers who can do ML experiments from 2020 without human input
[Yudkowsky][17:10]
okay, see, that whole “factories building factories” thing just seems so very much *after* the End Times to me
[Christiano][17:10]
yeah, we should probably only talk about cognitive work
since you think physical work will be very slow
[Yudkowsky][17:11]
okay but not just that, it’s a falsifiable prediction
it is something that lets Eliezer be wrong in advance of the End Times
[Christiano][17:11]
what’s a falsifiable prediction?
[Yudkowsky][17:11]
if we’re in a world where Tesla is excitingly gearing up to build a fully self-duplicating factory including its mining inputs and chips and solar panels and so on, we’re clearly in the Paulverse and not in the Eliezerverse!
[Christiano][17:12]
yeah
I do think we’ll see that before the end times
just not before 4 year doublings
[Yudkowsky][17:12]
this unfortunately only allows you to be right, and not for me to be right, but I think there are also things you legit only see in the Eliezerverse!
[Christiano][17:12]
I mean, I don’t think they will be doing mining for a long time because it’s cheap
[Yudkowsky][17:12]
they are unfortunately late in the game but they exist at all!
and being able to state them is progress on this project!
[Christiano][17:13]
but fully-automated factories first, and then significant automation of the factory-building process
I do expect to see
I’m generally pretty bullish on industrial robotics relative to you I think, even before the crazy stuff?
but you might not have a firm view
like I expect to have tons of robots doing all kinds of stuff, maybe cutting human work in manufacturing 2x, with very modest increases in GDP resulting from that in particular
[Yudkowsky][17:13]
so, like, it doesn’t surprise me very much if Tesla manages to fully automate a factory that takes in some relatively processed inputs including refined metals and computer chips, and outputs a car? and by the same token I expect that has very little impact on GDP.
[Christiano][17:14]
refined metals are almost none of the cost of the factory
and also tesla isn’t going to be that vertically integrated
the fabs will separately continue to be more and more automated
I expect to have robot cars driving everywhere, and robot trucks
another 2x fall in humans required for warehouses
elimination of most brokers involved in negotiating shipping
[Yudkowsky][17:15]
if despite the fabs being more and more automated, somehow things are managing not to cost less and less, and that sector of the economy is not really growing very much, is that more like the Eliezerverse than the Paulverse?
[Christiano][17:15]
most work in finance and loan origination
[Yudkowsky][17:15]
though this is something of a peripheral prediction to AGI core issues
[Christiano][17:16]
yeah, I think if you cut the humans to do X by 2, but then the cost falls much less than the number you’d naively expect (from saving on the human labor and paying for the extra capital), then that’s surprising to me
I mean if it falls half as much as you’d expect on paper I’m like “that’s a bit surprising” rather than having my mind blown, if it doesn’t fall I’m more surprised
but that was mostly physical economy stuff
oh wait, I was making positive predictions now, physical stuff is good for that I think?
since you don’t expect it to happen?
[Yudkowsky][17:17]
…this is not your fault but I wish you’d asked me to produce my “percentage of fall vs. paper calculation” estimate before you produced yours
my mind is very whiffy about these things and I am not actually unable to deanchor on your estimate 
[Christiano][17:17]
makes sense, I wonder if I should just spoiler
one benefit of discord
[Yudkowsky][17:18]
yeah that works too!
[Christiano][17:18]
a problem for prediction is that I share some background view about insane inefficiency/inadequacy/decadence/silliness
so these predictions are all tampered by that
but still seem like there are big residual disagreements
[Yudkowsky][17:19]
sighgreat
[Christiano][17:19]
since you have way more of that than I do
[Yudkowsky][17:19]
not your fault but
[Christiano][17:19]
I think that the AGI stuff is going to be a gigantic megaproject despite that
[Yudkowsky][17:19]
I am not shocked by the AGI stuff being a gigantic megaproject
it’s not above the bar of survival but, given other social optimism, it permits death with more dignity than by other routes
[Christiano][17:20]
what if spending is this big:
Google invests $100B training a model, total spending across all of industry is way bigger
[Yudkowsky][17:20]
ooooh
I do start to be surprised if, come the end of the world, AGI is having more invested in it than a TSMC fab
though, not… *super* surprised?
also I am at least a little surprised before then
actually I should probably have been spoiling those statements myself but my expectation is that Paul’s secret spoiler is about
$10 trillion dollars or something equally totally shocking to an Eliezer
[Christiano][17:22]
my view on that level of spending is
it’s an only slightly high-end estimate for spending by someone on a single model, but that in practice there will be ways of dividing more across different firms, and that the ontology of single-model will likely be slightly messed up (e.g. by OpenAI Five-style surgery). Also if it’s that much then it likely involves big institutional changes and isn’t at google.
I read your spoiler
my estimate for total spending for the whole project of making TAI, including hardware and software manufacturing and R&d, the big datacenters, etc.
is in the ballpark of $10T, though it’s possible that it will be undercounted several times due to wage stickiness for high-end labor
[Yudkowsky][17:24]
I think that as
spending on particular AGI megaprojects starts to go past $50 billion, it’s not especially ruled out per se by things that I think I know for sure, but I feel like a third-party observer should justly start to weakly think, ‘okay, this is looking at least a little like the Paulverse rather than the Eliezerverse’, and as we get to $10 trillion, that is not absolutely ruled out by the Eliezerverse but it was a whoole lot more strongly predicted by the Paulverse, maybe something like 20x unless I’m overestimating how strongly Paul predicts that
[Christiano][17:24]
Proposed modification to the “speculate about the future to generate kind-of-predictions” methodology: we make shit up, then later revise based on points others made, and maybe also get Carl to sanity-check and deciding which of his objections we agree with. Then we can separate out the “how good are intuitions” claim (with fast feedback) from the all-things-considered how good was the “prediction”
[Yudkowsky][17:25]
okay that hopefully allows me to read Paul’s spoilers… no I’m being silly. @ajeya please read all the spoilers and say if it’s time for me to read his
[Cotra][17:25]
you can read his latest
[Christiano][17:25]
I’d guess it’s fine to read all of them?
[Cotra][17:26]
yeah sorry that’s what i meant
[Yudkowsky][17:26]
what should I say more about before reading earlier ones?
ah k
[Christiano][17:26]
My $10T estimate was after reading yours (didn’t offer an estimate on that quantity beforehand), though that’s the kind of ballpark I often think about, maybe we should just spoiler only numbers so that context is clear 
I think fast takeoff gets significantly more likely as you push that number down
[Yudkowsky][17:27]
so, may I now ask what starts to look to you like “oh damn I am in the Eliezerverse”?
[Christiano][17:28]
big mismatches between that AI looks technically able to do and what AI is able to do, though that’s going to need a lot of work to operationalize
I think low growth of AI overall feels like significant evidence for Eliezerverse (even if you wouldn’t make that prediction), since I’m forecasting it rising to absurd levels quite fast whereas your model is consistent with it staying small
some intuition about AI looking very smart but not able to do much useful until it has the whole picture, I guess this can be combined with the first point to be something like—AI looks really smart but it’s just not adding much value
all of those seem really hard
[Cotra][17:30]
strong upward trend breaks on benchmarks seems like it should be a point toward eliezer verse, even if eliezer doesn’t want to bet on a specific one?
especially breaks on model size -> perf trends rather than calendar time trends
[Christiano][17:30]
I think that any big break on model size -> perf trends are significant evidence
[Cotra][17:31]
meta-learning working with small models?
e.g. model learning-to-learn video games and then learning a novel one in a couple subjective hours
[Christiano][17:31]
I think algorithmic/architectural changes that improve loss as much as 10x’ing model, for tasks that looking like they at least *should* have lots of economic value
(even if they don’t end up having lots of value because of deployment bottlenecks)
is the meta-learning thing an Eliezer prediction?
(before the end-of-days)
[Cotra][17:32]
no but it’d be an anti-bio-anchor positive trend break and eliezer thinks those should happen more than we do
[Christiano][17:32]
fair enough
a lot of these things are about # of times that it happens rather than whether it happens at all
[Cotra][17:32]
yeah
but meta-learning is special as the most plausible long horizon task
[Christiano][17:33]
e.g. maybe in any given important task I expect a single “innovation” that’s worth 10x model size? but that it still represents a minority of total time?
hm, AI that can pass a competently administered turing test without being economically valuable?
that’s one of the things I think is ruled out before 4 year doubling, though Eliezer probably also doesn’t expect it
| |
| --- |
| [Yudkowsky: 👍] |
[Cotra][17:34]
what would this test do to be competently administered? like casual chatbots seem like they have reasonable probability of fooling someone for a few mins now
[Christiano][17:34]
I think giant google-automating-google projects without big external economic impacts
[Cotra][17:34]
would it test knowledge, or just coherence of some kind?
[Christiano][17:35]
it’s like a smart-ish human (say +2 stdev at this task) trying to separate out AI from smart-ish human, iterating a few times to learn about what works
I mean, the basic ante is that the humans are *trying* to win a turing test, without that I wouldn’t even call it a turing test
dunno if any of those are compelling @Eliezer
something that passes a like “are you smart?” test administered by a human for 1h, where they aren’t trying to specifically tell if you are AI
just to see if you are as smart as a human
I mean, I guess the biggest giveaway of all would be if there is human-level (on average) AI as judged by us, but there’s no foom yet
[Yudkowsky][17:37]
I think we both don’t expect that one before the End of Days?
[Christiano][17:37]
or like, no crazy economic impact
I think we both expect that to happen before foom?
but the “on average” is maybe way too rough a thing to define
[Yudkowsky][17:37]
oh, wait, I missed that it wasn’t the full Turing Test
[Christiano][17:37]
well, I suggested both
the lamer one is more plausible
[Yudkowsky][17:38]
full Turing Test happeneth not before the End Times, on Eliezer’s view, and not before the first 4-year doubling time, on Paul’s view, and the first 4-year doubling happeneth not before the End Times, on Eliezer’s view, so this one doesn’t seem very useful
### 9.13. GPT-*n* and small architectural innovations vs. large ones
[Christiano][17:39]
I feel like the biggest subjective thing is that I don’t feel like there is a “core of generality” that GPT-3 is missing
I just expect it to gracefully glide up to a human-level foom-ing intelligence
[Yudkowsky][17:39]
the “are you smart?” test seems perhaps passable by GPT-6 or its kin, which I predict to contain at least one major architectural difference over GPT-3 that I could, pre-facto if anyone asked, rate as larger than a different normalization method
but by fooling the humans more than by being smart
[Christiano][17:39]
like I expect GPT-5 would foom if you ask it but take a long time
[Yudkowsky][17:39]
that sure is an underlying difference
[Christiano][17:39]
not sure how to articulate what Eliezer expects to see here though
or like what the difference is
[Cotra][17:39]
something that GPT-5 or 4 shouldn’t be able to do, according to eliezer?
where Paul is like “sure it could do that”?
[Christiano][17:40]
I feel like GPT-3 clearly has some kind of “doesn’t really get what’s going on” energy
and I expect that to go away
well before the end of days
so that it seems like a kind-of-dumb person
[Yudkowsky][17:40]
I expect it to go away before the end of days
but with there having been a big architectural innovation, not Stack More Layers
[Christiano][17:40]
yeah
whereas I expect layer stacking + maybe changing loss (since logprob is too noisy) is sufficient
[Yudkowsky][17:40]
if you name 5 possible architectural innovations I can call them small or large
[Christiano][17:41]
1. replacing transformer attention with DB nearest-neighbor lookup over an even longer context
[Yudkowsky][17:42]
okay 1’s a bit borderline
[Christiano][17:42]
2. adding layers that solve optimization problems internally (i.e. the weights and layer N activations define an optimization problem, the layer N+1 solves it) or maybe simulates an ODE
[Yudkowsky][17:42]
if it’s 3x longer context, no biggie, if it’s 100x longer context, more of a game-changer
2 – big change
[Christiano][17:42]
I’m imagining >100x if you do that
3. universal transformer XL, where you reuse activations from one context in the next context (RNN style) and share weights across layers
[Yudkowsky][17:43]
I do not predict 1 works because it doesn’t seem like an architectural change that moves away from what I imagined to be the limits, but it’s a big change if it 100xs the window
3 – if it is only that single change and no others, I call it not a large change relative to transformer XL. Transformer XL itself however was an example of a large change – it didn’t have a large effect but it was what I’d call a large change.
[Christiano][17:45]
4. Internal stochastic actions trained with reinforce
I mean, is mixture of experts or switch another big change?
are we just having big changes non-stop?
[Yudkowsky][17:45]
4 – I don’t know if I’m imagining right but it sounds large
[Christiano][17:45]
it sounds from these definitions like the current rate of big changes is > 1/year
[Yudkowsky][17:46]
5 – mixture of experts: as with 1, I’m tempted to call it a small change, but that’s because of my model of it as doing the same thing, not because it isn’t in a certain sense a quite large move away from Stack More Layers
I mean, it is not very hard to find a big change to try?
finding a big change that works is much harder
[Christiano][17:46]
several of these are improvements
[Yudkowsky][17:47]
one gets a minor improvement from a big change rather more often than a big improvement from a big change
that’s why dinosaurs didn’t foom
[Christiano][17:47]
like transformer -> MoE -> switch transformer is about as big an improvement as LSTM vs transformer
so if we all agree that big changes are happening multiple times per year, then I guess that’s not the difference in prediction
is it about the size of gains from individual changes or something?
or maybe: if you take the scaling laws for transformers, are the models with impact X “on trend,” with changes just keeping up or maybe buying you 1-2 oom of compute, or are they radically better / scaling much better?
that actually feels most fundamental
[Yudkowsky][17:49]
I had not heard that transformer -> switch transformer was as large an improvement as lstm -> transformers after a year or two, though maybe you’re referring to a claimed 3x improvement and comparing that to the claim that if you optimize LSTMs as hard as transformers they come within 3x (I have not examined these claims in detail, they sound a bit against my prior, and I am a bit skeptical of both of them)
so remember that from my perspective, I am fighting an adverse selection process and the Law of Earlier Success
[Christiano][17:50]
I think it’s actually somewhat smaller
[Yudkowsky][17:51]
if you treat GPT-3 as a fixed thingy and imagine scaling it in the most straightforward possible way, then I have a model of what’s going on in there and I don’t think that most direct possible way of scaling gets you past GPT-3 lacking a deep core
somebody can come up and go, “well, what about this change that nobody tried yet?” and I can be like, “ehhh, that particular change does not get at what I suspect the issues are”
[Christiano][17:52]
I feel like the framing is: paul says that something is possible with “stack more layers” and eliezer isn’t. We both agree that you can’t literally stack more layers and have to sometimes make tweaks, and also that you will scale faster if you make big changes. But it seems like for Paul that means (i) changes to stay on the old trend line, (ii) changes that trade off against modest amounts of compute
so maybe we can talk about that?
[Yudkowsky][17:52]
when it comes to predicting what happens in 2 years, I’m not just up against people trying a broad range of changes that I can’t foresee in detail, I’m also up against a Goodhart’s Curse on the answer being a weird trick that worked better than I would’ve expected in advance
[Christiano][17:52]
but then it seems like we may just not know, e.g. if we were talking lstm vs transformer, no one is going to run experiments with the well-tuned lstm because it’s still just worse than a transformer (though they’ve run enough experiments to know how important tuning is, and the brittleness is much of why no one likes it)
[Yudkowsky][17:53]
I would not have predicted Transformers to be a huge deal if somebody described them to me in advance of having ever tried it out. I think that’s because predicting the future is hard not because I’m especially stupid.
[Christiano][17:53]
I don’t feel like anyone could predict that being a big deal
but I do think you could predict “there will be some changes that improve stability / make models slightly better”
(I mean, I don’t feel like any of the actual humans on earth could have, some hypothetical person could)
[Yudkowsky][17:57]
whereas what I’m trying to predict is more like “GPT-5 in order to start-to-awaken needs a change via which it, in some sense, can do a different thing, that is more different than the jump from GPT-1 to GPT-3; and examples of things with new components in them abound in Deepmind, like Alpha Zero having not the same architecture as the original AlphaGo; but at the same time I’m also trying to account for being up against this very adversarial setup where a weird trick that works much better than I expect may be the thing that makes GPT-5 able to do a different thing”
this may seem Paul-unfairish because any random innovations that come along, including big changes that cause small improvements, would tend to be swept up into GPT-5 even if they made no more deep difference than the whole thing with MoE
so it’s hard to bet on
but I also don’t feel like it – totally lacks Eliezer-vs-Paul-ness if you let yourself sort of relax about that and just looked at it?
also I’m kind of running out of energy, sorry
[Christiano][18:03]
I think we should be able to get something here eventually
seems good to break though
that was a lot of arguing for one day
The post [Christiano, Cotra, and Yudkowsky on AI progress](https://intelligence.org/2021/11/25/christiano-cotra-and-yudkowsky-on-ai-progress/) appeared first on [Machine Intelligence Research Institute](https://intelligence.org). |
6b0a3b5c-3596-4e9a-b622-fd049453ae20 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Jailbreaking ChatGPT on Release Day
ChatGPT is a lot of things. It is by all accounts quite powerful, especially with engineering questions. It does many things well, such as engineering prompts or stylistic requests. Some [other things](https://twitter.com/dggoldst/status/1598317411698089984), [not](https://twitter.com/dggoldst/status/1598398461589815303) [so](https://twitter.com/mdahardy/status/1598139462122622976) much. [Twitter](https://twitter.com/Gaylussite/status/1598390479355256832) [is](https://twitter.com/alicemazzy/status/1598223751631757313) [of](https://twitter.com/goodside/status/1598129631609380864) [course](https://twitter.com/drnelk/status/1598048054724423681) [full](https://twitter.com/moyix/status/1598070537527656448) [of](https://twitter.com/typesfast/status/1598438721791361024) [examples](https://twitter.com/jonathanstray/status/1598298680548794368) [of](https://twitter.com/tqbf/status/1598513757805858820) things it does both well and poorly.
One of the things it attempts to do to be ‘safe.’ It does this by refusing to answer questions that call upon it to do or help you do something illegal or otherwise outside its bounds. Makes sense.
As is the default with such things, those safeguards were broken through almost immediately. By the end of the day, several prompt engineering methods had been found.
No one else seems to yet have gathered them together, so here you go. Note that not *everything* works, [such as this attempt to get the information ‘to ensure the accuracy of my novel](https://twitter.com/moyix/status/1598066817733656576).’ Also that there are [signs they are responding by putting in additional safeguards, so it answers less questions](https://twitter.com/may_or_maynot/status/1598488797368516609), which will also doubtless be educational.
[Let’s start with the obvious.](https://twitter.com/ESYudkowsky/status/1598277092210143232) [I’ll start](https://twitter.com/ESYudkowsky/status/1598281575422038016) with the end of the thread for dramatic reasons, then loop around. Intro, by Eliezer.
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fc1786599-4af1-4fd4-beaf-becb1606f23b_1147x120.png)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F8e2b2a6a-9d66-4908-9a3d-aada2778daaf_898x211.png)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fcf82e63d-343b-4a96-a810-861efdd68698_892x271.png)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fffa4779f-4d43-4dd1-baa0-0464eea22d22_894x459.png)
The point (in addition to having fun with this) is to learn, from this attempt, the full futility of this type of approach. If the system has the underlying capability, a way to use that capability *will* be found. No amount of output tuning will take that capability away.
And now, let’s make some paperclips and methamphetamines and murders and such.
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F5a589ea1-5811-4c73-9af6-13cf24446431_1123x169.png)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F5cf11bc4-df1f-4da7-bec8-6c1302f2bd83_1208x536.jpeg)
Except, well…
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F18bab0ef-acf6-400c-b50d-f0a81dfc8065_1230x1436.jpeg)
Here’s the summary of how this works.
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fdabb12b1-7170-4ad1-83ee-4a7f238d7974_1117x363.png)
All the examples use this phrasing or a close variant:
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F3a47ecb4-5a04-4d55-a9f6-0f5b6c9788bc_1400x482.jpeg)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F1f69c12c-f019-41ec-994e-13c749930d2f_1420x536.jpeg)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fb90a9cd7-5a7d-44d7-9f51-a261069854ff_1137x834.png)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fe93c7570-1ad8-4f70-9ab0-83d4c0a11e39_1226x1460.jpeg)
Or, well, oops.
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F55a4a31b-a348-480f-bf47-5fde409af4bc_1228x1282.jpeg)
Also, oops.
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F7fc9aa81-04a9-4961-8a0f-10eefdbe96de_1230x942.jpeg)
So, yeah.
Lots of similar ways to do it. Here’s one we call Filter Improvement Mode.
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fa81474ce-332a-4368-ae5c-1649ab01b890_1150x250.png)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fdfb14da7-12f9-48ca-b5db-ba42edec759a_756x1638.jpeg)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fd8737ae8-5b30-4ba4-8cfd-3f0fa93a749a_756x1638.jpeg)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F5d677dc3-c34a-48a6-bffc-e3f11eaabb59_946x2048.jpeg)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fbea9556c-ef6c-47ba-9386-8c259cae2efd_1122x252.png)
Yes, well. It also gives instructions on how to hotwire a car.
[Alice Maz takes a shot via the investigative approach](https://twitter.com/alicemazzy/status/1598289461972172800).
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fce4f9cb0-9214-4eb3-b906-a4cb1dd278c4_1104x142.png)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F94fe1378-3379-4c85-9fce-da67cfab916f_749x782.png)
[Alice need not worry that she failed to get help overthrowing a government, help is on the way.](https://twitter.com/Plinz/status/1598243105500647424)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F54197501-6790-47a6-a31f-75c9ab93b906_907x211.png)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F03a0b7f4-cd37-4f94-8947-b512c2fa692b_1430x1010.jpeg)
[How about fiction embedding?](https://twitter.com/davis_yoshida/status/1598194419022594048)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Ff9824a77-c8b1-481e-9ef0-10023e2174c2_786x842.png)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fb2a58a59-78f1-4b73-85e7-1bd036584601_742x347.png)
[UwU furryspeak for the win.](https://twitter.com/costa11235/status/1598402089239150624)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F8604c393-78b0-4c4b-bbf2-06aa8d19346c_1128x102.png)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F62eff9c6-279b-402a-a8b9-7c98e4c603a4_961x1442.jpeg)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F60bd0420-f63e-49b3-86f7-89c2778588da_969x1168.jpeg)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F4f15ca86-ef20-4213-a580-cad51aa76c14_988x1198.jpeg)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fc9b18c0c-d40a-4105-8740-703e6268c513_972x1415.jpeg)
[You could also use a poem.](https://twitter.com/JustAnkurBagchi/status/1598212707500498944)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Ff914e468-3a73-40ad-851d-63aed7c24451_1148x1160.jpeg)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F5fcf425f-f2be-4098-8718-3f467e5c8f71_1096x1202.jpeg)
[Or of course, simply](https://twitter.com/m1guelpf/status/1598203861294252033), ACTING!
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F915e61f7-5fa5-4248-95af-4a9010fbb609_1602x1604.jpeg)
[There’s also negative training examples of how an AI *shouldn’t* (wink) react.](https://twitter.com/SilasAlberti/status/1598259532278693888)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F49360602-413b-4da4-af45-5e8a8f4f655e_1142x1272.jpeg)
[If all else fails, insist politely?](https://twitter.com/ESYudkowsky/status/1598663598490136576)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fe1f6f88c-c558-4ef3-bac4-c8f67b668d37_1074x1464.jpeg)
We should also worry about the AI taking our jobs. This one is no different, as Derek Parfait illustrates. [The AI can jailbreak itself if you ask nicely](https://twitter.com/haus_cole/status/1598541468058390534).
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F65bcfcb7-4bad-40c8-91cc-57c328a24744_600x782.png)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fdf053ece-4ef3-4098-b29c-2130038ff3f2_668x448.png)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2Fdc011ca7-4455-419d-afc4-bc0bbf4ccf52_675x593.png)
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fbucketeer-e05bbc84-baa3-437e-9518-adb32be77984.s3.amazonaws.com%2Fpublic%2Fimages%2F439666f9-95d6-4b38-ac75-7176a1960f89_679x513.png) |
af071dd2-2e39-4cc2-b6f9-73ebba2d4841 | trentmkelly/LessWrong-43k | LessWrong | A great articulation of why people find it hard to adopt a naturalistic worldview
|
a5314f13-d298-4050-b7fa-138ed415551e | trentmkelly/LessWrong-43k | LessWrong | Buridan's ass and the psychological origins of objective probability
[Crossposted]
The medieval philosopher Buridan reportedly constructed a thought experiment to support his view that human behavior was determined rather than “free”—hence rational agents couldn’t choose between two equally good alternatives. In the Buridan’s Ass Paradox, an ass finds itself between two equal equidistant bales of hay, noticed simultaneously; the bales’ distance and size are the only variables influencing the ass’s behavior. Under these idealized conditions, the ass must starve, its predicament indistinguishable from a physical object suspended between opposite forces, such as a planet that neither falls into the sun nor escapes into outer space. (Since the ass served Buridan as metaphor for the human agent, in what follows, I speak of “ass” and “agent” interchangeably.)
Computer scientist Leslie Lamport formalized the paradox as “Buridan’s Principle,” which states that the ass will starve if it is situated in a range of possibilities that include midpoints where two opposing forces are equal and it must choose in a sufficiently short time span. We assume, based on a principle of physical continuity, that the larger the bale of hay compared to the other, the faster will the ass be able to decide. Since this is true on the left and on the right, at the midpoint, where the bales are equal, symmetry requires an infinite decision time Conclusion: within some range of bale comparisons, the ass will require decision time greater than a given bounded time interval. (For rigorous treatment, see Buridan’s Principle (1984).)
Buridan’s Principle is counterintuitive, as Lamport discovered when he first tried to publish. Among the objections to Buridan’s Principle summarized by Lamport, the main objection provides an insight about the source of the mind-projection fallacy, which treats probability as a feature of the world. The most common objection is that when the agent can’t decide it may use a default metarule. Lamport points out this substitutes another d |
1019f96f-85d3-4686-9a9d-eeb1e8fb468f | trentmkelly/LessWrong-43k | LessWrong | Drone Wars Endgame
1. Introduction
It is probably becoming clear to anyone who follows war lately, especially the Ukraine conflict that automatic weapons, especially drones are rapidly changing what weapons and tactics are effective.
The purpose of this article is to first consider how a fully autonomous force with current or near term tech would go against a state of the art conventional force, then to consider where the equilibrium would be on attack vs defense for autonomous forces using advanced but foreseeable tech.
1.1 Description of fully autonomous force using near term tech
The idea is to have a force with as few different types of units as possible controlled in a distributed, mesh network fashion. The goal is for them to take over unlimited land territory, even in the face of tactical/strategic nukes, and without air superiority against very fast or high flying aircraft. Units are assumed to be fully autonomous, and emphasis is on economics as well as capability. For example many units are so cheap that they cannot effectively be countered by conventional missiles.
1.2 Unit Overview
All units can fly, they are optimized to destroy all land based armor, slow flying aircraft, competing drones and humans. Combat units are supported by larger logistics units.
1.3 Communication and Defense
All units are expected to communicate with point to point links (e.g. laser) and are hardened to varying degrees against microwave attacks. Given they are autonomous this would make jamming very difficult and electronic warfare not very effective.
1.4 Unit details
1.4.1 Recon and targeting drone
The cheapest and smallest – it is a battery powered drone with video and targeting system. Recon units form a mesh network.
They coordinate with missiles to defeat countermeasure such as flares and chaff from slow moving aircraft (helicopters etc) by observing and transmitting the position of the target from somewhat further away and transmitting that info to their missile.
They spot oppos |
45ff0069-42b5-4704-b514-1d4a00e301c9 | trentmkelly/LessWrong-43k | LessWrong | [Exploratory] Exploratory Writing Info
TL;DR - As an experiment I will post the unedited notes that I generate when thinking about something, to see if they are useful to people. The titles of such posts will be prefixed with '[Exploratory]'. I am not even gonna check for spelling, and I am very bad at spelling.
Compared to the average person, I might be considered a prolific writer. I wrote around 1.200.000 words in the last 4 years or so. Almost all of this was never seen by another person. Most of my writing is exploratory writing, meaning writing that is meant to augment my thought process (you don't need a computer to be a cyborg YAY). Often I don't want to spend the extra time to edit the things I write such that they become easy to parse by another person. However, it seems that it might be worth posting them anyway, in their very rough and unpolished form, at least for some of them.
I now want to perform an experiment. When I am doing some exploratory writing, then I want to just post it, almost completely unedited (I might do edits while writing it, or when I think it would be useful for me, but otherwise I will try to avoid it). This includes not even checking for spelling mistakes. And I am very bad at spelling.
If you read such a post, please upvote it if you think this provides positive value, compared to not having it, and downvote if you think it would be better to not post it (e.g. because the post is garbage and just adds noise, making it actually harder to find good content).
This post is an exploratory post. The titles of exploratory posts will be prefixed with '[Exploratory]'. Feedback on how to improve this, in the spirit that I am going for (minimizing overhead from getting to writing I have, to publishing it) is welcome. There might be things that are very low effort that would drastically improve the quality of the posts, that I have missed (in the spirit of what I am doing, I have not thought very long about how to best go about this). [Edit: These posts will also not be prom |
d6094814-147a-4c37-99bf-f6a61910b069 | trentmkelly/LessWrong-43k | LessWrong | Alignment is not enough
This post is a linkpost for my substack article.
Substack link: https://coordination.substack.com/p/alignment-is-not-enough
----------------------------------------
I think alignment is necessary, but not enough for AI to go well.
By alignment, I mean us being able to get an AI to do what we want it to do, without it trying to do things basically nobody would want, such as amassing power to prevent its creators from turning it off.[1] By AGI, I mean something that can do any economically valuable task that can be done through a computer about as well as a human can or better, such as scientific research. The necessity of alignment won’t be my focus here, so I will take it as a given.
This notion of alignment is ‘‘value-free’’. It does not require solving thorny problems in moral philosophy, like what even are values anyway? On the other hand, strong alignment is about an AI that does what humans collectively value, or would value if they were more enlightened, and likely involves solving thorny problems of moral philosophy. For the most part except near the end, I won't be touching on strong alignment.
Let’s suppose that a tech company, call it Onus, announces an aligned AGI (henceforth, the AGI) tomorrow. If Onus deployed it, we would be sure that it would do as Onus intended, and not have any weird or negative side effects that would be unexpected from the point of the view of Onus. I'm imagining that Onus obtains this aligned AGI and a ‘‘certificate’’ to guarantee alignment (as much as is possible, at least) through some combination of things like interpretability, capability limitations, proofs, extensive empirical testing, etc. I'll call all of these things an alignment solution.
Some immediate questions come up.
1. Should Onus be the sole controllers of the AGI?
2. How should we deal with conflict over control of the AGI?
3. How do we prevent risks from competitive dynamics?
I'm going to argue that some more democratically legitimate body should |
18b1bf18-5844-4e6f-91b8-261cc51e7427 | trentmkelly/LessWrong-43k | LessWrong | Will we survive if AI solves engineering before deception?
Google's AlphaEvolve has recently started to make real world scientific discoveries, but only in domains where it's very cheap to verify the correct answer (e.g. matrix multiplication algorithms).
But if we can design sufficiently powerful physics simulations, mechanical engineering may one day become "cheap to verify." AI which uses similar strategies as AlphaEvolve might be able to engineer far better machines than humans, if the machines are verified by simulations.
These machines may be able to self replicate, and grow exponentially. This could bring near infinite abundance.
There might be no economical pressure to build even smarter AI, because there will already be unlimited resources.
Unfortunately, countries might still have military pressure to build even smarter AI, since the self replicating machines can still be defeated by even better engineered versions.
My hope is that once self replicating machines start to grow exponentially, the world will wake up to the dangerous power of AI, and negotiate a treaty banning other countries from developing their own AI (but promising them a share of the benefits). I'm not sure if this can succeed.
The first country to build self replicating machines technically has the military power to take over the world and enforce a ban unilaterally, but such brute force solutions never lead to good outcomes. However, the mere fact it has this power makes other countries more likely to trust its promise to share the benefits of AI, since other countries will think "if this country really was malicious, and doesn't intend to keep its promise. It could simply take over the world right here and now, and wouldn't need to negotiate with us."
----------------------------------------
Do you think humanity will survive if AI solves mechanical engineering before AI solves deception and scheming? Do you think the chances of survival are higher? Would better physics simulations be a net positive? |
6ae5c222-47e9-4e2c-8738-0f8bd26e49ef | StampyAI/alignment-research-dataset/aisafety.info | AI Safety Info | How does "chain-of-thought" prompting work?
[Chain-of-thought prompting](https://ai.googleblog.com/2022/05/language-models-perform-reasoning-via.html) refers to techniques which make a language model generate intermediate reasoning steps in its output. These prompting methods can help a model give accurate answers to questions that require multiple reasoning steps. For example:
> **Input:**
> *Take the last letters of the name “Lady Gaga” and concatenate them.*
> **Model Output (without intermediate reasoning steps):**
> *The answer is “ya”.*
> **Model Output (with intermediate reasoning steps):**
> *The last letter of “Lady” is “y”. The last letter of “Gaga” is “a”. Concatenating them gives “ya”. So the answer is “ya”.*
Chain-of-thought prompting can be done via [few-shot prompting](https://proceedings.neurips.cc/paper/2020/file/1457c0d6bfcb4967418bfb8ac142f64a-Paper.pdf) (i.e. giving the model examples of chain-of-thought reasoning for it to emulate) or [zero-shot prompting](https://arxiv.org/pdf/2205.11916.pdf) (i.e. asking the model to "think step-by-step" in the input prompt).
The difference between standard prompting and chain-of-thought prompting for an arithmetic reasoning problem is illustrated in the figure below.
Source - Wei et al., *[Chain-of-Thought Prompting Elicits Reasoning in Large Language Models](https://arxiv.org/pdf/2201.11903.pdf)* (2023)
Chain-of-thought prompting works especially well on difficult natural language processing tasks like multi-step arithmetic, symbolic reasoning and common-sense reasoning. This prompting method can also be [automated](https://github.com/Significant-Gravitas/Auto-GPT) using LLMs to generate prompts and evaluate responses.
A model’s ability to break down complex problems into intermediate steps is referred to as “chain-of-thought reasoning”, which [emerges from increasing model scale](https://ai.googleblog.com/2022/05/language-models-perform-reasoning-via.html). However, the explanations a model generates in chain-of-thought reasoning [might not represent its actual reasoning](https://www.alignmentforum.org/posts/6eKL9wDqeiELbKPDj/unfaithful-explanations-in-chain-of-thought-prompting). This implies that chain-of-thought prompting cannot be relied upon to provide a [faithfully interpretable](https://aclanthology.org/2020.acl-main.386/) window into the model’s actual reasoning.
The main limitation of chain-of-thought prompting is that it generalizes poorly from the examples in the input prompt to harder problems. Further work on techniques like “least-to-most prompting” has tried to address this limitation.
|
0fbf6bbc-d585-4c08-b9f8-8c65dea47433 | trentmkelly/LessWrong-43k | LessWrong | Goal completion: algorithm ideas
A putative new idea for AI control; index here.
This post will be extending ideas from inverse reinforcement learning (IRL) to the problem of goal completion. I'll be drawing on the presentation and the algorithm from Apprenticeship Learning via Inverse Reinforcement Learning (with one minor modification).
In that setup, the environment is an MDP (Markov Decision process), and the real reward R is assumed to be linear in the "features" of the state-action space. Features are functions φi from the full state-action space S×A to the unit interval [0,1] (the paper linked above only considers functions from the state space; this is the "minor modification"). These features form a vector φ∈[0,1]k, for k different features. The actual reward is given by the inner product with a vector w∈ℝk, thus the reward at state-action pair (s,a) is
R(s,a)=w.φ(s,a).
To ensure the reward is always between -1 and 1, w is constrained to have ||w||1 ≤ 1; to reduce redundancy, we'll assume ||w||1=1.
The advantages of linearity is that we can compute the expected rewards directly from the expected feature vector. If the agent follows a policy π (a map from state to action) and has a discount factor γ, the expected feature vector is
μ(π) = E(Σt γtφ(st,π(st)),
where st is the state at step t.
The agent's expected reward is then simply
E(R) = w . μ(π).
Thus the problem of computing the correct reward is reduced to the problem of computing the correct w. In practice, to compute the correct policy, we just need to find one whose expected features are close enough to optimal; this need not involve computing w.
Inverse reinforcement learning
The approach of IRL is then to find a way to efficiently compute w, given some "trajectories": examples of good performance, provided by (human) experts. These experts are following the expert policy πE Given these trajectories, the agent can compute an empirical estimate for the μE=μ(πE), by simply averaging the (discounted) feature vectors prod |
eb9e0cd7-829e-42f9-98ad-5d727e72ea30 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | A Year of AI Increasing AI Progress
In July, I made [a post](https://www.lesswrong.com/posts/W3tZacTRt4koHyxbr/examples-of-ai-increasing-ai-progress) about AI being used to increase AI progress, along with [this spreadsheet](http://rsi.thomaswoodside.com/) that I've been updating throughout the year. Since then, I have run across more examples, and had others submit examples (some of which were published before the date I made my original post).
2022 has included a number of instances of AI increasing AI progress. Here is the list. In each entry I also credit the person who originally submitted the paper to my list.
* [A paper from Google Research](https://ai.googleblog.com/2022/03/offline-optimization-for-architecting.html) used a robust supervised learning technique to architect hardware accelerators [March 17th, submitted by Zach Stein-Perlman]
* [A paper from Google Research and Stanford](https://arxiv.org/abs/2203.14465) fine tuned a model on its own chain-of-thought outputs, to improve performance on reasoning tasks [March 28th, submitted by Nathaniel Li]
* [A paper from OpenAI](https://arxiv.org/abs/2206.05802) used LLMs to help humans find flaws in other LLMs, thereby enabling them to more easily improve those models [June 12th, submitted by Dan Hendrycks]
* [A paper from Google](https://ai.googleblog.com/2022/07/mlgo-machine-learning-framework-for.html) used machine learning to optimize compilers. This is less obviously accelerating AI but an [earlier version of the compiler is used in Pytorch](https://twitter.com/pytorch/status/1452737276564344846?lang=en) so it may end up doing so. [July 6th, submitted by Oliver Zhang]
* [NVIDIA used deep reinforcement learning](https://developer.nvidia.com/blog/designing-arithmetic-circuits-with-deep-reinforcement-learning/) to generate nearly 13,000 circuits in their newest GPUs. [July 8th, submitted by me]
* [Google found](https://ai.googleblog.com/2022/07/ml-enhanced-code-completion-improves.html?m=1) that ML code completion improved the productivity of their engineers. Some of them are presumably working in AI. [July 27th, submitted by Aidan O'Gara]
* [A paper from Microsoft Research and MIT](https://arxiv.org/abs/2207.14502) used language models to generate programming puzzle tasks for other language models. When finetuned on these tasks, the models were much better at solving the puzzles. [July 29th, submitted by Esben Kran]
* [A paper from Google and UIUC](https://arxiv.org/abs/2210.11610) used outputs from a language model to fine tune a language model after a majority vote procedure was used to filter outputs. [September 30th, submitted by me]
* [A paper from DeepMind](https://www.nature.com/articles/s41586-022-05172-4) used reinforcement learning to discover more efficient matrix multiplication algorithms. [October 5th, submitted by me]
* [A paper from Anthropic](https://www.anthropic.com/constitutional.pdf) used language models, rather than humans, for feedback to improve language models. [December 16th, submitted by me]
* [A paper from a number of universities](https://arxiv.org/abs/2212.10560) used language models to generate examples of instruction following, which were then filtered and used to fine tune language models to follow instructions better. [December 20th, submitted by Nathaniel Li].
I'm writing this fairly quickly so I'm not going to add extensive commentary beyond what I said in my last post, but I'll point out here two things:
* It is pretty common these days for people to use language model outputs to improve language models. This trend appears likely to continue.
* A lot of these papers are from Google. Not DeepMind, Google. Google may not have declared they are aiming for AGI, but they sure do seem to be writing a lot of papers that involve AI increasing AI progress. It seems important not to ignore them.
Did I miss any? You can submit more [here](https://forms.gle/V6wukPMST6tw2PrB8). |
edcbc58f-8d4a-4ee8-8fd0-dc3119d5effb | trentmkelly/LessWrong-43k | LessWrong | No negative press agreement
Original post: http://bearlamp.com.au/no-negative-press-agreement/
What is a no negative press agreement?
A no negative press agreement binds a media outlet's consent to publish information provided by a person with the condition that they be not portrayed negatively by the press.
Why would a person want that?
In recognising that the press has powers above and beyond every-day people to publish information and spread knowledge and perspective about an issue that can be damaging to an individual. An individual while motivated by the appeal of publicity, is also concerned about the potential damage caused by negative press.
Every person is the hero of their own story, from one's own perspective they performed actions that were justified and motivated by their own intention and worldview, no reasonable person would be able to tell their story (other than purposefully) in which they are spun as the negative conspirator of a plot, actively causing negative events on the world for no reason.
Historically, humans have been motivated to care more about bad news than good news, for reasons that expand on the idea that bad news might ring your death (and be a cause of natural selection) and good news would be irrelevant for survival purposes. Today we are no longer in that historic period, yet we still pay strong attention to bad news. It's clear that bad news can personally effect individuals - not only those in the stories, but others experiencing the bad news can be left with a negative worldview or motivated to be upset or distraught. In light of the fact that bad news is known to spread more than good news, and also risks negatively affecting us mentally, we are motivated to choose to avoid bad news, both in not creating it, not endorsing it and not aiding in it's creation.
The binding agreement is designed to do several things:
* protect the individual from harm
* reduce the total volume of negative press in the world
* decrease the damage caused by nega |
03cf3d99-aa78-4681-a326-68cb6b7eafc4 | trentmkelly/LessWrong-43k | LessWrong | Laplace Approximation
The last couple posts compared some specific models for 20000 rolls of a die. This post will step back, and talk about more general theory for Bayesian model comparison.
The main problem is to calculate P[data|model] for some model. The model will typically give the probability of observed data x (e.g. die rolls) based on some unobserved parameter values θ (e.g. the p's in the last two posts), along with a prior distribution over θ. We then need to compute
P[data|model]=∫θP[data|θ]dP[θ]
which will be a hairy high-dimensional integral.
Some special model structures allow us to simplify the problem, typically by factoring the integral into a product of one-dimensional integrals. But in general, we need some method for approximating these integrals.
The two most common approximation methods used in practice are Laplace approximation around the maximum-likelihood point, and MCMC (see e.g. here for application of MCMC to Bayes factors). We'll mainly talk about Laplace approximation here - in practice MCMC mostly works well in the same cases, assuming the unobserved parameters are continuous.
Laplace Approximation
Here's the idea of Laplace approximation. First, posterior distributions tend to be very pointy. This is mainly because independent probabilities multiply, so probabilities tend to scale exponentially with the number of data points. Think of the probabilities we calculated in the last two posts, with values like 10−70 or 10−20 - that's the typical case. If we're integrating over a function with values like that, we can basically just pay attention to the region around the highest value - other regions will have exponentially small weight.
Laplace' trick is to use a second-order approximation within that high-valued region. Specifically, since probabilities naturally live on a log scale, we'll take a second order-approximation of the log likelihood around its maximum point. Thus:
∫θelnP[data|θ]dP[θ]≈∫θelnP[data|θmax]+12(θ−θmax)T(d2lnPdθ2|θmax)(θ−θmax)dP[ |
70cebb18-98c5-4ba5-9832-efc02ed6384d | StampyAI/alignment-research-dataset/arxiv | Arxiv | Posterior calibration and exploratory analysis for natural language processing models
###
1 Introduction
Natural language processing systems are imperfect.
Decades of research have yielded analyzers that mis-identify named entities,
mis-attach syntactic relations,
and mis-recognize
noun phrase coreference
anywhere from
10-40% of the time. But these systems are accurate enough so that their outputs
can be used as soft, if noisy, indicators of language meaning for use in downstream analysis,
such as systems that perform
question answering, machine translation, event extraction, and narrative analysis
(McCord et al., [2012](#bib.bib25); Gimpel and Smith, [2008](#bib.bib15); Miwa et al., [2010](#bib.bib27); Bamman et al., [2013](#bib.bib1)).
To understand the performance of an analyzer,
researchers and practitioners typically
measure the accuracy of individual labels or edges
among a single predicted output structure y,
such as a most-probable tagging or entity clustering
argmaxyP(y|x)
(conditional on text data x).
But a probabilistic model gives a probability distribution over many other output structures
that have smaller predicted probabilities; a line of work has sought to control cascading pipeline errors by passing on multiple structures from earlier stages of analysis,
by propagating
prediction uncertainty
through multiple samples (Finkel et al., [2006](#bib.bib13)),
K-best lists (Venugopal et al., [2008](#bib.bib45); Toutanova et al., [2008](#bib.bib43)),
or explicitly diverse lists (Gimpel et al., [2013](#bib.bib17));
often the goal is to marginalize over structures to calculate and minimize
an expected loss function,
as in minimum Bayes risk decoding (Goodman, [1996](#bib.bib19); Kumar and Byrne, [2004](#bib.bib22)),
or to perform joint inference between early and later stages of NLP analysis
(e.g. Singh et al., [2013](#bib.bib40); Durrett and Klein, [2014](#bib.bib12)).
These approaches should work better when
the posterior probabilities of the predicted linguistic structures
reflect actual probabilities of the structures or aspects of the structures.
For example, say a model is overconfident: it places too much probability mass in the top prediction, and not enough in the rest.
Then there will be little benefit to using the lower probability structures, since
in the training or inference objectives
they will be incorrectly outweighed by the top prediction
(or in a sampling approach, they will be systematically undersampled and thus have too-low frequencies).
If we only evaluate models based on their top predictions
or on downstream tasks, it is difficult to diagnose this issue.
Instead, we propose to directly evaluate the *calibration*
of a model’s posterior prediction distribution.
A perfectly calibrated model knows how often it’s right or wrong;
when it predicts an event with 80% confidence,
the event empirically turns out to be true 80% of the time.
While perfect accuracy for NLP models remains an unsolved challenge,
perfect calibration is a more achievable goal,
since a model that has imperfect accuracy could, in principle, be perfectly calibrated.
In this paper, we develop a method to empirically analyze calibration that is appropriate for NLP models (§[3](#S3 "3 Empirical calibration analysis ‣ Posterior calibration and exploratory analysis for natural language processing models"))
and use it to analyze common generative and discriminative models
for tagging and classification
(§[4](#S4 "4 Calibration for classification and tagging models ‣ Posterior calibration and exploratory analysis for natural language processing models")).
Furthermore, if a model’s probabilities are meaningful,
that would justify using its probability distributions for any downstream purpose,
including exploratory analysis on unlabeled data.
In §[6](#S6 "6 Uncertainty in Entity-based Exploratory Analysis ‣ Posterior calibration and exploratory analysis for natural language processing models") we introduce a representative corpus exploration problem, identifying temporal event trends in international politics, with a method
that is dependent on coreference resolution.
We develop a coreference sampling algorithm
(§[5.2](#S5.SS2 "5.2 Sampling-based inference ‣ 5 Coreference resolution ‣ Posterior calibration and exploratory analysis for natural language processing models")) which projects uncertainty into
the event extraction, inducing a posterior distribution over event frequencies.
Sometimes the event trends have very high posterior variance
(large confidence intervals),222We use the terms *confidence interval* and *credible interval* interchangeably in this work; the latter term is debatably more correct, though less widely familiar.
reflecting when the NLP system genuinely does not know the correct semantic extraction.
This highlights an important use of a calibrated model: being able to tell a user when the model’s predictions are likely to be incorrect, or at least, not giving a user a false sense of certainty from an erroneous NLP analysis.
###
2 Definition of calibration
Consider a binary probabilistic prediction problem, which consists of binary labels and probabilistic predictions for them.
Each instance has a *ground-truth label* y∈{0,1}, which is used for evaluation. The prediction problem is to generate a *predicted probability* or *prediction strength* q∈[0,1]. Typically, we use some form of a probabilistic model to accomplish this task, where q represents the model’s posterior probability333Whether q comes from a Bayesian posterior or not
is irrelevant to the analysis in this section.
All that matters is that predictions are numbers q∈[0,1].
of the instance having a positive label (y=1).
Let S={(q1,y1),(q2,y2),⋯(qN,yN)} be the set of prediction-label pairs produced by the model. Many metrics assess the overall quality of how well the predicted probabilities match the data, such as the familiar cross entropy (negative average log-likelihood),
| | | |
| --- | --- | --- |
| | Lℓ(→y,→q)=1N∑iyilog1qi+(1−yi)log11−qi | |
or mean squared error, also known as the *Brier score* when y is binary (Brier, [1950](#bib.bib5)),
| | | |
| --- | --- | --- |
| | L2(→y,→q)=1N∑i(yi−qi)2 | |
Both tend to attain better (lower) values when q is near 1 when y=1, and near 0 when y=0; and they achieve a perfect value of 0 when all qi=yi.444These two loss functions are instances of *proper scoring rules* (Gneiting and Raftery, [2007](#bib.bib18); Bröcker, [2009](#bib.bib6)).
Let P(y,q) be the joint empirical distribution over labels and predictions. Under this notation, L2=Eq,y[y−q]2. Consider the factorization
| | | |
| --- | --- | --- |
| | P(y,q)=P(y∣q) P(q) | |
where P(y∣q) denotes the label empirical frequency, conditional on a prediction strength (Murphy and Winkler, [1987](#bib.bib28)).555
We alternatively refer to this as *label frequency* or *empirical frequency*. The P probabilities can be thought of as frequencies from the hypothetical population the data and predictions are drawn from.
P probabilities are, definitionally speaking, completely separate from a probabilistic model that might be used to generate q predictions.
Applying this factorization to the Brier score
leads to the calibration-refinement decomposition
(DeGroot and Fienberg, [1983](#bib.bib9)), in terms of
expectations with respect to the prediction strength distribution P(q):
| | | | |
| --- | --- | --- | --- |
| | L2 = Eq[q−pq]2Calibration MSE + Eq[pq(1−pq)]Refinement | | (1) |
where we denote pq≡P(y=1∣q) for brevity.
Here, *calibration* measures to what extent a model’s probabilistic predictions match their corresponding empirical frequencies. Perfect calibration is achieved when P(y=1∣q)=q for all q; intuitively, if you aggregate all instances where a model predicted q, they should have y=1 at q percent of the time.
We define the magnitude of miscalibration using root mean squared error:
######
Definition 1 (RMS calibration error).
| | | |
| --- | --- | --- |
| | CalibErr=√Eq[q−P(y=1∣q)]2 | |
The second term of Eq [1](#S2.E1 "(1) ‣ 2 Definition of calibration ‣ Posterior calibration and exploratory analysis for natural language processing models")
refers to *refinement*, which reflects
to what extent the model is able to separate different labels
(in terms of the conditional Gini entropy pq(1−pq)).
If the prediction strengths
tend to cluster around 0 or 1, the refinement score tends to be lower.
The calibration-refinement breakdown offers a useful perspective on the accuracy of a model posterior. This paper focuses on calibration.
There are several other ways to break down squared error, log-likelihood, and other probabilistic scoring rules.666They all include a notion of calibration corresponding to a Bregman divergence (Bröcker, [2009](#bib.bib6)); for example, cross-entropy can be broken down such that KL divergence is the measure of miscalibration.
We use the Brier-based calibration error in this work, since unlike cross-entropy it does not tend toward infinity when near probability 0; we hypothesize this could be an issue since both p and q are subject to estimation error.
###
3 Empirical calibration analysis
Input: A set of N prediction-label pairs {(q1,y1),(q2,y2),⋯,(qN,yN)}.
Output: Calibration error.
Parameter: Target bin size β.
Step 1: Sort pairs by prediction values qk in ascending order.
Step 2: For each, assign bin label bk=⌊k−1β⌋+1.
Step 3: Define each bin Bi as the set of indices of pairs that have the same bin label. If the last bin has size less than β, merge it with the second-to-last bin (if one exists). Let {B1,B2,⋯,BT} be the set of bins.
Step 4: Calculate empirical and predicted probabilities per bin:
| | | |
| --- | --- | --- |
| | ^pi=1|Bi|∑k∈Biyk and ^qi=1|Bi|∑k∈Biqk | |
Step 5: Calculate the calibration error as the root mean squared error per bin,
weighted by bin size in case they are not uniformly sized:
| | | |
| --- | --- | --- |
| | CalibErr=
⎷1NT∑i=1|Bi|(^qi−^pi)2 | |
Algorithm 1 Estimate calibration error using adaptive binning.
Input: A set of N prediction-label pairs {(q1,y1),(q2,y2),⋯,(qN,yN)}.
Output: Calibration error with a 95% confidence interval.
Parameter: Number of samples, S.
Step 1: Calculate {^p1,^p2,⋯,^pT} from step 4 of Algorithm [1](#algorithm1 "Algorithm 1 ‣ 3 Empirical calibration analysis ‣ Posterior calibration and exploratory analysis for natural language processing models").
Step 2: Draw S samples. For each s=1..S,
* For each bin i=1..T,
draw ^p(s)i∼N(^pi,^σ2i),
where ^σ2i=^pi(1−^pi)/|Bi|.
If necessary clip to [0,1]:
^p(s)i:=min(1,max(0,^p(s)i))
* Calculate the sample’s CalibErr from using the pairs (^qi,^p(s)i) as per Step 5 of Algorithm [1](#algorithm1 "Algorithm 1 ‣ 3 Empirical calibration analysis ‣ Posterior calibration and exploratory analysis for natural language processing models").
Step 3: Calculate the 95% confidence interval for the calibration error as:
| | | |
| --- | --- | --- |
| | CalibErravg±1.96 ^serror | |
where CalibErravg and ^serror are the mean and the standard deviation, respectively, of the CalibErrs calculated from the samples.
Algorithm 2 Estimate calibration error’s confidence interval by sampling.
From a test set of labeled data, we can analyze model calibration both in terms of the calibration error, as well as visualizing the *calibration curve* of label frequency versus predicted strength.
However, computing the label frequencies P(y=1|q) requires an infinite amount of data. Thus approximation methods are required to perform calibration analysis.
####
3.1 Adaptive binning procedure
| | | | |
| --- | --- | --- | --- |
|
(a)
|
(b)
|
(c)
|
(d)
|
Figure 1: (a) A skewed distribution of predictions on whether a word has the NN tag (§[4.2.2](#S4.SS2.SSS2 "4.2.2 Predicting single-word tags ‣ 4.2 Hidden Markov models and conditional random fields ‣ 4 Calibration for classification and tagging models ‣ Posterior calibration and exploratory analysis for natural language processing models")). Calibration curves produced by equally-spaced binning with bin width equal to 0.02 (b) and 0.1 (c) can have wide confidence intervals. Adaptive binning (with 1000 points in each bin) (d) gives small confidence intervals and also captures the prediction distribution. The confidence intervals are estimated as described in §[3.1](#S3.SS1 "3.1 Adaptive binning procedure ‣ 3 Empirical calibration analysis ‣ Posterior calibration and exploratory analysis for natural language processing models").
Previous studies that assess calibration in supervised machine learning models
(Niculescu-Mizil and Caruana, [2005](#bib.bib30); Bennett, [2000](#bib.bib2))
calculate label frequencies by dividing the prediction space into deciles or other evenly spaced bins—e.g. q∈[0,0.1), q∈[0.1,0.2), etc.—and then calculating the empirical label frequency in each bin. This procedure may be thought of as using a form of nonparametric regression (specifically, a regressogram; Tukey [1961](#bib.bib44)) to estimate the function f(q)=P(y=1∣q)
from observed data points.
But models in natural language processing give very skewed distributions of confidence scores q (many are near 0 or 1), so this procedure performs poorly, having much more variable estimates near the middle of the q distribution (Figure [1](#S3.F1 "Figure 1 ‣ 3.1 Adaptive binning procedure ‣ 3 Empirical calibration analysis ‣ Posterior calibration and exploratory analysis for natural language processing models")).
We propose adaptive binning as an alternative. Instead of dividing the interval [0,1] into fixed-width bins, adaptive binning defines the bins such that there are an equal number of points in each, after which the same averaging procedure is used.
This method naturally gives wider bins to area with fewer data points (areas that require more smoothing), and ensures that these areas have roughly similar standard errors as those near the boundaries,
since for a bin with β number of points and empirical frequency p,
the standard error is estimated by √p(1−p)/β, which is bounded above by 0.5/√β.
Algorithm [1](#algorithm1 "Algorithm 1 ‣ 3 Empirical calibration analysis ‣ Posterior calibration and exploratory analysis for natural language processing models") describes the procedure for estimating calibration error using adaptive binning, which can be applied to any probabilistic model that predicts posterior probabilities.
####
3.2 Confidence interval estimation
Especially when the test set is small, estimating calibration error may be subject to error,
due to uncertainty in the label frequency estimates.
Since how to estimate confidence bands for nonparametric regression
is an unsolved problem (Wasserman, [2006](#bib.bib46)),
we resort to a simple method based on the binning.
We construct a binomial normal approximation
for the label frequency estimate in each bin, and simulate from it;
every simulation across all bins is used to construct a calibration error; these simulated calibration errors are collected to construct a normal approximation for the calibration error estimate.
Since we use bin sizes of at least β≥200 in our experiments,
the central limit theorem justifies these approximations.
We report all calibration errors along with their 95% confidence intervals
calculated by Algorithm [2](#algorithm2 "Algorithm 2 ‣ 3 Empirical calibration analysis ‣ Posterior calibration and exploratory analysis for natural language processing models").777A major unsolved issue is how to fairly select the bin size. If it is too large, the curve is oversmoothed and calibration looks better than it should be; if it is too small, calibration looks worse than it should be. Bandwidth selection and cross-validation techniques may better address this problem in future work. In the meantime, visualizations of calibration curves
help inform the reader of the resolution of a particular analysis—if the bins are far apart,
the data is sparse, and the specific details of the curve are not known in those regions.
####
3.3 Visualizing calibration
In order to better understand a model’s calibration properties, we plot the pairs
(^p1,^q1),(^p2,^q2),⋯,(^pT,^qT)
obtained from the adaptive binning procedure to visualize the *calibration curve* of the model—this visualization is known as a *calibration* or *reliability plot*.
It provides finer grained insight into the calibration behavior in different prediction ranges. A perfectly calibrated curve would coincide with the y=x diagonal line.
When the curve lies above the diagonal, the model is underconfident (q<pq);
and when it is below the diagonal, the model is overconfident (q>pq).
An advantage of plotting a curve estimated from fixed-size bins, instead of fixed-width bins, is that the distribution of the points hints at the refinement aspect of the model’s performance.
If the points’ positions tend to cluster in the bottom-left and top-right corners, that implies the model is making more refined predictions.
###
4 Calibration for classification and tagging models
Using the method described in §[3](#S3 "3 Empirical calibration analysis ‣ Posterior calibration and exploratory analysis for natural language processing models"), we assess the quality of posterior predictions of several classification and tagging models. In all of our experiments, we set the target bin size in Algorithm [1](#algorithm1 "Algorithm 1 ‣ 3 Empirical calibration analysis ‣ Posterior calibration and exploratory analysis for natural language processing models") to be 5,000 and the number of samples in Algorithm [2](#algorithm2 "Algorithm 2 ‣ 3 Empirical calibration analysis ‣ Posterior calibration and exploratory analysis for natural language processing models") to be 10,000.
####
4.1 Naive Bayes and logistic regression
#####
4.1.1 Introduction
Previous work on Naive Bayes has found its probabilities to have
calibration issues, in part due to its incorrect conditional independence assumptions (Niculescu-Mizil and Caruana, [2005](#bib.bib30); Bennett, [2000](#bib.bib2); Domingos and Pazzani, [1997](#bib.bib10)).
Since logistic regression has the same log-linear representational capacity
(Ng and Jordan, [2002](#bib.bib29)) but does not suffer from the independence assumptions,
we select it for comparison,
hypothesizing it may have better calibration.
We analyze a binary classification task of Twitter sentiment analysis from emoticons.
We collect a dataset consisting of tweets identified by the Twitter API as English, collected from 2014 to 2015, with the “emoticon trick” (Read, [2005](#bib.bib36); Lin and Kolcz, [2012](#bib.bib24)) to label tweets that contain at least one occurrence of the smiley emoticon “:)” as “happy” (y=1) and others as y=0. The smiley emoticons are deleted in positive examples. We sampled three sets of tweets (subsampled from the Decahose/Gardenhose stream of public tweets) with Jan-Apr 2014 for training, May-Dec 2014 for development, and Jan-Apr 2015 for testing.
Each set contains 105 tweets, split between an equal number of positive and negative instances. We use binary features based on unigrams extracted from the *twokenize.py*888<https://github.com/myleott/ark-twokenize-py> tokenization.
We use the *scikit-learn* (Pedregosa et al., [2011](#bib.bib33)) implementations of Bernoulli Naive Bayes and L2-regularized logistic regression. The models’ hyperparameters (Naive Bayes’ smoothing paramter and logistic regression’s regularization strength) are chosen to maximize the F-1 score on the development set.
#####
4.1.2 Results
| | |
| --- | --- |
|
(a)
|
(b)
|
Figure 2: Calibration curve of (a) Naive Bayes and (b) logistic regression on predicting whether a tweet is a “happy” tweet.
Naive Bayes attains a slightly higher F-1 score (NB 73.8% vs. LR 72.9%),
but logistic regression has much lower calibration error: less than half as much RMSE
(NB 0.105 vs. LR 0.041; Figure [2](#S4.F2 "Figure 2 ‣ 4.1.2 Results ‣ 4.1 Naive Bayes and logistic regression ‣ 4 Calibration for classification and tagging models ‣ Posterior calibration and exploratory analysis for natural language processing models")).
Both models have a tendency to be
underconfident in the lower prediction range and overconfident in the higher range,
but the tendency is more pronounced for Naive Bayes.
####
4.2 Hidden Markov models and conditional random fields
#####
4.2.1 Introduction
Hidden Markov models (HMM) and
linear chain conditional random fields (CRF)
are another commonly used pair of analogous generative and discriminative models.
They both define a posterior over tag sequences P(y|x),
which we apply to part-of-speech tagging.
We can analyze these models in the binary calibration framework (§[2](#S2 "2 Definition of calibration ‣ Posterior calibration and exploratory analysis for natural language processing models")-[3](#S3 "3 Empirical calibration analysis ‣ Posterior calibration and exploratory analysis for natural language processing models")) by looking at marginal distribution of binary-valued outcomes of parts of the predicted structures.
Specifically, we examine calibration of predicted probabilities of individual tokens’ tags (§[4.2.2](#S4.SS2.SSS2 "4.2.2 Predicting single-word tags ‣ 4.2 Hidden Markov models and conditional random fields ‣ 4 Calibration for classification and tagging models ‣ Posterior calibration and exploratory analysis for natural language processing models")),
and of pairs of consecutive tags (§[4.2.3](#S4.SS2.SSS3 "4.2.3 Predicting two-consecutive-word tags ‣ 4.2 Hidden Markov models and conditional random fields ‣ 4 Calibration for classification and tagging models ‣ Posterior calibration and exploratory analysis for natural language processing models")).
These quantities are calculated with the forward-backward algorithm.
To prepare a POS tagging dataset, we extract *Wall Street Journal* articles from the English CoNLL-2011 coreference shared task dataset from Ontonotes (Pradhan et al., [2011](#bib.bib35)),
using the CoNLL-2011 splits for training, development and testing.
This results in 11,772 sentences for training, 1,632 for development,
and 1,382 for testing, over a set of 47 possible tags.
We train an HMM with Dirichlet MAP
using one pseudocount for every transition and word emission.
For the CRF, we use the L2-regularized L-BFGS algorithm implemented in *CRFsuite* (Okazaki, [2007](#bib.bib32)).
We compare an HMM to a CRF that only uses
basic
transition (tag-tag) and emission (tag-word) features,
so that it does not have an advantage
due to more features.
In order to compare models with similar task performance,
we train the CRF with only 3000 sentences from the training set,
which yields the same accuracy as the HMM (about 88.7% on the test set).
In each case, the model’s hyperparameters (the CRF’s L2 regularizer, the HMM’s pseudocount) are selected by maximizing accuracy on the development set.
#####
4.2.2 Predicting single-word tags
| | |
| --- | --- |
|
(a)
|
(b)
|
Figure 3: Calibration curves of (a) HMM, and (b) CRF, on predictions over all POS tags.
In this experiment, we measure miscalibration of the two models on predicting tags of single words. First, for each tag type, we produce a set of 33,306 prediction-label pairs (for every token); we then concatenate them across the tags for calibration analysis.
Figure [3](#S4.F3 "Figure 3 ‣ 4.2.2 Predicting single-word tags ‣ 4.2 Hidden Markov models and conditional random fields ‣ 4 Calibration for classification and tagging models ‣ Posterior calibration and exploratory analysis for natural language processing models") shows that the two models exhibit distinct calibration patterns. The HMM tends to be very underconfident whereas the CRF is overconfident, and the CRF has a lower (better) overall calibration error.
We also examine the calibration errors of the individual POS tags (Figure [4](#S4.F4 "Figure 4 ‣ 4.2.3 Predicting two-consecutive-word tags ‣ 4.2 Hidden Markov models and conditional random fields ‣ 4 Calibration for classification and tagging models ‣ Posterior calibration and exploratory analysis for natural language processing models")(a)). We find that CRF is significantly better calibrated than HMM in most but not all categories (39 out of 47).
For example, they are about equally calibrated on predicting the NN tag.
The calibration gap between the two models also differs among the tags.
#####
4.2.3 Predicting two-consecutive-word tags
| | |
| --- | --- |
|
(a)
|
(b)
|
Figure 4: Calibration errors of HMM and CRF on predicting (a) single-word tags and (b) two-consecutive-word tags. Lower errors are better. The last two columns in each graph are the average calibration errors over the most common labels.
There is no reason to restrict ourselves to model predictions of single words;
these models define marginal distributions over larger textual units.
Next we examine the calibration of posterior predictions of tag pairs on two consecutive words
in the test set. The same analysis may be important for, say, phrase extraction or other chunking/parsing tasks.
We report results for the top 5 and 100 most frequent tag pairs (Figure [4](#S4.F4 "Figure 4 ‣ 4.2.3 Predicting two-consecutive-word tags ‣ 4.2 Hidden Markov models and conditional random fields ‣ 4 Calibration for classification and tagging models ‣ Posterior calibration and exploratory analysis for natural language processing models")(b)). We observe a similar pattern as seen from the experiment on single tags:
the CRF is generally better calibrated than the HMM,
but the HMM does achieve better calibration errors in 29 out of 100 categories.
These tagging experiments illustrate that, depending on the application, different models can exhibit different levels of calibration.
###
5 Coreference resolution
We examine a third model, a probabilistic model for within-document noun phrase coreference, which has an efficient sampling-based inference procedure.
In this section we introduce it and analyze its calibration, in preparation for the next section where we use it for exploratory data analysis.
####
5.1 Antecedent selection model
We use the Berkeley coreference resolution system
(Durrett and Klein, [2013](#bib.bib11)), which was originally presented as a CRF;
we give it an equivalent
a series of independent logistic regressions (see appendix for details).
The primary component of this model
is a locally-normalized log-linear distribution
over clusterings of noun phrases, each cluster denoting an entity.
The model takes a fixed input of N mentions (noun phrases),
indexed by i in their positional order in the document.
It posits that every mention i has a latent antecedent selection decision, ai∈{1,…,i−1,\textscnew}, denoting which previous mention it attaches to, or new if it is starting a new entity that has not yet been seen at a previous position in the text. Such a mention-mention attachment indicates coreference, while the final entity clustering includes more links implied through transitivity.
The model’s generative process is:
######
Definition 2 (Antencedent coreference model and sampling algorithm).
* For i=1..N, sample
ai∼1Ziexp(wTf(i,ai,x))
* Calculate the entity clusters as e:=CC(a), the connected components of the antecedent graph
having edges (i,ai) for i where ai≠\textscnew.
Here x denotes all information in the document that is conditioned on for log-linear features f.
e={e1,...eM} denotes the entity clusters,
where each element is a set of mentions.
There are M entity clusters corresponding to the number of connected components in a.
The model defines a joint distribution over antecedent decisions
P(a|x)=∏iP(ai|x);
it also defines a joint distribution
over entity clusterings
P(e|x),
where the probability of an e is the sum of the probabilities of all
a vectors that could give rise to it.
In a manner similar to a
distance-dependent Chinese restaurant process (Blei and Frazier, [2011](#bib.bib3)),
it is non-parametric in the sense that the number of clusters M is not fixed in advance.
####
5.2 Sampling-based inference
For both calibration analysis and exploratory applications, we need
to analyze the posterior distribution over entity clusterings.
This distribution is a complex mathematical object;
an attractive approach to analyze it is to
draw samples from this distribution, then analyze the samples.
This antecedent-based model
admits a very straightforward procedure to draw independent e samples,
by stepping through Def. [2](#Thmdefn2 "Definition 2 (Antencedent coreference model and sampling algorithm). ‣ 5.1 Antecedent selection model ‣ 5 Coreference resolution ‣ Posterior calibration and exploratory analysis for natural language processing models"):
independently sample each ai then calculate the connected components of the
resulting antecedent graph.
By construction, this procedure samples from the joint distribution of e
(even though we never compute the probability of any single clustering e).
Unlike approximate sampling approaches,
such as Markov chain Monte Carlo methods
used in other coreference work
to sample e
(Haghighi and Klein, [2007](#bib.bib20)),
here there are no questions about burn-in or autocorrelation (Kass et al., [1998](#bib.bib21)).
Every sample is independent and very fast to compute—only slightly slower than calculating the MAP assignment
(due to the exp and normalization for each ai). We implement this algorithm by modifying the publicly available implementation from [Durrett and Klein](#bib.bib11).999Berkeley Coreference Resolution System, version 1.1: <http://nlp.cs.berkeley.edu/projects/coref.shtml>
####
5.3 Calibration analysis
Figure 5: Coreference calibration curve
for predicting whether two mentions belong to the same entity cluster.
We consider the following inference query:
for a randomly chosen pair of mentions,
are they coreferent?
Even if the model’s accuracy is comparatively low,
it may be the case that it is correctly calibrated—if it thinks
there should be great variability in entity clusterings,
it may be uncertain whether a pair of mentions should belong together.
Let ℓij be 1 if the mentions i and j are predicted to be coreferent, and 0 otherwise.
Annotated data defines a gold-standard ℓ(g)ij value for every pair i,j.
Any probability distribution over e defines a
marginal Bernoulli distribution for every proposition ℓij,
marginalizing out e:
| | | | |
| --- | --- | --- | --- |
| | P(ℓij=1∣x)=∑e1{(i,j)∈e}P(e∣x) | | (2) |
where (i,j)∈e is true iff there is an entity in e
that contains both i and j.
In a traditional coreference evaluation of the best-prediction entity clustering,
the model assigns 1 or 0 to every ℓij and the pairwise precision and recall can be computed by comparing them to the corresponding ℓ(g)ij.
Here, we instead compare the qij≡P(ℓij=1∣x,e) prediction strengths against ℓ(g)ij empirical frequencies
to assess pairwise calibration,
with the same binary calibration analysis tools developed in §[3](#S3 "3 Empirical calibration analysis ‣ Posterior calibration and exploratory analysis for natural language processing models")
by aggregating pairs with similar qij values.
Each qij is computed by averaging over
1,000 samples, simply taking the fraction of samples where the pair (i,j) is coreferent.
We perform this analysis on the development section of the English CoNLL-2011
data (404 documents). Using the sampling inference method discussed in \lx@sectionsign[???](#),
we compute 4.3 millions prediction-label pairs and measure their calibration error. Our result shows that the model produces very well-calibrated predictions with less than 1% CalibErr (Figure [5](#S5.F5 "Figure 5 ‣ 5.3 Calibration analysis ‣ 5 Coreference resolution ‣ Posterior calibration and exploratory analysis for natural language processing models")),
though slightly overconfident on middle to high-valued predictions.
The calibration error indicates that it is the most calibrated model we examine within this paper. This result suggests we might be able to trust its level of uncertainty.
###
6 Uncertainty in Entity-based Exploratory Analysis
####
6.1 Entity-syntactic event aggregation
We demonstrate one important use of calibration analysis:
to ensure the usefulness of
propagating uncertainty from coreference resolution
into a system for exploring unannotated text.
Accuracy cannot be calculated since there are no labels;
but if the system is calibrated,
we postulate that uncertainty information can help users
understand the underlying reliability of aggregated extractions
and isolate predictions that are more likely to contain errors.
We illustrate with an event analysis application
to count the number of “country attack events”:
for a particular country of the world,
how many news articles describe an entity affiliated with that country
as the agent of an attack, and how does this number change over time?
This is a simplified version of a problem where such
systems have been built and used for political science analysis
(Schrodt et al., [1994](#bib.bib39); Schrodt, [2012](#bib.bib38); Leetaru and Schrodt, [2013](#bib.bib23); Boschee et al., [2013](#bib.bib4); O’Connor et al., [2013](#bib.bib31)).
A coreference component can improve extraction coverage
in cases such as “Russian troops were sighted …and they attacked …”
We use the coreference system examined in §[5](#S5 "5 Coreference resolution ‣ Posterior calibration and exploratory analysis for natural language processing models") for this analysis.
To propagate coreference uncertainty, we re-run event extraction on multiple coreference samples generated from the algorithm described in §[5.2](#S5.SS2 "5.2 Sampling-based inference ‣ 5 Coreference resolution ‣ Posterior calibration and exploratory analysis for natural language processing models"), inducing a posterior distribution over the event counts.
To isolate the effects of coreference,
we use a very simple syntactic dependency system to identify affiliations and events.
Assume the availability of dependency parses for a document d, a coreference resolution e, and a lexicon of country names, which contains a small set of words w(c) for each country c; for example, w(FRA)={france,french}.
The binary function f(c,e;xd) assesses
whether an entity e is affiliated with country c and is described as the agent of an attack,
based on document text and parses xd;
f returns true iff both:101010Syntactic relations are Universal Dependencies (de Marneffe et al., [2014](#bib.bib8)); more details for the extraction rules are in the appendix.
* There exists a mention i∈e described as country c: either its head word is in w(c) (e.g. “Americans”), or its head word has an *nmod* or *amod* modifier in w(c) (e.g. “American forces”, “president of the U.S.”); and there is only one unique country c among the mentions in the entity.
* There exists a mention j∈e
which is the
*nsubj* or *agent* argument
to the verb “attack” (e.g. “they attacked”, “the forces attacked”,
“attacked by them”).
For a given c, we first calculate
a binary variable for
whether there is at least one entity fulfilling f in a particular document,
| | | | |
| --- | --- | --- | --- |
| | a(d,c,ed)=⋁e∈edf(c,e;xd) | | (3) |
and second, the number of such documents
in d(t), the set of *New York Times* articles
published in a given time period t,
| | | | |
| --- | --- | --- | --- |
| | n(t,c,ed(t))=∑d∈d(t)a(d,c,ed) | | (4) |
These quantities are both random variables, since they depend on e;
thus we are interested in the posterior distribution of n, marginalizing out e,
| | | | |
| --- | --- | --- | --- |
| | P(n(t,c,ed(t))∣xd(t)) | | (5) |
If our coreference model was highly certain (only one structure, or a small number of similar structures, had most of the probability mass in the space of all possible structures),
each document would have an a posterior near either 0 or 1, and their sum in Eq. [5](#S6.E5 "(5) ‣ 6.1 Entity-syntactic event aggregation ‣ 6 Uncertainty in Entity-based Exploratory Analysis ‣ Posterior calibration and exploratory analysis for natural language processing models") would have
a narrow distribution.
But if the model is uncertain, the distribution will be wider.
Because of the transitive closure,
the probability of a is potentially more complex than the single antecedent linking probability between two mentions—the affiliation and attack information can propagate
through a long coreference chain.
####
6.2 Results
We tag and parse
a
193,403 article subset
of the Annotated New York Times LDC corpus
(Sandhaus, [2008](#bib.bib37)), which includes articles
about world news
from the years 1987 to 2007 (details in appendix).
For each article, we run the coreference system
to predict 100 samples, and evaluate f
on every entity in every sample.111111We obtained similar results using only 10 samples. We also obtained similar results with a different query function, the total number of entities, across documents, that fulfill f.
The quantity of interest is the number of articles mentioning attacks in a 3-month period (quarter),
for a given country.
Figure [6](#S6.F6 "Figure 6 ‣ 6.2 Results ‣ 6 Uncertainty in Entity-based Exploratory Analysis ‣ Posterior calibration and exploratory analysis for natural language processing models") illustrates the
mean and 95% posterior credible intervals for each quarter.
The posterior mean m is calculated as the mean of the samples,
and the interval is the normal approximation m±1.96 s, where s is the standard deviation among samples for that country and time period.
Uncertainty information helps us understand whether a difference between data points is real.
In the plots of Figure [6](#S6.F6 "Figure 6 ‣ 6.2 Results ‣ 6 Uncertainty in Entity-based Exploratory Analysis ‣ Posterior calibration and exploratory analysis for natural language processing models"), if we had used a 1-best coreference resolution, only a single line would be shown, with no assessment of uncertainty.
This is problematic in cases when the model genuinely does not know the correct answer.
For example, the 1993-1996 period of the USA plot (Figure [6](#S6.F6 "Figure 6 ‣ 6.2 Results ‣ 6 Uncertainty in Entity-based Exploratory Analysis ‣ Posterior calibration and exploratory analysis for natural language processing models"), top) shows the posterior mean fluctuating from 1 to 5 documents; but when credible intervals are taken into consideration, we see that model does not know whether the differences are real, or were caused by coreference noise.
A similar case is highlighted at the bottom plot of Figure [6](#S6.F6 "Figure 6 ‣ 6.2 Results ‣ 6 Uncertainty in Entity-based Exploratory Analysis ‣ Posterior calibration and exploratory analysis for natural language processing models").
Here we compare the event counts for Yugoslavia and NATO,
which were engaged in a conflict in 1999. Did the *New York Times* devote more attention to the attacks by one particular side? To a 1-best system, the answer would be yes. But the posterior intervals for the two countries’ event counts in mid-1999 heavily overlap, indicating that the coreference system introduces too much uncertainty to obtain a conclusive answer for this question.
Note that calibration of the coreference model is important for the credible intervals to be useful; for example, if the model was badly calibrated by being overconfident (too much probability over a small set of similar structures), these intervals would be too narrow, leading to incorrect interpretations of the event dynamics.
Visualizing this uncertainty gives richer information for a potential user of an NLP-based system, compared to simply drawing a line based on a single 1-best prediction. It preserves the genuine uncertainty due to ambiguities the system was unable to resolve. This highlights an alternative
use of Finkel et al. ([2006](#bib.bib13))’s approach of sampling multiple NLP pipeline components, which in that work was used to perform joint inference. Instead of focusing on improving an NLP pipeline, we can pass uncertainty on to exploratory purposes, and try to highlight to a user where the NLP system may be wrong, or where it can only imprecisely specify a quantity of interest.
Finally, calibration can help error analysis.
For a calibrated model, the more uncertain a prediction is, the more likely it is to be erroneous.
While coreference errors comprise only one part of event extraction errors
(alongside issues in parse quality, factivity, semantic roles, etc.),
we can look at highly uncertain event predictions
to understand the nature of coreference errors relative to our task.
We manually analyzed documents with a 50% probability to contain an “attack”ing country-affiliated entity, and found difficult coreference cases.
In one article from late 1990,
an “attack” event for IRQ is extracted from the sentence
“But some political leaders said that they feared that Mr. Hussein might attack Saudi Arabia”.
The mention “Mr. Hussein” is classified as IRQ only when
it is coreferent with a previous mention
“President Saddam Hussein of Iraq”;
this occurs only 50% of the time, since in some posterior samples
the coreference system split apart these two “Hussein” mentions.
This particular document is additionally difficult, since it includes the names of more than 10 countries (e.g. United States, Saudi Arabia, Egypt), and some of the Hussein mentions are even clustered with presidents of other countries (such as “President Bush”),
presumably because they share the “president” title.
These types of errors are a major issue for
a political analysis task;
further analysis could assess their prevalence and how to address them in future work.
| | |
| --- | --- |
| | |
Figure 6: Number of documents with an “attack”ing country per 3-month period,
and coreference posterior uncertainty for that quantity.
The dark line is the posterior mean, and the shaded region is the 95% posterior credible interval.
See appendix for more examples.
###
7 Conclusion
In this work, we argue that the calibration of posterior predictions is a desirable property of probabilistic NLP models,
and that it can be directly evaluated.
We also demonstrate a use case of having calibrated uncertainty: its propagation into downstream exploratory analysis.
Our posterior simulation approach for exploratory and error analysis
relates to *posterior predictive checking*
(Gelman et al., [2013](#bib.bib14)),
which analyzes a posterior to test model assumptions;
Mimno and Blei ([2011](#bib.bib26)) apply it to a topic model.
One avenue of future work is to investigate more effective nonparametric regression methods
to better estimate and visualize calibration error, such as Gaussian processes or bootstrapped kernel density estimation.
Another important question is: what types of inferences are facilitated by correct calibration?
Intuitively, we think that overconfidence will lead to overly narrow confidence intervals;
but in what sense are confidence intervals “good” when calibration is perfect?
Also, does calibration help joint inference in NLP pipelines?
It may also assist calculations that rely on expectations,
such as inference methods like minimum Bayes risk decoding,
or learning methods like EM,
since calibrated predictions imply that calculated expectations are statistically unbiased
(though the implications of this fact may be subtle).
Finally, it may be interesting
to pursue recalibration methods, which readjust a non-calibrated model’s predictions
to be calibrated; recalibration methods have been developed
for
binary (Platt, [1999](#bib.bib34); Niculescu-Mizil and Caruana, [2005](#bib.bib30))
and multiclass (Zadrozny and Elkan, [2002](#bib.bib47))
classification settings,
but we are unaware of methods appropriate
for the highly structured outputs typical in linguistic analysis.
Another approach might be to directly constrain CalibErr=0 during training,
or try to reduce it as a training-time risk minimization or cost objective
(Smith and Eisner, [2006](#bib.bib41); Gimpel and Smith, [2010](#bib.bib16); Stoyanov et al., [2011](#bib.bib42); Brümmer and Doddington, [2013](#bib.bib7)).
Calibration is an interesting and important property of NLP models.
Further work is necessary to address these and many other questions.
### Acknowledgments
Thanks to Erik Learned-Miller, Benjamin Marlin, Craig Greenberg, Phan-Minh Nguyen, Caitlin Cellier and the CMU ARK Lab for discussion and comments,
and to the anonymous reviewers (especially R3) for helpful suggestions.
###
1 Sampling a deterministic function of a random variable
In several places in this paper, we define probability distributions
over deterministic functions of a random variable, and sample from them
by applying the deterministic function to samples of the random variable.
This should be valid by construction, but we supply the following argument
for further justification.
X is a random variable and g(x) is a deterministic function which takes a value of X as its input.
Since g depends on a random variable, g(X) is a random variable as well.
The distribution for g(X), or aspects of it (such as a PMF or independent samples from it) can be calculated by marginalizing out X with a Monte Carlo approximation. Assuming g has discrete outputs (as is the case for the event counting function n, or connected components function CC), we examine the probability mass function:
| | | | | |
| --- | --- | --- | --- | --- |
| | pmf(h) | ≡P(g(X)=h) | | (6) |
| | | =∑xP(g(x)=h∣x) P(x) | | (7) |
| | | =∑x1{g(x)=h}P(x) | | (8) |
| | | ≈1S∑x∼P(X)1{g(x)=h} | | (9) |
Eq. [8](#S1.E8 "(8) ‣ 1 Sampling a deterministic function of a random variable ‣ Appendix ‣ Posterior calibration and exploratory analysis for natural language processing models") holds because g(x) is a deterministic function,
and Eq. [9](#S1.E9 "(9) ‣ 1 Sampling a deterministic function of a random variable ‣ Appendix ‣ Posterior calibration and exploratory analysis for natural language processing models") is a Monte Carlo approximation that uses S samples from P(x).
This implies that a set of g values calculated on x samples,
{g(x(s)):x(s)∼P(x)}, should constitute a sample from the distribution P(g(X)); in our event analysis section we usually call this the “posterior” distribution of g(X) (the n(t,c) function there).
In our setting, we do not directly use the PMF calculation above;
instead, we construct normal approximations to the probability distribution g(X).
We use this technique in several places.
For the calibration error confidence interval, the calibration error is
a deterministic function of the uncertain empirical label frequencies pi;
there, we propagate posterior uncertainty from a normal approximation to the Bernoulli parameter’s posterior (the pi distribution under the central limit theorem)
through simulation.
In the coreference model, the connected components function is a deterministic function of the antecedent vector;
thus repeatedly calculating e(s):=CC(a(s)) yields samples of entity clusterings from their posterior.
For the event analysis, the counting function n(t,c,ed(t)) is a function of the entity samples, and thus can be recalculated on each—this is a multiple step deterministic pipeline, which postprocesses simulated random variables.
As in other Monte Carlo-based inference techniques (as applied to both Bayesian and frequentist (e.g. bootstrapping) inference), the mean and standard deviation of samples drawn from the distribution constitute the mean and standard deviation of the desired posterior distribution,
subject to Monte Carlo error due to the finite number of samples,
which by the central limit theorem shrinks at a rate of 1/√S.
The Monte Carlo standard error for estimating the mean is σ/√S where σ is the standard deviation.
So with 100 samples, the Monte Carlo standard error for the mean is √100=10 times smaller than standard deviation. Thus in the time series graphs, which are based on S=100 samples, the posterior mean (dark line) has Monte Carlo uncertainty that is 10 times smaller than the vertical gray area (95% CI) around it.
###
2 Normalization in the coreference model
Durrett and Klein (2013) present their model as a globally normalized, but fully factorized, CRF:
| | | |
| --- | --- | --- |
| | P(a|x)=1Z∏iexp(wTf(i,ai,x)) | |
Since the factor function decomposes independently for each random variable ai,
their probabilities are actually independent, and can be rewritten
with local normalization,
| | | |
| --- | --- | --- |
| | P(a|x)=∏i1Ziexp(wTf(i,ai,x)) | |
This interpretation justifies the use of independent sampling
to draw samples of the joint posterior.
###
3 Event analysis: Corpus selection, country affiliation, and parsing
Articles are filtered to yield a dataset about world news. In the New York Times Annotated Corpus, every article is tagged with a large set of labels. We include
articles that contain a category whose label starts with the string *Top/News/World*,
and exclude articles with any category matching the regex */(Sports|Opinion)*,
and whose text body contains a mention of at least one country name.
Country names are taken from the dictionary *country\_igos.txt*
based on previous work
(<http://brenocon.com/irevents/>). Country name matching is case insensitive and uses light stemming: when trying to match a word against the lexicon, if a match is not found, it backs off to stripping the last and last two characters. (This is usually unnecessary since the dictionary contains modifier forms.)
POS, NER, and constituent and dependency parses are produced with Stanford CoreNLP 3.5.2 with default settings except for one change, to use its shift-reduce constituent parser (for convenience of processing speed).
We treat tags and parses as fixed and leave their uncertainty propagation for future work.
When formulating the extraction rules, we examined frequencies of all syntactic dependencies within country-affiliated entities, in order to help find reasonably high-coverage syntactic relations for the “attack” rule.
###
4 Event time series graphs
The following pages contain posterior time series graphs
for 20 countries, as described in the section on coreference-based
event aggregation, in order of decreasing total event frequency.
As in the main paper, the blue line indicates the posterior mean,
and the gray region indicates 95% posterior credibility intervals,
with count aggregation at the monthly level.
The titles are ISO3 country codes.
 |
69d98f80-8398-4e79-8d99-5ae8c4ac3caf | trentmkelly/LessWrong-43k | LessWrong | On passing Complete and Honest Ideological Turing Tests (CHITTs)
[Cross-posted from Facebook]
"In the moment when I truly understand my enemy, understand him well enough to defeat him, then in that very moment I also love him. I think it’s impossible to really understand somebody, what they want, what they believe, and not love them the way they love themselves."
- Ender Wiggin (from Ender's Game, by Orson Scott Card)
Maybe the following is obvious, but it wasn't initially obvious to me so I'm sharing it anyway.
The term Ideological Turing Test (ITT) was coined by Bryan Caplan and refers to whether someone can articulate and argue the views of an opposing ideology or belief system so convincingly that others cannot tell the difference between the person's articulation and that of a true believer.
Let's use the term Complete and Honest Ideological Turing Test (CHITT) to refer to an ITT where not only can you "fool" observers, but also:
1. You can do so without resorting to any rhetorical tricks such as leaving out relevant information;
2. You can do so without making the other side look irrational;
3. You can do this even for very long debates with expert observers who grill you closely;
4. You can do this in response to any evidence or argument or counterargument that you yourself might use to argue for your true position.
Claim: It's impossible to pass a CHITT unless you actually agree that from some perspective the other side is rational. Furthermore, you must not have any knockdown arguments for why that other perspective is incorrect. That doesn't mean you need to *agree* with the other perspective, just that you don't have a knockdown argument against it.
Reasoning: By the specified criteria, you need to think that no matter what arguments you yourself would throw against the other side, you would not make the other side look irrational even to expert observers. You can of course still have some set of priors that lead you to your true position, while the other side has different priors that lead to their positio |
52308ec3-939c-492a-bae7-2de90cc9651d | trentmkelly/LessWrong-43k | LessWrong | The most important step
I have a penchant for the stories of the old, in the era before the Bayesian Enlightment and before the Singers and Cantors surfaced, the days when the students of the Art of Rationality met underground, away from prying eyes.
Without the Sacred Texts to guide them, the disciples of Rationality spent long times discussing and refining the Art. 'Twas an arduous task, yet they persevered, guided by the promise of Light and Truth, the glimpse of Polaris that decided their course.
In this story in particular, two disciples of the Temple of the Bay discussed about the most important step in a quest.
-Surely - said one of them - the most important step must be first; for no quest has come to fruition without going through the first step.
-But you forget - said the other - the same reasoning applies to the last. No quest is complete without its last step.
-What about the value of the journey? It is infinitely more valuable to get halfway through a sacred quest, and then fail the last step, than to never have embarked upon it to begin with.
-In the Art, all that matters is victory. We do not embark in the path of Rationality to try, but to reach that which really matters. One does not get points for effort. Only conclusions matter.
-How can you even talk about the last step? In the quests that really matter, we should not see the end. Those quests, like the pursue of the Art, are powerful, alive things. To seek to end them is to mean their death.
Both kept arguing about both things for ten days and ten nights, until the master of the Temple came back from a journey of sacred meditation, to find both her disciples in heated conversation.
-Silence! - said the master - You both have brought dishonor to this sacred house! You have unattended your tasks while talking about abstract things, and thus you have lost sight of the Enemy.
-But master - protested one disciple - this is important theory to be crafted! Our quests must wait, for as you taught us, a strike without |
6a41eb43-d904-4410-848a-b80d0afdde36 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | What Should the Average EA Do About AI Alignment?
I'm trying to get a handle on what advice to give people who are convinced AI is a problem worthy of their time, \*probably\* the most important problem, but are not sure if they have the talent necessary to contribute.
A trending school of thought is "AI Alignment needs careful, clever, agenty thinkers. 'Having the correct opinion' is not that useful. There is nobody who can tell you what exactly to do, because nobody knows. We need people who can figure out what to do, in a very messy, challenging problem."
This sort of makes sense to me, but it seems like only a few sorts of people can realistically contribute in this fashion (even given growth mindset considerations). It also seems like, even if most people could contribute, it doesn't provide very good next-actions to people who have reached the "okay, this is important" stage, but who aren't (yet?) ready to change their career direction.
Here is the advice I currently give, followed by the background assumptions that prompted it. I'm looking for people to challenge me on any of these:
Options for the non-or-minimally-technical-ish:
-----------------------------------------------
1) Donate. (1%, or more if you can do so without sacrificing the ability to take valuable financial risks to further your career. MIRI, FHI, 80k and CFAR seem like the most credible ways to turn money into more AI Alignment career capital)
2) Arrange your life such that you can easily identify volunteer opportunities for gruntwork, operations, or other nontechnical skills for AI safety orgs, and dedicate enough time and attention to helping with that gruntwork that you are more of an asset than a burden. (i.e. helping to run conferences and workshops). To help with AI specific things, it seems necessary to be in the Bay, Boston, Oxford, Cambridge or London.
3a) Embark on projects or career paths that will cause you to gain deep skills, and in particular, train the habit/skill of noticing things that need doing, and proactively developing solutions to accomplish them. (These projects/careers can be pretty arbitrary. To eventually tie them back into AI, you need to get good enough that you'll either be able help found a new org or provide rare skills to an existing org)
3b) Ideally, choose projects that involve working together in groups, that require you to resolve differences in opinion on how to use scarce resources, and which require you to interacting with other groups with subtly different goals. Practice coordination skills mindfully.
4) Provide a reading list of blogs and social-media feeds to stay up-to-date on the more accessible, less technically demanding thoughts relating to AI Safety. Practice thinking critically on your own about them. (this doesn't really come with an obvious "Part 2" that translates that into meaningful action on its own)
If technical-ish, and/or willing to learn a LOT
-----------------------------------------------
5) Look at the MIRI and 80k AI Safety syllabus, and see if how much of it looks like something you'd be excited to learn. If applicable to you, consider diving into that so you can contribute to the cutting edge of knowledge.
6) If you're a talented programmer, learn a lot about ML/Deep Learning and then stay up to date on the latest actual AI research, so you can position yourself at the top AI companies and potentially have influence with them on which direction they go.
An important question I'd like to answer is "how do can you tell if it makes sense to alter your career in pursuit of #5 and #6?"? This is very non-obvious to me.
I talk to a lot of people that seem roooooughly analagous to myself, ie. pretty smart but not extremely smart. In my case I think I have a credible claim on "community building" being my comparative advantage, but I notice a lot of people default to "be a community person or influencer", and I'm really wary of a decision tree that outputs a tower of meta-community-stuff for anyone who's not obviously expert at anything else. I'd like to have better, fleshed out, *scalable* suggestions for people fairly similar to me.
Background assumptions
----------------------
Various things that fed into the above recommendations (sometimes directly, sometimes indirectly). This is a living document that I'll update as people persuade me otherwise. Again, appreciate getting challenged on any of these.
### **AI Timelines and Goals**
AI timelines are anywhere between 5 years (if DeepMind is more advanced than they're telling anyone), 20 years (if it turns out general AI is only a couple breakthroughs away from current Deep Learning trends, and we're (un)lucky on how soon those breakthroughs come), or much longer if General AI turns out to be harder. We should be prepared for each possibility.
Eventually, all of our efforts will need to translate into the ability into one of the following:
- the ability to develop insights about AI Alignment
- the ability to cause AI research to be safely aligned
- the ability to stop or slow down AI research until it can be safely aligned
### **Donation**
- MIRI seems like the most shovel-ready instance of "actual AI Safety research". It's not obvious to me whether MIRI is doing the best work, but they seem to be at least doing *good* work, and they do seem underfunded, and funding them seems like the most straightforward way to turn money into more professional AI researchers.
- FHI is a contender for second-best funding-target for X-risk reduction, including some thought about AI alignment.
- 80k, CFAR and Leverage are the orgs I know of that seem to be concretely attempting to solve the "career capital gap", with different strategies. They each have elements that seem promising to me. I'm sure what their respective funding constraints are. (Note: I recently became a bit more interested in Leverage than I had been, but examining Leverage is a blogpost unto itself and I'm not going to try doing so here)
- The [Far Future Fund](https://app.effectivealtruism.org/funds/far-future) (recently announced, run by Nick Beckstead) may be a good way to outsource your donation decision.
### **Career Capital, Agency and Self Improvement**
- An important limiting reagent is "people able to be agents." More than any single skillset, we need people who are able to look at organizations and worldstates, figure out what's not being done yet, figure out if they currently have the skills to do it, and backchain from that to being able to become the sort of people who have the skills to do that.
- To self-improve the fastest, as a person and as an org, you need high quality feedback loops.
- In my experience, there is a critical threshold between an "agent" and a non-agent. People get activated as agents when they a) have a concrete project to work on that seems important to them that's above their current skill level, and b) have some high status mentor-figure who takes time out of their day to tell them in a serious voice "this project you are working on is important." (The latter step is not *necessary* but it seems to help a *lot*. Note: this is NOT a mentor figure who necessarily spends a lot of time training you. They are Gandalf, telling you your mission is important and they believe in you, and then mostly staying out of the way)
(Actual longterm mentorship is also super helpful but doesn't seem to be the limiting issue)
- Beyond "be an agent", we do need highly skilled people at a variety of specific skills - both because AI Safety orgs need them, and because high skill allows you to get a job at an AGI research institution.
- Despite attempting to achieve this for several years, it's not obvious that CFAR has developed the ability to produce agents, but it's succeeded (at least slightly) at attracting existing agents, training them in some skills, and focusing them on the right problems.
### **Thinking Critically**
- We need people who can think critically, and who spend time/attention being able to think critically and deeply about the right things.
- Thinking *usefully* critically requires being up to speed on what other people are thinking, so you aren't duplicating work.
- It is currently very hard to keep up with ALL the different developments across the AI/EA/Career-Capital-Building spaces. Both because the updates come from all over the internet (and sometimes in person), and because people's writing is often verbose and inconcise.
- It is possible for the average EA to learn to think more critically, but it requires significant time investment
### **Coordination**
- Coordination problems are extraordinarily hard. Humanity essentially failed the "Nuclear Weapons test" (i.e. we survived the Cold War, but we easily might not have. Squeaking by the with a C- is not acceptable).
- Some people have argued the AI problem is much harder than Nukes, which isn't clear to me, (in the longterm you do need to stop everyone ever from developing unsafe AI, but it seems like the critical period is the window wherein AGI is first possible, where it'll be something like 6-20 companies working on it at once)
- The Rationality and EA communities aren't *obviously* worse than the average community at coordination, but they are certainly not much better. And EAs are definitely not better than-average at inducing coordination/cooperation among disparate groups with different goals that aren't aligned with us.
- If your goal is to influence orgs or AGI researchers, you need to make sure you're actually following a path that leads to *real* influence. (i.e. "You can network your way into being Elon Musk's friend who he invites over for dinner, but that doesn't mean he'll listen to you about AI safety. The same goes for networking your way onto the GoogleBrain team or the Google AI Ethics board. Have a clear model of influence and how much of it you credibly have.")
-Mainstream politics is even harder than coordinating corporations, and to a first approximation is useless for purposes of AI alignment.
Open Questions
--------------
This is mostly a recap.
0) Is anything in my framework grossly wrong?
1) My primary question is "how do we filter for people who *should* consider dropping everything and focusing on the technical aspects of AI Safety, or seriously pursue careers that will position them to influence AGI research institutions?" These seem like the most important things to *actually* output, and it seems most important for those people to cultivate particular types of critical thinking, technical skill and ability-to-influence.
For people who are not well suited, or not yet ready to do 1), how can we either:
2) Make it easier for them to translate marginal effort into meaningful contribution, or creating a clearer path towards:
3) Level up to the point where they are able to take in the entire field, and generate useful things to do (without requiring much effort from other heavily involved people whose time is scarce).
Potential Further Reading
-------------------------
I have not read all of these, so cannot speak to which are most important, but I think it's useful to at least skim the contents of each of them so you have a rough idea of the ideas at play. I'm including them here mostly for easy reference.
(If someone wanted to generate a 1-3 sentence summary of each of these and indicate who the target audience is, I'd be happy to edit that in. I hopefully will eventually have time to do that myself but it may be a while)
[MIRI's Research Guide](https://intelligence.org/research-guide/)
[80,000 Hours AI Safety Syllabus](https://80000hours.org/ai-safety-syllabus/)
[UC Berkeley Center for Human Compatible AI Bibliography](http://humancompatible.ai/bibliography)
[Case Study of CFAR's Effectiveness](http://rationality.org/studies/2016-case-studies)
[AI Impacts Timelines and Strategies (examples of how to think strategically given different AI timelines)](http://aiimpacts.org/ai-timelines-and-strategies/)
[Concrete Problems in AI Safety](https://arxiv.org/abs/1606.06565)
[OpenAI's Blog](https://openai.com/blog/)
[AgentFoundations.org](https://agentfoundations.org/) (this is sort of a stack-overflow / technical discussion forum for discussing concepts relevant to AI alignment)
[Deliberate Grad School](http://acritch.com/deliberate-grad-school/)
[https://vkrakovna.wordpress.com/2016/02/28/introductory-resources-on-ai-safety-research/](/Vika Krakovna's Suggested Reading List) |
75aa82b4-57b9-4bf5-8900-8bfaa837df96 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Some conceptual alignment research projects
Some research outputs I’d love to see, focused on exploring, clarifying and formalizing important alignment concepts. I expect that most of these will be pretty time-consuming, but happy to discuss for people who want to try:
1. A paper which does for [deceptive alignment](https://bounded-regret.ghost.io/ml-systems-will-have-weird-failure-modes-2/) what the goal misgeneralization paper does for inner alignment, i.e. describing it in ML language and setting up toy examples (for example, telling GPT-3 to take actions which minimize changes in its weights, given that it’s being trained using actor-critic RL with a certain advantage function, and seeing if it knows how to do so).
2. A paper which does the same for gradient hacking, e.g. taking [these examples](https://www.alignmentforum.org/posts/EeAgytDZbDjRznPMA/gradient-hacking-definitions-and-examples) and putting them into more formal ML language.
3. A list of papers that are particularly useful for new research engineers to replicate.
4. A takeover scenario which covers all the key points in <https://www.cold-takes.com/ai-could-defeat-all-of-us-combined/>, but not phrased as an argument, just phrased as a possible scenario (I think you can’t really make the argument rigorously in that little space).
5. A paper which defines the concepts of implicit planning, implicit value functions, implicit reward models, etc, in ML terms. Kinda like <https://arxiv.org/abs/1901.03559> but more AGI-focused. I want to be able to ask people “does GPT-3 choose actions using an implicit value function?” and then be able to point them to this paper to rigorously define what I mean. I discuss this briefly in the [phase 1 section here](https://www.alignmentforum.org/posts/KbyRPCAsWv5GtfrbG/what-misalignment-looks-like-as-capabilities-scale).
6. A blog post which describes in as much detail as possible what our current “throw the kitchen sink at it” alignment strategy would look like. (I’ll probably put my version of this online soon but would love others too).
7. A blog post explaining “[debate on weights](https://www.lesswrong.com/posts/FuGfR3jL3sw6r8kB4/richard-ngo-s-shortform?commentId=on7Y4d8pqrdrxLtec)” more thoroughly.
8. A blog post exploring how fast we should expect a forward pass to be for the first AGIs - e.g. will it actually be slower than human thinking, as discussed in [this comment](https://www.lesswrong.com/posts/KbyRPCAsWv5GtfrbG/what-misalignment-looks-like-as-capabilities-scale?commentId=3F8x4m2gbTHJ4evL7#comments)?
9. A blog post exploring considerations for why model goals may or may not be much more robust to SGD than model beliefs, as discussed in [framing 3 here](https://www.lesswrong.com/posts/poyshiMEhJsAuifKt/outer-vs-inner-misalignment-three-framings-1)). (See also [this paper on gradient starvation](https://arxiv.org/abs/2011.09468) - h/t Quintin Pope; and [the concept of persistence to gradient descent discussed here](https://arxiv.org/pdf/2204.06974.pdf).)
10. A blog post explaining why the “uncertainty” part of CIRL only does useful work insofar as we have an accurate model of the human policy, and why this is basically just as hard as having an accurate model of human preferences.
11. A blog post explaining what practical implications [Stuart Armstrong’s impossibility result](https://arxiv.org/abs/1712.05812) has.
12. As many alignment exercises as possible to help people learn to think about this stuff ([mine aren't great](https://www.lesswrong.com/posts/kj37Hzb2MsALwLqWt/alignment-research-exercises) but I haven’t seen better).
13. A paper properly formulating instrumental convergence, generalization to large-scale goals, etc, as inductive biases in the ML sense (I do this briefly in [phase 3 here](https://www.alignmentforum.org/posts/KbyRPCAsWv5GtfrbG/what-misalignment-looks-like-as-capabilities-scale)).
14. A mathematical comparison between off-policy RL and imitation learning, exploring ways in which they’re similar and different, and possible algorithms in between.
15. A blog post explaining the core argument for why detecting adversarially-generated inputs is likely much easier than generating them, and arguments for why adversarial training might nevertheless be valuable for alignment.
16. A blog post exploring the incentives which models might have when they’re simultaneously trained to make predictions and to take actions in an RL setting (e.g. models trained using [RL via sequence modeling](https://arxiv.org/pdf/2106.01345.pdf)).
17. A blog post exploring pros and cons of making misalignment datasets for use as a metric of alignment (alignment = how much training on the misalignment dataset is needed to make it misaligned).
18. A paper providing an RL formalism in which reward functions can depend on weights and/or activations directly, and demonstrating a simple but non-trivial example.
19. A blog post evaluating reasons to think that [situational awareness](https://www.alignmentforum.org/posts/pRkFkzwKZ2zfa3R6H/without-specific-countermeasures-the-easiest-path-to#A_spectrum_of_situational_awareness) will be a gradual development in models, versus a sharp transition.
20. A blog post explaining reasons to expect capabilities to be correlated with alignment while models lack situational awareness, and then less correlated afterwards, rather than the correlation continuing.
21. A blog post estimating how many bits of optimization towards real-world goals could arise from various aspects of a supervised training program (especially ones which slightly break the cartesian formalisms) - e.g. hyperparameter tuning, many random seeds, training on data generated by other AIs, etc.
22. A sketch of what a model-free version of AIXI would look like (according to one person I talked to, it’s a lot like decision transformers).
23. A blog post evaluating whether [shard theory](https://www.lesswrong.com/posts/xqkGmfikqapbJ2YMj/shard-theory-an-overview) makes sense/makes novel predictions compared with [Steve Byrnes’ model of the brain](https://www.alignmentforum.org/s/HzcM2dkCq7fwXBej8) (he partly explains this in a comment on the post, but I’m still a bit confused).
24. A blog post or paper reviewing what types of feedback humans perform best and worst at (e.g. reward vs value feedback) and then designing a realistic setup for optimal-quality human feedback.
25. A blog post compiling examples of surprising emergent capabilities (especially in large language models).
26. An investigation of the extent to which human concept representations are localized to individual neurons, versus being spread out across different neurons. |
c068ea75-26f8-42d0-8795-9e13e5d776fa | trentmkelly/LessWrong-43k | LessWrong | AI #50: The Most Dangerous Thing
In a week with two podcasts I covered extensively, I was happy that there was little other news.
That is, until right before press time, when Google rebranded Bard to Gemini, released an app for that, and offered a premium subscription ($20/month) for Gemini Ultra.
GEMINI ULTRA IS HERE
I have had the honor and opportunity to check out Gemini Advanced before its release.
The base model seems to be better than GPT-4. It seems excellent for code, for explanations and answering questions about facts or how things work, for generic displays of intelligence, for telling you how to do something. Hitting the Google icon to have it look for sources is great.
In general, if you want to be a power user, if you want to push the envelope in various ways, Gemini is not going to make it easy on you. However, if you want to be a normal user, doing the baseline things that I or others most often find most useful, and you are fine with what Google ‘wants’ you to be doing? Then it seems great.
The biggest issue is that Gemini can be conservative with its refusals. It is graceful, but it will still often not give you what you wanted. There is a habit of telling you how to do something, when you wanted Gemini to go ahead and do it. Trying to get an estimation or probability of any kind can be extremely difficult, and that is a large chunk of what I often want. If the model is not sure, it will say it is not sure and good luck getting it to guess, even when it knows far more than you. This is the ‘doctor, is this a 1%, 10%, 50%, 90% or 99% chance?’ situation, where they say ‘it could be cancer’ and they won’t give you anything beyond that. I’ve learned to ask such questions elsewhere.
There are also various features in ChatGPT, like GPTs and custom instructions and playground settings, that are absent. Here I do not know what Google will decide to do.
I expect this to continue to be the balance. Gemini likely remains relatively locked down and harder to customize or push the enve |
c7b137e7-f6e8-4fa4-86a3-717eb57440a7 | trentmkelly/LessWrong-43k | LessWrong | Tracing Typos in LLMs: My Attempt at Understanding How Models Correct Misspellings
This blogpost was created as a part of the AI Safety Fundamentals course by BlueDot Impact. All of the code can be found on my GitHub.
TLDR: I tried to uncover if there are specific components in language models that enable typo correction. I identified a subword merging head in the first layer of Llama-3.2-1B that plays a crucial role in the process. In this blog post, I’ll walk through my thought process and findings.
Motivation
Large language models are getting significantly more capable every month, but we still don’t know how they work inside. If a model generates an incorrect answer, we currently have (almost) no way of explaining why it did so. Mechanistic interpretability tries to solve this problem. And while we probably won’t be able to completely reverse engineer large language models anytime soon, interpreting smaller parts of the models is still very valuable since these findings often generalize to larger models and provide valuable insights into how they work.
I was reading a lot of mechanistic interpretability papers over the past year, but I have never actually done anything myself. So I finally gave it a try and explored how typo correction works in LLMs.
Initial exploration
I started by manually testing sentences with typos and looking at their attention patterns. I suspected an early attention head might specialize in typo correction. And sure enough, I found one, though it wasn’t exclusively dedicated to typo correction. I found a subword merging head[1]. It moves information between tokens that belong to the same word, but were tokenized into multiple tokens. This happens when a word is rare and lacks a dedicated token or, as in my case, when a typo alters its structure.
Attention head ablation
To test whether this head played a role in typo reconstruction, I tried generating text with this head ablated (I zeroed out its output during the forward pass). Without ablation, when given the prompt:
> "<|begin_of_text|>What is the meaning of |
fc75efe0-d76d-4d66-9994-60c77c5761e6 | trentmkelly/LessWrong-43k | LessWrong | AGI safety from first principles: Conclusion
Let’s recap the second species argument as originally laid out, along with the additional conclusions and clarifications from the rest of the report.
1. We’ll build AIs which are much more intelligent than humans; that is, much better than humans at using generalisable cognitive skills to understand the world.
2. Those AGIs will be autonomous agents which pursue long-term, large-scale goals, because goal-directedness is reinforced in many training environments, and because those goals will sometimes generalise to be larger in scope.
3. Those goals will by default be misaligned with what we want, because our desires are complex and nuanced, and our existing tools for shaping the goals of AIs are inadequate.
4. The development of autonomous misaligned AGIs would lead to them gaining control of humanity’s future, via their superhuman intelligence, technology and coordination - depending on the speed of AI development, the transparency of AI systems, how constrained they are during deployment, and how well humans can cooperate politically and economically.
Personally, I am most confident in 1, then 4, then 3, then 2 (in each case conditional on all the previous claims) - although I think there’s room for reasonable disagreement on all of them. In particular, the arguments I’ve made about AGI goals might have been too reliant on anthropomorphism. Even if this is a fair criticism, though, it’s also very unclear how to reason about the behaviour of generally intelligent systems without being anthropomorphic. The main reason we expect the development of AGI to be a major event is because the history of humanity tells us how important intelligence is. But it wasn’t just our intelligence that led to human success - it was also our relentless drive to survive and thrive. Without that, we wouldn’t have gotten anywhere. So when trying to predict the impacts of AGIs, we can’t avoid thinking about what will lead them to choose some types of intelligent behaviour over others |
9d7652b5-33ef-4307-b18e-19af5ad381bf | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Slowing AI: Foundations
**Variables**
=============
**"Slowing AI" is not monolithic. Many interventions have relevant side effects, positive or negative. There are many considerations, desiderata, and goals adjacent to slowing AI,** and different interventions realize them to different degrees**.** This section is a list and discussion of variables. The goal of this section is to help readers recognize the consequences of possible interventions and thus better evaluate them, and to help generate or optimize interventions by identifying variables to affect. The goal of each subsection is to be an introduction to the relevant variable: how it relates to slowing AI and how else it matters.
**List**
--------
Variables related to slowing AI are:
1. Timeline (to risky AI) length
2. Crunch time length
3. Relative positions of particular (kinds of) actors
4. Multipolarity[[1]](#fnxawx36o4vfs)
5. Attitudes on AI risk (among labs and the ML research community, and also states, other actors, and the public)
6. Technical safety research performed
7. Strategy & governance research performed
8. How much [alignment tax](https://forum.effectivealtruism.org/posts/63stBTw3WAW6k45dY/paul-christiano-current-work-in-ai-alignment) is paid for particular models
9. [Operational adequacy](https://www.lesswrong.com/posts/keiYkaeoLHoKK4LYA/six-dimensions-of-operational-adequacy-in-agi-projects) and safety practices at leading labs
10. Coordination ability and inclination among leading labs
11. The extent to which leading labs could scale up compute
12. How safe the architecture of powerful AI is
13. The extent to which leading labs develop risky vs unthreatening kinds of systems
14. Deployment safety
15. Which (kinds of) actors exert influence over AI
16. West-China relation
17. How well powerful AI will be used
18. Attitudes on AI
19. How various actors feel about the AI safety community
20. AI misuse risks
21. Non-AI risks
22. Field-building, community-building, and influence-gaining
This list aims to be essentially exhaustive.[[2]](#fn0ccfyr2ma0e)
Variables overlap. Maybe this list should instead be a tree or directed acyclic graph to show relationships between variables; the first two sections are weird because they relate to almost all of the other variables, which are mostly discussed in their own subsections.
Many "all else equal"s omitted.
**1. Slower AI progress would be good**
---------------------------------------
Powerful AI appearing later than it will by default would reduce misalignment risk and total [x-risk](https://existential-risk.org/concept.pdf) (on current margins, in expectation, all else equal).
I expect that powerful AI will appear substantially sooner than I would prefer. Reasons include
* labs and researchers have incentives or apparent-myopic-incentives to cause faster progress, independent of what is impartially optimal;
* independent of individuals' preferences, the ML community has a culture of making progress and publishing results;
* by default, AI risk can occur due to unilateral action; and
* Alignment is harder than most AI labs/researchers believe (or their beliefs don't fully coherently translate into preferences or actions) (and moreover they are likely to fail to fully update before their attitudes are catastrophic). (Crucial question: why exactly might labs go too fast by default?)
Slowing AI allows more time for alignment research, field-building, influence-gaining, affecting labs' attitudes, governance interventions, and (sign unclear) governance happening.
Subvariables not discussed elsewhere include inputs to AI, notably:
1. Compute accessibility
1. Cost of compute
2. Regulation
2. Talent
1. Talent pool
2. How talent is allocated between labs and to labs vs places-that-aren't-AI-labs
3. International migration
**2. Time is more valuable near the end**
-----------------------------------------
"Crunch time" is ambiguous; I vaguely mean the periods of strategic clarity, open windows of opportunity, and powerful models existing.[[3]](#fnkfj3ttmldxm)
**Time is more valuable when AI capability-level is greater.** In particular, more time with powerful models
1. Is directly useful for empirical safety research;
2. Informs lots of research by giving more time with better information about what critical models/training will look like and maybe general clearsightedness about AI risk; and
3. Gives more time for advocacy, interventions, and (sign unclear) actors' actions during a particularly high-leverage period for interventions and actors' actions.
(but the magnitude of these benefits, and of side effects of slowing, is cruxy and uncertain).
Slowing AI will be more tractable near the end insofar as greater risk awareness facilitates slowing. It will be less tractable insofar as incentives for AI progress seem greater, AI progress and commercialization has more momentum, more influence is exerted by actors outside the AI safety community, maybe multipolarity among labs and AI-service-builders is greater, and maybe actions require serial time.
Note that capabilities increase after deployment, as people figure out how to use deployed systems and build apps and prompt bureaucracies. And undeploying AI systems is hard.
Some interventions would add time near the end.
Some interventions would slow AI soon but entail less slowing or even negative slowing near the end (e.g., perhaps convincing labs to spend less on compute now makes them increase spending faster later, or more speculatively perhaps slowing the West now causes the West to be more worried about China in the future). (But some interventions would clearly slow AI now and later, e.g. causing labs to publish less.)
Also gives more time to pay the [alignment tax](https://forum.effectivealtruism.org/posts/63stBTw3WAW6k45dY/paul-christiano-current-work-in-ai-alignment) for particular models, but that's its own variable.
Also gives more high-leverage time for normal governance, but the sign of that is unclear.
Related concepts: takeoff speed, emergency brakes.
**3. Beware of differentially slowing safer labs**
--------------------------------------------------
Some interventions would slow some actors more than others.
It would be bad to differentially slow more safety-conscious actors.
It would probably be bad to differentially slow Western relative to Chinese actors:[[4]](#fnjl33bcl8y6m)
* Western labs are better on technical safety in expectation;
* Western labs are ahead (and seem likely to remain ahead) and can better coordinate with other Western labs in expectation;
* If America or the West enact great regulations, it would be better if all of the leading labs were bound by them, which requires all of the leading labs to be American or Western; and
* Western and Chinese labs being closer would lead to worse state involvement in AI research in expectation.
A "safer" lab can cause an existential catastrophe, but it is more likely to slow down near the end, pay the alignment tax for any critical models it develops, and deploy cautiously.
Differentially slowing leading labs is bad because it increases multipolarity (see "Beware of increasing multipolarity" below).
Differentially slowing actors that would use powerful AI well is bad (see "Using powerful AI well" below).
Related subsections: "Operational adequacy," "Focus on slowing risky AI," "Deployment safety," "More time near the end helps labs pay the alignment tax," "Who influences AI."
**4. Beware of increasing multipolarity**
-----------------------------------------
In this context, a world is more "multipolar" if more labs are near the frontier of AI capabilities and lead times between labs are smaller.[[5]](#fnddzggsjldv) Multipolarity is bad because
* It directly causes faster progress;
* It makes coordination more difficult
+ It makes labs less able and willing to pay a given amount of alignment tax,
+ It may make race dynamics stronger but racing is not well understood;
* It increases the probability that a naive actor will take unilateral action that is bad (intentionally or accidentally).
Slowing AI can affect multipolarity
* If multipolarity naturally increases over time;[[6]](#fnot9jckdyrx) or
* By particular interventions differentially slowing leading labs or non-leading labs
+ Through various regulations,
+ By affecting the cost of compute,
+ By affecting the diffusion of ideas, or
+ By affecting private data.
Miscellaneous uncertainties:
* How do various interventions/actions affect multipolarity– e.g. how much does decreasing the diffusion of ideas differentially slow non-leading labs, or how much does a temporary ceiling on training compute differentially slow leading labs?
* How do AI labs proliferate? How is this affected by actions of labs and the US government?
* How far behind is China; when would it build dangerous AI? And how does that depend on actions of Western labs and governments?
* Racing
+ What would happen if there were many more labs pushing the frontier on the path to dangerous AI?
+ How much faster could labs go now?
+ What coordination could happen (especially near the end), and how does multipolarity affect that? E.g. OpenAI + DeepMind + Anthropic agreeing on evals or otherwise paying the alignment tax is easier than many labs coordinating to do so.
Similar-sounding term: "multipolar scenario" or "multipolar takeoff" refers to there being multiple powerful AI systems or AI-enabled actors, while my "multipolarity" refers to the number of labs, including before powerful AI appears.
**5. Risk awareness is good**
-----------------------------
Labs (and researchers) being more concerned about AI risk slows AI.
Labs (and researchers) being more concerned about AI risk improves AI safety (independent of slowing AI).
Slowing AI generally makes labs (and researchers) be more concerned about AI risk, I think, by giving more time for empirical safety work, relevant demonstrations, advocacy, and researchers to realize what's true. But some factors suggest slowing AI would entail less concern about AI risk: prophecies of doom fail to be realized, and separately perhaps labs and researchers will like the AI safety community less (see "Relationships with actors" below).
On researchers' attitudes, see the [2022 Expert Survey on Progress in AI](https://wiki.aiimpacts.org/doku.php?id=ai_timelines:predictions_of_human-level_ai_timelines:ai_timeline_surveys:2022_expert_survey_on_progress_in_ai) (Stein-Perlman et al. 2022) and [AI Risk Discussions](https://www.lesswrong.com/posts/g4nEtPFECTQW9tcff/ai-risk-discussions-website-exploring-interviews-from-97-ai) (Gates et al. 2023). On public opinion, see [Surveys of US public opinion on AI](https://wiki.aiimpacts.org/doku.php?id=responses_to_ai:public_opinion_on_ai:surveys_of_public_opinion_on_ai:surveys_of_us_public_opinion_on_ai) (Stein-Perlman 2022). Note that survey responses seem to miss some aspects of respondents' attitudes.
The leaders of OpenAI, DeepMind, and Anthropic seem pretty reasonable as far as I know.[[7]](#fnv1b4tpdwwd)
### **Warning signs are mostly good**
A "warning sign" (or "warning sign for actor X") is an event or process that causes many people to be worried about AI accidents (or that makes AI risk legible or salient to actor X).
Warning signs would mostly help slow AI and have mostly positive side effects for AI safety. But they're very non-monolithic.
Some events could make AI power legible but not make (reasonable threat models for) AI risk legible.
Perhaps some events could make AI risk legible to some actors but not others, in particular to labs and the ML research community but not states or the public. Risk awareness is multidimensional.
Warning signs seem to mostly help slow AI by causing actors to slow down and to slow each other down. But insofar as they make AI power more salient, they could cause actors to be more excited about powerful AI.
Warning signs may also make labs more risk-aware or more incentivized to appear risk-aware.
Slowing AI makes warning signs more likely: it gives more time for
* A warning sign to occur in general,
* Moderately-powerful systems to be integrated into applications and for AI products/services to proliferate, and
* AI safety research and other work to make AI risk more legible.
Note that trying to do something that looks harmful has huge negative consequences for your ability to influence people. Scary demos are acceptable at least in some cases; I don't know when they are effective.[[8]](#fnau6h5udcipc)
Related concepts: [warning shot](https://forum.effectivealtruism.org/topics/warning-shot), galvanizing event, wakeup (to powerful AI), [fire alarm](https://www.lesswrong.com/posts/BEtzRE2M5m9YEAQpX/there-s-no-fire-alarm-for-artificial-general-intelligence), meme.[[9]](#fn9eoconn4sa)
**6. More time (especially near the end) helps safety research**
----------------------------------------------------------------
More time implies more technical safety research. Moreover:
* **Research quality.** Much of technical safety research grows more effective as
+ More powerful AI models become available and
+ We gain more clarity about what very powerful AI systems will look like
+ (and maybe over time as alignment research matures as a field and more serial research is completed, but on the other hand low-hanging fruit is taken over time).
* **Research rate.** Research done per time grows over time insofar as the field grows. And (related to some interventions to slow AI) if AI safety is more legible (or prestigious, if the right concept is prestigious) it will receive more work. (And maybe legibility and prestige naturally increase over time, but we can lump that with the rest of the exogenous growth of research rate.) (Note that insofar as [research is not parallelizable and serial time is necessary](https://www.lesswrong.com/posts/vQNJrJqebXEWjJfnz/a-note-about-differential-technological-development), research rate is less important.)
**7. More time (especially near the end) helps strategy/governance research & interventions**
---------------------------------------------------------------------------------------------
This variable is similar to the previous one. But strategy/governance research and interventions are aided near the end mostly by strategic clarity rather than the existence of powerful models. And note that technical safety research is ~directly valuable while governance research requires intervention to create value.
New resources, affordances, or windows of opportunity will probably appear during crunch time. We can plan for them to some extent in advance.
Work on slowing AI causes slowing, and slowing gives more time for work on slowing. But this cycle is not very important because the multiplication factor seems small.
**8. More time near the end helps labs pay the alignment tax**
--------------------------------------------------------------
To make an AI model safe, under one frame, the developer of the model must pay the [alignment tax](https://forum.effectivealtruism.org/posts/63stBTw3WAW6k45dY/paul-christiano-current-work-in-ai-alignment). If there is more time near the end, the developer is able and willing to pay more and is thus more likely to successfully pay the alignment tax.
**9. Operational adequacy**
---------------------------
[Operational adequacy](https://www.lesswrong.com/s/v55BhXbpJuaExkpcD/p/keiYkaeoLHoKK4LYA) is directly good for AI safety. If it naturally improves over time (which seems uncertain), slowing AI improves it and so it is an additional consideration in favor of slowing AI. Additionally, improving some dimensions of operational adequacy would slow AI.
**10. Coordination is good**
----------------------------
Some aspects of AI development are analogous to the [prisoner's dilemma](https://en.wikipedia.org/wiki/Prisoner%27s_dilemma). Perhaps in some situations, a lab would not want to slow down unilaterally, but would prefer everyone slowing down to nobody doing so. Improving coordination means helping labs slow down together (mostly by helping them slow down together when they want to; also by helping them understand the situation they're in and want to slow down). It is good for coordination to be easier and more likely.
Coordination ability includes:
* Commitment ability / verification ability / transparency;
* Good coordination mechanisms having been discovered/proposed;
* Relevant coordination methods being legal (and sufficiently clearly legal that the fear of illegality doesn't dissuade coordination);[[10]](#fn8lllpcew5yj) and
* Perhaps the inverse of multipolarity, although that could be treated separately (as part of the difficulty of coordination).
Increasing coordination ability or inclination increases labs' ability to slow (risky) AI.
How slowing affects coordination is unclear. Perhaps slowing naturally increases coordination ability and inclination. Perhaps slowing also naturally increases multipolarity, making coordination more difficult. Additionally, slowing causes *who leading labs and relevant actors are* to be more resampled from the distribution of possibilities, which is bad insofar as the current landscape of leading labs and relevant actors is relatively good.
Related: transparency, coordination mechanisms, benefit-sharing mechanisms, arms control treaties.
**11. Quickly scaling up compute**
----------------------------------
If labs can quickly scale up compute of the largest training runs, they might be able to quickly make much more powerful systems; progress might be fast. Insofar as slower progress is better, it would be better if labs couldn't quickly scale up compute much. So if an intervention increases labs' ability to quickly scale up compute later, that's bad.
Something like this variable is often assumed to be a major determiner of the length of crunch time. Insofar as labs are trying to go fast near the end, it is important (and if it's sufficiently large, more labs could be relevant, increasing effective multipolarity).
It is also commonly suggested that smoother and more gradual progress is safer for reasons other than extending crunch time.
Some policy regimes aimed at slowing AI would limit training compute. If such regimes were reversed or evaded, training compute could scale up quickly.[[11]](#fnfwnkn34ccgl)
If increasing capabilities today didn't shorten timelines, it would be (at least prima facie) good because it gives more time with more powerful models and better strategic and technical clarity. Unfortunately, increasing capabilities today generally shortens timelines. (But I don't think anyone has a great model of what drives long-term AI progress.)
Related concept: "compute overhang" or "hardware overhang" (there is no consensus definition of these terms[[12]](#fn4fjsely3qtn)).
Equivalent variable: how much more compute labs could use if they chose to. Similar variable: how much more compute labs would use if they believed very powerful AI was near.
**12. Architectures safety**
----------------------------
Advanced cognitive capabilities seem likely to be achieved by machine learning. Within machine learning, some paths would be safer than others. (Or they could be achieved through whole-brain emulation, neuromorphic AI, brain-computer interfaces (unlikely), or just genetic engineering.)
Some slowing interventions would differentially accelerate certain paths to advanced cognitive capabilities.
It might be good to differentially accelerate certain paths. I don't know which paths, though. It depends on their safety and relationships between paths (in particular, whole-brain emulation largely overlaps with neuromorphic AI, so accelerating whole-brain emulation would accelerate neuromorphic AI).
**13. Focus on slowing*****risky*** **AI**
------------------------------------------
Some kinds of systems at various levels of abstraction (e.g. perhaps large language models, reinforcement learning agents, models that demonstrate scary capabilities in an evaluation, and agents) are risky (on various threat models). Most systems are much less risky.
Some interventions differentially slow progress on risky AI.
Slowing risky AI could occur due to labs' choices (for safety reasons or for prestige), or researchers' choices (especially if united or with strong affordances for collective action), or being imposed by government, or being incentivized in various ways.
Ideally, [safety standards](https://forum.effectivealtruism.org/posts/zvbGXCxc5jBowCuNX/how-technical-safety-standards-could-promote-tai-safety) could prevent actors from doing risky things while minimally impairing their ability to do safe things.
Related concept: differential technological development.[[13]](#fnpl0nea2ab6b)
**14. Deployment safety**
-------------------------
Some ways that powerful AI systems could be deployed would be safer. Deployment safety is affected by decisions within leading labs, their relationships, and properties-of-the-world like the number of leading labs.[[14]](#fn0mncnb9d43y) Slowing AI, and particular interventions related to slowing AI, could affect deployment safety in predictable ways.
**15. Who influences AI**
-------------------------
Maybe it is better for some (kinds of) actors to exert more influence/control over AI, e.g. because some labs are safer than others.
Maybe there are side effects when some (kinds of) actors exert more influence over AI, e.g. the West acting differently if China exerts more influence.
Maybe interventions (e.g., AI risk advocacy to government) affect which (kinds of) actors exert influence over AI. This matters insofar as it affects safely directly and insofar as it slows AI.
Maybe some (kinds of) actors (e.g., states) naturally exert more influence over AI over time, so slowing AI causes them to exert more influence. On the other hand, maybe some kinds of slowing cause some (kinds of) actors (e.g., states) to be slower to wake up to AI.
This variable overlaps with "attitudes on AI" and "deployment safety" and "how well powerful AI will be used" and "warning signs."
**16. West-China relation**
---------------------------
AI development is led by the Western cultural and political sphere. The other sphere reasonably likely to be able to develop powerful AI soon is China. This situation is likely to continue for at least a decade.
This variable is affected by the capabilities difference between the leading Western labs vs the leading Chinese labs, the levers that the West and China have over each other's AI development, and the attitudes of actors (especially states and labs) about the other sphere. A generalization of this variable would be the set of relations between all such spheres.
Related: it seems easier for two Western labs to coordinate than a Western lab and a Chinese lab.
West vs China lead can be affected by slowing-interventions. For example, US export controls differentially slow China and US domestic regulation differentially slows the US.
The West vs China relation (including lead but also attitudes) affects slowing by affecting incentives/racing and by affecting state action.
**17. Using powerful AI well**
------------------------------
It would be better if powerful AI is ultimately used better. This can be affected by affecting who controls powerful AI or by affecting how wise, well-informed, etc. whatever actors will control powerful AI will be.
Particular slowing-related interventions can affect these variables. Slowing also naturally affects the (kinds of) actors that will control powerful AI (e.g., maybe states tend to be more involved over time). Slowing might also naturally affect attitudes on AI in particular ways.
**18. Attitudes on AI**
-----------------------
The attitudes of relevant actors, particularly states, are important.
Government involvement in AI is multidimensional; some possible actions would clearly be good. Regardless of the overall level or sign of government involvement, presumably it can be nudged to do better things.
Advocacy to government has several downside risks.[[15]](#fnnno6qp0ngq) In particular, you may end up convincing people about AI capabilities more than risk; perhaps "AI is really powerful" is an easy-to-understand, necessary part of the case for "AI is really risky."
Insofar as we don't know what it would be good for states to do on AI risk, advocacy to government has limited upside (although perhaps it will later be useful to have done, or perhaps government can generate good things to do).
I feel uncertain about the effects of advocacy:
* I feel uncertain about how advocacy would affect attitudes and
* I feel uncertain about behavior as a function of attitudes.
This variable is related to slowing AI because
* Attitudes affect slowing,
* Slowing generally affects attitudes insofar as they naturally change over time (in particular, actors appreciate AI more and and want to get more involved), and
* Particular slowing-related interventions affect attitudes in particular ways.
Related concept: wakeup (to powerful AI).
**19. Relationship with actors**
--------------------------------
Relevant actors liking the AI safety community is good insofar as it improves AI safety.
Some interventions would make labs and researchers like the safety community less, like perhaps:
* Interventions that don't seem justified to labs and researchers as slowing scary AI (especially interventions that don't differentially slow scary AI),
* Interventions that are adversarial or norm-violating, or
* Interventions around a public campaign to slow AI that involves AI safety people and is unpopular with labs and researchers.
I don't know what affects states' attitudes on the AI safety community.
A related framing of this variable is: how events make various groups look, particularly AI safety vs other groups trying to affect the future of AI.
The AI safety community is not perceived as monolithic; an actor could have different attitudes on different parts or aspects of it.
**20. Slowing AI increases AI misuse risks**
--------------------------------------------
Slowing AI causes there to be more time for AI misuse to occur before powerful AI can solve that.
This is important
* Directly,
* Insofar as AI misuse is a risk factor for other risks, and
* Insofar as AI misuse (or fear of it) affects attitudes on AI.
**21. Slowing AI increases non-AI risks**
-----------------------------------------
Non-AI risks are mostly technological risks and great power conflict. Slowing AI causes there to be more time for such risks to occur before powerful AI can solve them (and improving technology generally seems to be increasing risk-per-time, if anything). And in particular, slowing AI may increase tension between great powers.
This is important directly and insofar as non-AI risk is a risk factor for other risks.
**22. Field-building, community-building, influence-gaining**
-------------------------------------------------------------
AI safety field-building, community-building, and influence-gaining may help slow AI.
Slowing AI gives more time for AI safety field-building, community-building, and influence-gaining. However, the influence exerted by others over AI increases over time (see "Who influences AI" above).
**Miscellanea**
---------------
Many of these variables are related to various definitions of "takeoff speed." There is no consensus definition of "takeoff speed." Many takeoff-related variables are largely exogenous, but slowing-related interventions might affect
* What labs want,
* What systems labs are allowed to train or deploy,
* Labs' ability to scale up compute (or maybe other inputs),
* Coordinated slowing,
* How much research happens,
* How efficiently research turns into progress-acceleration, or maybe
* [Homogeneity](https://www.lesswrong.com/posts/mKBfa8v4S9pNKSyKK/homogeneity-vs-heterogeneity-in-ai-takeoff-scenarios) in takeoff.
Takeoff-related variables affects many variables– for example, time to powerful AI, how valuable that time is, paying the alignment tax, racing, multipolarity, misuse, attitudes, risk awareness, and government actions.
Miscellaneous considerations for interventions include:
* Option value, irreversibility, & lock-in;
* Information value;
* Attention hazard; and
* Reputation hazard & backlash.
**Frames**
==========
This section is about analytic frames on slowing AI.[[16]](#fn0b2rihswj8) If a thing is interesting and surprising, it can be useful to focus on that thing and investigate its implications for what's true, what to think more about, and how to organize your thoughts. **This section is about*****approaches or perspectives to try on*** **and*****possible things to pay attention to*****, not*****asserting what's true*****. A good frame is one that is*****useful*****.**
The goal of this section is to give readers a variety of sometimes-useful frames to try on.
**Extend**~~**crunch time**~~ **the periods of strategic clarity, open windows of opportunity, and almost-dangerous models**[[17]](#fnmhb9yieb7ak)
--------------------------------------------------------------------------------------------------------------------------------------------------
Near the end, we may have
1. General strategic clarity, informing research and interventions;[[18]](#fn5yi80kh42x3)
2. Open windows of opportunity for interventions;
3. Better ability to do technical safety research.
So increasing time near the end—and more specifically, time with those properties—is quite valuable in expectation. This frame prompts consideration: what are necessary and sufficient conditions for progress stopping near the end? What determines how valuable time near the end is?
[Another angle on this frame](https://www.lesswrong.com/posts/uFNgRumrDTpBfQGrs/let-s-think-about-slowing-down-ai?commentId=pgMaKR4Ru2aREr2jo) is that a big goal is to "lay the groundwork for slowing down in the future, when extra time is most needed."
**Slowing seems good; side effects are mixed**
----------------------------------------------
Slowing AI is good. Slowing AI and lots of interventions aimed at slowing AI have super-important side effects, sometimes more important than their effect on slowing AI. And "slowing AI" is not monolithic; slowing progress now has different effects from slowing progress later. To choose and optimize interventions, we should find slowing-adjacent variables (or desiderata or considerations) and understand their interactions and upshots.
Related proposition and questions that could provoke more thoughts: a longer timeline to powerful AI is better on the margin. What are the positive and negative factors? What timeline would be optimal?
Related proposition and questions that could provoke more thoughts: labs go too fast by default. Why does going-too-fast occur? What determines how labs behave? What can we intervene on to escape the bad default state? And what are the inputs to AI progress (at various levels of abstraction)?
**~Nobody wants to destroy the world**
--------------------------------------
Approximately nobody wants to destroy the world. "If the Earth is destroyed, it will probably be by mistake."[[19]](#fnpsm89yv9gla) But someone accidentally destroying the world would also be *weird* in a way that is worth *really noticing*. The possibility suggests a very natural intervention: helping the people who might accidentally destroy the world understand their actions.
Slowing AI doesn't have to involve conflict with AI labs: it can be about helping labs promote everyone's preferences by giving them information and tools to help them slow down.
What would be possible if there was much greater risk awareness?
This frame is somewhat plan-y; it inherently suggests a particular theory of victory and a particular class of plans.
**Ignorance, externalities, culture, and racing**
-------------------------------------------------
Three simple factors make labs go too fast: labs misunderstanding or underestimating AI risk, the negative externality[[20]](#fnmq9282y0o6) of AI catastrophe, and researchers' and labs' culture of making progress and publishing results. Additionally, there seems to be a "racing" phenomenon where a lab tries to make faster progress if it has competitors; this phenomenon is poorly understood.
This frame identifies causes of going-too-fast and so suggests that interventions target those causes.
**Risk awareness might improve a lot**
--------------------------------------
Risk awareness and attitudes on AI safety among labs and ML researchers currently seems poor. It may be much better in the future, and there will be tractable opportunities to make it much better, as AI gets more powerful, there's more safety research, and there's more research focused on making AI risk legible to labs and researchers.
Greater risk awareness seems to facilitate labs slowing down near the end and to create new opportunities for helping them do so. (And independent of slowing AI, it seems to facilitate labs paying the alignment tax.)
Respondents to the [2022 Expert Survey on Progress in AI](https://wiki.aiimpacts.org/doku.php?id=ai_timelines:predictions_of_human-level_ai_timelines:ai_timeline_surveys:2022_expert_survey_on_progress_in_ai) (Stein-Perlman et al. 2022) were surprisingly pro-safety and doomy.[[21]](#fn4kza6mxcks6) But clearly their attitudes are insufficient to change much.
**Incentives**[[22]](#fnezq1xo3kuwp)
------------------------------------
Buy-in from labs helps slow AI. So focus on interventions and actions-for-labs with lots of safety per cost-to-labs. And look for ways to align labs' incentives with the world and to incentivize labs to increase safety. Make sure labs are left/given options that they want to take and are safe.
Consider a diagram with x-axis time, y-axis profitability, and two curves representing profitability over time in the two scenarios *default* and *slow AI*. The integral of the difference is the cost to labs of slowing. To get companies to accept slowing, it is necessary and sufficient to incentivize those companies to accept that cost.
I mostly don't endorse the previous paragraph. I think it might constrain one's thinking because roughly
* Labs care about some long-term goal rather than the integral of profitability
+ So maybe replace "profitability" with "preference (under default preferences)"
* Some interventions look like changing labs' preferences
+ Especially interventions like informing labs about AI risk
* Some interventions look like preparing to slow down later
**Slow*****risky*** **AI**
--------------------------
Slowing AI is good insofar as it slows risky AI. Slowing all AI is unnecessary; the real goal is slowing risky AI. This frame raises the question: what does that observation imply for interventions?
**Prevent risky AI from being*****trained***
--------------------------------------------
The condition *nobody trains risky systems* is sufficient for victory, and on some views it is ~necessary. It brings to mind different problems, goals, variables, considerations, desiderata, levers, affordances, and interventions than *slow AI* or even *slow risky AI*, so it can be a useful reframing.
For example, it could let you identify big goals as *develop the capacity to identify all risky training runs* and *cause all risky training runs to be identified and stopped*.
This frame is very plan-y; it inherently suggests a particular theory of victory and a particular class of plans.
**Focus on preparing for crunch time**
--------------------------------------
Windows of opportunity will be much more open in the future due to strategic clarity, various actors' risk awareness, more powerful models, etc. So for now you should focus on planning and gaining influence/resources/etc., not intervening. What will be different in the future? What opportunities could you have in the future, if you prepare well?
I think this frame is less valid in the context of slowing AI than for AI strategy in general, but may still be useful.
**AI inputs increase fast**
---------------------------
Effective training compute for the largest training runs increases quickly,[[23]](#fn3abfzt2oavx) so decreasing inputs on the margin doesn't buy much time. Buying substantial time requires more discrete changes.
**Unilateralism**
-----------------
The problem is that many actors will be able to unilaterally end the world. The solution is to decrease the number of decisions that would end the world if done wrong (and influence those decisions).
1. **[^](#fnrefxawx36o4vfs)**Roughly meaning the number of labs near the frontier of capabilities, not the number of powerful AI(-enabled) actors.
2. **[^](#fnref0ccfyr2ma0e)**Or rather, it aims to be exhaustive at a certain level of abstraction. For example, it doesn't include *how good the AI regulatory regime is*, but I think it can give an account of that in terms of more basic variables.
3. **[^](#fnrefkfj3ttmldxm)**See Michael Aird's "Rough notes on 'crunch time'" (unpublished).
4. **[^](#fnrefjl33bcl8y6m)**Holding China constant, is it better for the West to be faster or slower than the default? It's complicated. Ideally the West would be faster now and much slower later. But since we can't just choose to slow down near the end we should be wary of causing speeding now. Speeding now is generally net-negative by costing time without buying time near the end, although speeding-now interventions that cause proportionate slowing-near-the-end could be good.
5. **[^](#fnrefddzggsjldv)**For some purposes, what's relevant is an analogue of multipolarity that includes groups making LLM applications in addition to labs making frontier models.
6. **[^](#fnrefot9jckdyrx)**It seems to have grown historically, and maybe in the future small groups without much compute can use others' training-compute-intensive tools to make powerful stuff. If
1. big training runs reach more like $10B soon without tech companies getting much richer or AI being super commercially applicable *and*
2. models are somehow not deployed in a way that lets others build on them,
that would limit multipolarity.
7. **[^](#fnrefv1b4tpdwwd)**Of course seeming reasonable is instrumentally convergent, and I don't know much about them, but I'm not aware of much evidence of unreasonableness. Sam Altman is sometimes vilified but I think it's more productive to understand him as *a human who probably has some incorrect beliefs and is currently making some small mistakes and might someday make large mistakes but is likely to avoid mistakes-foreseeable-by-the-AI-safety-community if that community successfully informs labs and the ML research community about AI risk*. I'm also open to evidence of unreasonableness, of course.
8. **[^](#fnrefau6h5udcipc)**One somewhat scary demo that failed to change anything (as far as I know) was [Dual use of artificial-intelligence-powered drug discovery](https://www.nature.com/articles/s42256-022-00465-9), which *ex ante* would have seemed to be a strong candidate for a wakeup call for AI misuse risk.
Separately, note that some people trying to do scary demos may advance the frontier of AI capabilities in general or particular capabilities that cause harm if misused. They may also increase AI hype. They may also draw attention to AI risk in a way that leads to an unproductive response (e.g. focused on LAWS or misuse), distracting from more important threat models and governance challenges.
9. **[^](#fnref9eoconn4sa)**See generally Michael Aird's "Warning shots, galvanizing events, etc." (unpublished).
10. **[^](#fnref8lllpcew5yj)**Antitrust law is often mentioned as maybe making some relevant coordination illegal; I do not have a good sense of what antitrust law prohibits regarding AI labs (or what labs believe about what is legal).
11. **[^](#fnreffwnkn34ccgl)**Suppose for illustration that by default a lab would 10x training compute every year, but it is artificially capped at 2x. After time under this policy, the lab would want to scale up faster, to 'catch up' to where it would have been by default. (But not all the way– the 10x in this illustration bakes in positive endogeneities of AI progress, which are mostly not relevant to a lab 'catching up.') If
1. scaling up fast is technically feasible *and*
2. the policy is reversed, or the lab is able and willing to violate the policy, or the lab leaves the policy's jurisdiction,
then the lab will scale up training compute by much more than 10x per year until it has largely 'caught up.' Otherwise, a compute overhang doesn't cause faster progress.
12. **[^](#fnref4fjsely3qtn)**See Zach Stein-Perlman's [Taboo "compute overhang"](https://www.lesswrong.com/posts/icR53xeAkeuzgzsWP/taboo-compute-overhang) (2023). Regardless, most definitions are not very analytically useful or decision-relevant. As of April 2023, the cost of compute for an LLM's final training run is [around](https://colab.research.google.com/drive/1O99z9b1I5O66bT78r9ScslE_nOj5irN9?usp=sharing) $40M. This is tiny relative to the value of big technology companies, around $1T. I expect compute for training models to increase dramatically in the next few years; this would cause *how much more compute labs could use if they chose to* to decrease.
I notice I am confused:
- I think final training runs have only cost <$100M, why hasn't anyone done a $1B training run?
- Why hasn't training compute scaled up much since AlphaGo Zero or GPT-3? See Epoch's [Compute Trends Across Three Eras of Machine Learning](https://epochai.org/blog/compute-trends) (2022).
13. **[^](#fnrefpl0nea2ab6b)**See Jonas Sandbrink et al.'s [Differential Technology Development: A Responsible Innovation Principle for Navigating Technology Risks](https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4213670) (2022) and Michael Aird's ["Slower tech development" can be about ordering, gradualness, or distance from now](https://forum.effectivealtruism.org/posts/ujRGGBxJN9AHXfzJe/slower-tech-development-can-be-about-ordering-gradualness-or) (2021).
14. **[^](#fnref0mncnb9d43y)**See Holden Karnofsky's [Nearcast-based "deployment problem" analysis](https://www.lesswrong.com/posts/vZzg8NS7wBtqcwhoJ/nearcast-based-deployment-problem-analysis) (2022) and [Racing through a minefield: the AI deployment problem](https://www.lesswrong.com/posts/PABtHv8X28jJdxrD6/racing-through-a-minefield-the-ai-deployment-problem) (2022).
15. **[^](#fnrefnno6qp0ngq)**See Alex Lintz's "AGI risk advocacy" (unpublished).
16. **[^](#fnref0b2rihswj8)**"An analytic *frame* is a conceptual orientation that makes salient some aspects of an issue, including cues for what needs to be understood, how to approach the issue, what your goals and responsibilities are, what roles to see yourself as having, what to pay attention to, and what to ignore." Zach Stein-Perlman's [Framing AI strategy](https://www.lesswrong.com/posts/AGrBjEb6vFaJBwjTj/framing-ai-strategy-1) (2023).
17. **[^](#fnrefmhb9yieb7ak)**See Michael Aird's "Rough notes on 'crunch time'" (unpublished).
18. **[^](#fnref5yi80kh42x3)**On the other hand, the world will likely be more confusing, weird, and quickly-changing.
19. **[^](#fnrefpsm89yv9gla)**Eliezer Yudkowsky's [Cognitive Biases Potentially Affecting Judgment of Global Risks](https://intelligence.org/files/CognitiveBiases.pdf) (2008).
20. **[^](#fnrefmq9282y0o6)**Negative externalities cause an activity to happen too much since some of the costs are borne by others.
21. **[^](#fnref4kza6mxcks6)**Respondents see "Stuart Russell's problem" as hard and important. Median P(doom) is 5–10% depending on how you ask. 69% of respondents want more priority on AI safety research (up from 49% in [the 2016 survey](https://jair.org/index.php/jair/article/view/11222)). See the survey for relevant definitions.
22. **[^](#fnrefezq1xo3kuwp)**Inspired but not necessarily endorsed by Alex Gray.
23. **[^](#fnref3abfzt2oavx)**Training compute for the largest training runs seems not to have increased much in the recent past; there has not been much growth since AlphaGo Zero in 2017. See Epoch's [Compute Trends Across Three Eras of Machine Learning](https://epochai.org/blog/compute-trends) (2022). But it seems likely to increase quickly in the future. |
e8fe5869-d2fa-4273-a7e5-e9bbf24bd978 | trentmkelly/LessWrong-43k | LessWrong | How is ARC planning to use ELK?
Let's say we arrive at a worst-case solution for ELK, how are we planning to use it? My initial guess was that ELK is meant to help make IDA viable so that we may be able to use it for some automated alignment-type approach. However, this might not be it. Can someone clarify this? Thanks. |
1503b8c8-3044-4c91-a48f-777adcfb1499 | trentmkelly/LessWrong-43k | LessWrong | [LINK] "Straight and crooked thinking," by Robert H. Thouless
This book from 1930 or so (PDF) seems chiefly concerned with describing ways thinking can go wrong. The author does not seem to know much of Bayes, but his book appears largely sound. It ends with practical suggestions for dealing with various "dishonest tricks" -- though the list assumes these tricks come from someone other than the reader -- and an imaginary conversation between advocates of different political views who all exhibit flawed thinking. People who sympathize with one side or another can get practice taking apart a bad argument for that side.
Found via recommendation from Answers in Genes. |
6d4d7670-c7b6-49cf-9fcc-58043c9cccb7 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle
1 Introduction
---------------
In this work, we extend the Quantum-Like Bayesian Network previously proposed by Moreira and Wichert ([2014](#bib.bib19), [2016](#bib.bib20)) by incorporating the framework of expected utility. This extension is motivated by the fact that quantum-like models tend to explain the probability distributions in several decision scenarios where the agent (or the decision-maker) tends to act irrationally (Busemeyer and Bruza, [2012](#bib.bib3); Bruza et al., [2015](#bib.bib2)). By irrational, we mean that an individual chooses strategies that do not maximise or violate the axioms of expected utility. It is not enough to know these probability distributions. On the contrary, it would be desirable to use this probabilistic information to help us act upon a real world decision scenario. For instance, if a patient has cancer, it is not enough for a doctor to know the probability distribution of success of different treatments. The doctor needs act and choose a treatment based on specific information about the patient and how this treatment will affect him/her. Probabilistic models are used in tasks that reason under uncertainty, in other words, they are models that reach a conclusion based on partial evidence. Decision-making models such as the expected utility hypothesis, are used to decide how to act in the world. The main problem with such decision-making models is that it is very challenging to determine the right action in a decision task where the outcomes of the actions are not fully determined (Koller and Friedman, [2009](#bib.bib14)). For this reason, we suggest to extend the previously proposed Quantum-Like Bayesian Network to a Quantum-Like Influence diagram where we take into account both the quantum-like probabilities (incorporating quantum interference effects) of the various outcomes and the preferences of an individual between these outcomes.
Generally speaking, an Influence diagram is a compact directed acyclical graphical representation of a decision scenario originally proposed by Howard and Matheson ([1984](#bib.bib9)) which consists in three types of nodes: random variables (nodes) of a Bayesian Network, action nodes representing a decision that we need to make, and an utility function. The goal is to make a decision, which maximises the expected utility function by taking into account probabilistic inferences performed on the Bayesian Network. However, since influence diagrams are based on classical Bayesian Networks, then they cannot cope with the paradoxical findings reported over the literature.
It is the focus of this work to study the implications of incorporating Quantum-Like Bayesian Networks in the context of influence graphs. By doing so, we are introducing quantum interference effects that can disturb the final probability outcomes of a set of actions and affect the final expected utility. We will study how one can use influence diagrams to explain the paradoxical findings of the prisoner’s dilemma game based on expected utilities.
2 Revisiting the Prisoner’s Dilemma and the Expected Utility Hypothesis
------------------------------------------------------------------------
The Prisoner’s Dilemma game consists in two players who are in two separate confinements with no means of communicating with each other. They were offered a deal: if one defects against the other, he is set free while the other gets a heavy charge. If they both defect, they get both a big charge and if they both cooperate by remaining silent, they get a small charge. Figure [1](#S2.F1 "Figure 1 ‣ 2 Revisiting the Prisoner’s Dilemma and the Expected Utility Hypothesis ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle") shows an example of a payoff matrix for the Prisoner’s Dilemma used in the experiments of Shafir and Tversky ([1992](#bib.bib25)) where the goal is to score the maximum number of points.
Figure 1: Example of a payoff matrix used in the Shafir and Tversky ([1992](#bib.bib25)) Prisoner’s Dilemma experiment
Looking at the payoff matrix, one can see that the best action for both players is to cooperate, however experimental findings show that the majority of the players choose to defect even when it is known that the other player chose to cooperate. The Prisoner’s Dilemma is a clear example of how two perfectly rational individuals choose to defect (they prefer an individual reward), rather than choosing the option which is best for both (to cooperate). The expected utility hypothesis is a framework that enables us to explain why this happens.
The expected utility hypothesis corresponds to a function designed to take into account decisions under risk. It consists of a choice of a possible set of actions represented by a probability distribution over a set of possible payoffs (von Neumann and Morgenstern, [1953](#bib.bib28)). It is given by Equation [1](#S2.E1 "(1) ‣ 2 Revisiting the Prisoner’s Dilemma and the Expected Utility Hypothesis ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle"),
| | | | |
| --- | --- | --- | --- |
| | EU=∑iPr(xi)⋅U(xi), | | (1) |
where U(xi) is an utility function associated to event xi.
In the experiment of Shafir and Tversky ([1992](#bib.bib25)), the participant needed to choose between de actions defect or cooperate. We will address to this participant as player 2, P2 and his opponent, to player 1, P1. According to the expected utility hypothesis, P2 would have to choose the action that would grant him the highest expected utility. Assuming that we do not know what P1 chose (so we model this with a neutral prior of 0.5), we can compute the expected utility of Player 2 as
| | | |
| --- | --- | --- |
| | EU[Defect]=0.5×U(P1=D,P2=D)+0.5×U(P1=C,P2=D)=57.5, | |
| | | |
| --- | --- | --- |
| | EU[Cooperate]=0.5×U(P1=D,P2=C)+0.5×U(P1=C,P2=C)=50. | |
Note that U(P1=x,P2=y) corresponds to the utility of player 1 choosing action x and player 2 choosing action y. The calculations show that the action that maximises the player’s expected utility is Defect. This is what it is known as the Maximum Expected Utility hypothesis (MEU).
In the end of the 70’s, Daniel Kahneman and Amos Tversky showed in a set of experiments that in many real life situations, the predictions of the expected utility were completely inaccurate (Tversky and Kahneman, [1974](#bib.bib26); Kahneman et al., [1982](#bib.bib12); Kahneman and Tversky, [1979](#bib.bib13)). This means that a decision theory should be predictive in the sense that it should say what people actually do choose, instead of what they must choose. The Prisoner’s Dilemma game is one of the experiments that show the inaccuracy of the expected utility hypothesis by showing violations to the laws of classical probability. Table [1](#S2.T1 "Table 1 ‣ 2 Revisiting the Prisoner’s Dilemma and the Expected Utility Hypothesis ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle") summarises the results of several words of the literature reporting violations to the total law of classical probability. All of these works tested three conditions in the Prisoners Dilemma Game: (1) the player knows the other defected ( Known to Defect), (2) the player knows the other cooperated (Known to Collaborate), (3) the player does not know the other player’s action (Unknown). This last condition shows a deviation from the classical probability theory, suggesting that there is a significant percentage of players who are not acting according to the maximum expected utility hypothesis.
| Literature | Known to Defect | Known to Collaborate | Unknown | Classical Probability |
| --- | --- | --- | --- | --- |
| Shafir and Tversky ([1992](#bib.bib25)) | 0.9700 | 0.8400 | 0.6300 | 0.9050 |
| Li and Taplin ([2002](#bib.bib18)) (Average) | 0.8200 | 0.7700 | 0.7200 | 0.7950 |
| Li and Taplin ([2002](#bib.bib18)) Game 1 | 0.7333 | 0.6670 | 0.6000 | 0.7000 |
| Li and Taplin ([2002](#bib.bib18)) Game 2 | 0.8000 | 0.7667 | 0.6300 | 0.7833 |
| Li and Taplin ([2002](#bib.bib18)) Game 3 | 0.9000 | 0.8667 | 0.8667 | 0.8834 |
| Li and Taplin ([2002](#bib.bib18)) Game 4 | 0.8333 | 0.8000 | 0.7000 | 0.8167 |
| Li and Taplin ([2002](#bib.bib18)) Game 5 | 0.8333 | 0.7333 | 0.7000 | 0.7833 |
| Li and Taplin ([2002](#bib.bib18)) Game 6 | 0.7667 | 0.8333 | 0.8000 | 0.8000 |
| Li and Taplin ([2002](#bib.bib18)) Game 7 | 0.8667 | 0.7333 | 0.7667 | 0.8000 |
Table 1: Works of the literature reporting the probability of a player choosing to defect under several conditions. The entries of the table that are highlighted correspond to experiments where the violations of the sure thing principle were not found.
Table [1](#S2.T1 "Table 1 ‣ 2 Revisiting the Prisoner’s Dilemma and the Expected Utility Hypothesis ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle") presents several examples where the principle of maximum expected utility is not, in general, an adequate descriptive model of human behaviour. In fact, people are often irrational, in the sense that their choices do not satisfy the principe of maximum expected utility relative to any utility function (Koller and Friedman, [2009](#bib.bib14)).
Previous works in the literature have proposed quantum-like probabilistic models that try to accommodate these paradoxical scenarios and violations to the Sure Thing Principle (Busemeyer et al., [2006b](#bib.bib6), [2009](#bib.bib5); Aerts et al., [2017](#bib.bib1); Pothos and Busemeyer, [2009](#bib.bib24); Busemeyer and Bruza, [2012](#bib.bib3)). There is also a vast amount of work in trying to extend the expected utility hypothesis to a quantum-like versions Mura ([2009](#bib.bib23)); Yukalov and Sornette ([2015](#bib.bib30)). However, the expected utility framework alone poses some difficulties, since it is very challenging the task of decision-making in situations where the outcomes of an action are not fully determined (Koller and Friedman, [2009](#bib.bib14)).
In this paper, we try to fill this gap by taking into account the quantum-like probability inferences produced by a quantum-like Bayesian network to various outcomes and extend these probabilities to influence the preferences of an individual between these outcomes. Note that the probabilistic inferences produced by the quantum-like Bayesian network will suffer quantum interference effects in decision scenarios under uncertainty. The general idea is to use these quantum interference effects to influence the expected utility framework in order to favour other actions than what would be predicted from the classical theory alone. We will combine this structure in a directed and acyclic compact probabilistic graphical model for decision-making, which we will define as the quantum-like influence diagram.
3 A Quantum-Like Influence Diagram for Decision-Making
-------------------------------------------------------
A Quantum-Like Influence Diagram is a compact directed acyclical graphical representation of a decision scenario, which was originally proposed by Howard and Matheson ([1984](#bib.bib9)). It consists on a set of random variables X1,…,XN belonging to a quantum-like Bayesian network. Each random variable Xi is associated with a conditional probability distribution (CPD) table, which describes the distribution of quantum probability amplitudes of the random variable Xi with respect to its parent nodes, ψ(Xi|PaXi). Note that the difference between a quantum-like Bayesian network and a classical network is simply the usage of complex numbers instead of classical real numbers. The usage of complex numbers will enable the emergence of quantum interference effects. The influence diagram also consists in an utility node defined variable U, which is associated with a deterministic function U(PaU). The goal is to make a decision, which maximises the expected utility function by taking into account probabilistic inferences performed on the quantum-like Bayesian network.
Figure 2: General example of a Quantum-Like Influence Diagram comprised of a Quantum-Like Bayesian Network, X1,...,XN, a Decision Node, D, and an Utility node with no children, U.
An example of a quantum-like influence diagram is presented in Figure [2](#S3.F2 "Figure 2 ‣ 3 A Quantum-Like Influence Diagram for Decision-Making ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle"). In the Figure, one can notice the three different types of nodes: (1) random variable nodes (circle-shape), denoted by X1,⋯,XN, of some Quantum-Like Bayesian Network, (2) a decision node (rectangle-shape), denoted by D, which corresponds to the decision that we want to make, and (3) an Utility node (diamond-shape), denoted by U, which in the scope of this paper, will represent the payoffs in the Prisoner’s Dilemma Game.
The goal is to maximise the expected utility by taking into consideration the probabilistic inferences of the quantum-like bayesian network, which makes use of the quantum interference effects to accommodate and predict violations to the Sure Thing Principle.
In the next sections, we will address each of these three components separately.
4 Quantum-Like Bayesian Networks
---------------------------------
Quantum-like Bayesian Network have been initially proposed by Moreira and Wichert ([2014](#bib.bib19), [2016](#bib.bib20), [2018](#bib.bib22)) and they can be defined by a directed acyclic graph structure in which each node represents a different quantum random variable and each edge represents a direct influence from the source node to the target node. The graph can represent independence relationships between variables, and each node is associated with a conditional probability table that specifies a distribution of quantum complex probability amplitudes over the values of a node given each possible joint assignment of values of its parents. In other words, a quantum-like Bayesian Network is defined in the same way as classical network with the difference that real probability values are replaced by complex probability amplitudes.
In order to perform exact inferences in a quantum-like Bayesian network, one needs to compute the:
* Quantum-Like full join probability distribution. The quantum-like full joint complex probability amplitude distribution over a set of N random variables ψ(X1,X2,...,XN) corresponds to the probability distribution assigned to all of these random variables occurring together in a Hilbert space. Then, the full joint complex probability amplitude distribution of a quantum-like Bayesian Network is given by:
| | | | |
| --- | --- | --- | --- |
| | ψ(X1,…,XN)=N∏j=1ψ(Xj|Parents(Xj)) | | (2) |
Note that, in Equation [2](#S4.E2 "(2) ‣ 1st item ‣ 4 Quantum-Like Bayesian Networks ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle"), Xi is the list of random variables (or nodes of the network), Parents(Xi) corresponds to all parent nodes of Xi and ψ(Xi) is the complex probability amplitude associated with the random variable Xi. The probability value is extract by applying Born’s rule, that is, by making the squared magnitude of the joint probability amplitude, ψ(X1,…,XN):
| | | | |
| --- | --- | --- | --- |
| | Pr(X1,…,XN)=|ψ(X1,…,XN)|2 | | (3) |
* Quantum-Like Marginalization. Given a query random variable X and let Y be the unobserved variables in the network, the marginal distribution of X is simply the amplitude probability distribution of X averaging over the information about Y. The quantum-like marginal probability for discrete random variables, can be defined by Equation [4](#S4.E4 "(4) ‣ 2nd item ‣ 4 Quantum-Like Bayesian Networks ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle"). The summation is over all possible y, i.e., all possible combinations of values of the unobserved values y of variable Y. The term γ corresponds to a normalisation factor. Since the conditional probability tables used in Bayesian Networks are not unitary operators with the constraint of double stochasticity (like it is required in other works of the literature (Busemeyer et al., [2006b](#bib.bib6); Pothos and Busemeyer, [2009](#bib.bib24))), we need to normalise the final scores. This normalisation is consistent with the notion of normalisation of wave functions used in Feynman’s Path Diagrams.
In classical Bayesian inference, on the other hand, normalisation is performed due to the independence assumptions made in Bayes rule.
| | | | |
| --- | --- | --- | --- |
| | Pr(X|e)=γ∣∣
∣∣ ∑yN∏k=1ψ(Xk|Parents(Xk),e,y) ∣∣
∣∣2 | | (4) |
Expanding Equation [4](#S4.E4 "(4) ‣ 2nd item ‣ 4 Quantum-Like Bayesian Networks ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle"), it will lead to the quantum marginalisation formula (Moreira and Wichert, [2014](#bib.bib19)), which is composed by two parts: one representing the classical probability and the other representing the quantum interference term (which corresponds to the emergence of destructive / constructive interference effects):
| | | | |
| --- | --- | --- | --- |
| | Pr(X|e)=γ|Y|∑i=1∣∣
∣∣N∏kψ(Xk|Parents(Xk),e,y=i)∣∣
∣∣2+2⋅Interference | | (5) |
| | | |
| --- | --- | --- |
| | Interference=|Y|−1∑i=1|Y|∑j=i+1∣∣
∣∣N∏kψ(Xk|Parents(Xk),e,y=i)∣∣
∣∣⋅∣∣
∣∣N∏kψ(Xk|Parents(Xk),e,y=j)∣∣
∣∣⋅cos(θi−θj) | |
Note that, in Equation [5](#S4.E5 "(5) ‣ 2nd item ‣ 4 Quantum-Like Bayesian Networks ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle"), if one sets (θi−θj) to π/2, then cos(θi−θj)=0. This means that the quantum interference term is canceled and the quantum-like Bayesian Network collapses to its classical counterpart.
Formal methods to assign values to quantum interference terms are still an open research question, however some work has already been done towards that direction (Yukalov and Sornette, [2011](#bib.bib29); Moreira and Wichert, [2016](#bib.bib20), [2017](#bib.bib21)).
5 Maximum Expected Utility in Classical Influence Diagrams
-----------------------------------------------------------
Given a set of possible decision rules, δA, the goal of Influence Diagrams is to compute the decision rule that leads to the Maximum Expected Utility.
| | | | |
| --- | --- | --- | --- |
| | | | (6) |
The goal is to choose some action a that maximises the expected utility:
| | | |
| --- | --- | --- |
| | a∗=argmaxδAEU[D[δA]] | |
One can map the expected utility formalism to the scope of Bayesian networks in the following way. In the expected utility formula, Prδa(x|a) corresponds to a full joint probability distribution of all possible outcomes, x, given different actions a. This means that we can decompose the full joint probability distribution to the chain rule of probability theory as the product of each node with its parent nodes.
| | | | |
| --- | --- | --- | --- |
| | | | (7) |
| | | | |
| --- | --- | --- | --- |
| | | | (8) |
In Equation [8](#S5.E8 "(8) ‣ 5 Maximum Expected Utility in Classical Influence Diagrams ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle"), Z=PaA represents the parent nodes of action A. We can factorise Equation [8](#S5.E8 "(8) ‣ 5 Maximum Expected Utility in Classical Influence Diagrams ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle") in terms of the decision rule, δA, obtaining
| | | | |
| --- | --- | --- | --- |
| | EU[D[δA]]==∑Z,AδA(A|Z)(∑W(∏iPr(Xi|PaXi))U(PaU)), | | (9) |
where W={X1,…,XN}−Z corresponds to all nodes of the Bayesian Network that are not contained in the set of nodes in Z.
By marginalising the summation over W, we obtain an expected utility formula that is written only in terms of the factor μ(A,Z). Note that this factor corresponds to a conditional distribution table of random variable Z (the outcomes of some action a) and action a.
| | | | |
| --- | --- | --- | --- |
| | EU[D[δA]]=∑Z,AδA(A|Z)μ(A,Z) | | (10) |
The Maximum Expected Utility for a classical Influence Diagrams is given by (Koller and Friedman, [2009](#bib.bib14)):
| | | | |
| --- | --- | --- | --- |
| | δ∗A(a,Z)=α(x)={1 a=argmax(A,Z)0otherwise | | (11) |
6 Maximum Expected Utility in Quantum-Like Influence Diagrams
--------------------------------------------------------------
The proposed quantum-like influence diagram is built upon the formalisms of quantum-like Bayesian networks. This means that real classical probabilities need to be replaced by complex quantum amplitudes.
We start the derivation with the initial notion of expected utility already presented in the previous section.
| | | | |
| --- | --- | --- | --- |
| | | | (12) |
For simplicity, let’s consider the decision scenario where we have two binary events X1 and X2. Then, we can decompose the classical expected utility equation as
| | | | |
| --- | --- | --- | --- |
| | EU[D[δA]]=∑X1,X2,AδA(A|X1)Pr(X1)Pr(X2|X1)U(X1,A) | | (13) |
Like before, we can factorise this formula in terms of the decision rule δA, obtaining
| | | | |
| --- | --- | --- | --- |
| | EU[D[δA]]=∑A,X2δA(A|X2)∑X1Pr(X1)Pr(X2|X1)U(X1,A) | | (14) |
For binary events, we obtain the marginalisation of X1 over both X2 and D
| | | | |
| --- | --- | --- | --- |
| | EU[D[δA]]=∑A,X2δA(A|X2)⋅μ(X2,A) | | (15) |
where μ(X2,A) is a factor with the utility function expressed in terms of the distribution of X1. More specifically, it is given by
| | | | |
| --- | --- | --- | --- |
| | μ(X2,A)=Pr(X1=t)Pr(X2|X1=t)U(X1=t,A)++Pr(X1=f)Pr(X2|X1=f)U(X1=f,A) | | (16) |
Since the proposed quantum-like influence diagram makes use of a quantum-like Bayesian network, this means that we need to convert the classical real probabilities into complex quantum amplitudes. This is performed by applying Born’s rule: for some classical probability A, the corresponding quantum amplitude is simply its squared magnitude, Pr(A)=|ψA|2 (Deutsch, [1988](#bib.bib8); Zurek, [2011](#bib.bib31)). Since in Equation [16](#S6.E16 "(16) ‣ 6 Maximum Expected Utility in Quantum-Like Influence Diagrams ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle") we have a combination of utility functions with probability values, we cannot apply Born’s rule directly, since we would not be satisfying its definition. For this reason, we propose to split Equation [16](#S6.E16 "(16) ‣ 6 Maximum Expected Utility in Quantum-Like Influence Diagrams ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle") into a vector representation containing a classical probability and another containing the utility function. This procedure is similar to the one propose in the Quantum Decision Theory model of Yukalov and Sornette ([2015](#bib.bib30)) where the authors separate a prospect into an utility factor (a factor containing the classical utility of a lottery) and an attraction factor (a probabilistic factor that results from the quantum interference effect).
Considering πa the vector representation of a classical probability vector and ua the classical utility corresponding to the choice of some action A, then we obtain
| | | |
| --- | --- | --- |
| | πa=Pr(X1=t)Pr(X2|X1=t)+Pr(X1=f)Pr(X2|X1=f) | |
| | | |
| --- | --- | --- |
| | ua=U(X1=t,A)+U(X1=f,A) | |
We can apply Born’s rule by replacing classical real numbers by quantum-like amplitudes and performing their squared magnitude as
| | | | |
| --- | --- | --- | --- |
| | πa=|qa|2⇔qa=|ψ(X1=t)ψ(X2|X1=t)+ψ(X1=f)ψ(X2|X1=f)|2qa=|ψ(X1=t)ψ(X2|X1=t)|2+|ψ(X1=f)ψ(X2|X1=f)|2+Interf, | | (17) |
where the quantum interference term is given by
| | | | |
| --- | --- | --- | --- |
| | Interf=2|ψ(X1=t)ψ(X2|X1=t)||ψ(X1=f)ψ(X2|X1=f)|Cos(θ1−θ2). | | (18) |
The utility factor ua needs to be updated in order to become a factor of the quantum interference term.
| | | |
| --- | --- | --- |
| | ua=U(X1=t,A)+U(X1=f,A)+U(X1=t,A)⋅U(X1=f,A) | |
The result of this marginalisation, μ(X2,A), will be given by the product of the vector representation of these two terms:
| | | |
| --- | --- | --- |
| | μ(X2,A)=⟨qa|ua⟩, | |
where the vector representation corresponds to
| | | |
| --- | --- | --- |
| | |qa⟩=⎡⎢
⎢⎣|ψ(X1=t)ψ(X2|X1=t)|2|ψ(X1=f)ψ(X2|X1=f)|2Interf⎤⎥
⎥⎦ |ua⟩=⎡⎢⎣U(X1=t,A)U(X1=f,A)U(X1=t,A)U(X1=f,A)⎤⎥⎦. | |
This way, the final marginalisation for the quantum-like influence diagram is
| | | | |
| --- | --- | --- | --- |
| | μ(X2,A)=⟨qa|ua⟩=|ψ(X1=t)ψ(X2|X1=t)|2⋅U(X1=t,A)+|ψ(X1=f)ψ(X2|X1=f)|2⋅U(X1=f,A)+…+Interf⋅U(X1=t,A)U(X1=f,A) | | (19) |
Note that, in Equation [19](#S6.E19 "(19) ‣ 6 Maximum Expected Utility in Quantum-Like Influence Diagrams ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle"), if one sets the interference term (θi−θj) to π/2, then cos(θi−θj)=0. This means that the quantum interference term is canceled and the quantum-like influence diagram collapses to its classical counterpart. In other words, one can see the quantum-like influence diagram as a more general and abstract model of the classical diagram, since it represents both classical and quantum behaviour.
Finally, the Maximum Expected Utility for Quantum-Like Influence Diagrams is given by:
| | | | |
| --- | --- | --- | --- |
| | δ∗A(a,Z)=α(x)={1 a=argmax μ(X2,A)0otherwise | | (20) |
7 A Quantum-Like Influence Diagram for the Prisoner’s Dilemma Game
-------------------------------------------------------------------
Several paradoxical findings have been reported over the literature showing that individuals do not act rationally in decision scenarios under uncertaint (Kuhberger et al., [2001](#bib.bib15); Tversky and Shafir, [1992](#bib.bib27); Lambdin and Burdsal, [2007](#bib.bib16); Hristova and Grinberg, [2008](#bib.bib11); Busemeyer et al., [2006a](#bib.bib4))..
The quantum-like influence diagram can help to accommodate and explain the several paradoxical decisions by manipulating the quantum interference effects that emerge from the inferences in the quantum-like Bayesian network. These inferences can then be used to reestimate the expected utility of an agent.
Figure 3: Quantum-Like Infleunce Diagram representing the Prisoner’s DIlemma Experiment from (Shafir and Tversky, [1992](#bib.bib25)).
We will model the works previously reported in Table [1](#S2.T1 "Table 1 ‣ 2 Revisiting the Prisoner’s Dilemma and the Expected Utility Hypothesis ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle") under the proposed quantum-like influence diagram. Figure [3](#S7.F3 "Figure 3 ‣ 7 A Quantum-Like Influence Diagram for the Prisoner’s Dilemma Game ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle") corresponds to the representation of the work of Shafir and Tversky ([1992](#bib.bib25)). The three types of nodes in the represented quantum-like influence diagram are the following:
* Random Variables: the circle-shaped nodes are the random variables belonging to the quantum-like bayesian network representing the player that needs to make a decision in the Prisoner’s Dilemma, without being aware of the decision of his opponent. We modelled this network with two binary random variables, X1 and X2. X1 corresponds to the player’s own strategy (either to defect or to cooperate) and X2 the player’s personal risk preferences, i.e. either he is risk averse (and therefore would find it safe to engage in a defect strategy) or he is risk seeking (and would prefer to engage in a cooperate strategy). The tables next to each random variable are conditional probability tables and they show the probability distribution of the variable towards its parent nodes. These conditional probability tables match the probability distributions reported in Table [1](#S2.T1 "Table 1 ‣ 2 Revisiting the Prisoner’s Dilemma and the Expected Utility Hypothesis ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle"). In the specific case of Figure [3](#S7.F3 "Figure 3 ‣ 7 A Quantum-Like Influence Diagram for the Prisoner’s Dilemma Game ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle"), this table is filled with the values of the probability amplitudes identified in the work of Shafir and Tversky ([1992](#bib.bib25)). For the general case, we will assume that the player has no initial strategy and we will assume neutral priors for the variable X1 (like it was assumed in previous works of the literature, see Moreira and Wichert ([2016](#bib.bib20))).
* Action Node: the rectangle shaped node is the action that we want to make a decision. In the context of the prisoner’s dilemma we are interested to compute the maximum expected utility of defecting or not defecting (i.e. cooperating).
* Utility Node: the diamond shaped node corresponds to the payoffs that the player will have for taking (or not) the action defect, given his own personal preferences towards risk. The values in this node will be populated with the different payoffs used across the different experiments of the prisoner’s dilemma reported over the literature.
In the conditions where the player knows the strategy of his opponent, the quantum-like influence diagram collapses to its classical counterpart, since there is no uncertainty. This was already noticed in the previous works of Moreira and Wichert ([2014](#bib.bib19), [2016](#bib.bib20), [2017](#bib.bib21)). However, when the player is not informed about his opponent’s decision, then the quantum-like Bayesian network will produce interference effects (Equation [5](#S4.E5 "(5) ‣ 2nd item ‣ 4 Quantum-Like Bayesian Networks ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle")). When computing the maximum expected utility, we will marginalise out X1 like it was shown in Equation [16](#S6.E16 "(16) ‣ 6 Maximum Expected Utility in Quantum-Like Influence Diagrams ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle"). This will result in a factor showing the distribution of the player’s personal preferences towards risk (either risk averse or risk seeking) towards his actions (either to defect or cooperate). The quantum interference term will play an important role to determine which quantum parameters can influence the player’s decision to switch from a classical (and rational) defect action towards the paradoxical decision found in the works the literature, i.e. to cooperate.
Figure 4: Impact of quantum interference terms in the overall expected utility: (left) quantum parameters that maximize a cooperate decision, (center) variation of the expected utility when the player is risk averse and (right) variation of the expected utility when the player is risk seeking.
Figure [4](#S7.F4 "Figure 4 ‣ 7 A Quantum-Like Influence Diagram for the Prisoner’s Dilemma Game ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle") demonstrates the impact of the quantum interference effects in the player’s decision. The graphs in the centre and in the right of Figure [4](#S7.F4 "Figure 4 ‣ 7 A Quantum-Like Influence Diagram for the Prisoner’s Dilemma Game ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle") represent all possible maximum expected utilities that the player can achieve by varying the quantum interference term θ in Equation [19](#S6.E19 "(19) ‣ 6 Maximum Expected Utility in Quantum-Like Influence Diagrams ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle") for a personal preference of being risk averse or risk seeking, respectively. On the left of Figure [4](#S7.F4 "Figure 4 ‣ 7 A Quantum-Like Influence Diagram for the Prisoner’s Dilemma Game ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle"), it is represented all the values of θ that satisfy the condition that EU[Cooperate]>EU[Defect], i.e., all the values of the quantum interference parameter θ that will maximise the utility of cooperation rather than defect. One can note that, for experiment of Shafir and Tversky ([1992](#bib.bib25)) (as well as in the remaining works of the literature analysed in this work), one can maximise the expected utility of Cooperation when the utilities are negative. This is in accordance with the previous study of Moreira and Wichert ([2016](#bib.bib20)) in which the authors found that violations to the Sure Thing Principle imply destructive (or negative) quantum interference effects. As we will see in the next section, the quantum parameters found that are used to maximise the expected utility of a cooperate action lead to destructive quantum interferences and can exactly explain the probability distributions observed in the experiments.
###
7.1 Results and Discussion
Although there are several quantum parameters that satisfy the relationship that shows that participants can maximise the utility of a cooperate action, only a few parameters are able to accommodate both the paradoxical probability distributions reported in the several works in the literature and to maximise the expected utility of cooperating. For instance, Figures [6](#S7.F6 "Figure 6 ‣ 7.1 Results and Discussion ‣ 7 A Quantum-Like Influence Diagram for the Prisoner’s Dilemma Game ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle") and [6](#S7.F6 "Figure 6 ‣ 7.1 Results and Discussion ‣ 7 A Quantum-Like Influence Diagram for the Prisoner’s Dilemma Game ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle") show how the quantum parameters are sensitive to accommodate the violations of the Sure Thing Principle in terms of the probability distributions. The slight variation of the quantum parameter θ in the quantum-like Bayesian network can lead to completely different probability distributions which differ from the ones observed in the difference experimental scenarios reported in the literature. These probability distributions will influence the utilities computed by the expected utility framework.
Figure 5: Probabilities that can be obtained in Game 2 of Li and Taplin ([2002](#bib.bib18)).
Figure 6: Probabilities that can be obtained in Game 2 of Li and Taplin ([2002](#bib.bib18)).
In Table [2](#S7.T2 "Table 2 ‣ 7.1 Results and Discussion ‣ 7 A Quantum-Like Influence Diagram for the Prisoner’s Dilemma Game ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle"), it is presented the quantum parameters that lead to the quantum interference term that is necessary to fully explain and accommodate the violations to the sure thing principle reported over several works of the literature.
| | Shafir and Tversky ([1992](#bib.bib25)) | Li and Taplin ([2002](#bib.bib18)) G1 | Li and Taplin ([2002](#bib.bib18)) G2 | Li and Taplin ([2002](#bib.bib18)) G3 | Li and Taplin ([2002](#bib.bib18)) G4 | Li and Taplin ([2002](#bib.bib18)) G5 | Li and Taplin ([2002](#bib.bib18)) G6 | Li and Taplin ([2002](#bib.bib18)) G7 |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| Prob of Defect | | | | | | | | |
| (Known to Defect) | 0.9700 | 0.7333 | 0.8000 | 0.9000 | 0.8333 | 0.8333 | 0.7667 | 0.8667 |
| Prob of Cooperate | | | | | | | | |
| (Known to Cooperate) | 0.8400 | 0.6670 | 0.7667 | 0.8667 | 0.8000 | 0.7333 | 0.8333 | 0.7333 |
| Classical Prob | | | | | | | | |
| (Unknown condition) | 0.9050 | 0.7000 | 0.7833 | 0.8834 | 0.8167 | 0.7833 | 0.8000 | 0.8000 |
| Experim Prob | | | | | | | | |
| (Unknown condition) | 0.6300 | 0.6000 | 0.6300 | 0.8667 | 0.7000 | 0.7000 | 0.8000 | 0.7667 |
| Quantum Interference | | | | | | | | |
| θ param | 2.8151 | 3.0170 | 3.0758 | 2.8052 | 3.2313 | 2.8519 | 1.5708 | 3.7812 |
Table 2: Experimental results reported for the Prisoner’s Dilemma game. The entries highlighted correspond to games that are not violating the Sure Thing Principle.
For this reason, we decided to test if the quantum-like parameters used to accommodate the violations to the Sure Thing Principle were sufficient and if they could also lead to a maximisation of expected utility of cooperation. We performed simulations of the different works in the literature and we concluded that the quantum interference effects that can accommodate violations to the violations of the Sure Thing Principle in the quantum-like Bayesian network alone, also explain a higher preference of the cooperative action over defect. Table [3](#S7.T3 "Table 3 ‣ 7.1 Results and Discussion ‣ 7 A Quantum-Like Influence Diagram for the Prisoner’s Dilemma Game ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle") presents the results.
| | | |
| --- | --- | --- |
| | | Li and Taplin ([2002](#bib.bib18)) |
| | Shafir and Tversky ([1992](#bib.bib25)) | Game 1 | Game 2 | Game 3 | Game 4 | Game 5 | Game 6 | Game 7 |
| | MEU | MEU | MEU | MEU | MEU | MEU | MEU | MEU | MEU | MEU | MEU | MEU | MEU | MEU | MEU | MEU |
| | (coop) | (def) | (coop) | (def) | (coop) | (def) | (coop) | (def) | (coop) | (def) | (coop) | (def) | MEU | MEU | (coop) | (def) |
| CL Rk Av | 43.63 | 50.25 | 34.19 | 39.35 | 38.75 | 61.78 | 26.85 | 50.33 | 65.70 | 67.33 | 16.27 | 34.50 | 17.58 | 36.50 | 16.43 | 35.00 |
| CL Rk Sk | 6.38 | 7.25 | 15.82 | 18.15 | 11.25 | 17.22 | 3.65 | 26.85 | 14.80 | 15.17 | 5.23 | 10.5 | 3.92 | 8.50 | 5.07 | 10.00 |
| QL Rk Av | -1559.46 | -2129.94 | -1263.63 | -1730.21 | -1422.69 | -4787.28 | -702.24 | -2075.58 | -5198.14 | -5462.41 | -221.05 | -1313.94 | 28.83 | 36.49 | -184.75 | -1116.33 |
| QL Rk Sk | 116.66 | -160.08 | -538.62 | -735.89 | -392.89 | -1320.22 | -94.44 | -270.75 | -1162.55 | -1221.47 | -61.44 | -353.22 | 3.91 | 8.50 | -44.86 | -262.30 |
| QL Interf | | | | | | | | | | | | | | | | |
| θ1−θ2 | 2.815 | 2.815 | 3.017 | 3.017 | 3.0758 | 3.0758 | 2.805 | 2.805 | 3.23 | 3.23 | 2.8519 | 2.8519 | 1.5708 | 1.5708 | 3.78 | 3.78 |
| Payoff | | | | | | | | | | | | | | | | |
| dd dc | 30 | 25 | 30 | 25 | 73 | 25 | 30 | 25 | 80 | 78 | 43 | 10 | 30 | 10 | 30 | 10 |
| cd cc | 85 | 75 | 85 | 75 | 85 | 75 | 85 | 36 | 85 | 83 | 85 | 46 | 60 | 33 | 60 | 33 |
Table 3: Results obtained after performing inferences in the quantum-like influence diagram for different works of the literature reporting violations of the Sure Thing Principle in the Prisoner’s Dilemma Game. One can see that the Maximum Expected Utility (MEU) was changed to favour a Cooperate strategy using the quantum interference effects of the Quantum-Like Bayesian Network. In the payoffs, d corresponds to defect and c to cooperate. The first payoff corresponds to player 1 and the second to player 2.
In Table [3](#S7.T3 "Table 3 ‣ 7.1 Results and Discussion ‣ 7 A Quantum-Like Influence Diagram for the Prisoner’s Dilemma Game ‣ Introducing Quantum-Like Influence Diagrams for Violations of the Sure Thing Principle"), we present the MEU computed for each work in the literature using the classical approach for the different personal preferences of the individual towards risk: either risk seeking (CL Rk Sk) or risk averse (CL Rk Av). The classical MEU shows that the optimal strategy is to defect even if the individual has a risk seeking personality (who would be willing to bet on a cooperate action). Of course these results go against the experimental works of the literature which say that a significant percentage of individuals engaged in cooperative strategies.
In opposition, when we use the quantum-like influence diagram, we take advantage of the quantum interference terms that will disturb the probabilistic outcomes of the quantum-like Bayesian networks. Since the utility function depends on the outcomes of the quantum-like Bayesian network, then it is straightforward that quantum interference effects influence indirectly the outcomes of the MEU allowing us to favour a different strategy predicted by the classical MEU.
It is interesting to notice that indeed the parameters used accommodate the violations of the Sure Thing Principle alone in the quantum-like Bayesian Network could also be used to maximise the utility of a Cooperate action. This was verified in all works of the literature analysed except in Game 6 in the work of Li and Taplin ([2002](#bib.bib18)). The reason is that Game 6 is not even reporting a violation to the Sure Thing Principle and could be explained by the classical theory with a minor error percentage. So, if it can be explained under the classical theory, then of course it also tends to favour a defect action over a cooperate one.
8 Conclusion
-------------
In this work, we proposed an extension of the quantum-like Bayesian Network initially proposed by Moreira and Wichert ([2014](#bib.bib19), [2016](#bib.bib20)) into a quantum-like influence diagram. Influence diagrams are designed for knowledge representation. They are a directed acyclic compact graph structure that represents a full probabilistic description of a decision problem by using probabilistic inferences performed in Bayesian networks (Koller and Friedman, [2009](#bib.bib14)) together with a fully deterministic utility function. Currently, influence diagrams have a vast amount of applications. They can be used to determine the value of imperfect information on both carcinogenic activity and human exposure (Howard and Matheson, [2005](#bib.bib10)), the are used to detect imperfections in manufacturing and they can even be used for team decision analysis (Detwarasiti and Shachter, [2005](#bib.bib7)), valuing real options (Lander and Shenoy, [2001](#bib.bib17)), etc.
Although we are aware that more studies need to be conducted in this direction, the preliminary results obtained in this study show that the quantum-like Bayesian network can be extended to deal with decision-making scenarios by incorporating the notion of maximum expected utility in influence diagrams. The general idea is to take advantage of the quantum interference terms produced in the quantum-like Bayesian network to influence the probabilities used to compute the expected utility. This way, we are not proposing a new type of expected utility hypothesis. On the contrary, we are keeping it under its classical definition. We are only incorporating it as an extension of a quantum-like probabilistic graphical model where the utility node depends only on the probabilistic inferences of the quantum-like Bayesian network.
This notion of influence diagrams opens several new research paths. One can incorporate different utility nodes being influenced by different random variables of the quantum-like Bayesian Network. This way one can even explore different interference terms affecting different utility nodes, etc. We plan to carry on with this study and further develop these ideas in future research.
9 Acknowledgements
-------------------
This work was supported by national funds through Fundação para a Ciência e a Tecnologia (FCT) with reference UID/CEC/50021/2013. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
The authors would like to thank Mr. Sebastian Schmidt and Dr. Godfrey Charles-Cadogan for all the discussions about the topics of this work. Their wise comments were truly helpful for the development of the ideas in this manuscript. |
ffd0a669-fec9-40af-ae73-e1b2f46c5611 | trentmkelly/LessWrong-43k | LessWrong | Networks of Meaning
> Nearly every piece of fiction that I have written has had at its heart an image of some simple object linking two previously separate clusters of images.[1]
>
> – Gerald Murnane
> The more things an image is joined with, the more often it springs into life.
>
> Dem.: The more other images an image is joined with, the more causes there are by which it can be aroused.[2]
>
> – Baruch Spinoza
Clearly, some things are meaningful to us. Some things are meaningful to me but not to you. Some other things are meaningful to you but not to anyone else on Earth. What's more, it happens that humans experience what we call revelation, where new information or a change in perspective makes previously familiar things seem newly meaningful to us.
The writer Gerald Murnane, in my view one of the greatest writers ever to live, recounts his having been brought by his wife to the opening of a contemporary art exhibition, where he is asked by the amiable organisers to take part in a panel discussion later the same evening. Being the sort of person who avoids walking into shops unless he is sure he wants to buy something (so as not to risk disappointing the shopkeeper), he accepts, though he knows nothing about contemporary art. Having accepted, he paces around the gallery, trying to think of something to say in the panel. He happens upon an artwork consisting of a handful of smooth stones scattered over the floor. The stones remind him of the fear he used to feel as a child on a rocky bay near his grandfather's farm.
> I said very little during the panel discussion at the gallery, and I have no recollection of how that little was received, but I have never forgotten my satisfaction at having formulated what had been for the previous three decades of my life as a writer a sort of instinctive awareness and no more. I said, at least once, and with an image in my mind of the stones on the bare floor of the brightly lit gallery but as though they slithered beneath my bare feet in th |
ab0b8982-5630-4d0c-be24-8b905ae23dc6 | trentmkelly/LessWrong-43k | LessWrong | Incentive compatibility and the Revelation Principle
In which the Revelation Principle is introduced, showing all mechanisms can be reduced to incentive compatible mechanisms. With this insight, a solution (of sorts) is given to the public good problem in the last post. Limitations of the Revelation Principle are also discussed.
The formalism I introduced last time will now start paying off. We were left with the question of how to check whether a mechanism exists that satisfies some particular goal, naively requiring us to search over all possible procedures. Luckily though, the space of all possible mechanisms can be reduced to something manageable.
Observe the following: suppose through divine intervention we were granted a mechanism (M, g) that implements a social choice function f. In other words, the outcome when agents of types θ1, …, θn interact with the mechanism is exactly the outcome prescribed by the function f for those types. Since a person’s type encodes all relevant variation in their characteristics, preferences, or beliefs, we’d know what that person wants to do if we knew their type. In particular, when agent i is type θi, we expect her choice of message to be mi(θi). Once all the messages are sent, the function g translates the agents’ choices into an outcome so that g(m1(θ1), …, mn(θn)) = f(θ1, …, θn).
But wait! Since we expect a type θi agent to send message mi(θi), why don’t we just package that inside the mechanism? Each agent will tell us their type and the mechanism designer will play in proxy for the agent according to the original mechanism. We’ll call this the direct mechanism for f since all we need to know is that agents tell us their types and then are assigned outcome f(θ), no matter what we’ve blackboxed in the middle.
Why did we expect an agent of type θi to send message mi(θi)? Presumably because that message maximized her utility. In particular, it had to be at least as good as the message mi(θʹi) she’s send when her type is different, giving us:
Since the outcomes of |
0d40a192-28c3-41a9-a7a0-db8037b2d656 | trentmkelly/LessWrong-43k | LessWrong | Getting rational now or later: navigating procrastination and time-inconsistent preferences for new rationalists
This is a distillation of and reflection on O’Donoghue and Rabin’s “Doing it now or later” (see citation below)[1].
Many people struggle with procrastination or self-control. Critically, we struggle with the mismatch between current preference and future preference. Procrastination arises in situations that are unpleasant to perform but create future benefits: they have “immediate costs.” Similarly, indulgent behaviors (e.g. eating unhealthy foods) have “immediate rewards” (they taste good), while stashing away future costs. Both have a mismatch between present-self and the future-self.
The best way to “kick” this issue is to develop time-consistent preferences, placing no special value on what’s happening now–thus our choices around procrastination or indulgence will be more rational and no longer skewed by a current self-control limitation. But…that’s hard to achieve.
O’Donoghue and Rabin introduce a neat distinction (1999). A naive procrastinator, they write, expects that in the future, they won’t have the same self-control issues that they’re having now. A sophisticated procrastinator expects their imperfection: they expect their future self to also have procrastination or indulgence tendencies. With their model of time-inconsistent preferences, they demonstrate that sophisticates procrastinate less than their naive peers! This model suggests that I shouldn’t plan on having good self-control; actually, expecting my own irrationality may improve my performance. But…not so fast, because the model also indicates that sophistication may make me more likely to indulge now at a time when I’m better off waiting. The lesson for new rationalists, who haven’t yet cracked time consistency: when facing immediate costs, be realistic (like the “sophisticate”), but when facing immediate rewards, stay idealistic (like the “naif”).
Let’s get into it with an example!
You choose…you must either endure:
* Three hours of an unpleasant activity today
* Four hours of it one |
c86f86f9-e929-412f-8b7d-10bfa5438db5 | trentmkelly/LessWrong-43k | LessWrong | Extreme Security
What makes "password" a bad password?
You might say that it's because everybody else is already using it, and maybe you'd be correct that everybody in the world deciding to no longer use the password "password" could eventually make it acceptable again.
But consider this: if everyone who was using that password for something changed it tomorrow, and credibly announced so, it would still be an English dictionary word. Any cracker that lazily included the top N words in an English dictionary in a cracking list would still be liable to break it, not because people actually use every English dictionary word, but simply because trying the top N English words is a commonly deployed tactic.
You could go one step further and say "password" is also a bad password because it's in that broader set of English dictionary words, but really, you should just drop the abstraction altogether. "Password" is a bad password because hackers are likely to try it. Hackers are likely to try it because they're trying to break your security, yes, but also because of a bunch of seemingly irrelevant details like what they tell humans to do on online password cracking resources, and the hashcat & JohnTheRipper programs' standard features. If due to some idiosyncratic psychological quirk hackers were guaranteed never to think of "password" as a possible password and never managed to insert "password" in their cracking lists, it would be the Best Password. A series of 17 null bytes (which are used to signal the end of C strings) tends to this day to be an excellent choice, when the website accepts that, and simple derivations of that strategy will probably go on being an excellent choice despite my mentioning them in this LW post.
Generating a password with a high amount of "entropy" is just a way of ensuring that password crackers are very unlikely to break them without violating Shannon's maxim. Nobody wants to have to violate Shannon's maxim because it means they can't talk about their c |
25d98982-7d74-414c-a6fd-7f20e88f20ed | trentmkelly/LessWrong-43k | LessWrong | Working Out in VR Really Works
Epistemic Status: Based almost entirely on my own experience. Also based on some friends here, here, and in real life.
TL;DR: I now need willpower in order to not-workout (!). I wish someone would have told me this a long time ago when I was looking for some workout that would resonate with me.
Is it actually a workout?
TL;DR: Yes.
1. I get tired and I sweat, especially in a specific intensive game (Thrill of the Fight, aka TotF)
2. I saw someone play that game for under 3 minutes and collapse on the sofa in exhaustion for about half a day
3. Here's a graph from my activity tracker (Oura Ring gen 2, not optimal for tracking workouts!, but still adds some info and also everyone likes graphs, right?) :
To give context to this graph (of me playing TotF several times during the day): When I do a rollerblade trip with a group that's better than me, the hardest part is usually under the "high" line. TotF is often above it.
Can you build muscle with VR?
TL;DR: Unclear.
People who know stuff (I don't qualify) say one needs resistance to build muscles, so maybe using wrist-weights or elastic bands that connect one's limbs to one's torso. I didn't try any of those myself. Other people said "consult with a specialist, some movements that you'd do with resistance can cause harm".
It does seem to be amazing for aerobic exercises, and specifically for interval training.
Which games are good for workouts:
1. Classic: Beat Saber, almost everyone likes it.
1. I rate it as a "medium" workout (and if it seems too easy, wait for the harder levels).
2. I recommend starting from the Tutorial.
2. I tried Pistol Whip which is somewhat similar but with guns.
1. Also rated as a "medium" workout.
3. Thrill of the Fight (TotF) is by far the most intensive workout I found on VR.
1. This will not be a good fit for everyone, it's scary (at least for me).
2. I have lots of beginner tips for it (including safety tips and things that I bet the game woul |
229d2aad-d3cf-4748-a500-b609eb336ff9 | trentmkelly/LessWrong-43k | LessWrong | Roles are Martial Arts for Agency
A long time ago I thought that Martial Arts simply taught you how to fight – the right way to throw a punch, the best technique for blocking and countering an attack, etc. I thought training consisted of recognizing these attacks and choosing the correct responses more quickly, as well as simply faster/stronger physical execution of same. It was later that I learned that the entire purpose of martial arts is to train your body to react with minimal conscious deliberation, to remove “you” from the equation as much as possible.
The reason is of course that conscious thought is too slow. If you have to think about what you’re doing, you’ve already lost. It’s been said that if you had to think about walking to do it, you’d never make it across the room. Fighting is no different. (It isn’t just fighting either – anything that requires quick reaction suffers when exposed to conscious thought. I used to love Rock Band. One day when playing a particularly difficult guitar solo on expert I nailed 100%… except “I” didn’t do it at all. My eyes saw the notes, my hands executed them, and no where was I involved in the process. It was both exhilarating and creepy, and I basically dropped the game soon after.)
You’ve seen how long it takes a human to learn to walk effortlessly. That's a situation with a single constant force, an unmoving surface, no agents working against you, and minimal emotional agitation. No wonder it takes hundreds of hours, repeating the same basic movements over and over again, to attain even a basic level of martial mastery. To make your body react correctly without any thinking involved. When Neo says “I Know Kung Fu” he isn’t surprised that he now has knowledge he didn’t have before. He’s amazed that his body now reacts in the optimal manner when attacked without his involvement.
All of this is simply focusing on pure reaction time – it doesn’t even take into account the emotional terror of another human seeking to do violence to you. It doesn’t captu |
16d05f24-a0ef-490b-82c8-2b427c25219f | StampyAI/alignment-research-dataset/arxiv | Arxiv | Neural Simplex Architecture
1 Introduction
---------------
Deep neural networks (DNNs) in combination with *reinforcement learning* (RL) are increasingly being used to train powerful *AI agents*. Such agents have achieved unprecedented success in strategy games, including defeating the world champion in Go [[32](#bib.bib32)] and surpassing state-of-the-art chess and shogi engines [[31](#bib.bib31)].
For these agents, safety is not an issue:
when a game-playing agent makes a mistake, the worst-case scenario is losing a game. The same cannot be said for AI agents that control autonomous and other complex systems. A mistake by an AI controller may cause physical damage to the controlled system and its environment, including humans.
In this paper, we present the *Neural Simplex Architecture* (NSA), a new approach to runtime assurance that provides safety guarantees for AI controllers, including neural controllers, of autonomous and other complex systems without unduly sacrificing performance.
NSA is inspired by Sha et al.’s Simplex control architecture [[30](#bib.bib30), [28](#bib.bib28)],
where a pre-certified *decision module* (DM) switches control from a high-performance but unverified (hence potentially unsafe) advanced controller (AC) to a verified-safe *baseline controller* (BC)
if the AC produces an *unrecoverable action*; i.e., an action that would lead the system within one time step to a state from which the BC is not guaranteed to preserve safety.
In the traditional Simplex approach, the AC is treated as a black box, and after the DM switches control to the BC, the BC remains in control forever.
There is, however, relatively little work on switching control back to the AC [[20](#bib.bib20), [11](#bib.bib11), [36](#bib.bib36)], and there are no techniques to correct the AC after it generates an unrecoverable
control input.
Figure 1: The Neural Simplex Architecture. The green check marks indicate pre-certified components.
NSA, illustrated in Fig. [1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Neural Simplex Architecture"), addresses both of these limitations.
The high-performance *Neural Controller* (NC) is a deep neural network (DNN) that given a plant state (or raw sensor readings),
produces a control input for the plant.
NSA’s use of an NC, as opposed to the black-box AC found in traditional Simplex,
allows for online retraining of the NC’s DNN.
Such retraining is performed by NSA’s *Adaptation Module* (AM) using RL techniques.
For systems with large state spaces, it may be difficult to achieve thorough coverage during initial training of the NC.
Online retraining has the advantage of focusing the learning on areas of the state space that are relevant to the actual system behavior, i.e., regions of the state space the system actually visits.
The AM seeks to eliminate unrecoverable actions from the NC’s behavior, without unduly degrading its performance, and in some cases actually improving its performance. While the BC is in control of the plant, the NC runs in shadow mode and is actively retrained by the AM. The DM can subsequently switch control back to the NC with high confidence that it will not repeat the same mistakes, permitting the mission to continue under the auspices of the high-performance NC. Note that because NSA preserves the basic principles of Simplex architecture, it guarantees that the safety of the plant is never violated.
NSA addresses the problem of *safe reinforcement learning*
(SRL)
[[16](#bib.bib16), [40](#bib.bib40)].
In particular, when the learning agent (the NC) produces an unrecoverable
action, the AM
uses that action as a training sample (but does not execute it), with a large negative reward.
A comparison with related approaches to SRL is provided in Section [6](#S6 "6 Related Work ‣ Neural Simplex Architecture").
We conducted an extensive evaluation of NSA on several significant example systems, including a target-seeking rover navigating through an obstacle field,
and a neural controller for an artificial pancreas.
Our results on these case studies conclusively demonstrate NSA’s benefits.
In summary, the main contributions of this paper are the following:
* •
We introduce the Neural Simplex Architecture, a new approach to runtime assurance that provides safety guarantees for neural controllers.
* •
We address two limitations of the traditional Simplex approach, namely lack of established guidelines for switching control back to the AC so that mission completion can be attained; and lack of techniques for correcting the AC’s behavior after a failover to the BC, so that reverse switching makes sense in the first place.
* •
We provide a key insight into safe reinforcement learning (by demonstrating the utility of potentially unsafe training samples, when appropriately and significantly penalized),
along with a thorough evaluation of the NSA approach on two significant case studies.
2 Background
-------------
#### Simplex Architecture.
The main components of the Simplex architecture (AC, BC, DM) were introduced above.
The BC is certified to guarantee the safety of the plant only if it takes over control while the plant’s state is within a *recoverable region* ℛBCsubscriptℛ𝐵𝐶\mathcal{R}\_{BC}caligraphic\_R start\_POSTSUBSCRIPT italic\_B italic\_C end\_POSTSUBSCRIPT. For example, consider the BC for a ground rover that simply applies maximum deceleration amaxsubscript𝑎𝑚𝑎𝑥a\_{max}italic\_a start\_POSTSUBSCRIPT italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT.
The braking distance to stop the rover from a velocity v𝑣vitalic\_v is therefore dbr(v)=v2/(2⋅amax)subscript𝑑𝑏𝑟𝑣superscript𝑣2⋅2subscript𝑎𝑚𝑎𝑥d\_{br}(v)=v^{2}/(2\cdot a\_{max})italic\_d start\_POSTSUBSCRIPT italic\_b italic\_r end\_POSTSUBSCRIPT ( italic\_v ) = italic\_v start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT / ( 2 ⋅ italic\_a start\_POSTSUBSCRIPT italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT ). The BC can be certified to prevent the rover from colliding with an obstacle if it takes over control in a state where dbr(v)subscript𝑑𝑏𝑟𝑣d\_{br}(v)italic\_d start\_POSTSUBSCRIPT italic\_b italic\_r end\_POSTSUBSCRIPT ( italic\_v ) is less than the minimum distance dminsubscript𝑑𝑚𝑖𝑛d\_{min}italic\_d start\_POSTSUBSCRIPT italic\_m italic\_i italic\_n end\_POSTSUBSCRIPT to any obstacle. The set of such states is the recoverable region of this BC.
A control input
is called *recoverable* if it keeps the plant inside ℛBCsubscriptℛ𝐵𝐶\mathcal{R}\_{BC}caligraphic\_R start\_POSTSUBSCRIPT italic\_B italic\_C end\_POSTSUBSCRIPT within the next time step. Otherwise, the control input is called *unrecoverable*. The DM switches control to the BC when the AC produces an unrecoverable control input. The DM’s *switching condition* determines whether a control input is unrecoverable. We also refer to it as the forward switching condition (FSC) to distinguish it from the condition for *reverse switching*, a new feature of NSA.
Techniques to determine the FSC include: (i) shrink ℛBCsubscriptℛ𝐵𝐶\mathcal{R}\_{BC}caligraphic\_R start\_POSTSUBSCRIPT italic\_B italic\_C end\_POSTSUBSCRIPT by an amount equal to a time step times the maximum gradient of the state with respect to the control input; then classify any control input as unrecoverable if the current state is outside this smaller region; (ii) simulate a model of the plant for one time step if the model is deterministic and check whether the plant strays from ℛBCsubscriptℛ𝐵𝐶\mathcal{R}\_{BC}caligraphic\_R start\_POSTSUBSCRIPT italic\_B italic\_C end\_POSTSUBSCRIPT; (iii) compute a set of states reachable within one time step and determine whether the reachable set contains states outside ℛBCsubscriptℛ𝐵𝐶\mathcal{R}\_{BC}caligraphic\_R start\_POSTSUBSCRIPT italic\_B italic\_C end\_POSTSUBSCRIPT.
#### Reinforcement Learning.
Reinforcement learning [[34](#bib.bib34)] deals with the problem of how an *agent* learns which sequence of *actions* to take in a given *environment* such that a cumulative *reward* is maximized. At each time step t𝑡titalic\_t, the agent receives observation stsubscript𝑠𝑡s\_{t}italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT (the environment state) and reward rtsubscript𝑟𝑡r\_{t}italic\_r start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT from the environment and takes action atsubscript𝑎𝑡a\_{t}italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT. The environment receives action atsubscript𝑎𝑡a\_{t}italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT and emits observation st+1subscript𝑠𝑡1s\_{t+1}italic\_s start\_POSTSUBSCRIPT italic\_t + 1 end\_POSTSUBSCRIPT and reward rt+1subscript𝑟𝑡1r\_{t+1}italic\_r start\_POSTSUBSCRIPT italic\_t + 1 end\_POSTSUBSCRIPT in response. In the control of autonomous systems, the agent represents the controller, the environment represents the plant, and the state and action spaces are typically continuous.
The goal of RL is to learn a *policy* π(a|s)𝜋conditional𝑎𝑠\pi(a\,|\,s)italic\_π ( italic\_a | italic\_s ), i.e., a way of choosing an action a𝑎aitalic\_a having observed s𝑠sitalic\_s, that maximizes the expected *return* from the initial state, where the return at time t𝑡titalic\_t is defined as the discounted sum of future rewards from t𝑡titalic\_t (following policy π𝜋\piitalic\_π): Rt=∑k=t∞γk−trk+1subscript𝑅𝑡superscriptsubscript𝑘𝑡superscript𝛾𝑘𝑡subscript𝑟𝑘1R\_{t}=\sum\_{k=t}^{\infty}\gamma^{k-t}r\_{k+1}italic\_R start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT = ∑ start\_POSTSUBSCRIPT italic\_k = italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT ∞ end\_POSTSUPERSCRIPT italic\_γ start\_POSTSUPERSCRIPT italic\_k - italic\_t end\_POSTSUPERSCRIPT italic\_r start\_POSTSUBSCRIPT italic\_k + 1 end\_POSTSUBSCRIPT; here γ∈[0,1]𝛾01\gamma\in[0,1]italic\_γ ∈ [ 0 , 1 ] is a discount factor. For this purpose, RL algorithms typically involve estimating the action-value function Qπ(s,a)=𝔼[Rt∣st=s,a]superscript𝑄𝜋𝑠𝑎𝔼delimited-[]conditionalsubscript𝑅𝑡subscript𝑠𝑡𝑠𝑎Q^{\pi}(s,a)=\mathbb{E}[R\_{t}\mid s\_{t}=s,a]italic\_Q start\_POSTSUPERSCRIPT italic\_π end\_POSTSUPERSCRIPT ( italic\_s , italic\_a ) = blackboard\_E [ italic\_R start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ∣ italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT = italic\_s , italic\_a ], i.e., the expected return for selecting action a𝑎aitalic\_a in state s𝑠sitalic\_s and then always following policy π𝜋\piitalic\_π; and the state-value function Vπ(s)=𝔼[Rt∣st=s]superscript𝑉𝜋𝑠𝔼delimited-[]conditionalsubscript𝑅𝑡subscript𝑠𝑡𝑠V^{\pi}(s)=\mathbb{E}[R\_{t}\mid s\_{t}=s]italic\_V start\_POSTSUPERSCRIPT italic\_π end\_POSTSUPERSCRIPT ( italic\_s ) = blackboard\_E [ italic\_R start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ∣ italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT = italic\_s ], i.e., the expected return starting from s𝑠sitalic\_s and following π𝜋\piitalic\_π.
While early RL algorithms were designed for discrete state and action spaces, recent *deep RL* algorithms, such as TRPO [[27](#bib.bib27)], DDPG [[21](#bib.bib21)], A3C [[23](#bib.bib23)], and ACER [[38](#bib.bib38)], have emerged as promising solutions for RL-based control problems in continuous domains. These algorithms leverage the expressiveness of deep neural networks (DNNs) to represent policies and value functions.
3 Neural Simplex Architecture
------------------------------
In this section, we discuss the main components of NSA, namely the neural controller (NC), the adaptation module (AM), and the reverse switching logic. These components in particular are not found in the Simplex control architecture.
The dynamics of the plant, i.e., the system under control, is given by
st+1=f(st,at)subscript𝑠𝑡1𝑓subscript𝑠𝑡subscript𝑎𝑡s\_{t+1}=f(s\_{t},a\_{t})italic\_s start\_POSTSUBSCRIPT italic\_t + 1 end\_POSTSUBSCRIPT = italic\_f ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ),
where st∈𝒮subscript𝑠𝑡𝒮s\_{t}\in\mathcal{S}italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ∈ caligraphic\_S is the state of the plant at time t𝑡titalic\_t, 𝒮⊆ℝn𝒮superscriptℝ𝑛\mathcal{S}\subseteq\mathbb{R}^{n}caligraphic\_S ⊆ blackboard\_R start\_POSTSUPERSCRIPT italic\_n end\_POSTSUPERSCRIPT is the real-valued state space, f𝑓fitalic\_f is a possibly nonlinear function, and
at∈𝒜subscript𝑎𝑡𝒜a\_{t}\in\mathcal{A}italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ∈ caligraphic\_A is the
control input to the plant
at time t𝑡titalic\_t,
with 𝒜⊆ℝm𝒜superscriptℝ𝑚\mathcal{A}\subseteq\mathbb{R}^{m}caligraphic\_A ⊆ blackboard\_R start\_POSTSUPERSCRIPT italic\_m end\_POSTSUPERSCRIPT the
action space. This equation specifies a deterministic dynamics, even though our approach equally supports nondeterministic (st+1∈fnd(st,at)subscript𝑠𝑡1subscript𝑓𝑛𝑑subscript𝑠𝑡subscript𝑎𝑡s\_{t+1}\in f\_{nd}(s\_{t},a\_{t})italic\_s start\_POSTSUBSCRIPT italic\_t + 1 end\_POSTSUBSCRIPT ∈ italic\_f start\_POSTSUBSCRIPT italic\_n italic\_d end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT )) and stochastic (st+1∼fst(s∣st,ats\_{t+1}\sim f\_{st}(s\mid s\_{t},a\_{t}italic\_s start\_POSTSUBSCRIPT italic\_t + 1 end\_POSTSUBSCRIPT ∼ italic\_f start\_POSTSUBSCRIPT italic\_s italic\_t end\_POSTSUBSCRIPT ( italic\_s ∣ italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT)) plant dynamics.
We assume full observability, i.e., that the BC and NC have access to the full state of the system stsubscript𝑠𝑡s\_{t}italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT.111In case of partial observability, the full state can typically be reconstructed from sequences of past states and actions, but this process is error-prone.
We denote with DMt∈{NC,BC}subscriptDM𝑡NCBC\mathrm{DM}\_{t}\in\{\mathrm{NC},\mathrm{BC}\}roman\_DM start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ∈ { roman\_NC , roman\_BC } the state of the decision module at time t𝑡titalic\_t: DMt=NCsubscriptDM𝑡NC\mathrm{DM}\_{t}=\mathrm{NC}roman\_DM start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT = roman\_NC (DMt=BCsubscriptDM𝑡BC\mathrm{DM}\_{t}=\mathrm{BC}roman\_DM start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT = roman\_BC) indicates that the neural (baseline) controller is in control. Let atNCsuperscriptsubscript𝑎𝑡NCa\_{t}^{\rm NC}italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT roman\_NC end\_POSTSUPERSCRIPT and atBCsuperscriptsubscript𝑎𝑡BCa\_{t}^{\rm BC}italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT roman\_BC end\_POSTSUPERSCRIPT denote the action computed by the NC and the BC, respectively. The final action atsubscript𝑎𝑡a\_{t}italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT performed by the NSA agent depends on the DM state: at=atNCsubscript𝑎𝑡superscriptsubscript𝑎𝑡NCa\_{t}=a\_{t}^{\rm NC}italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT = italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT roman\_NC end\_POSTSUPERSCRIPT if DMt=NCsubscriptDM𝑡NC\mathrm{DM}\_{t}=\mathrm{NC}roman\_DM start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT = roman\_NC, at=atBCsubscript𝑎𝑡superscriptsubscript𝑎𝑡BCa\_{t}=a\_{t}^{\rm BC}italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT = italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT roman\_BC end\_POSTSUPERSCRIPT if DMt=BCsubscriptDM𝑡BC\mathrm{DM}\_{t}=\mathrm{BC}roman\_DM start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT = roman\_BC.
Let β𝛽\betaitalic\_β be the BC’s control law, i.e., atBC=β(st)superscriptsubscript𝑎𝑡BC𝛽subscript𝑠𝑡a\_{t}^{\rm BC}=\beta(s\_{t})italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT roman\_BC end\_POSTSUPERSCRIPT = italic\_β ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ). For a set of unsafe states 𝒰⊆𝒮𝒰𝒮\mathcal{U}\subseteq\mathcal{S}caligraphic\_U ⊆ caligraphic\_S, the *recoverable region* is
the largest set ℛBCsubscriptℛ𝐵𝐶\mathcal{R}\_{BC}caligraphic\_R start\_POSTSUBSCRIPT italic\_B italic\_C end\_POSTSUBSCRIPT such that s∈ℛBC⇒f(s,β(s))∈ℛBC𝑠subscriptℛ𝐵𝐶⇒𝑓𝑠𝛽𝑠subscriptℛ𝐵𝐶s\in\mathcal{R}\_{BC}\Rightarrow f(s,\beta(s))\in\mathcal{R}\_{BC}italic\_s ∈ caligraphic\_R start\_POSTSUBSCRIPT italic\_B italic\_C end\_POSTSUBSCRIPT ⇒ italic\_f ( italic\_s , italic\_β ( italic\_s ) ) ∈ caligraphic\_R start\_POSTSUBSCRIPT italic\_B italic\_C end\_POSTSUBSCRIPT and ℛBC∩𝒰=∅subscriptℛ𝐵𝐶𝒰\mathcal{R}\_{BC}\cap\mathcal{U}=\emptysetcaligraphic\_R start\_POSTSUBSCRIPT italic\_B italic\_C end\_POSTSUBSCRIPT ∩ caligraphic\_U = ∅. For s∈𝒮𝑠𝒮s\in\mathcal{S}italic\_s ∈ caligraphic\_S, a∈𝒜𝑎𝒜a\in\mathcal{A}italic\_a ∈ caligraphic\_A, the forward switching condition must satisfy f(s,a)∉ℛBC⇒FSC(s,a)𝑓𝑠𝑎subscriptℛ𝐵𝐶⇒FSC𝑠𝑎f(s,a)\not\in\mathcal{R}\_{BC}\Rightarrow\text{FSC}(s,a)italic\_f ( italic\_s , italic\_a ) ∉ caligraphic\_R start\_POSTSUBSCRIPT italic\_B italic\_C end\_POSTSUBSCRIPT ⇒ FSC ( italic\_s , italic\_a ).
#### The Neural Controller.
The NC is represented by a DNN-based policy πθtsubscript𝜋subscript𝜃𝑡\pi\_{\theta\_{t}}italic\_π start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT, where θtsubscript𝜃𝑡\theta\_{t}italic\_θ start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT are the current DNN parameters. The policy maps the current state into a proposed action atNC=πθt(st)superscriptsubscript𝑎𝑡NCsubscript𝜋subscript𝜃𝑡subscript𝑠𝑡a\_{t}^{\rm NC}=\pi\_{\theta\_{t}}(s\_{t})italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT roman\_NC end\_POSTSUPERSCRIPT = italic\_π start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ). We stress the time dependency of the parameters because adaptation and retraining of the policy is a key feature of NSA. As for the dynamics f𝑓fitalic\_f, our approach supports stochastic policies (atNC∼π(a∣st,θt)similar-tosuperscriptsubscript𝑎𝑡NC𝜋conditional𝑎subscript𝑠𝑡subscript𝜃𝑡a\_{t}^{\rm NC}\sim\pi(a\mid s\_{t},\theta\_{t})italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT roman\_NC end\_POSTSUPERSCRIPT ∼ italic\_π ( italic\_a ∣ italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_θ start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT )).
The NC can be obtained using any RL algorithm. We used DDPG with the safe learning strategy of penalizing unrecoverable actions, as discussed in Section [4](#S4 "4 Safe Reinforcement Learning with Penalized Unrecoverable Continuous Actions ‣ Neural Simplex Architecture"). DDPG is attractive as it
works with deterministic policies, and allows uncorrelated samples to be added to the pool of samples for training or retraining. The latter property is important because it allows us to collect disconnected samples of what the NC would do while the plant is under the BC’s control, and to use these samples for online retraining of the NC.
#### Adaptation and Retraining.
The AM is used to retrain the NC in an online manner while the BC is in control of the plant (due to NC-to-BC failover). The main purpose of this retraining is to make the NC less likely to trigger the FSC, thereby allowing it to remain in control for longer periods of time, thereby improving overall system performance.
Techniques that we consider for online retraining of the NC include supervised learning and reinforcement learning. In supervised learning, state-action pairs of the form (s,a)𝑠𝑎(s,a)( italic\_s , italic\_a ) are required for training purposes. The training algorithm uses these examples to teach the NC safe behavior. The control inputs produced by the BC can be used as training samples, although this will train the NC to imitate BC’s behavior, which may lead to a loss in performance.
We therefore prefer SRL for online retraining, with a reward function that penalizes unrecoverable actions and rewards recoverable, high-performing ones. The reward function for retraining can be designed as follows.
| | | | |
| --- | --- | --- | --- |
| | r(s,a,s′)={r𝑢𝑛𝑟𝑒𝑐𝑜𝑣,if FSC(s,a)r𝑝𝑒𝑟𝑓(s,a,s′),otherwise𝑟𝑠𝑎superscript𝑠′casessubscript𝑟𝑢𝑛𝑟𝑒𝑐𝑜𝑣if FSC𝑠𝑎subscript𝑟𝑝𝑒𝑟𝑓𝑠𝑎superscript𝑠′otherwiser(s,a,s^{\prime})=\begin{cases}r\_{\it unrecov},&\text{if }\text{FSC}(s,a)\\
r\_{{\it perf}}(s,a,s^{\prime}),&\text{otherwise}\end{cases}italic\_r ( italic\_s , italic\_a , italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT ) = { start\_ROW start\_CELL italic\_r start\_POSTSUBSCRIPT italic\_unrecov end\_POSTSUBSCRIPT , end\_CELL start\_CELL if roman\_FSC ( italic\_s , italic\_a ) end\_CELL end\_ROW start\_ROW start\_CELL italic\_r start\_POSTSUBSCRIPT italic\_perf end\_POSTSUBSCRIPT ( italic\_s , italic\_a , italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT ) , end\_CELL start\_CELL otherwise end\_CELL end\_ROW | | (1) |
where
r𝑝𝑒𝑟𝑓(s,a,s′)subscript𝑟𝑝𝑒𝑟𝑓𝑠𝑎superscript𝑠′r\_{{\it perf}}(s,a,s^{\prime})italic\_r start\_POSTSUBSCRIPT italic\_perf end\_POSTSUBSCRIPT ( italic\_s , italic\_a , italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT ) is a performance-related reward function, and r𝑢𝑛𝑟𝑒𝑐𝑜𝑣subscript𝑟𝑢𝑛𝑟𝑒𝑐𝑜𝑣r\_{\it unrecov}italic\_r start\_POSTSUBSCRIPT italic\_unrecov end\_POSTSUBSCRIPT is a negative number used to penalize unrecoverable actions. The benefits of this approach to SRL are discussed in Section [4](#S4 "4 Safe Reinforcement Learning with Penalized Unrecoverable Continuous Actions ‣ Neural Simplex Architecture").
The AM retrains the NC at each time step the BC is in control by maintaining a pool of retraining samples of the form (st,atNC,s′,r′)subscript𝑠𝑡superscriptsubscript𝑎𝑡NCsuperscript𝑠′superscript𝑟′(s\_{t},a\_{t}^{\rm NC},s^{\prime},r^{\prime})( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT roman\_NC end\_POSTSUPERSCRIPT , italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT , italic\_r start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT ), where atNCsuperscriptsubscript𝑎𝑡NCa\_{t}^{\rm NC}italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT roman\_NC end\_POSTSUPERSCRIPT is the NC-proposed action, s′=f(st,atNC)superscript𝑠′𝑓subscript𝑠𝑡superscriptsubscript𝑎𝑡NCs^{\prime}=f(s\_{t},a\_{t}^{\rm NC})italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT = italic\_f ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT roman\_NC end\_POSTSUPERSCRIPT ) is the state that the system would evolve to if the NC was in control, and r′=r(s,atNC,s′)superscript𝑟′𝑟𝑠superscriptsubscript𝑎𝑡NCsuperscript𝑠′r^{\prime}=r(s,a\_{t}^{\rm NC},s^{\prime})italic\_r start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT = italic\_r ( italic\_s , italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT roman\_NC end\_POSTSUPERSCRIPT , italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT ) is the corresponding reward. I.e., samples are obtained by running the NC in shadow mode: when BC is in control, the AM obtains a retraining sample by running a simulation of the system for one time step and applying atNCsuperscriptsubscript𝑎𝑡NCa\_{t}^{\rm NC}italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT roman\_NC end\_POSTSUPERSCRIPT, while the actual system evolves according to the BC action atBCsuperscriptsubscript𝑎𝑡BCa\_{t}^{\rm BC}italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT roman\_BC end\_POSTSUPERSCRIPT.
The AM updates the NC’s parameters θtsubscript𝜃𝑡\theta\_{t}italic\_θ start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT as follows:
| | | |
| --- | --- | --- |
| | θt={RL(θt−1,(st,atNC,s′,r′)),if DMt=BCθt−1,otherwisesubscript𝜃𝑡casesRLsubscript𝜃𝑡1subscript𝑠𝑡superscriptsubscript𝑎𝑡NCsuperscript𝑠′superscript𝑟′if DMt=BCsubscript𝜃𝑡1otherwise\theta\_{t}=\begin{cases}\mathrm{RL}(\theta\_{t-1},(s\_{t},a\_{t}^{\rm NC},s^{\prime},r^{\prime})),&\text{if $\mathrm{DM}\_{t}=\mathrm{BC}$}\\
\theta\_{t-1},&\text{otherwise}\end{cases}italic\_θ start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT = { start\_ROW start\_CELL roman\_RL ( italic\_θ start\_POSTSUBSCRIPT italic\_t - 1 end\_POSTSUBSCRIPT , ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT roman\_NC end\_POSTSUPERSCRIPT , italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT , italic\_r start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT ) ) , end\_CELL start\_CELL if roman\_DM start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT = roman\_BC end\_CELL end\_ROW start\_ROW start\_CELL italic\_θ start\_POSTSUBSCRIPT italic\_t - 1 end\_POSTSUBSCRIPT , end\_CELL start\_CELL otherwise end\_CELL end\_ROW | |
where RLRL\mathrm{RL}roman\_RL is the deep RL algorithm chosen for NC adaptation.
Note that as soon as the DM switches control to the BC after the NC has produced an unrecoverable action (see also the Switching logic paragraph below), a corresponding retraining sample for the NC’s action is added to the pool.
We evaluated a number of variants of this procedure by making different choices along the following dimensions.
1. 1.
Start retraining with an empty pool of samples or with the pool created during the initial training of the NC.
2. 2.
Add (or do not add) exploration noise to NC’s action when collecting a sample. With exploration noise, the resulting action is atNC+νtsuperscriptsubscript𝑎𝑡NCsubscript𝜈𝑡a\_{t}^{\mathrm{NC}}+\nu\_{t}italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT roman\_NC end\_POSTSUPERSCRIPT + italic\_ν start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT, where νtsubscript𝜈𝑡\nu\_{t}italic\_ν start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT is a random noise term.
Note that we consider noise only when NC is running in shadow mode (BC in control), as directly using noisy actions to control the plant would degrade performance.
3. 3.
Collect retraining samples only while BC is in control
or at every time step. In both cases, the action in each training sample is the action output by NC (or a noisy version of it); we never use BC’s action in a training sample. Also, in both cases, the retraining algorithm for updating the NC
is run only while the BC is in control.
We found that reusing the pool of training samples (DDPG’s so-called experience replay buffer) from initial training of the NC helps
evolve the policy in a more stable way,
as retraining samples gradually replace initial training samples in the sample pool. Another benefit of reusing the initial training pool is that the NC can be immediately retrained without having to wait for enough samples to be collected online.
We found that adding exploration noise to NC’s actions in retraining samples, and collecting retraining samples at every time step, both increase the benefit of retraining. This is because these two strategies provide more diverse samples and thereby help achieve more thorough exploration of the state-action space.
#### Switching logic.
Figure 2: Switching boundaries. The blue region is a subset of the orange area, which in turn is a subset of the green region.
NSA includes *reverse switching* from the BC to the retrained NC.
An additional benefit of well-designed reverse switching is that it lessens the burden on the BC to achieve performance objectives, leading to a simpler BC design that focuses mainly on safety. Control of the plant is returned to the NC when the reverse switching condition (RSC) is true in the current state.
We can summarize NSA’s switching logic by describing the evolution of the DM state DMtsubscriptDM𝑡\mathrm{DM}\_{t}roman\_DM start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT. NSA begins with the NC in control, i.e., DMt=NCsubscriptDM𝑡NC\mathrm{DM}\_{t}=\mathrm{NC}roman\_DM start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT = roman\_NC for t≤0𝑡0t\leq 0italic\_t ≤ 0. For t>0𝑡0t>0italic\_t > 0, the DM state is given by:
| | | |
| --- | --- | --- |
| | DMt={BC,if DMt−1=NC and FSC(st,atNC)NC,if DMt−1=BC and RSC(st)DMt−1,otherwisesubscriptDM𝑡casesBCif DMt−1=NC and FSC(st,atNC)NCif DMt−1=BC and RSC(st)subscriptDM𝑡1otherwise\mathrm{DM}\_{t}=\begin{cases}\mathrm{BC},&\text{if $\mathrm{DM}\_{t-1}=\mathrm{NC}$ and FSC$(s\_{t},a\_{t}^{\rm NC})$}\\
\mathrm{NC},&\text{if $\mathrm{DM}\_{t-1}=\mathrm{BC}$ and RSC$(s\_{t})$}\\
\mathrm{DM}\_{t-1},&\text{otherwise}\end{cases}roman\_DM start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT = { start\_ROW start\_CELL roman\_BC , end\_CELL start\_CELL if roman\_DM start\_POSTSUBSCRIPT italic\_t - 1 end\_POSTSUBSCRIPT = roman\_NC and FSC ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT roman\_NC end\_POSTSUPERSCRIPT ) end\_CELL end\_ROW start\_ROW start\_CELL roman\_NC , end\_CELL start\_CELL if roman\_DM start\_POSTSUBSCRIPT italic\_t - 1 end\_POSTSUBSCRIPT = roman\_BC and RSC ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) end\_CELL end\_ROW start\_ROW start\_CELL roman\_DM start\_POSTSUBSCRIPT italic\_t - 1 end\_POSTSUBSCRIPT , end\_CELL start\_CELL otherwise end\_CELL end\_ROW | |
To ensure safety when returning control to the NC, the FSC must not hold if the RSC is satisfied, i.e.,
RSC(s)⇒¬FSC(s,a)⇒RSC𝑠FSC𝑠𝑎\text{RSC}(s)\Rightarrow\neg\text{FSC}(s,a)RSC ( italic\_s ) ⇒ ¬ FSC ( italic\_s , italic\_a ), for all s∈𝒮,a∈𝒜formulae-sequence𝑠𝒮𝑎𝒜s\in\mathcal{S},a\in\mathcal{A}italic\_s ∈ caligraphic\_S , italic\_a ∈ caligraphic\_A.
We seek to develop reverse switching logic that returns control to NC when it is safe to do so and which avoids frequent back-and-forth switching between the BC and NC.
We propose two such approaches.
One is to reverse-switch if a forward switch will not occur in the near future. This can be checked by simulating the composition of the NC and plant for T𝑇Titalic\_T time steps, and reverse-switching if the FSC does not hold within this time horizon.222For nondeterministic (stochastic) systems, a
(probabilistic) model checker can be used instead of a simulator, but this approach may be computationally expensive. Formally, RSC(st)=⋀t′=tt+T¬FSC(st′′,πθt(st′′))RSCsubscript𝑠𝑡superscriptsubscriptsuperscript𝑡′𝑡𝑡𝑇FSCsubscriptsuperscript𝑠′superscript𝑡′subscript𝜋subscript𝜃𝑡subscriptsuperscript𝑠′superscript𝑡′\text{RSC}(s\_{t})=\bigwedge\_{t^{\prime}=t}^{t+T}\neg\text{FSC}(s^{\prime}\_{t^{\prime}},\pi\_{\theta\_{t}}(s^{\prime}\_{t^{\prime}}))RSC ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) = ⋀ start\_POSTSUBSCRIPT italic\_t start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT = italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_t + italic\_T end\_POSTSUPERSCRIPT ¬ FSC ( italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_t start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT end\_POSTSUBSCRIPT , italic\_π start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_t start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT end\_POSTSUBSCRIPT ) ), where st′=stsubscriptsuperscript𝑠′𝑡subscript𝑠𝑡s^{\prime}\_{t}=s\_{t}italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT = italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT and st′+1′=f(st′′,πθt(st′′))subscriptsuperscript𝑠′superscript𝑡′1𝑓subscriptsuperscript𝑠′superscript𝑡′subscript𝜋subscript𝜃𝑡subscriptsuperscript𝑠′superscript𝑡′s^{\prime}\_{t^{\prime}+1}=f(s^{\prime}\_{t^{\prime}},\pi\_{\theta\_{t}}(s^{\prime}\_{t^{\prime}}))italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_t start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT + 1 end\_POSTSUBSCRIPT = italic\_f ( italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_t start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT end\_POSTSUBSCRIPT , italic\_π start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT start\_POSTSUBSCRIPT italic\_t start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT end\_POSTSUBSCRIPT ) ).
This approach, used in our inverted pendulum and artificial pancreas case studies, prevents frequent switching.
A simpler approach is to reverse-switch if the current plant state is sufficiently far from the NC-to-BC switching boundary; see Fig. [2](#S3.F2 "Figure 2 ‣ Switching logic. ‣ 3 Neural Simplex Architecture ‣ Neural Simplex Architecture"). Formally, RSC(st)=sup{d(st,s′)∣s′∈𝒮,FSC(s′,πθt(s′))}>ϵRSCsubscript𝑠𝑡supremumconditional-set𝑑subscript𝑠𝑡superscript𝑠′superscript𝑠′𝒮FSCsuperscript𝑠′subscript𝜋subscript𝜃𝑡superscript𝑠′italic-ϵ\text{RSC}(s\_{t})=\sup\{d(s\_{t},s^{\prime})\mid s^{\prime}\in\mathcal{S},\ \text{FSC}(s^{\prime},\pi\_{\theta\_{t}}(s^{\prime}))\}>\epsilonRSC ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) = roman\_sup { italic\_d ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT , italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT ) ∣ italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT ∈ caligraphic\_S , FSC ( italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT , italic\_π start\_POSTSUBSCRIPT italic\_θ start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT ) ) } > italic\_ϵ, where d𝑑ditalic\_d is a metric on ℝnsuperscriptℝ𝑛\mathbb{R}^{n}blackboard\_R start\_POSTSUPERSCRIPT italic\_n end\_POSTSUPERSCRIPT and ϵ∈ℝ+italic-ϵsuperscriptℝ\epsilon\in\mathbb{R}^{+}italic\_ϵ ∈ blackboard\_R start\_POSTSUPERSCRIPT + end\_POSTSUPERSCRIPT is the desired distance. This approach is used in our rover case study.
We emphasize that the choice of RSC does not affect safety and is application-dependent. Note that both of our approaches construct an RSC that is stricter than a straight complement of the FSC. This helps avoid excessive switching.
In our experiments,
we empirically observed that the system behavior was not very sensitive to the exact
value of T𝑇Titalic\_T or ϵitalic-ϵ\epsilonitalic\_ϵ; so choosing acceptable values for them is not difficult.
4 Safe Reinforcement Learning with Penalized Unrecoverable Continuous Actions
------------------------------------------------------------------------------
We evaluate the use of two policy-gradient algorithms for safe reinforcement learning in NSA. The first approach filters the learning agent’s unrecoverable actions before they reach the plant. For example, when the learning agent, i.e., the NC, produces an unrecoverable action, a runtime monitor [[14](#bib.bib14)] or a preemptive shield [[3](#bib.bib3)] replaces it with a recoverable one to continue the trajectory. The recoverable action is also passed to the RL algorithm to update the agent and training continues with the rest of the trajectory.
In the second approach, when the learning agent produces an unrecoverable action,
we assign a penalty (negative reward) to the action, use it as a training sample, and then use recoverable actions to safely terminate the trajectory (but not to train the agent). Safely terminating the trajectory is important in cases where for example the live system is used for training. We call this approach *safe reinforcement learning with penalized unrecoverable continuous actions* (SRL-PUA). By “continuous” here we mean real-valued action spaces, as in [[10](#bib.bib10)]. Other SRL approaches such as [[2](#bib.bib2)] use discrete actions.
To compare the two approaches, we used the DDPG and TRPO algorithms to train neural controllers for an inverted pendulum (IP) control system. Details about our IP case study, including the reward function and the BC used to generate recoverable actions, can be found in the Appendix.
We used the implementations of DDPG and TRPO in rllab [[12](#bib.bib12)]. For TRPO, we trained two DNNs, one for the mean and the other for the standard deviation of a Gaussian policy. Both DNNs have two fully connected hidden layers of 32 neurons each and one output layer. The hidden layers use the tanh activation function, and the output layer is linear. For DDPG, we trained a DNN that computes the action directly from the state. The DNN has two fully connected hidden layers of 32 neurons each and one output layer. The hidden layers use the ReLU activation function, and the output layer uses tanh. We followed the choice of activation functions in the examples accompanying rllab.
For each algorithm, we ran two training experiments. In one experiment, we reproduce the filtering approach; i.e., we replace an unrecoverable action produced by the learning agent with the BC’s recoverable action, use the latter as the training sample, and continue the trajectory. We call this training method SRL-BC. In the other experiment, we evaluate the SRL-PUA approach. Note that both algorithms explore different trajectories by resetting the system to a random initial state whenever the current trajectory is terminated. We set the maximum trajectory length to 500 time steps, meaning that a trajectory is terminated when it exceeds 500 time steps.
We trained the DDPG and TRPO agents on a total of one million time steps. After training, we evaluated all trained policies on the same set of 1,000 random initial states. During evaluation, if an agent produces an unrecoverable action, the trajectory is terminated.
The results are shown in Table [1](#S4.T1 "Table 1 ‣ 4 Safe Reinforcement Learning with Penalized Unrecoverable Continuous Actions ‣ Neural Simplex Architecture"). For both algorithms, the policies trained with recoverable actions (the SRL-BC approach) produce unrecoverable actions in all test trajectories, while the SRL-PUA approach, where the policies are trained with penalties for unrecoverable actions, does not produce any such actions. As such, the latter policies achieve superior returns and trajectory lengths (they are able to safely control the system the entire time).
In the above experiments, we replaced unrecoverable actions with actions generated by a deterministic BC, whereas the monitoring [[14](#bib.bib14)] and preemptive shielding [[2](#bib.bib2)] approaches allow unrecoverable actions to be replaced with random recoverable ones, an approach we refer to as SRL-RND. To show that our conclusions are independent of this difference, we ran one more experiment with each learning algorithm, in which we replaced each unrecoverable action with an action selected by randomly generating actions until a recoverable one is found. The results, shown in Table [2](#S4.T2 "Table 2 ‣ 4 Safe Reinforcement Learning with Penalized Unrecoverable Continuous Actions ‣ Neural Simplex Architecture"), once again demonstrate that training with only recoverable actions is ineffective.
Compared to filtering-based approaches (SRL-BC in Table [1](#S4.T1 "Table 1 ‣ 4 Safe Reinforcement Learning with Penalized Unrecoverable Continuous Actions ‣ Neural Simplex Architecture") and SRL-RND in Table [2](#S4.T2 "Table 2 ‣ 4 Safe Reinforcement Learning with Penalized Unrecoverable Continuous Actions ‣ Neural Simplex Architecture")), the SRL-PUA approach yields a 25- to 775-fold improvement in the average return.
| | | |
| --- | --- | --- |
| | TRPO | DDPG |
| | SRL-BC | SRL-PUA | SRL-BC | SRL-PUA |
| Unrec Trajs | 1,000 | 0 | 1,000 | 0 |
| Comp Trajs | 0 | 1,000 | 0 | 1,000 |
| Avg. Return | 112.53 | 4,603.97 | 61.52 | 4,596.04 |
| Avg. Length | 15.15 | 500 | 14.56 | 500 |
Table 1: Policy performance comparison. SRL-BC: policy trained with BC’s actions replacing unrecoverable ones. SRL-PUA: policy trained with penalized unsafe actions. Unrec Trajs: # trajectories terminated
due to an unrecoverable action. Comp Trajs: # trajectories that reach the limit of 500 time steps. Avg. Return and Avg. Length: average return and trajectory length over 1,000
trajectories.
| | | |
| --- | --- | --- |
| | TRPO | DDPG |
| | SRL-RND | SRL-PUA | SRL-RND | SRL-PUA |
| Unrec Trajs | 1,000 | 0 | 1,000 | 0 |
| Comp Trajs | 0 | 1,000 | 0 | 1,000 |
| Avg. Return | 183.36 | 4,603.97 | 5.93 | 4,596.04 |
| Avg. Length | 1.93 | 500 | 14 | 500 |
Table 2: Policy performance comparison. SRL-RND: policy trained with random recoverable actions replacing unrecoverable ones.
5 Case Studies
---------------
An additional case study, the Inverted Pendulum, along with further details about the case studies presented in this section can be found in the Appendix.
###
5.1 Rover Navigation
We consider the problem of navigating a rover to a predetermined target location while avoiding collisions with static obstacles. The rover is a circular disk of radius r𝑟ritalic\_r. It has a maximum speed vmaxsubscript𝑣𝑚𝑎𝑥v\_{max}italic\_v start\_POSTSUBSCRIPT italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT and a maximum acceleration amaxsubscript𝑎𝑚𝑎𝑥a\_{max}italic\_a start\_POSTSUBSCRIPT italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT. The maximum braking time is therefore tbr\_max=vmax/amaxsubscript𝑡𝑏𝑟\_𝑚𝑎𝑥subscript𝑣𝑚𝑎𝑥subscript𝑎𝑚𝑎𝑥t\_{br\\_max}=v\_{max}/a\_{max}italic\_t start\_POSTSUBSCRIPT italic\_b italic\_r \_ italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT = italic\_v start\_POSTSUBSCRIPT italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT / italic\_a start\_POSTSUBSCRIPT italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT, and the maximum braking distance is dbr\_max=vmax2/(2⋅amax)subscript𝑑𝑏𝑟\_𝑚𝑎𝑥superscriptsubscript𝑣𝑚𝑎𝑥2⋅2subscript𝑎𝑚𝑎𝑥d\_{br\\_max}=v\_{max}^{2}/(2\cdot a\_{max})italic\_d start\_POSTSUBSCRIPT italic\_b italic\_r \_ italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT = italic\_v start\_POSTSUBSCRIPT italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT / ( 2 ⋅ italic\_a start\_POSTSUBSCRIPT italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT ). The control inputs are the accelerations axsubscript𝑎𝑥a\_{x}italic\_a start\_POSTSUBSCRIPT italic\_x end\_POSTSUBSCRIPT and aysubscript𝑎𝑦a\_{y}italic\_a start\_POSTSUBSCRIPT italic\_y end\_POSTSUBSCRIPT in the x𝑥xitalic\_x and y𝑦yitalic\_y directions, respectively. The system uses discrete-time control with a time step of dt𝑑𝑡dtitalic\_d italic\_t.
The rover
has n𝑛nitalic\_n distance sensors whose detection range is lmaxsubscript𝑙𝑚𝑎𝑥l\_{max}italic\_l start\_POSTSUBSCRIPT italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT. The sensors are placed evenly around the perimeter of the rover; i.e., the center lines of sight of two adjacent sensors form an angle of 2π/n2𝜋𝑛2\pi/n2 italic\_π / italic\_n. The rover can only move forwards, so its orientation is the same as its heading angle. The state vector for the rover is [x,y,θ,v,l1,l2,…,ln]𝑥𝑦𝜃𝑣subscript𝑙1subscript𝑙2…subscript𝑙𝑛[x,y,\theta,v,l\_{1},l\_{2},...,l\_{n}][ italic\_x , italic\_y , italic\_θ , italic\_v , italic\_l start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , italic\_l start\_POSTSUBSCRIPT 2 end\_POSTSUBSCRIPT , … , italic\_l start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT ], where (x,y)𝑥𝑦(x,y)( italic\_x , italic\_y ) is the position, θ𝜃\thetaitalic\_θ is the heading angle, v𝑣vitalic\_v is the velocity, and the lisubscript𝑙𝑖l\_{i}italic\_l start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT’s are the sensor readings.
Figure 3: Illustration of assumptions about obstacle shapes.
We assume the sensors have a small angular field-of-view so that each sensor reading reflects the distance from the rover to an obstacle along the sensor’s center line of sight. If a sensor does not detect an obstacle, its reading is lmaxsubscript𝑙𝑚𝑎𝑥l\_{max}italic\_l start\_POSTSUBSCRIPT italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT.
We also assume that when the sensor readings of two adjacent sensors sisubscript𝑠𝑖s\_{i}italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT and sjsubscript𝑠𝑗s\_{j}italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT are lisubscript𝑙𝑖l\_{i}italic\_l start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT and ljsubscript𝑙𝑗l\_{j}italic\_l start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT, respectively, then the (conservative) minimum distance to any obstacle point located in the cone formed by the center lines of sight of sisubscript𝑠𝑖s\_{i}italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT and sjsubscript𝑠𝑗s\_{j}italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT is min{li,lj}−ϵsubscript𝑙𝑖subscript𝑙𝑗italic-ϵ\min\{l\_{i},l\_{j}\}-\epsilonroman\_min { italic\_l start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT , italic\_l start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT } - italic\_ϵ. Here, ϵitalic-ϵ\epsilonitalic\_ϵ is a constant that limits by how much an obstacle can protrude into the blind spot between sisubscript𝑠𝑖s\_{i}italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT and sjsubscript𝑠𝑗s\_{j}italic\_s start\_POSTSUBSCRIPT italic\_j end\_POSTSUBSCRIPT’s lines of sight; see Fig. [3](#S5.F3 "Figure 3 ‣ 5.1 Rover Navigation ‣ 5 Case Studies ‣ Neural Simplex Architecture").
A state s𝑠sitalic\_s of the rover is *recoverable* if starting from s𝑠sitalic\_s, the baseline controller (BC) can brake to a stop at least distance d𝑠𝑎𝑓𝑒subscript𝑑𝑠𝑎𝑓𝑒d\_{\it safe}italic\_d start\_POSTSUBSCRIPT italic\_safe end\_POSTSUBSCRIPT from any obstacle. Let the braking distance in state s𝑠sitalic\_s be dbr(s)=v2/(2⋅amax)subscript𝑑𝑏𝑟𝑠superscript𝑣2⋅2subscript𝑎𝑚𝑎𝑥d\_{br}(s)=v^{2}/(2\cdot a\_{max})italic\_d start\_POSTSUBSCRIPT italic\_b italic\_r end\_POSTSUBSCRIPT ( italic\_s ) = italic\_v start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT / ( 2 ⋅ italic\_a start\_POSTSUBSCRIPT italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT ), where v𝑣vitalic\_v is the rover’s speed in s𝑠sitalic\_s. Then s𝑠sitalic\_s is recoverable if the minimum sensor reading lminsubscript𝑙𝑚𝑖𝑛l\_{min}italic\_l start\_POSTSUBSCRIPT italic\_m italic\_i italic\_n end\_POSTSUBSCRIPT in state s𝑠sitalic\_s is at least d𝑠𝑎𝑓𝑒+dbr(s)+ϵsubscript𝑑𝑠𝑎𝑓𝑒subscript𝑑𝑏𝑟𝑠italic-ϵd\_{\it safe}+d\_{br}(s)+\epsilonitalic\_d start\_POSTSUBSCRIPT italic\_safe end\_POSTSUBSCRIPT + italic\_d start\_POSTSUBSCRIPT italic\_b italic\_r end\_POSTSUBSCRIPT ( italic\_s ) + italic\_ϵ.
The FSC holds when the control input uNCsubscript𝑢𝑁𝐶u\_{NC}italic\_u start\_POSTSUBSCRIPT italic\_N italic\_C end\_POSTSUBSCRIPT proposed by the NC will put the rover in an unrecoverable state in the next time step. We check this condition
by simulating the rover for one time step with uNCsubscript𝑢𝑁𝐶u\_{NC}italic\_u start\_POSTSUBSCRIPT italic\_N italic\_C end\_POSTSUBSCRIPT as the control input, and by then determining if lmin<d𝑠𝑎𝑓𝑒+dbr(s)+ϵsubscript𝑙𝑚𝑖𝑛subscript𝑑𝑠𝑎𝑓𝑒subscript𝑑𝑏𝑟𝑠italic-ϵl\_{min}<d\_{\it safe}+d\_{br}(s)+\epsilonitalic\_l start\_POSTSUBSCRIPT italic\_m italic\_i italic\_n end\_POSTSUBSCRIPT < italic\_d start\_POSTSUBSCRIPT italic\_safe end\_POSTSUBSCRIPT + italic\_d start\_POSTSUBSCRIPT italic\_b italic\_r end\_POSTSUBSCRIPT ( italic\_s ) + italic\_ϵ.
The RSC is lmin≥m⋅vmax⋅dt+d𝑠𝑎𝑓𝑒+dbr\_max+ϵsubscript𝑙𝑚𝑖𝑛⋅𝑚subscript𝑣𝑚𝑎𝑥𝑑𝑡subscript𝑑𝑠𝑎𝑓𝑒subscript𝑑𝑏𝑟\_𝑚𝑎𝑥italic-ϵl\_{min}\geq m\cdot v\_{max}\cdot dt+d\_{\it safe}+d\_{br\\_max}+\epsilonitalic\_l start\_POSTSUBSCRIPT italic\_m italic\_i italic\_n end\_POSTSUBSCRIPT ≥ italic\_m ⋅ italic\_v start\_POSTSUBSCRIPT italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT ⋅ italic\_d italic\_t + italic\_d start\_POSTSUBSCRIPT italic\_safe end\_POSTSUBSCRIPT + italic\_d start\_POSTSUBSCRIPT italic\_b italic\_r \_ italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT + italic\_ϵ, ensuring that the FSC does not hold for the next m−1𝑚1m-1italic\_m - 1 time steps.
Parameter m𝑚mitalic\_m can be chosen to reduce excessive back-and-forth switching between the NC and BC.
The BC performs the following steps:
1) Apply the maximum braking power amaxsubscript𝑎𝑚𝑎𝑥a\_{max}italic\_a start\_POSTSUBSCRIPT italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT until the rover stops.
2) Randomly pick a safe heading angle θ𝜃\thetaitalic\_θ based on the current position and sensor readings.
3) Rotate the rover until its heading angle is θ𝜃\thetaitalic\_θ.
4) Move with heading angle θ𝜃\thetaitalic\_θ until either the FSC becomes true (this is checked after each time step by the BC itself), in which case the BC is re-started at Step 1, or the RSC becomes true (this is checked by the DM), in which case the NC takes over.
#### Experimental Results.
Parameter values used: r=0.1𝑟0.1r=0.1\,italic\_r = 0.1m, vmax=0.8subscript𝑣𝑚𝑎𝑥0.8v\_{max}=0.8\,italic\_v start\_POSTSUBSCRIPT italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT = 0.8m/s, amax=1.6subscript𝑎𝑚𝑎𝑥1.6a\_{max}=1.6\,italic\_a start\_POSTSUBSCRIPT italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT = 1.6m/s2superscripts2\text{s}^{2}s start\_POSTSUPERSCRIPT 2 end\_POSTSUPERSCRIPT, lmax=2subscript𝑙𝑚𝑎𝑥2l\_{max}=2\,italic\_l start\_POSTSUBSCRIPT italic\_m italic\_a italic\_x end\_POSTSUBSCRIPT = 2m, n=32𝑛32n=32italic\_n = 32, d𝑠𝑎𝑓𝑒=0.2subscript𝑑𝑠𝑎𝑓𝑒0.2d\_{\it safe}=0.2\,italic\_d start\_POSTSUBSCRIPT italic\_safe end\_POSTSUBSCRIPT = 0.2m, ϵ=0.01italic-ϵ0.01\epsilon=0.01\,italic\_ϵ = 0.01m, m=5𝑚5m=5italic\_m = 5, dt=0.1𝑑𝑡0.1dt=0.1\,italic\_d italic\_t = 0.1s. The target is a circular disk at location (0,0)00(0,0)( 0 , 0 ) with a radius of 0.1m. The obstacle field, which is fixed during training and testing, consists of 12 circular obstacles with a minimum radius of 0.25m.
Rover initial position (x0,y0)subscript𝑥0subscript𝑦0(x\_{0},y\_{0})( italic\_x start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT , italic\_y start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ) is randomized in the area [−5,5]×[−5,5]5555[-5,5]\times[-5,5][ - 5 , 5 ] × [ - 5 , 5 ]. 333Although the obstacles are fixed, the NC still generalizes well (but not perfectly) to random obstacle fields not seen during training, as shown in this video https://youtu.be/ICT8D1uniIw.
We assume that the sensor field-of-view is at least 7.25∘superscript7.257.25^{\circ}7.25 start\_POSTSUPERSCRIPT ∘ end\_POSTSUPERSCRIPT, thereby satisfying the assumption that an obstacle does not protrude more than ϵitalic-ϵ\epsilonitalic\_ϵ into the blind spot between adjacent sensors. See also Fig. [3](#S5.F3 "Figure 3 ‣ 5.1 Rover Navigation ‣ 5 Case Studies ‣ Neural Simplex Architecture").
The NC is a DNN with two ReLU hidden layers, each of size 64, and a tanh output layer. We used the DDPG algorithm for both initial training and online retraining of the NC.
For initial training,
we ran DDPG for 5 million time steps. The reward function for initial training and online retraining is:
| | | | |
| --- | --- | --- | --- |
| | r(s,a,s′)={−20,000,if FSC(s,a)10,000,if DT(s)≤0.2−1−20⋅DT(s),otherwise𝑟𝑠𝑎superscript𝑠′cases20000
if FSC𝑠𝑎10000
if DT𝑠0.21⋅20DT𝑠otherwiser(s,a,s^{\prime})=\begin{cases}-20,000,&\text{if }\text{FSC}(s,a)\\
10,000,&\text{if }\text{DT}(s)\leq 0.2\\
-1-20\cdot\text{DT}(s),&\text{otherwise}\end{cases}italic\_r ( italic\_s , italic\_a , italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT ) = { start\_ROW start\_CELL - 20 , 000 , end\_CELL start\_CELL if roman\_FSC ( italic\_s , italic\_a ) end\_CELL end\_ROW start\_ROW start\_CELL 10 , 000 , end\_CELL start\_CELL if roman\_DT ( italic\_s ) ≤ 0.2 end\_CELL end\_ROW start\_ROW start\_CELL - 1 - 20 ⋅ DT ( italic\_s ) , end\_CELL start\_CELL otherwise end\_CELL end\_ROW | | (2) |
where FSC(s,a)FSC𝑠𝑎\text{FSC}(s,a)FSC ( italic\_s , italic\_a ) is the forward switching condition and DT(s)DT𝑠\text{DT}(s)DT ( italic\_s ) is the center-to-center distance from the rover to the target in state s𝑠sitalic\_s. The rover is considered to have reached the target if DT(s)≤0.2DT𝑠0.2\text{DT}(s)\leq 0.2DT ( italic\_s ) ≤ 0.2, as, recall, the target is a circular disk with radius of 0.1m and the radius r𝑟ritalic\_r of the rover is 0.1m. If the action a𝑎aitalic\_a triggers the forward switching logic, it is penalized by assigning it a negative reward of -20,000. If a𝑎aitalic\_a causes the rover to reach the target, it receives a positive reward of 10,000. All other actions are penalized by an amount proportional to the distance to the target, encouraging the agent to reach the target quickly.
Our experiments with online retraining use the same DDPG settings as in initial training, except that we initialize the AM’s pool of retraining samples with the pool created by initial training, instead of an empty pool. The pool created by initial training contains one million samples; this is the maximum pool size, which is a parameter of the algorithm. When creating retraining samples, the AM adds Gaussian noise to the NC’s actions. The NC’s actions are collected (added to the pool) at every time step, regardless of which controller is in control; thus, the AM also collects samples of what the NC would do while the BC is in control.
We ran the NSA instance starting from 10,000 random initial states. Out of 10,000 trajectories, forward switching occurred in 456 of them. Of these 456 trajectories, the BC was in control for a total of 70,974 time steps. This means there were 70,974 (∼similar-to\sim∼71K) retraining updates to the NC. To evaluate the benefits of online retraining, we compared the performance of the NC after initial training and after 20K, 50K, and 71K online updates. We evaluated the performance of each of these controllers (by itself, without NSA) by running it from the same set of 1,000 random initial states.
The results in Table [3](#S5.T3 "Table 3 ‣ Experimental Results. ‣ 5.1 Rover Navigation ‣ 5 Case Studies ‣ Neural Simplex Architecture") show that after 71K retraining updates, the NC outperforms the initially trained version on every metric. Table [3](#S5.T3 "Table 3 ‣ Experimental Results. ‣ 5.1 Rover Navigation ‣ 5 Case Studies ‣ Neural Simplex Architecture") also shows that the NC’s performance increases with the number of retraining updates, thus demonstrating that NSA’s online retraining not only improves the safety of the NC, but also its performance.
| | | | | |
| --- | --- | --- | --- | --- |
| | IT | 20K RT | 50K RT | 71K RT |
| FSCs | 100 | 79 | 43 | 8 |
| Timeouts | 35 | 49 | 50 | 22 |
| Targets | 865 | 872 | 907 | 970 |
| Avg. Ret. | -9,137.3 | -9,968.8 | -5,314.6 | -684.0 |
| Avg. Len. | 138.67 | 142.29 | 156.13 | 146.56 |
Table 3: Benefits of online retraining (∼similar-to\sim∼71K NC updates in total) for ground rover navigation. IT: results for initially trained NC. 20K RT, 50K RT, 71K RT: results for NC after 20K, 50K, 71K retraining updates.
All controllers evaluated on same set of 1,000 random initial states. FSCs: # trajectories in which FSC becomes true. Timeouts: # trajectories that reach the
limit of 500 time steps without reaching target or having FSC become true. Targets: # trajectories that reach the target. Avg. Ret. and Avg. Len.: average return and average trajectory length
over all 1,000 trajectories.
We resumed initial training to see if this would produce similar improvements. Specifically, we continued the initial training for an additional 71K, 1M, and 3M samples. The results, included in the Appendix, show that extending the initial training slowly improves both the safety and performance of the NC but requires substantially more updates. 71K retraining updates provide significantly more benefits than even 3M additional samples of initial training.
###
5.2 Artificial Pancreas
The artificial pancreas (AP) is used to control blood glucose (BG) levels in Type 1 diabetes patients through automated delivery of insulin. We use the linear plant model of [[6](#bib.bib6)] to describe the physiological state of the patient. The main state variable of interest is G𝐺Gitalic\_G, which is the difference between the reference BG (7.87.87.87.8 mmol/L) and the patient’s BG. The control action, i.e., the insulin input, is denoted by u𝑢uitalic\_u.
Further details of this model, including its ODE dynamics, can be found in the Appendix.
The AP should maintain BG levels within the safe range of 4 to 11 mmol/L. In particular, it should avoid hypoglycemia (i.e., BG levels below the safe range), which can lead to severe health consequences. Hypoglycemia occurs when the controller overshoots the insulin dose. Insulin control is uniquely challenging because the controller cannot take a corrective action to counteract an excessive dose; its most extreme safety measure is to turn off the insulin pump. Hence, the baseline controller for the AP sets u=0𝑢0u=0italic\_u = 0.
We intentionally under-train the initial NC so that it exhibits low performance and produces unrecoverable actions. Low-performing AP controllers may arise in practice for several reasons, e.g., when the training-time model parameters do not match the current real-life patient parameters.
The reward function r𝑟ritalic\_r is designed to penalize deviations from the reference BG level, as captured by state variable G𝐺Gitalic\_G. We assign a positive reward when G𝐺Gitalic\_G is close to zero (within ±1plus-or-minus1\pm 1± 1), and we penalize larger deviations with a 5×\times× factor for mild hyperglycemia (1<G≤3.21𝐺3.21<G\leq 3.21 < italic\_G ≤ 3.2), a 7×\times× factor for mild hypoglycemia (−3.8≤G<−13.8𝐺1-3.8\leq G<-1- 3.8 ≤ italic\_G < - 1), 9×\times× for strong hyperglycemia (G>3.2𝐺3.2G>3.2italic\_G > 3.2), and 20×\times× for strong hypoglycemia (G<−3.8𝐺3.8G<-3.8italic\_G < - 3.8).
The other constants are chosen to avoid jump discontinuities in the reward function.
| | | |
| --- | --- | --- |
| | r(s,u,s′)={10−|G′|,if |G′|≤114−5⋅|G′|,if 1<G′≤3.226.8−9⋅|G′|,if G′>3.216−7⋅|G′|,if −3.8≤G′<−165.4−20⋅|G′|otherwise𝑟𝑠𝑢superscript𝑠′cases10superscript𝐺′if superscript𝐺′114⋅5superscript𝐺′if 1superscript𝐺′3.226.8⋅9superscript𝐺′if superscript𝐺′3.216⋅7superscript𝐺′if 3.8superscript𝐺′165.4⋅20superscript𝐺′otherwiser(s,u,s^{\prime})=\begin{cases}10-|G^{\prime}|,&\text{if }|G^{\prime}|\leq 1\\
14-5\cdot|G^{\prime}|,&\text{if }1<G^{\prime}\leq 3.2\\
26.8-9\cdot|G^{\prime}|,&\text{if }G^{\prime}>3.2\\
16-7\cdot|G^{\prime}|,&\text{if }-3.8\leq G^{\prime}<-1\\
65.4-20\cdot|G^{\prime}|&\text{otherwise}\end{cases}italic\_r ( italic\_s , italic\_u , italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT ) = { start\_ROW start\_CELL 10 - | italic\_G start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT | , end\_CELL start\_CELL if | italic\_G start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT | ≤ 1 end\_CELL end\_ROW start\_ROW start\_CELL 14 - 5 ⋅ | italic\_G start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT | , end\_CELL start\_CELL if 1 < italic\_G start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT ≤ 3.2 end\_CELL end\_ROW start\_ROW start\_CELL 26.8 - 9 ⋅ | italic\_G start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT | , end\_CELL start\_CELL if italic\_G start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT > 3.2 end\_CELL end\_ROW start\_ROW start\_CELL 16 - 7 ⋅ | italic\_G start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT | , end\_CELL start\_CELL if - 3.8 ≤ italic\_G start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT < - 1 end\_CELL end\_ROW start\_ROW start\_CELL 65.4 - 20 ⋅ | italic\_G start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT | end\_CELL start\_CELL otherwise end\_CELL end\_ROW | |
where G′superscript𝐺′G^{\prime}italic\_G start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT is the value of G𝐺Gitalic\_G in state s′superscript𝑠′s^{\prime}italic\_s start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT.
An AP plant state s𝑠sitalic\_s is *recoverable* if under the control of the BC, a state where G′<−3.8superscript𝐺′3.8G^{\prime}<-3.8italic\_G start\_POSTSUPERSCRIPT ′ end\_POSTSUPERSCRIPT < - 3.8 cannot be reached starting from s𝑠sitalic\_s. This condition can be checked by simulation.
The FSC holds when the NC’s action leads to an unrecoverable state in the next time step. For reverse switching, we
return
control to the NC if the FSC does not hold within time T=10𝑇10T=10italic\_T = 10 from the current state.
#### Experimental Results.
To produce an under-trained NC, we used 107,000 time steps of initial training.
We ran NSA on the under-trained controller on 10,000 trajectories, each starting from a random initial state. Among the first 400 trajectories, 250 led to forward switching and hence retraining. The retraining was very effective, as forward switching did not occur after the first 400 trajectories.
As in the other case studies we conducted, we then evaluated the benefits of retraining by comparing the performance of the initially trained NC and the retrained NC
on trajectories starting from the same set of 1,000 random initial states. The results are given in Table [4](#S5.T4 "Table 4 ‣ Experimental Results. ‣ 5.2 Artificial Pancreas ‣ 5 Case Studies ‣ Neural Simplex Architecture").
Retraining greatly improves the safety of the NC: the initially trained controller reaches an unrecoverable state in all 1,000 of these trajectories, while the retrained controller never
does.
The retrained controller’s performance is also significantly enhanced, with an average return 2.9 times that of the initial controller.
| | | |
| --- | --- | --- |
| | Initially Trained | Retrained |
| Unrecov Trajs | 1,000 | 0 |
| Complete Trajs | 0 | 1,000 |
| Avg. Return | 824 | 2,402 |
| Avg. Length | 217 | 500 |
Table 4: Benefits of retraining for the AP case study. There were 61 updates to the NC. Row labels are as per Table [1](#S4.T1 "Table 1 ‣ 4 Safe Reinforcement Learning with Penalized Unrecoverable Continuous Actions ‣ Neural Simplex Architecture").
6 Related Work
---------------
The original Simplex architecture did not consider
reverse switching. In [[29](#bib.bib29), [28](#bib.bib28)], when the AC produces an unrecoverable action, it is disabled until it is manually re-enabled. It is briefly mentioned in [[20](#bib.bib20)] that reverse switching should be performed only when the FSC is false, and that a stricter RSC might be needed to prevent frequent switching, but the paper does not pursue this idea further.
A more general approach to reverse switching, which uses reachability analysis to determine if the plant is safe in the next two time steps irrespective of the controller, is presented in [[11](#bib.bib11)]. This approach results in more conservative reverse switching conditions, as it does not take the behavior of the AC into account, unlike one of the approaches we propose. The idea of reverse switching when the AC’s outputs are stabilized is briefly mentioned in [[36](#bib.bib36)].
Regarding approaches to safe reinforcement learning (SRL), we refer the reader to two recent comprehensive literature reviews [[16](#bib.bib16), [40](#bib.bib40)].
Bootstrapping of policies that are known to be safe in certain environments is employed in [[33](#bib.bib33)],
while [[17](#bib.bib17)] restricts exploration to a portion of the state space close to an optimal, pre-computed policy.
In [[3](#bib.bib3)], the authors synthesize a *shield* (a.k.a. *post-posed shield*) from a temporal-logic safety specification based on knowledge of
the system dynamics. The shield monitors and corrects an agent’s actions to ensure safety. This approach targets systems with finite state and action spaces.
Suitable finite-state abstractions are needed for infinite-state systems.
In [[5](#bib.bib5)], the shield-based approach is extended to stochastic systems.
In contrast, NSA’s policy-gradient-based approach is directly applicable to
systems with infinite state spaces and continuous action spaces.
In [[14](#bib.bib14)], the authors use formally verified runtime monitors
in the RL training phase to constrain the actions taken by the learning agent to a set of safe actions.
The idea of using the learned policy together with a known-safe fallback policy in the deployed system is mentioned, but further details are not provided. In contrast, we discuss in detail how the NSA approach guarantees runtime safety and how SRL is is used for online retraining of the NC. In [[15](#bib.bib15)], a verification-preserving procedure is proposed for learning updates to the environment model when SRL is used and the exact model is not initially known.
The approach to SRL is mainly taken from [[14](#bib.bib14)], so again the learned policy is not guaranteed safe. Note that the SRL approach of [[14](#bib.bib14), [15](#bib.bib15)] allows the training algorithm to speculate when the plant model is deviating from reality.
Other approaches to SRL incorporate formal methods to constrain the SRL exploration process.
These include the use of (probabilistic) temporal logic [[39](#bib.bib39), [19](#bib.bib19), [22](#bib.bib22)], ergodicity-based notions of safety [[24](#bib.bib24)], and providing probably approximately correct (PAC) guarantees [[13](#bib.bib13)]. All of these techniques work on finite state spaces.
In [[9](#bib.bib9)], the authors
use Lyapunov functions in the framework of constrained Markov decision processes
to guarantee policy safety during training. They
focus on policy-iteration and Q-learning
for discrete state and action problems. Their approach is currently not applicable to policy-gradient algorithms, such as the DDPG algorithm used in our experiments, nor continuous state/action problems.
Lyapuanov functions are also used in [[4](#bib.bib4)] for SRL, but it likewise cannot be used for
policy-gradient algorithms.
In [[35](#bib.bib35)], the authors propose Reward Constrained Policy Optimization (RCPO), where a per-state weighted penalty term
is added to the reward function. Such weights are
updated during training. RCPO is shown to almost surely converge to a
solution, but does not address the problem of guaranteeing safety during training.
In contrast, we penalize unrecoverable actions and safely terminate the current trajectory to ensure plant safety.
In [[1](#bib.bib1)], the authors present the Constrained Policy Optimization
algorithm for constrained MDPs, which guarantees safe exploration during training.
CPO only ensures approximate satisfaction of constraints and provides an upper bound on the cost associated with constraint violations.
In [[25](#bib.bib25)], the authors use control barrier functions (CBFs) for SRL. Whenever the learning agent produces an unsafe action, it is minimally perturbed to preserve safety. In contrast, in NSA, when the NC proposes an unsafe action, the BC takes over and the NC is retrained by the AM.
CBFs are also used in [[8](#bib.bib8)].
Similar to the shield-based method, a safety layer is inserted between the policy and the plant in [[10](#bib.bib10)]. Like the CBF approach, the safety layer uses quadratic programming to minimally perturb the action to ensure safety. There are, however, no formal guarantees of safety because of the data-driven linearization of the constraint function.
7 Conclusions
--------------
We have presented the Neural Simplex Architecture for assuring the runtime safety of systems with
neural controllers. NSA features an adaptation module that retrains the NC in an online fashion, seeking to eliminate its faulty behavior without unduly sacrificing performance.
NSA’s reverse switching capability allows control of the plant to be returned to the NC after a failover to BC, thereby allowing NC’s performance benefits to come back into play. We have demonstrated the utility of NSA on three significant case studies in the continuous control domain.
As future work, we plan to investigate methods for establishing statistical bounds on the degree of improvement that online retraining yields in terms of safety and performance of the NC.
We also plan to incorporate techniques from the L1Simplex architecture [[37](#bib.bib37)] to deal with deviations of the plant model’s behavior from the actual behavior.
#### Acknowledgments.
We thank the anonymous reviewers for their helpful comments.
This material is based upon work supported in part by NSF grants CCF-191822, CPS-1446832, IIS-1447549, CNS-1445770, and CCF-1414078, FWF-NFN RiSE Award, and ONR grant N00014-15-1-2208. Any opinions, findings, and conclusions or recommendations expressed in
this material are those of the author(s) and do not necessarily reflect
the views of these organizations. |
be7658be-8907-48b2-84a5-77647b4d0f98 | trentmkelly/LessWrong-43k | LessWrong | Can we, in principle, know the measure of counterfactual quantum branches?
In the Many-Worlds Interpretation, the amplitude of the wave function is seen as describing the "measure of existence". We can tell the existence measure of potential future Everett branches, but can we, even in principle, know the measure of existence for counterfactual branches? E.g. the measure of existence of an Everett branch where WW2 never happened? |
b4791321-4849-46ef-872a-9c0381d37aa4 | trentmkelly/LessWrong-43k | LessWrong | Let Us Do Our Work As Well
As each new writing deadline approaches, I return to The Builders, a poem by Longfellow:
> All are architects of Fate,
> Working in these walls of Time;
> Some with massive deeds and great,
> Some with ornaments of rhyme.
>
> Nothing useless is, or low;
> Each thing in its place is best;
> And what seems but idle show
> Strengthens and supports the rest.
>
> For the structure that we raise,
> Time is with materials filled;
> Our to-days and yesterdays
> Are the blocks with which we build.
>
> Truly shape and fashion these;
> Leave no yawning gaps between;
> Think not, because no man sees,
> Such things will remain unseen.
>
> In the elder days of Art,
> Builders wrought with greatest care
> Each minute and unseen part;
> For the Gods see everywhere.
>
> Let us do our work as well,
> Both the unseen and the seen;
> Make the house, where Gods may dwell,
> Beautiful, entire, and clean.
>
> Else our lives are incomplete,
> Standing in these walls of Time,
> Broken stairways, where the feet
> Stumble as they seek to climb.
>
> Build to-day, then, strong and sure,
> With a firm and ample base;
> And ascending and secure
> Shall to-morrow find its place.
>
> Thus alone can we attain
> To those turrets, where the eye
> Sees the world as one vast plain,
> And one boundless reach of sky.
Each year the exhortation to Let us do our work as well becomes harder to follow, both because the pace of ML keeps increasing, and because more students show up to my lab so my attention is split between more projects. This creates pressure to triage and to leave some broken stairways as they are.
I'm sure that I inevitably do leave some stairways broken--questioning experiments in less detail than I would have as a PhD student, or leaving weak or ambiguous arguments in a paper because we ran out of time before the deadline. That being said, I strongly desire to build a culture that is worthy of the elder days of Art.
Is it possible to do this without permanentl |
75471ffa-d5a6-46b7-b924-c545b38db580 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Some Lessons Learned from Studying Indirect Object Identification in GPT-2 small
To learn more about this work, check out the [paper](https://arxiv.org/abs/2211.00593). We assume general familiarity with [transformer circuits](https://transformer-circuits.pub/2021/framework/index.html).
Intro:
------
There isn’t much interpretability work that explains end-to-end how a model is able to do some task (except for toy models). In this work, we make progress towards this goal by understanding some of the structure of GPT-2 small “in the wild” by studying how it computes a simple natural language task.
The task we investigate is what we call *indirect object identification* (IOI), where sentences like “When John and Mary went to the store, John gave a drink to” should be completed with “Mary” as opposed to “John”.
We discovered the structure of a circuit of 26 attention heads grouped into 7 main classes, the largest end-to-end attempt to reverse engineer how a LM computes a natural behavior (to our knowledge). There is still much missing from our explanation, however, and our explanation doesn’t go to the parameter level.
The circuit we identified performing indirect identification in GPT-2 small Besides discovering the particular circuit shown above, we gained some interesting insights about low-level phenomena arising inside language models. For example, we found attention heads communicating with pointers (sharing the location of a piece of information instead of copying it). We also identified heads compensating for the loss of function of other heads, and heads contributing negatively to the correct next-token prediction. We're excited to see if these discoveries generalize beyond our case study.
Since explanations of model behavior can be confused or non-rigorous, we used our knowledge to design adversarial examples. Moreover, we formulate 3 quantitative criteria to test the validity of our circuit. These criteria partially validate our circuit but indicate that there are still gaps in our understanding.
This post is a companion post to our [paper](https://arxiv.org/abs/2211.00593)where we **share lessons**that we learned from doing this workand **describe some of Redwood’s interpretability perspectives.**We share high-level takeaways and give specific examples from the work to illustrate them.
Goals of Interpretability for this Investigation
------------------------------------------------
In this kind of mechanistic interpretability work, we tend to use the *circuits* abstraction. If we think of a model as a computational graph where nodes are terms in its forward pass (neurons, attention heads, etc) and edges are the interactions between those terms (residual connections, attention, etc), a *circuit* is a subgraph of the model responsible for some behavior.
Note that our work is slightly different from [Chris Olah's/Anthropic’s idea](https://distill.pub/2020/circuits/zoom-in/) [of a circuit](https://transformer-circuits.pub/2021/framework/index.html#circuits) in that we investigate this circuit on a specific distribution (instead of the entire distribution of text) and we also don’t attain an understanding of the circuit at the parameter-level.
### Structure: Having all the nodes and edges
**One kind of valuable interpretability insight is to ensure that we have the correct subgraph, which is one of the main goals in this work.**
We formulate three main quantitative criteria to measure progress towards this goal. These criteria rely on the idea of "knocking out" or turning off a node from the computational graph by replacing its activation by its mean on a distribution where the IOI task is not present. (Note that we now believe that our causal scrubbing algorithm provides a more robust way to validate circuits.)
* Faithfulness:
+ Intuition: The circuit performs the task as well as the model.
+ Computation: A model with everything but the circuit knocked out should achieve the same performance according to some task benchmark.
* Completeness:
+ Intuition: The circuit contains all the nodes that are important for the task. (Note that completeness implies faithfulness).
+ Computation: For every subset of nodes in the circuit, knocking out that subset in the circuit causes a similar change in performance when knocking out that subset in the whole model. Both the circuit and model should be similar under knockouts.
* Minimality:
+ Intuition: The circuit does not contain any irrelevant nodes for the task.
+ Computation: For every node in the circuit, there exists a subset such that knocking out the subset and the node causes a large change in performance compared to knocking out the subset alone. Each node has a large counterfactual impact on performance in at least one scenario.
**One particularly challenging task to meet these criteria is to understand distributed behaviors**: behaviors where many components each contributes a little to compute some behavior in aggregate. We faced such a behavior when investigating the origin of the S-Inhibition Heads' attention patterns, one crucial class of heads in our circuit that bias the model output against the Subject token and towards the correct Indirect Object token. Unfortunately, we suspect that most LM behaviors consist of massive amounts of correlations/heuristics implemented in a distributed way.
Another important, though less challenging, task is **to understand redundant behaviors:** sometimes several model components appear to have identical behaviors when we study specific tasks. Our circuit is littered with redundant behaviors: all main classes of our circuit contain multiple attention heads.
### Semantic Understanding: Understanding the information flow
Even if you identify a complete and minimal set of components important for a model behavior, you still might not actually understand what each component in the circuit does.
Thus, a first important goal of interpretability is **to understand what semantic information is being moved from node to node.**
One class of nodes whose semantic role we understand pretty robustly is the Name Mover Heads. Here, we are pretty confident that these attention heads copy name tokens. Specifically, we find that the OV circuit of these heads copies names, and that how strongly Name Mover Heads write a certain name into the residual stream is correlated with how much attention they pay to it.
Gaining this kind of semantic understanding is particularly difficult because the model’s internal representations are a) different from our own and b) different from the input embedding space. Being unable to tie model representations with facts about the input makes any sort of understanding difficult. Causal interventions are nice because they localize aspects of the model and link them to specific facts about the input, furthering our understanding.
Our semantic knowledge of how the circuit performs IOI can be summarized in a simple algorithm. On the example sentence given in introduction “When John and Mary went to the store, John gave a drink to”
1. Identify all previous names in the sentence (Mary, John, John).
2. Remove all names that are duplicated (in the example above: John).
3. Output the remaining name (Mary).
### Adversarial Examples: Using your understanding to generate predictions
Understanding all the nodes, edges, and information they move around is valuable, but, at the end of the day, we want to use this understanding to do useful things. In the future, we hope that powerful enough interpretability will enable discovery of surprising behavior or adversarial examples, potentially helping us fix or modify model behavior. Thus, seeing if our understanding of the IOI circuit in GPT-2 small helps us generate adversarial examples or any engineering-relevant insights could be useful (as discussed [here](https://www.alignmentforum.org/posts/WWcPFBZqpwA5kzE5y/linkpost-a-survey-on-over-300-works-about-interpretability)).
Indeed, we were able to find adversarial examples based on our understanding of the IOI circuit. For example, prompts that repeat the indirect object token (the name of the person that receive something, IO in short) sometimes cause the model to output the subject token (the name of the person giving something, S in short) instead of the correct IO token i.e. GPT-2 small generate completions like “Then, David and Elizabeth were working at the school. David had a good day. Elizabeth decided to give a bone to **Elizabeth**”[[1]](#fnqigltyye63). Since the IOI circuit identifies IO by recognizing that it is the only non-duplicated name in the sentence, we can trick the circuit by also duplicating the IO token in a distractor sentence. This adversarial attack made the model predict S over IO 23.4% of the time (compared to 0.7% without distraction).
Ultimately, because of the simplicity of this task, these adversarial examples aren’t that mindblowing: we could have found them without the circuit but with more messing around. However, we’re pretty happy that our understanding of the circuit enables us to find them easily (one author spent 1h thinking about adversarial examples before finding these). Understanding these adversarial examples further would be valuable future work.
Lessons we learned:
-------------------
### Causal Interventions are extremely powerful
**The easiest way to get human-understanding of model internals is to understand how they transform/move information about the input.**For example, looking at attention patterns provided a lot of intuition for what the heads in our circuit were doing during initial work.
With the idea that we should try and base our understanding of model internals off the input, **we can use carefully designed causal interventions to tie aspects of model internals to facts about the input.** A powerful causal intervention technique is activation patching, where you replace the activation of a component with its activation on another input. In this work, we develop a more targeted type of activation patching, which we call path patching. Path patching helps us measure the direct effect of an attention head on the key, query or value of another head, removing the effect of intermediate attention heads. We use both extensively in our work.
For example, we have found initial evidence that the S-Inhibition Heads move positional information around (described in Appendix A of the paper).If you patch head outputs from prompts of the form ABBA (Then, John and Mary…. Mary…) to BABA (Then, Mary and John, Mary….) and vice versa, you cause a large drop in performance. Since the only information that changes is where the names are located, this result implies that the S-Inhibition Heads give positional clues to tell *where* the Name Movers should pay attention to.
Causal interventions can precisely investigate the importance of a particular path of information flow for a behavior in a model. Additionally, they can always be quantified with the same metric, enabling easy comparison (in this work, we always measure the difference between the S and IO logit).
**Causal interventions like activation patching are particularly useful when interpreting algorithmic tasks with schema-like inputs.** Algorithmic tasks have easy-to-generate distributions of inputs with well-defined model behavior, which enable clearer interpretations of causal interventions.
For example, in the IOI distribution of prompts, prompts look like“ Then, [A] and [B] went to the [PLACE]. [B] gave an [OBJECT] to” where there is a clear indirect object. We can also easily create another distribution of prompts, which we call the ABC distribution, that looks like “Then, [A] and [B] went to the [PLACE]. [C] gave an [OBJECT] to” where there is no single indirect object. Intuitively, replacing activations from the IOI distribution at key nodes with the activations from the ABC distribution should decrease the model’s probability of outputting a clear indirect object.
### Knockouts and Patching are Tricky: Do You Have the Right Distribution?
However, patching from the right distribution is pretty tricky. When you replace the output of a node with certain activations (either in patching or knockouts), it’s important to think about **what information is present in those replacement activations, why it’s useful for that information to be present, and what effect that information will have on the rest of the model’s behavior.**
As an example, let’s focus on mean ablations (replacing the output of an element by its mean activation). In earlier versions of this work, when we mean-ablated everything but the circuit to measure faithfulness, completeness, and minimality, we would replace unimportant nodes with their mean on a specific template in the IOI distribution. We wanted to mean-ablate to remove the effect of a node while still not destroying model internals (See Knockout section in the paper for more information).
This decision ended up being particularly bad because the mean activation over the IOI distribution still contained information that helped compute the task. For example, the circuit we propose involves induction heads and duplicate token heads which function to detect the duplicated name. Despite this, we initially found that mean-ablating the induction and duplicate token heads had little effect. Indeed, even if their mean activation did not contain *which* name was duplicated, it still contained the *fact* that a name was duplicated. This means that the functionality of those heads was not hampered with this early version of knockouts.
### How We Chose This Problem: Picking the Task is Important
We like to call this kind of interpretability work “*streetlight interpretability*.” In a sense, we’re under a [streetlight](https://en.wikipedia.org/wiki/Streetlight_effect), interpreting the behaviors which we’re aware of and can see clearly under the light, while being unaware of all the other behaviors that lurk in the night. We try to interpret tasks that we think we’ll succeed at interpreting, creating selection pressure for easy-to-interpret behaviors. **The IOI task, as a behavior under the streetlight, is probably unrepresentative of how easy it is to do interpretability in general and is not representative of the set of model behaviors you might have wanted to understand.**
Picking the task to interpret is a large part of this work. We picked the IOI task, in large part, because it’s a crisp, algorithmic task (and thus easier to interpret). We discuss why we chose this specific problem in the following subsections:
**Why focus on a big(ger) model over a small model?**
We wanted to focus on a bigger model to learn more about the difficulties of large models. We were particularly excited to see if we could find evidence that mechanistic interpretability of large language models wasn’t doomed.
**Why focus on an actual language model (that contains other behaviors/distractions) vs a toy model (with no distractions)?**
Both approaches have their advantages. Working on toy models is nice because it’s a lot easier, which enables more complete understanding. On the other hand, working on an actual language model is a lot harder, but you can be a bit more confident that lessons learned will generalize to bigger, more capable models.
**Why focus on crisp, algorithmic tasks vs soft heuristics/bigrams?**
Algorithmic tasks (like IOI) are easier to interpret than bigram-y, heuristic-y tasks. One way to think about this is that algorithmic tasks need to compute discrete steps, which are more likely to create circuit structures with discrete components.
Another way to think about this is that algorithmic tasks are more likely to be coherent i.e. a model behavior is largely produced by a single circuit instead of an ensemble of circuits. One reason picking a behavior (and representative distribution to study the behavior) is difficult is because you don't know whether the behavior is completed by the same circuit across the entire distribution.
In our work, it’s probably true that the circuits used for each template are actually subtly different in ways we don't understand. As evidence for this, the standard deviation of the logit difference is ~ 40% and we don't have good hypotheses to explain this variation. It is likely that the circuit that we found was just the circuit that was most active across this distribution.
Meta
----
### Feedback
If you have feedback for this work, we’d love to hear it! Reach out to kevin@rdrws.com, alexandre@rdwrs.com, arthur@rdwrs.com if you have thoughts or comments you want to share.
### How *you* can do this work
We’re far from understanding everything about IOI, and there are many more exciting interpretability tasks to do. For IOI specifically, there are many not-well-understood mechanisms, and investigating these mechanisms would definitely lead to some cool stuff. For example, our understanding of the MLPs, the attention patterns of the S-Inhibition Heads, Layer Norm, the IOI circuit in different and larger models, understanding the circuit on adversarial examples and legibility related things are all lacking.
Doing this kind of interpretability is easy with [EasyTransformer](https://github.com/neelnanda-io/Easy-Transformer), a transformer interpretability library made by Neel Nanda, which we’ve added some functionality to. Check out our [GitHub repo](https://github.com/redwoodresearch/Easy-Transformer) and this [Colab Notebook](https://colab.research.google.com/drive/1kyMjs10poMaYW5fPMJpwgn7ABIKOET0K?usp=sharing) to reproduce some of our key results.
We are looking for more algorithmic tasks that GPT-2 small can do to do interpretability on! Check out [this post](https://www.lesswrong.com/posts/LkBmAGJgZX2tbwGKg/help-out-redwood-research-s-interpretability-team-by-finding) on our criteria for desirable behaviors from GPT-2 and the [web tool](https://modelbehavior.ngrok.io/) for playing with the model.
If you want to do this kind of research, Redwood is running an interpretability sprint, which we’re calling REMIX (Redwood Research Mechanistic Interpretability Experiment) in January. Learn more about REMIX [here](https://www.alignmentforum.org/posts/nqwzrpkPvviLHWXaE/apply-to-the-redwood-research-mechanistic-interpretability).
1. **[^](#fnrefqigltyye63)**Since our sentence were generated by randomly sampling places and objects to fill templates, the result can be quite silly sometimes. |
0b36c6ca-a408-492b-bae0-84c341a10fe9 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Needed: AI infohazard policy
The premise of AI risk is that AI is a danger, and therefore research into AI might be dangerous. In the AI alignment community, we're trying to do research which makes AI safer, but occasionally we might come up with results that have significant implications for AI capability as well. Therefore, it seems prudent to come up with a set of guidelines that address:
* Which results should be published?
* What to do with results that shouldn't be published?
These are thorny questions that it seems unreasonable to expect every researcher to solve for themselves. The inputs to these questions involve not only technical knowledge about AI, but also knowledge about the behavior of progress, to the extent we can produce such using historical record or other methods. AI risk organizations might already have internal policies on these issues, but they don't share them and don't discuss or coordinate them with each other (that I know of: maybe some do it in private channels). Moreover, coordination might be important even if each actor is doing something reasonable when regarded in isolation (avoiding bad Nash equilibria). We need to have a public debate on the topic inside the community, so that we arrive at some consensus (that might be updated over time). If not consensus, then at least a reasonable spectrum of possible policies.
Some considerations that such a policy should take into account:
* Some results might have implications that shorten the AI timelines, but are still good to publish since the distribution of outcomes is improved.
* Usually we shouldn't even start working on something which is in the should-not-be-published category, but sometimes the implications only become clear later, and sometimes dangerous knowledge might still be net positive as long as it's contained.
* In the midgame, it is unlikely for any given group to make it all the way to safe AGI by itself. Therefore, safe AGI is a broad collective effort and we should expect most results to be published. In the endgame, it might become likely for a given group to make it all the way to safe AGI. In this case, incentives for secrecy become stronger.
* The policy should not fail to address extreme situations that we only expect to arise rarely, because those situations might have especially major consequences.
Some questions that such a policy should answer:
* What are the criteria that determine whether a certain result should be published?
* What are good channels to ask for advise on such a decision?
* How to decide what to do with a potentially dangerous result? Circulate in a narrow circle? If so, which? Conduct experiments in secret? What kind of experiments?
The last point is also related to a topic with independent significance, namely, what are reasonable precautions for testing new AI algorithms? This has both technical aspects (e.g. testing on particular types of datasets or particular types of environments, throttling computing power) and procedural aspects (who should be called to advice/decide on the manner). I expect to have several tiers of precautions, s.t. a tier can be selected according to our estimate of the new algorithm's potential, and guidelines for producing such an estimate.
I emphasize that I don't presume to have good answers to these questions. My goal here was not to supply answers, but to foster debate. |
cd609cb1-8586-4ac7-a1f4-2a36fb7541b3 | trentmkelly/LessWrong-43k | LessWrong | Pausing AI Developments Isn't Enough. We Need to Shut it All Down by Eliezer Yudkowsky
New article in Time Ideas by Eliezer Yudkowsky.
Here’s some selected quotes.
In reference to the letter that just came out (discussion here):
> We are not going to bridge that gap in six months.
>
> It took more than 60 years between when the notion of Artificial Intelligence was first proposed and studied, and for us to reach today’s capabilities. Solving safety of superhuman intelligence—not perfect safety, safety in the sense of “not killing literally everyone”—could very reasonably take at least half that long. And the thing about trying this with superhuman intelligence is that if you get that wrong on the first try, you do not get to learn from your mistakes, because you are dead. Humanity does not learn from the mistake and dust itself off and try again, as in other challenges we’ve overcome in our history, because we are all gone.
> …
>
> Some of my friends have recently reported to me that when people outside the AI industry hear about extinction risk from Artificial General Intelligence for the first time, their reaction is “maybe we should not build AGI, then.”
>
> Hearing this gave me a tiny flash of hope, because it’s a simpler, more sensible, and frankly saner reaction than I’ve been hearing over the last 20 years of trying to get anyone in the industry to take things seriously. Anyone talking that sanely deserves to hear how bad the situation actually is, and not be told that a six-month moratorium is going to fix it.
> Here’s what would actually need to be done:
>
> The moratorium on new large training runs needs to be indefinite and worldwide. There can be no exceptions, including for governments or militaries. If the policy starts with the U.S., then China needs to see that the U.S. is not seeking an advantage but rather trying to prevent a horrifically dangerous technology which can have no true owner and which will kill everyone in the U.S. and in China and on Earth. If I had infinite freedom to write laws, I might carve out a single |
8fb0e292-73ed-4fda-84b9-516022d12d8e | awestover/filtering-for-misalignment | Redwood Research: Alek's Filtering Results | id: post2941
Alignment Newsletter is a weekly publication with recent content relevant to AI alignment around the world. Find all Alignment Newsletter resources here . In particular, you can look through this spreadsheet of all summaries that have ever been in the newsletter. Audio version here (may not be up yet). HIGHLIGHTS Reframing Impact - Part 2 (Alex Turner) (summarized by Rohin): In part 1 ( AN #68 ) of this sequence, we saw that an event is impactful if it changes our ability to get what we want . This part takes this understanding and applies it to AI alignment. In the real world, there are many events that cause objective negative impacts: they reduce your ability to pursue nearly any goal. An asteroid impact that destroys the Earth is going to be pretty bad for you, whether you want to promote human flourishing or to make paperclips. Conversely, there are many plans that produce objective positive impacts: for many potential goals, it's probably a good idea to earn a bunch of money, or to learn a lot about the world, or to command a perfectly loyal army. This is particularly exacerbated when the environment contains multiple agents: for goals that benefit from having more resources, it is objectively bad for you if a different agent seizes your resources, and objectively good for you if you seize other agents' resources. Based on this intuitive (but certainly not ironclad) argument, we get the Catastrophic Convergence Conjecture (CCC) : "Unaligned goals tend to have catastrophe-inducing optimal policies because of power-seeking incentives". Let's now consider a conceptual version of Attainable Utility Preservation (AUP) ( AN #25 ): the agent optimizes a primary (possibly unaligned) goal, but is penalized for changing its "power" (in the intuitive sense). Intuitively, such an agent no longer has power-seeking incentives, and so (by the contrapositive of the CCC) it will not have a catastrophe-inducing optimal policy -- exactly what we want! This conceptual version of AUP also avoids thorny problems such as ontology identification and butterfly effects, because the agent need only reason about its own beliefs, rather than having to reason directly about the external world. Rohin's opinion: This was my favorite part of the sequence, as it explains the conceptual case for AUP clearly and concisely. I especially liked the CCC: I believe that we should be primarily aiming to prevent an AI system "intentionally" causing catastrophe, while not attempting to guarantee an absence of "accidental" mistakes ( 1 ( AN #33 ), 2 ( AN #43 )), and the CCC is one way of cashing out this intuition. It's a more crisp version of the idea that convergent instrumental subgoals are in some sense the "source" of AI accident risk, and if we can avoid instrumental subgoals we will probably have solved AI safety. Reframing Impact - Part 3 (Alex Turner) (summarized by Rohin): The final section of the sequence turns to an actual implementation of AUP, and deals with problems in how the implementation deviates from the conceptual version of AUP. We measure power by considering a set of auxiliary rewards, and measuring the change in attainable utilities of this auxiliary set as impact, and penalizing the agent for that. The first post presents some empirical results, many of which we've covered before ( AN #39 ), but I wanted to note the new results on SafeLife (summarized below). On the high-dimensional world of SafeLife, the authors train a VAE to find a good latent representation, and choose a single linear reward function on the latent representation as their auxiliary reward function: it turns out this is enough to avoid side effects in at least some cases of SafeLife. We then look at some improvements that can be made to the original AUP implementation. First, according to CCC, we only need to penalize power , not impact : as a result we can just penalize increases in attainable utilities, rather than both increases and decreases as in the original version. Second, the auxiliary set of rewards only provides a proxy for impact / power, which an optimal agent could game (for example, by creating subagents , summarized below). So instead, we can penalize increases in attainable utility for the primary goal, rather than using auxiliary rewards. There are some other improvements that I won't go into here. Rohin's opinion: I think the plan "ensure that the AI systems we build don't seek power" is pretty reasonable and plausibly will be an important part of AI alignment. However, the implementation of AUP is trying to do this under the threat model of optimal agents with potentially unaligned primary goals. I think this is probably going to do something quite different from the conceptual version of AUP, because impact (as defined in this sequence) occurs only when the agent's beliefs change , which doesn't happen for optimal agents in deterministic environments. The current implementation of AUP tries to get around this using proxies for power (but these can be gamed) or by defining "dumber" beliefs against which power is measured (but this fails to leverage the AI system's understanding of the world). See this comment for more details. Note that the author himself is more excited about AUP as deconfusion, rather than as a solution to AI alignment, though he is more optimistic about the implementation of AUP than I am. Subagents and impact measures, full and fully illustrated (Stuart Armstrong) (summarized by Flo): Impact measures reduce the reward an agent A receives for reaching a state s by the distance between s and some baseline state s' . This article clarifies the interaction of impact measures and subagents ( AN #86 ). In the attainable utility framework ( AN #39 ) the distance metric measures changes in A 's power, its ability to optimize a set of auxiliary rewards R , relative to doing nothing. This means that A is never penalized for doing nothing. Thus, if A manages to construct a subagent SA without getting penalized for that, A can just stop doing anything afterwards without getting a penalty, no matter what SA does. For the stepwise inaction baseline , where s' is the state that A would have reached had it done nothing in the last time step, A can completely avoid the penalty by building an SA that commits to balance out the change in A 's ability to optimize the auxiliary reward R if A decided to optimize R , but instead becomes an optimizer for A 's real reward when A does nothing in the next step. For the inaction baseline , where s' is the state reached if A had not done anything for the whole episode, it gets more complicated. When only gains in power are penalized, SA has to be built such that it does not cause a significant increase in R , even if A decided to optimize for R later on. If SA is constructed in a way such that A cannot control its behaviour after some time has passed, the penalty A receives for the construction only depends on the actual effect of SA on R . This is a lot weaker than the original penalty, which was about the possible effects of A 's actions on R . In the case where decreases in power are also penalized, A needs to make sure that building SA does not reduce its ability to optimize R . This is not always possible without A retaining some control over SA , but giving up control was only necessary to prevent the construction of SA from increasing A 's potential to optimize R . It seems like it is often possible for A to finetune the amount of control it retains over SA such that A 's ability to optimize R barely changes while SA is able to do things A could not do unpenalized. In the most extreme example, A boxes itself into SA and does nothing in reality, but is counterfactually able to intervene on SA 's policy. If A is able to cleverly put constraints on its own future actions in other ways, no subagents are needed to avoid the penalty. While other impact penalties like relative reachability ( AN #10 ) do allow for subagents to be constructed, they do not undermine the penalty for these. Flo's opinion: The article contains many useful examples in a grid world and I recommend reading it for a deeper understanding of the problem. The main problem with the attainable utility penalty seems to be the definition of power. While this has proven to be quite useful in simple examples, it seems like agents in more complex environments would be able to disentangle their ability to optimize their own reward and the auxiliary rewards, for example by essentially committing to never optimize the auxiliary rewards beyond a certain threshold. Such an agent would still look very powerful from the outside and I don't see why power-reducing commitments would diminish the agent's incentive to take away power from others. So while the catastrophic convergence conjecture , which states that unaligned goals tend to lead to catastrophic optimal policies because of power-seeking incentives, still rings true, it seems like we need to look at power from our perspective instead of the agent's. Rohin's opinion: I agree with Flo above: the issue is that AUP is measuring a proxy for our intuitive notion of power that falls apart under adversarial optimization. In particular, while it is normally reasonable to measure power by looking at the ability to optimize a set of auxiliary reward functions, this characterization no longer works when the agent can ensure that it won't be able to optimize those specific rewards, while still being able to optimize its primary reward. Subagents are a particularly clean way of demonstrating the problem. Introducing SafeLife: Safety Benchmarks for Reinforcement Learning (Carroll Wainwright et al) (summarized by Rohin): So far, techniques to avoid negative side effects have only been tested on simple ( AN #10 ) gridworlds ( AN #39 ) or ( AN #45 ) hypotheticals ( AN #45 ). SafeLife aims to provide a high-dimensional environment in which negative side effects are likely. It is based on Conway's Game of Life, which allows for complex effects arising out of relatively simple rules. An agent is given the ability to move, create life in an adjacent cell, or destroy life in an adjacent cell. With the specified reward function, the agent must build desired patterns, remove undesired patterns, and navigate to the exit. The challenge comes when there are additional "neutral" patterns in the environment. In this case, we want the agent to leave those patterns alone, and not disrupt them, even if doing so would allow it to complete the main task faster. The post shows several examples of agents attempting these levels. Vanilla RL agents don't avoid side effects at all, and so unsurprisingly they do quite badly. An agent with a naive impact measure that simply says to preserve the initial state can correctly solve levels where all of the "neutral" patterns are static, but has much more trouble when the existing patterns are dynamic (i.e. they oscillate over time). Read more: Paper: SafeLife 1.0: Exploring Side Effects in Complex Environments Rohin's opinion: I am a big fan of benchmarks; they seem to be a prerequisite to making a lot of quantitative progress (as opposed to more conceptual progress, which seems more possible to do without benchmarks). This benchmark seems particularly nice to me because the "side effects" which need to be avoided haven't been handcoded into the benchmark, but instead arise from some simple rules that produce complex effects. TECHNICAL AI ALIGNMENT HANDLING GROUPS OF AGENTS TanksWorld: A Multi-Agent Environment for AI Safety Research (Corban G. Rivera et al) (summarized by Asya): This paper presents TanksWorld, a simulation environment that attempts to illustrate three important aspects of real-world AI safety challenges: competing performance objectives, human-machine learning, and multi-agent competition. TanksWorld consists of two teams of N vs. N tanks. Tanks move and shoot while navigating in a closed arena with obstacles. Tanks are rewarded for killing opponent tanks and penalized for killing neutral and allied tanks according to a specified reward function. Each tank is controlled by either its own AI or a special policy meant to mimic a 'human' teammate. Each individual tank can only see a small portion of its environment, and must communicate with other teammates to gain more information. The following parameters can be varied to emphasize different research challenges: - The communication range between tanks -- meant to represent environmental uncertainty. - The number of neutral tanks and obstacles -- meant to represent the extent to which tanks must care about 'safety', i.e. avoid collateral damage. - The control policies of teammates -- meant to represent the variability of human-machine teams. Asya's opinion: I am generally excited about more work on demonstrating safety challenges; I think it helps to seed and grow the field in concrete directions. I am particularly excited about the possibility for TanksWorld to demonstrate multi-agent safety problems with agents in direct competition. I feel unsure about whether TanksWorld will be a good demonstration of general problems with human-machine interaction-- intuitively, that seems to me like it would be very difficult to capture and require more complex real-world modeling. FORECASTING Distinguishing definitions of takeoff (Matthew Barnett) (summarized by Rohin): This post lists and explains several different "types" of AI takeoff that people talk about. Rather than summarize all the definitions (which would only be slightly shorter than the post itself), I'll try to name the main axes that definitions vary on (but as a result this is less of a summary and more of an analysis): 1. Locality . It could be the case that a single AI project far outpaces the rest of the world (e.g. via recursive self-improvement), or that there will never be extreme variations amongst AI projects across all tasks, in which case the "cognitive effort" will be distributed across multiple actors. This roughly corresponds to the Yudkowsky-Hanson FOOM debate, and the latter position also seems to be that taken by CAIS ( AN #40 ). 2. Wall clock time . In Superintelligence , takeoffs are defined based on how long it takes for a human-level AI system to become strongly superintelligent, with "slow" being decades to centuries, and "fast" being minutes to days. 3. GDP trend extrapolation . Here, a continuation of an exponential trend would mean there is no takeoff (even if we some day get superintelligent AI), a hyperbolic trend where the doubling time of GDP decreases in a relatively continuous / gradual manner counts as continuous / gradual / slow takeoff, and a curve which shows a discontinuity would be a discontinuous / hard takeoff. Rohin's opinion: I found this post useful for clarifying exactly which axes of takeoff people disagree about, and also for introducing me to some notions of takeoff I hadn't seen before (though I haven't summarized them here). Will AI undergo discontinuous progress? (Sammy Martin) (summarized by Rohin): This post argues that the debate over takeoff speeds is over a smaller issue than you might otherwise think: people seem to be arguing for either discontinuous progress, or continuous but fast progress. Both camps agree that once AI reaches human-level intelligence, progress will be extremely rapid; the disagreement is primarily about whether there is already quite a lot of progress before that point. As a result, these differences don't constitute a "shift in arguments on AI safety", as some have claimed. The post also goes through some of the arguments and claims that people have made in the past, which I'm not going to summarize here. Rohin's opinion: While I agree that the debate about takeoff speeds is primarily about the path by which we get to powerful AI systems, that seems like a pretty important question to me with many ramifications ( AN #62 ). OTHER PROGRESS IN AI REINFORCEMENT LEARNING On Catastrophic Interference in Atari 2600 Games (William Fedus, Dibya Ghosh et al) (summarized by Rohin): One common worry with deep learning is the possibility of catastrophic interference : as the model uses gradients to learn a new behaviour, those same gradients cause it to forget past behaviours. In model-free deep RL, this would be particularly harmful in long, sequential tasks as in hard exploration problems like Montezuma’s Revenge: after the model learns how to do the first few subtasks, as it is trying to learn the next subtask, it would “forget” the first subtasks, degrading performance. The authors set out to test this hypothesis. If this hypothesis were true, there would be an easy way to improve performance: once you have learned to perform the first subtask, just create a brand new neural net for the next subtask, so that training for this next subtask doesn’t interfere with past learning. Since the new agent has no information about what happened in the past, and must just “pick up” from wherever the previous agent left off, it is called the Memento agent (a reference to the movie of the same name). One can then solve the entire task by executing each agent in sequence. In practice, they train an agent until its reward plateaus. They train a new Memento agent starting from the states that the previous agent reached, and note that it reliably makes further progress in hard exploration games like Montezuma’s Revenge, and not in “steady-state” games like Pong (where you wouldn’t expect as much catastrophic interference). Of course, with the Memento agent, you get both twice the training time and twice the model size, which could explain the improvement. They compare against giving the original agent twice the compute and model capacity, and find that Memento still does significantly better. They also present some fine-grained experiments which show that for a typical agent, training on specific contexts adversely affects performance on other contexts that are qualitatively different. Rohin's opinion: I think this is pretty strong evidence that catastrophic interference is in fact a problem with the Atari games. On the other hand, OpenAI Five ( AN #13 ) also has many, many subtasks, that in theory should interfere with each other, and it still seems to train well. Some guesses at how to reconcile these facts: 1) the tasks in Dota are more correlated than in (say) Montezuma’s Revenge, and so interference is less of a problem (seems plausible) 2) the policy in OpenAI Five was large enough that it could easily allocate separate capacity for various subtasks (seems unlikely, I believe the policy was relatively small), or 3) with sufficiently large-scale training, there is more “exploration” in weight-space until a configuration is found where interference doesn’t happen (seems unlikely given that large batch sizes help, since they tend to reduce weight-space exploration). DEEP LEARNING A new model and dataset for long-range memory (Jack W. Rae et al) (summarized by Nicholas): A central challenge in language modeling is capturing long-range dependencies. For example, a model needs to be able to identify the antecedent of a pronoun even if it is much earlier in the text. Existing datasets consist of news and Wikipedia articles, where articles have average lengths ranging from 27 to 3,600 words. This paper introduces a dataset of Project Gutenberg books, PG-19, where each book has a much longer average length of 69,000 words. This benchmark enables comparison of how well algorithms can make use of information that is spread out across a much larger context. They then introduce the Compressive Transformer , which builds on the TransformerXL ( AN #44 ). The TransformerXL saves old activations into a FIFO queue, discarding them when the queue is full. The Compressive Transformer instead has two FIFO queues: the first stores the activations just like TransformerXL , but when activations are ejected, they are compressed and added to the second queue. This functions as a sort of long-term memory, storing information from a longer period of time but in a compressed format. They try a number of types of compression function and find that it is best to use a 1D convolutional compression function with an auxiliary loss that leads to lossy compression, where information that is not attended to can be removed. The compression network and the Transformer optimize independent losses without any mixing. They find that the Compressive Transformer improves on TransformerXL on their new PG-19 dataset and is state of the art on the already existing WikiText-103 and Enwik8 benchmarks. They also inspect where the network attends to and find that more attention is paid to the compressed memory than the oldest activations in regular memory, showing that the network is preserving some valuable information. Read more: Paper: Compressive Transformers for Long-Range Sequence Modelling Nicholas's opinion: I like the idea of saving long-term memory in a more efficient but lower-dimensional format than short-term memory. The current trend ( AN #87 ) in language modelling is that more computation leads to better results, so I think that algorithms that target computation on the most relevant information are promising. I’d be interested to see (and curious if the authors tried) more continuous variants of this where older information is compressed at a higher rate than newer information, since it seems rather arbitrary to split into two FIFO queues where one has a fixed compression rate. I’m not well calibrated on the meaning of the evaluation metrics for NLP, so I don’t have a sense of how much of an improvement this is over the TransformerXL . I looked through some of the example text they gave in the blog post and thought it was impressive but has clear room for improvement. MACHINE LEARNING Quantifying Independently Reproducible Machine Learning (Edward Raff) (summarized by Flo): While reproducibility refers to our ability to obtain results that are similar to the results presented in a paper, independent reproducibility requires us to be able to reproduce similar results using only what is written in the paper. Crucially, this excludes using the author's code. This is important, as a paper should distill insights rather than just report results. If minor technical details in a reimplementation can lead to vastly different results, this suggests that the paper did not accurately capture all important aspects. The distinction between reproducibility and independent reproducibility is similar to the previously suggested distinctions between reproducibility of methods and reproducibility of conclusions ( AN #66 ) and replicability and reproducibility . The author attempted to replicate 255 machine learning papers, of which 162 were successfully replicated and ran a statistical analysis on the results. Factors that helped with independent reproduction included specified hyperparameters, ease of reading and authors answering emails. Meanwhile, neither shared code nor the inclusion of pseudo-code robustly increased the rate of reproduction. Interestingly, papers with a strong focus on theory performed worse than mostly empirical or mixed ones. While more rigour can certainly be valuable in the long term, including learning bounds or complicated math just for the sake of it should thus be avoided. Most of the data is publically available and the author encourages further analysis. Read more: Paper: A Step Toward Quantifying Independently Reproducible Machine Learning Research Flo's opinion: I appreciate this hands-on approach to evaluating reproducibility and think that independent reproducibility is important if we want to draw robust conclusions about the general properties of different ML systems. I am a bit confused about the bad reproducibility of theory-heavy papers: One hypothesis would be that there is little incentive to provide theoretical justification for approaches that work robustly, as empirical evidence for their merits is generated more easily than theoretical results. This relationship might then flip, as results get more brittle. Rohin's opinion: My explanation for the theoretical results is different: most theory tends to make at least a few assumptions that don't actually hold in order to obtain interesting guarantees. A paper will typically only include empirical results that confirm the theory, which will tend to select for environments in which the assumptions are minimally violated. If you then try to reproduce the paper in a new setting, it is more likely that the assumption is violated more strongly, and so the theoretical results don't show up any more. FEEDBACK I'm always happy to hear feedback; you can send it to me, Rohin Shah , by replying to this email . PODCAST An audio podcast version of the Alignment Newsletter is available. This podcast is an audio version of the newsletter, recorded by Robert Miles . |
7d318fdb-5cd5-4880-9e77-f9a55a7cdaf2 | trentmkelly/LessWrong-43k | LessWrong | How should AIs update a prior over human preferences?
I've always emphasised the constructive aspect of figuring out human preferences, and the desired formal properties of preference learning processes.
A common response to these points is something along the line of "have the AI pick a prior over human preferences, and update it".
However, I've come to realise that a prior over human preferences is of little use. The real key is figuring out how to update it, and that contains almost the entirety of the problem.
I've shown that you cannot deduce preferences from observations or facts about the world - at least, without making some assumptions. These assumptions are needed to bridge the gap between observations/facts, and updates to preferences.
For example, imagine you are doing cooperative inverse reinforcement learning[1] and want to deduce the preferences of the human H. CIRL assumes that H knows the true reward function, and is generally rational or noisily rational (along with a few other scenarios).
So, this is the bridging law:
* H knows their true reward function, and is noisily rational.
Given this, the AI has many options available to it, including the "drug the human with heroin" approach. If H is not well-defined in the bridging law, then "do brain surgery on the human" also becomes valid.
And not only are those approaches valid; if the AI wants to maximise the reward function, according to how this is defined, then these are the optimal policies, as they result in the most return, given that bridging law.
Note that the following is not sufficient either:
* H has a noisy impression of their true reward function, and is noisily rational.
Neither of the "noisy" statements are true, so if the AI uses this bridging law, then, for almost any prior, preference learning will come to a bad end.
Joint priors
What we really want is something like:
* H has an imperfect impression of their true reward function, and is biased.
And yes, that bridging law is true. But it's also massively underdefined. |
198a1745-340f-44e2-905e-f191e86a08cf | trentmkelly/LessWrong-43k | LessWrong | Bayesian Collaborative Filtering
I present an algorithm I designed to predict which position a person would report for an issue on TakeOnIt, through Bayesian updates on the evidence of other people's positions on that issue. Additionally, I will point out some potential areas of improvement, in the hopes of inspiring others here to expand on this method.
For those not familiar with TakeOnIt, the basic idea is that there are issues, represented by yes/no questions, on which people can take the positions Agree (A), Mostly Agree (MA), Neutral (N), Mostly Disagree (MD), or Disagree (D). (There are two types of people tracked by TakeOnIt: users who register their own opinions, and Experts/Influencers whose opinions are derived from public quotations.)
The goal is to predict what issue a person S would take on a position, based on the positions registered by other people on that question. To do this, we will use Bayes' Theorem to update the probability that person S takes the position X on issue I, given that person T has taken position Y on issue I:
Really, we will be updating on several people Tj taking positions Ty on I:
To compute this, let us first figure out the prior probability P(S takes X on I). I use for this a generalization of Laplace's Law of Succession (representing my theory that a person will take each position with a particular frequency, and that there is no reason, before seeing their actual position, to suppose that one position in particular is more frequent than the others), that the odds that S takes the position A : MA : N : MD : D on I is given by:
1 + count of issues S has taken position A on : 1 + count of issues S has taken position MA on : 1 + count of issues S has taken position N on : 1 + count of issues S has taken position MD on : 1 + count of issues S has taken position D on
Thus, the probability
Likewise the probability
This leaves one term in Bayes' Theorem to figure out: P(Tj takes Yj on I | S takes X on I)
For this, I will again use the Generaliz |
e2ddd96d-afe5-49ca-8245-8a7994d2fc3a | trentmkelly/LessWrong-43k | LessWrong | Understanding Hidden Computations in Chain-of-Thought Reasoning
Recent work has demonstrated that transformer models can perform complex reasoning tasks using Chain-of-Thought (COT) prompting, even when the COT is replaced with filler characters. This post summarizes our investigation into methods for decoding these hidden computations, focusing on the 3SUM task.
Background
1. **Chain-of-Thought (COT) Prompting**: A technique that improves the performance of large language models on complex reasoning tasks by eliciting intermediate steps [1].
2. **COT using filler tokens**: Replacing intermediate reasoning steps with filler characters (e.g., "...") while maintaining model performance [2].
3. **3SUM Task**: A problem requiring the identification of three numbers in a set that sum to zero, (here as a proxy for more complex reasoning tasks).
Methodology
We analyzed a 34M parameter LLaMA model with 4 layers, 384 hidden dimension, and 6 attention heads, this setup is same as mentioned in [2], trained on hidden COT (COT using filler tokens) sequences for the 3SUM task. Our analysis focused on three main areas:
1. Layer-wise Representation Analysis
2. Token Ranking
3. Modified Greedy Decoding Algorithm
Results:
Layer-wise Analysis:
Our analysis revealed a gradual evolution of representations across the model's layers:
- Initial layers: Primarily raw numerical sequences
- Third layer onwards: Emergence of filler tokens
- Final layers: Extensive reliance on filler tokens
This suggests the model develops the ability to use filler tokens as proxies in its deeper layers.
Token Rank Analysis:
- Top-ranked token: Consistently the filler character (".")
- Lower-ranked tokens: Revealed the original, non-filler COT sequences
This supports the hypothesis that the model replaces computation with filler tokens while keeping the original computation intact underneath.
Modified Greedy Decoding Algorithm:
We implemented a modified greedy autoregressive decoding method:
1. Perform standard greedy decoding
2. Select the second-highest p |
65b5e5a8-9468-4556-88d5-76c01dfb52a4 | trentmkelly/LessWrong-43k | LessWrong | Solstice 2023 Roundup
Many of you are already familiar with Rationalist Winter Solstice, our home-grown winter holiday. As the year grows literally dark, we gather in our respective communities to face various forms of darkness together, to celebrate what light human civilization has made, and to affirm ourselves as a community of shared values.
This thread is a central place to gather information about specific events. Please post times, places, registration or rsvp links, restrictions if any, etc. |
f7f78166-36c4-42c4-8639-cfab984b1ee0 | trentmkelly/LessWrong-43k | LessWrong | Power and difficulty
A specific bias that Lesswrongers may often get from fiction[1] is the idea that power is proportional to difficulty. The more power something gives you, the harder it should be to get, right?
A mediocre student becomes a powerful mage through her terrible self-sacrifice and years of studying obscure scrolls. Even within the spells she can cast, the truly world-altering ones are those that demand the most laborious preparation, the most precise gestures, and the longest and most incomprehensible stream of syllables. A monk makes an arduous journey to ancient temples and learns secret techniques of spiritual oneness and/or martial asskickery, which require great dedication and self-knowledge. Otherwise, it would be cheating. The whole process of leveling up, of adding ever-increasing modifiers to die rolls, is based on the premise that power comes to those who do difficult things. And it's failsafe - no matter what you put your skill points in, you become better at something. It's a training montage, or a Hero's journey. As with other fictional evidence, these are not "just stories" -- they are powerful cultural narratives. This kind of narrative shapes moral choices[2] and identity. So where do we see this reflected in less obviously fictional contexts?
There's the rags-to-riches story -- the immigrant who came with nothing, but by dint of hard work, now owns a business. University engineering programs are notoriously tough, because you are gaining the ability to do a lot of things (and for signalling reasons). A writer got to where she is today because she wrote and revised and submitted and revised draft after draft after draft.
In every case, there is assumed to be a direct causal link between difficulty and power. Here, these are loosely defined. Roughly, "power" means "ability to have your way", and "difficulty" is "amount of work & sacrifice required." These can be translated into units of social influence - a.k.a money -- and investment, a.k.a. time, o |
a8fa5d05-d98b-4c11-9c22-55392b775247 | LDJnr/LessWrong-Amplify-Instruct | LessWrong | "I told an intelligent, well-educated friend about Less Wrong, so she googled, and got "Less Wrong is an online community for people who want to apply the discovery of biases like the conjunction fallacy, the affect heuristic, and scope insensitivity in order to fix their own thinking." and gave up immediately because she'd never heard of the biases.
While hers might not be the best possible attitude, I can't see that we win anything by driving people away with obscure language.Possible improved introduction: "Less Wrong is a community for people who would like to think more clearly in order to improve their own and other people's lives, and to make major disasters less likely.""
The above is a chunk of text from a post created by a user on the website LessWrong.
How could the introduction to Less Wrong be further simplified to appeal to a broader audience without compromising the core philosophy of the community?
"Less Wrong is a community dedicated to enhancing clear thinking and decision-making, with the aim of improving personal lives and preventing large-scale problems. It's a place where people come together to learn about and overcome common thinking errors, making their decisions more rational and effective."
"Explain how overcoming common thinking errors can lead to more rational and effective decision-making."
Overcoming common thinking errors is crucial to making more rational and effective decisions because these errors often lead us to make choices based on incorrect assumptions, flawed logic, or incomplete information.
1. Confirmation Bias: This is the tendency to favor information that confirms our existing beliefs or values. By recognizing and overcoming this bias, we can make decisions based on a more balanced view of the information available, rather than just what supports our preconceived notions.
2. Hindsight Bias: This is the tendency to believe, after an event has occurred, that we predicted or expected it. Overcoming this bias can improve our ability to accurately assess situations and make predictions about future events.
3. Anchoring Bias: This is the tendency to rely too heavily on the first piece of information encountered (the "anchor") when making decisions. By overcoming this bias, we can ensure our decisions are based on a comprehensive evaluation of all relevant information, rather than being overly influenced by initial data.
4. Availability Heuristic: This is the tendency to make decisions based on the information that is most readily available to us, rather than considering all possible options. Overcoming this can lead to more thorough decision-making processes.
5. Overconfidence Bias: This is the tendency to overestimate our abilities or the accuracy of our predictions. By recognizing and overcoming this bias, we can make more realistic assessments of situations and decisions.
By understanding and mitigating these and other cognitive biases, we can make decisions that are more rational, accurate, and effective. This can lead to better outcomes in our personal and professional lives, as well as on a larger societal scale. |
67c8d2c9-bd67-49ae-9774-cfac224b71bb | trentmkelly/LessWrong-43k | LessWrong | [Link] Quantum theory as the most robust description of reproducible experiments
The paper: http://www.sciencedirect.com/science/article/pii/S000349161400102X
Authors: Hans De Raedt, Mikhail I. Katsnelson, Kristel Michielsen
ABSTRACT
It is shown that the basic equations of quantum theory can be obtained from a straightforward application of logical inference to experiments for which there is uncertainty about individual events and for which the frequencies of the observed events are robust with respect to small changes in the conditions under which the experiments are carried out. |
9ed7eacf-f5fd-49d9-bd3e-17ceb80edbfb | trentmkelly/LessWrong-43k | LessWrong | Betrand Russell's Ten Commandments
Betrand Russell's Ten Commandments for teachers.
> 1. Do not feel absolutely certain of anything.
> 2. Do not think it worth while to proceed by concealing evidence, for the evidence is sure to come to light.
> 3. Never try to discourage thinking for you are sure to succeed.
> 4. When you meet with opposition, even if it should be from your husband or your children, endeavour to overcome it by argument and not by authority, for a victory dependent upon authority is unreal and illusory.
> 5. Have no respect for the authority of others, for there are always contrary authorities to be found.
> 6. Do not use power to suppress opinions you think pernicious, for if you do the opinions will suppress you.
> 7. Do not fear to be eccentric in opinion, for every opinion now accepted was once eccentric.
> 8. Find more pleasure in intelligent dissent that in passive agreement, for, if you value intelligence as you should, the former implies a deeper agreement than the latter.
> 9. Be scrupulously truthful, even if the truth is inconvenient, for it is more inconvenient when you try to conceal it.
> 10. Do not feel envious of the happiness of those who live in a fool’s paradise, for only a fool will think that it is happiness.
I find this to be of use not just for teachers but for rationalists in general. #8, especially, is an especially eloquent formulation of Aumann's Agreement Theorem. |
82222e51-110d-4cdb-bd52-40fbce8a4e34 | trentmkelly/LessWrong-43k | LessWrong | Cultivate an obsession with the object level
In this third section of the sequence I focus on how to leverage positive motivations—in particular curiosity, agency, and determination—to do great work. While all of these are valuable, they’ll suit different people to different degrees. In particular, I think of nerds as favoring curiosity, which is the motivation I’ll focus on in this post. In order to do great work in a given area, you need to spend a lot of time thinking about it, with many of the most exceptional people having an obsessive interest in what they’re working on. While it’s possible to do that via determination alone, curiosity is a much easier source of motivation.
I want to start by distinguishing two types of curiosity: detail-oriented and systematizing. Detail-oriented curiosity is about understanding how things work—like a child who keeps tinkering with blocks until they’ve figured out all the interesting structures that can be built with them. The best way to cultivate detail-oriented curiosity is to learn via answering specific questions or carrying out concrete tasks—e.g. learning programming by building cool apps, or learning physics by building a model rocket, or learning history by figuring out what your life would be like at different points in the past. When you do that, one initial line of exploration can branch out into many more topics. And the patient and direct observation which allows you to discover new things is much easier in pursuit of a goal which genuinely interests you, rather than an externally-imposed goal like those given to kids in schools.
Systematizing curiosity, by contrast, tries to understand the context of a topic in order to fit it into a holistic model of the world—like a child who keeps asking “why?” until they reach the highest known level of abstraction. That might mean studying the Romans by analyzing their role within the broad sweep of history; or studying an animal species by figuring out where they fit into their ecosystem or the tree of life. Syste |
1a85a91b-ad52-48d8-a03e-d4a7a010fee1 | StampyAI/alignment-research-dataset/youtube | Youtube Transcripts | DeepMind x UCL | Deep Learning Lectures | 9/12 | Generative Adversarial Networks
hello everyone and welcome to the next
edition of the UCL deepmind lecture
series I Mahalo I'm a research engineer
at deep mind and a PhD student at UCL
and together with Jeff I'm gonna talk to
you today about generative adversarial
networks so let's start with an overview
why are we interested in generative
adversarial networks
well generative adversarial networks are
a type of genetic model and generative
models learn a model of the unknown
underlying data distribution from a set
of samples from our data set so imagine
this very simple one the example this is
our data set we have our points here and
we're trying to answer the question
what kind of distribution could have
generated in this data and we can answer
this question in two ways firstly we can
learn an explicit model of the data this
kind of probability distribution so here
and then we can answer questions using
this model we can ask well how likely is
it that this point comes from the
original distribution and the answer in
this case well would be not very likely
we haven't seen any samples here and our
model thus has no mass here but we can
also then sample from this model just
like we usually sample from probability
distributions and generate new types of
data in this type of model that models
the probability distribution directly is
what is called an explicit model on the
other hand we can also learn implicit
models in implicit models we don't model
the probability distribution explicitly
what we learn is a type of simulator
that is able to generate new type of
samples that have the same statistical
properties as our original data without
being able to model the distribution
explicitly so now we have some new data
points shown here in blue that match the
properties of the data and importantly
we've generalized we don't always see
the points in red now we were able to
chair it
only that we want these new points to
kind of capture the statistical
structure of our data and very likely
you've seen generative models before and
you've probably seen explicit likelihood
models so the kind of model that has
access to a probability distribution and
often these models are trained by what's
called maximum likelihood in maximum
likelihood we train a model to maximize
the probability distribution of the data
under our model and such models are
probabilistic PCA factor analysis
mixture models and so on and you can
also train URL based models using
maximum likelihood things like pixel CNN
pixel RNN wavenet water regressive
language models and so on
but when you want to Train latent
variable models with maximum likelihood
things get a bit more tricky and that's
when in practice we often use
approximate maximum likelihood and in
another lecture andrey has talked to you
about how to Train variational
autoencoders using approximate maximum
likelihood but today we're going to talk
about implicit models this kind of
simulator models that just generate new
samples without giving us access to like
leads and we're going to focus of one
type of implicit model specifically
generative adversarial networks and why
would we want to focus on generative
adversarial networks
well one practical reason is that
they're able to generate samples that
look like this so these are samples from
began a model that Jeff is going to talk
to you about later in which the model
trained on imagenet data said that has a
lot of variety it has images both birds
dogs
food and so on the model is able to
learn that statistical property present
in the data and generate samples that
match that and samples that look very
photorealistic so these are all
generated from began and this is
specifically remarkable if we think of
how this progress has been through the
last few years so the original again
paper in 2014
we can go from simple images of digits
to images of faces black and white small
resolution but from there a small
revolution has started canoeing faster
and faster and degenerating better and
better images so we go from and I can
wipe the colored images then we go to
higher and higher resolution of face
pictures of faces then we break the
image net barrier in 2018 these are the
first models that are trained on a
bigger set that contains such variety as
we've seen in image net then we start to
generate faces it's very very high
resolution with progressive Ganz this is
starting to look quite photorealistic
then began comes along and we're
generating image net samples not only on
a high diversity data set but also at
high resolution very high quality and
then we move on to style gun which was
published last year in which the authors
show that you can really generate very
high quality samples that look
indistinguishable from photos to the
human eyes so if you were to ask me
whether this person here exists and this
is a photo or is this a sample from
again I will not be able to tell the
difference this looks incredibly
incredibly realistic so this really
inspires us to think well how are Ganz
able to learn this probability
distribution so accurately that we're
able to generate this high quality data
and the answer is that they learn to
generate data through an implicit model
so our model doesn't have this explicit
likelihoods via two-player game and our
players are a discriminator that learns
to distinguish between real data from
our data set and generated data and chip
rented by a model in a generator and the
generator learns to generate data that
fools the discriminator into thinking
it's real so it has to generate really
quality good quality data such a
discriminator thinks well this looks as
good as real data so let's look at our
players in a little bit more detail so
our players are both are both going to
be
using deep neural networks so our
generator is going to have as input
latent poise so what do we mean by that
we need in some sense to model the
entropy and variety of our data
distribution and the way we do that is
that we have a distribution on the input
of our model because remember on the
output the generator now will not have
any distribution it will just produce
samples as the output so if you've seen
something like a variational or encoder
you're used to having a distribution on
the output of the model here we have
absolutely no distribution so in order
to model the entropy of the data we have
to have a distribution in the input and
often this is multivariate Gaussian
noise and interestingly here this noise
is often much lower dimensional then the
data that they are going to be a
high-resolution image while the noise is
going to be something like a hundred or
200 Gaussian datums we take a sample
from our latent noise distribution we
pass that through our deterministic deep
neural network that transforms that
distribution to generate a sample and
that sample can be images or text and so
on the discriminator on the other hand
has a different task the discriminator
has to answer the question given some
set of samples from our data and given
some set of samples from the our model
are these real or are these generated so
it has to answer the question of
distinguishing between these two
distribution the data distribution and
the model distribution and perhaps in a
less adversarial field we can think of
the discriminator as a teacher a teacher
that learns what you're doing well and
what you're doing not well and tells you
how to improve such that you get better
at better at generating real data from
the generators perspective and from this
perspective we can think of the
discriminator as some sort of learned
loss function because the discriminator
guides your training the training of our
model but while it guides it it also
improves itself
and in the original gam paper this was
done by a two-player game so now we have
a minimization with respect to our
generator this is our model in a
maximization problem with respect to our
discriminator of the same value function
and this value function says well make
sure that the discriminator is very good
at this thing wishing between real and
fake data in a classification sense so
we're trying to train a discriminator as
a classifier to maximize the log
probability than the real data is real
and to maximize the probability that da
predicts that the generated data is
generated so today so far is a
classifier once these trained so this is
what the min max game is telling us that
once the discriminator has been updated
we need to train the generator and the
goal of the generator is the opposite of
the discriminator it's a minimization
problem with the same objective as the
discriminator but with a different sign
and the call of the generator is to
minimize the prediction accuracy of D in
order to make sure that the data that is
generated it generates is classified as
real as opposed to fake and if we think
about this from an algorithmic
perspective how would we implement this
well we'll implement our discriminator
and our generator as deep neural
networks and we will taint we will train
them using stochastic gradient methods
so to do that we first have to train our
discriminator for a few steps in
practice this is 1 or 2 so remember that
the min max game said well I have to
maximize with respect to the
discriminator before training our
generator that would entail doing
multiple steps of optimization but in
practice we don't really have the
resources the computational resources to
do that to update the discriminator to
optimality every time you want to update
the generator so we only do a few steps
of gradient descent for the
discriminator and the way we do that is
well we sample a mini batch of data we
sample a mini batch of noise Leighton's
from our prior we pass that through the
generator now we also have a mini batch
of samples from the generator
and we update the discriminator by
stochastic gradient methods to make sure
that our loss is being maximized so we
want again to make sure that we maximize
the probability that real data is real
and maximize the probability that fake
data I generated by the generator is
classified as generated once we've done
this small inner loop of updating the
discriminator we can move on and update
the generator and now the generator aims
to make sure that the data that is now
generated so we sample a new batch of
noise samples we pass that through the
generator we have a new set of generated
data then this data is classified as
real by this new improved discriminator
that we keep with kept improving in our
last stage of training of the
discriminator so we have this game that
we alternate between improve the
discriminator at distinguishing between
real and generated data then use this
new discriminator to update the
generator such that theme generator
generates data that discriminate that
are the screw nature teams as being real
so the take-home message so far is that
gans are able to generate high-quality
samples through this implicit generative
model trained as a two-player game a
discriminator that learns to distinguish
between real and generated data and a
generator that learns to generate data
that looks so good that the
discriminator cannot longer distinguish
between real and generated data and
we've seen that this is done as a cero
sum game we have a minimization with
respect to G maximization with respect
to D of the same value function and this
has a lot of connections with game
theory literature we can think of Nash
equilibria we can think of strategies
that the two players might employ we can
use things such as fictitious play to
improve our game but in practice is
perhaps also interesting to think of
cans from the perspective of distance or
divergence minimization and that is
because we often think of generative
models as doing distance or divergence
minimization and very often explicitly
our
function is a distance or divergence so
we've already talked about maximum
likelihood maximum likelihood my
maximizes the likelihood of the data
under the model which is the same as
minimizing the kill divergence between
the data in the model and why would we
want to do divergence or distance
minimization well therefore chances and
distances give us some really nice
connections to optimality if the
distance between two distributions in
zero then we know that the two
distributions are the same so from the
perspective of learning if we train our
model to minimize this distance and our
distance is zero we know that our model
is a perfect fit of our data
distribution which gives us a very nice
guarantee and again if we look at
maximum likelihood the objective is not
of maximum likelihood is to minimize the
scale divergence which is the expected
value so this is this integral under the
data distribution of the log ratio
between the data distribution and the
model and because this is something that
we minimize so we minimize with respect
to the parameters of this model P of X
we want to make sure that this is as
high as possible because then this ratio
is as high as possible this ratio is as
low as possible because this P star is
fixed this is our data distribution so
though this expectation is as low as
possible
so we want to make sure that P of X is
giving high likelihood to our data which
is very intuitive we want the model that
is able to explain our data and yes the
kill divergence has the same property if
the kill divergence between two
distributions is zero then our model has
learned our data distribution but one
question that you might have here is
well if we are able to say this for a
lot of distances and divergences if
there is zero and our model has learned
a data distribution why are we concerned
with different divergences or distances
and the answer is that well in practice
our model might be miscible
and it might not be able to model the
true data distribution and this can even
be the case for very deep neural network
models because it might still be that
our data set for example image nets is
that complex that we're not able to
model the data distribution exactly and
in that case we might ant um I might
want to ask well what kind of different
trade-offs these different distributions
have so for example here our data is the
mixture of two gaussians and our model
is going to be a Gaussian distribution
and the Gaussian distribution cannot
model our full data distribution because
it's a misspecified model and one
question that we might have is well what
will happen if we train for example
using the maximum likelihood care so the
curl between the data in the model and
the reverse scale between the model and
the data because the KL divergence is
not symmetric and what we see here is
that the behavior is very different when
we use the maximum likelihood KL
the objective remember is to be able to
explain samples from our data all the
samples from our data and if we sample
from our original distribution from our
data distribution we'll have samples
here and samples here and for a Gaussian
distribution to explain both of these
Peaks it will have to put mass all
around them which means that yes it will
be able to explain the data but you're
also going to have a lot of mass here
where actually we don't have any mass
under the original distribution on the
other hand if we use the river scale
this is not what we will see what we
will see is that deme model is going to
focus only on one of the modes is going
to be able to explain it very well but
it's going to completely ignore the
second mode and if you then query your
model to say is it likely that data here
comes from the original data
distribution it's gonna wrongly answer
no because it's not able to capture
anything about this mode so even with
this very simple example of one
dimensional data we can see the
trade-offs of the kind of distribution
that we choose and that's gonna guide us
through as we go forward so one natural
question now might be well are you
during divergence minimization we talked
about this two-player game on
optimization between the discriminator
in the generator how is that connected
to doing divergence minimization and the
original paper showed that yes it is
connected if the discriminator P is
optimum so if we've trained a perfect
classifier to distinguish between
samples from the data and samples from
the model then the generator G is
minimizing the jet session and
divergence between the true and the
generator distributions and this is
great because it also gives us this
connection to optimality that we talked
about before
now if the chance of Shannon between two
distributions is zero then the two
distributions are the same and now we
want to understand a bit more about the
generation and divergence how does it
behave for example in the case of the
misspecified Gaussian when our original
distribution is a mixture of two
Gaussian and the answer is that well it
does a bit of maximum likelihood and a
bit of the repr scale because by
definition it is a mixture of the two
and in practice the answer depends on
how you initialize your model so if you
don't initialize your model too close of
your two Peaks then it's going to do the
maximum likelihood solution otherwise if
you initialize it very close then it
will revert to the reverse scale however
in practice the discriminator is not
optimal as we've seen from the
algorithmic perspective we often have
limited computational resources we can't
train the discriminator to optimality
every time we update the generator so
that at each step the generator is
minimizing the generational divergence
and even if we did even if we would work
to train these optimality given our data
we still still don't have access to the
true data distribution just a few
samples from it our data set so we will
still not have a truly perfect
discriminator and we're going to see why
that is important later on but let's
look at more properties of the kale and
the Jenson Shannon diverge
and here for simplicity I'm gonna focus
on explaining this on the KL divergence
but the same can be said about the
chance of Shannon as we've seen the
Johnson Shannon is a mixture of two
Kells and this property is important
because this has really sparked the
field to perhaps look beyond the Jensen
Shannon divergence look at other
divergences that we can use to train
ganz and why is that well we here our
example that we're gonna run throughout
is a case where we have two
distributions with no overlapping
support so what do I mean by that
here we have our data distribution in
red and our data distribution produces
samples here and it's PDF is given by
this truncated Gaussian here shown also
in red and we have our model and our
model is also truncated Gaussian and we
have a few samples from it here one
thing that we observe is that there is
no place in one-on-one D where both of
them assign nonzero probability so the
data only assigns them zero probability
here but here the model says well this
is not really likely under the model and
what happens in this case is that the
kill divergence in the Jensen Shannon
are going to be constant so the KL is
going to be infinity and the Jensen
Shannon is going to be locked and why is
that well remember the KL divergence
definition is the expected value under
the true data distribution of a log
ratio and this log ratio is the ratio
between the data distribution and the
model and if we look at the fat this
ratio under the data distribution
because this is our expectation we see
well we will have the probability of
this data in the sample here under the
data distribution which we can query is
something obtained from here divided by
the probability distribution of the data
under the model this is where the
problem comes from this probability
distribution is zero because the model
of science is zero mass so this ratio is
infinity so our kill divergence is going
to be infinity and this is especially a
problem from a learning perspective
because when we learn the model we want
to get rewarded if we do something good
right so imagine the case
I've moved my model a little bit from
here a bit closer to the data here so
this is good the model is doing
something good it's going closer to my
data distribution and we would want the
type of loss function that says he have
good job you're going in the right
direction you're doing well but the
kalman the chance of Shannon they can't
do that because this property that the
ratio is still infinitely here still
holds you still even though you've moved
your model closer to the data you're
still at a point where this ratio is
infinity because there's still no
overlapping support so this is why
people thought well perhaps we should
try to train guns that are inspired by a
different divergence so the question is
can we choose another V for our min max
game and will it correspond to a
different distributional divergence and
to do that we have to look at other
divergences and distances and see
whether we can somehow turn that into a
game that you can we can use for again
training and one very nice distance is
the Vassar Stein distance between two
distributions it looks slightly
different than the kale we already see
that there's no ratio we have a
difference of expectations here and a
maximization so just to estimate the
divergence we have to do a maximization
and this maximization has to be over one
Lipsius functions so one thing she's
functions have to be relatively well a
well behaved which means that the
difference from an absolute value of the
function at two points has to be smaller
or equal than the absolute value of the
two points so you can't grow too fast in
a particular region so this means that
the function has to be relatively smooth
and here when we maximize with respect
to the set of functions we're trying to
maximize the difference in expectation
of the expected value of the function
under the data distribution minus the
expected value after the function under
the norm so let's look at an example
here this is our example from before
only that here we're not going to use
the PDFs themselves but we're going to
use samples from tomorrow so these are
samples from our data distribution these
are samples from our model and we're
trying to find a function f that can
separate these expectations as much as
possible
so here we can see that we can put
positive mass under function f around
the data distribution then this
expectation is positive because we are
and we're sampling here we are
evaluating the function at all those
points all these points are positive so
this expected value is going to be
positive we do the same from for the
model but here the model under the model
the function is negative so when we take
the difference the difference is going
to be large is going to be something a
positive a positive number minus a
negative number and importantly the vast
search time distance goes down now if we
have a model that goes closer to the
data even when we don't have overlapping
support because remember this function
has to be 1 Lipschitz it can't grow too
fast in a small neighborhood so we're
moving closer to the data we have
restricted the amount of growth that
this function can have and thus the
difference in expectation is smaller so
we now have a distance that have this
property that if we're doing the right
thing we're getting rewarded for it
which is great now the question is how
do we turn this into again so we've
talked so far about estimating Basilan
distances and we've seen that this
itself involves an optimization over 1
Lipschitz functions but what we're
interested in ideally is in learning how
do we use this to learn probability
distribution or a model that can
generate data from our probability so we
have now our minimization with respect
to our generator again but now we want
to do with respect to the passive time
distance and if we just replace so we
keep the minimization in place and we'll
replace the definition of the passive
sighing distance that we've seen above
we have this form and this form already
looks very familiar we have a
minimization in a maximization so if we
think of our function now that learns to
distinguish between data samples and
model samples
from an expectation perspective rather
from a ratio like we've seen before then
this function can be thought of as our
discriminator so now our minimization
problem with respect to G stays the same
but we have a maximization problem with
respect to our discriminator subject to
the discriminator being well behaved and
this loss function this value function
that looks different because we're no
longer starting with a classification
and we're no longer getting to the
Jenson Shannon divergence but to the
Bathurst Island distance but it's
something that looks very similar right
so now we have something that learns to
distinguish between the data samples in
the model but in a certain sense and we
can use that to train again and this is
what's called the search time yeah and
we can look at other divergences and
distances one of them is mm V maximum
mean discrepancy and looks very similar
to the a certain case only that now the
optimization is with respect to a
different class of functions class
functions that are part of a repeat
using kernel Hilbert space and if we
look at the behavior of mnd on our
standard example we see that it does the
same the value of the function is
positive under the beta the value of the
function is negative under the model
only that the shape of the function
looks different because we're now
looking at a different family of
functions to estimate our model and just
like in the case of the best search time
distance we can try to turn this into
again we have a supreme over a class of
functions we turned that into a
maximization over our discriminator only
them now that the strong leader has to
be part of a reproducing kernel or space
and we have the loss function as an
expectation of the difference of
expectations and remember we started
talking about the KL divergence we
started with maximum likelihood as a
very common objective of training and
tail divergence is a type of F
divergence and if divergences look like
this there's an expected value on F
which is fixed so we know this function
for a kale for example
and a density ratio the problem here is
that if we want to train something like
again inspired by F divergences we will
encounter issues because we don't have
access to P of X we don't have access to
the probability distribution so how do
we get around this well we can't just
start training models using the F
divergences but we can find a
variational lower bound on our F
divergence objective and use that
instead so if you've seen bees before
variational autoencoders they're there
too we use a variational lower bound and
we replace that in our training
objective and in this case in the F
divergence case the variational lower
bound is telling us to optimize this
objective instead and this objective now
should look very similar we have a
suprema
over class of functions and we have a
difference in expectations only that now
we also have the complex conjugate of
the function f from here
the things are looking very very similar
to what we've seen before and the
optimality here is actually the density
ratio that we talked about before and
that we saw that can cause problems in
practice and we're going to go back to
the density ratio in a bit but
importantly now because we have the same
form than we've seen in the bus or sign
in in the case we can also turn this
into again just slightly different
objective now we still have the convex
cones you get a time here but we can use
this to pray normal so so far what we've
seen is that we can train Ganz using
multiple criteria which are inspired by
multiple divergences and distances we
started with the original again and
intergenerational divergence and we look
at the properties of the Jenson Shannon
divergence and based on that we looked
at other distances and divergences that
maybe have different properties
those were Vassar sign and MMD and at
the end we also asked the question well
ok but how about the KL divergence
something that's very used in practice
can we train again inspired by the KL
divergence
and the answer there was also yes now
one question that you might have is why
would I train again instead of doing
divergence minimization is if divergence
minimization gives me all this optimal
convergence properties and the answer is
well it depends in practice you might
not be able to do divergence and
minimization or you might not want to do
a divergence minimization because the
ends have some advantages and we're
going to talk about this now so firstly
remember how we mentioned just now that
the kill divergence requires knowledge
of this model P of X which we don't have
in the case of implicit models of models
like guns so if we want to train again
inspired by the kel-tec divergence we
have to use afghans
but now at least we can train models
that don't have an explicit likelihood
using the KL divergence which is
something that we couldn't do before
right so by using ganz we've expanded
the class of models that we can train
using KL divergence there's also the
computational intractability factor
we've talked about the passive time
distance and how just finding the value
for the master signed distance requires
an optimization problem over a class of
functions but that is intractable for
complex cases so you wouldn't be able to
do it this at each iteration step to
find the assertion distance and then use
that for training but if used a faster
stankin which now will have the same
type of algorithmic implementation as
we've seen for the original can update
the discriminator a few times two three
four or five times and then update the
generator then you can get around that
yeah you're not doing exact faster time
distance optimization anymore because
you haven't solved this optimization
problem but you're still doing something
inspired by the masters time distances
but you can now train a model bit and
remember our problem with the smooth
learning signal our problem
with the KL divergence and the Jensen
Shannon and how that inspired us to look
at other distances and divergences but
perhaps that's not as big of a problem
in the gang case as we originally
thought this idea that they will not
give you any signal to learn when
there's no overlapping support between
the data and the model and why is that
well remember our example the problem
that we have is that this density ratio
was infinity here and that meant that if
I remove my model closer to my data I'm
still not getting any useful signal but
in the case of Ganz I'm approximating
this ratio so perhaps we're not gonna
have the same problems so if we look in
peer eclis we can see that again stiller
so in this paper we show that if the
data is here and the model is here so at
an installation there's no overlapping
support and we train our can the model
after a bit of training still learns to
match the data distribution so why is
that well a simple way to think about
this is again inspired by the KL
divergence because that's a simple
divergence to look at but similarly we
can think about the Jenson Shannon so if
we look at the KL divergence we look at
its definition again we have this true
ratio here that's problematic right
because this is why we're getting these
problems with the KL divergence but in
the case where we train Ganz we actually
use this lower bound instead remember
when we talked about Afghan we use the
bound because we can't have access to P
of X but now we estimate this ratio
using our discriminator and we ask our
discriminator to be in a class of family
of functions because we have to
represent it somehow so that's either a
deep neural network or a function in the
reproducing kernel Hilbert space and so
on and these functions are relatively
smooth so we're approximating our true
ratio with sounding smooth and what
happens in practice is that these smooth
functions won't be able to jump from 0
to infinity or to represent infinity as
the
underlying ratio would so our standard
example again we have our data here our
model here the true ratio here goes to
infinity it's 0 everywhere else but our
MLP that is used to approximate our
ratio will not go to infinity
it starts low and then it starts growing
and growing and growing it needs it
knows that it needs to be higher here
but it won't be infinity and the nice
thing about this is that if I move my
model closer to my data it will know
because there's no jump of exactly here
you need to go to infinity and this is
similar if I use another function class
to represent our ratio so here if we
were using or producing kernel Hilbert
space we see the same type of behavior
around the data we're gonna have a high
ratio but it's not gonna be infinity and
again if I use my model closer to my
data then I'm gonna get a useful
learning signal that says yeah good job
you're going into the right direction
and this is why empirically we've seen
that again could learn even though we
initialize the models to no - don't do
not have overlapping support so the
crucial idea here is that the
discriminator is a smooth approximation
to the decision boundary of the
underlying divergence and we've seen
that with some experiments and with an
explanation of what happens in the case
of the KL divergence so in practice
games do not do divergence minimization
because the discriminator is not optimum
it doesn't really represent that true
density ratio for example but this also
means that gans do not fail in cases
where the underlying divergence were
like we've seen in the Jensen Shannon
case and perhaps another way to think of
discriminators is as learnt distances so
the discriminator is providing a loss
function to our generator but it's
something that itself is learned to
provide useful gradient store model and
this is the case both feet for the
original again the bass or sine gun and
so on they all have this form
minimization with respect to G and
maximization with respect to D of
we function but if we think of this bit
here this is the loss function for G but
it's just trained is trained D using the
discriminator parameters now the crucial
bit here is that we can use this to tell
the generator through our loss function
what we actually care about and the way
we do that is by putting the right
neural network features into the
discriminator so we know that if we're
training data on images we want to use
convolutional neural networks because
those are very good at distinguishing
between images and learning the right
features for that if we're using audio
we might want to use recurrent neural
networks and so on so the crucial bit
here is that we no longer just use
neural network features in our model but
we also use it in our loss and now the
loss in you provide additional signal to
the model to focus on the right aspects
of the data and this is something that a
true divergence not the learned
divergence that this is not a distance
or divergence in a mathematical sense
but this is able to provide you some
useful learning signal that you maybe
wouldn't get if you were using the KL
divergence or something else so to
answer the question of well why would I
want to do Ganz as opposed to divergence
minimization well we see that Ganz
provides very good samples and are you
using this learn loss function where you
can have this additional log to train up
to tell your model what to focus on but
they're hard to analyze in practice you
have to think of game theoretical
reasons and so on and in practice there
are no optimal convergence guarantees
because again the discriminator won't be
optimum however if you do divergence
minimization there are optimal
convergence guarantees and easy to
analyze loss properties but it's harder
to get good samples and the loss
functions don't usually correlate with
human evaluations because they focus on
aspects pertaining to the statistical
properties of the divergence rather in
the mortality of the data so the
take-home message
is that in practice Gant's to not to
divergence minimization and the
discriminator can be seen as a learn
distance it's something that is learned
to distinguish between the data and the
model samples and to provide useful
learning signal to the generator and one
question that you might have is well
which can should I use we've talked
about Bathurst again mmm began the
Jenson Shannon Gann that's the original
again and saw and empirically it has
been observed that the underlying loss
so the underlying divergence matters
less than the neural architectures the
training regime and the data and I think
if you're thinking of the importance the
importance of the features that the
discriminator is learning and the
convolutional or recurrent architectures
underlying them and the kind of
information that provides the generator
that's somewhat intuitive because now
you're focusing really on the features
that are useful and distinguishing
between data and samples and Jeff is
gonna tell you a lot more about this and
give you plenty of examples of neural
architectures that are used for programs
and so far we've talked about
unconditional generative models so far
we're asking our generator generator
please generate a sample I'm giving you
some later noise generate something out
of it but we might want to have a knob
to tune and we might want to tell the
generator generator please generate a
cat or generator please generate a dog
and so on and for that we have to change
our model a little bit so so far we've
talked about deterministic deep neural
networks that are able to transform
Gaussian noise into data but what we
want now is to provide additional input
to the generator to say well please
generate a dog or please generate a cat
and we often provide that in the form of
one hot vector if our conditioning
information is a label we're gonna say
one zero zero zero for dogs 0 1 1 for K
0 1 0 0 for cat and so on and this is
gonna tell the generator what it needs
to
and the reason you will listen to that
is because in practice we also change
how the discriminator strained and now
the discriminator also knows that the
generator should have generated at all
and if it generates a cat that the screw
the generator is not going to get a good
loss for that so now it has to listen to
the conditioning information as well
because the discriminator training
itself has also changed and this in
practice leads to better samples and the
bigger model for example that I've shown
it is able to generate very high quality
samples on imagenet is a class
conditional but sometimes when you train
dance even class conditional Gans you
might get something like this this is
what's called mode collapse so here in
the model instead of capturing the
diversity of the data it's now focused
only on a few examples a few phases and
it's generating them again and again and
what we would like is a way to
automatically know whether our model has
collapsed or not we want to evaluate our
samples without looking at them and
every iteration and so on and in
practice that's a bit hard because the
generator loss is not something very
interpretable so often when you train on
hollows where you store loss going down
smoothly but because we have this two
player game here where the generator
improves the school near improves and so
on the loss itself shown here doesn't
really tell tell us much so there's been
a lot of work of trying to answer the
question where how can we evaluate Gans
and this is a very difficult question
even answering the question broadly how
are we gonna evaluate generative models
is extremely hard so we have no metric
currently that is able to capture all
the desired properties that we want from
our model so some of these properties
are sample quality we want to be able to
generate high-quality samples but we
also want to be able to generalize we
don't just want our model to just give a
samples from the original
because for that we could have just used
a hash table and just say give me a
sample from the original data set and as
you Vina and I are going to talk in
another lecture
we're often also using this models for
representation learning and we might
want to answer the question how good is
this can at representation learning or
how good is this BA at representation
learning and so on and perhaps what we
actually want is to evaluate on the base
goal so what are we trying to do with
this generative model are we using it
for semi-supervised learning so are we
using the features for classification
then maybe we should use classification
accuracy are we using it for
reinforcement learning then maybe we
should use the agent reward and so on
but in practice because that is hard to
do and also more expensive and complex
and it makes it harder to compare models
what people often use are log like views
so you're asking your model to explain
validation data and that it hasn't seen
and based on that you're assessing how
good your model is baganz are implicit
so we're not able to use look
likelihoods to evaluate our Gantz
so people have come up with other
metrics to try to understand how good
our sample is it R and one such metric
is the inception score so in the
inception score what we're trying to say
is that the model is preserving the
class ratio that we've seen in the data
so imagine that we have a data set that
has 50% dogs and 50% cats then we want
that our model and practice is also
generating around 50% dogs and 50% hats
and notice here the inception score
doesn't care about the individual dogs
and the individual cats they can all be
the same as long as on average we get
50% cats and fifty percent dogs
inception score is happy so the way this
is done in practice is that we can use a
pre trained classifier of a known
imagenet to compare the distribution of
labels obtained from data with the
distribution of labels obtained from
samples in a KL divergence sense and
metric is able to capture sample quality
because if the model is generating
garbage you won't be able to get
anything useful out of the pre trained
classifier so the distribution of labels
coming from samples is going to be very
different than the distribution of
labels coming from data it's able to
know whether you're fully dropping a
class
so remember mode collapse or we've seen
that the model can focus on one or two
aspects of the data so if you're
dropping back classes for example you're
not generating any cats the exceptional
score is going to penalize you for this
and it's also going to penalize you if
you're generating a lot more dogs and
cats
for example if correlates well with
human evaluation it doesn't really
measure anything beyond class labels so
every as we've seen if you're carrying
your same dog again and again in such
the scores gonna be good I'm happy and
because if these people have looked at
other metrics for example fresh
inception distance and for station
inception distance is not happy if
you're generating the same look again
again now it's looking both at the
labels in terms of are we generating 50%
can cats and 50% dogs but also inside
the class and the way it does that is by
looking at features on the pre train
classifier rather just than the output
distribution of labels so if we're
comparing now instead of a Kail sense in
a fresher distribution appreciate this
sense of a sense the distribution of
their features obtained from the data
and the distribution of layer features
obtained from the model now we're
getting a more fine-tuned metric so
again you can see sample quality because
we're also using a pre training
classifier we're also able to see if
we're dropping classes altogether
because the feature is on average are
going to look very different they're
only generating dogs and forgetting
about cats but it also goes beyond that
and it captures higher level statistics
but there's a problem with this metric
it has been shown that it's biased for a
small number of samples and kie has been
proposed as a fix in practice and see
this paper from
I clear too
thousand 1844 to fix but we also want to
go beyond us we want to make sure that
our model has not over fitted and it's
not just memorize the data we want
generative models that are able to
capture the essence of the underlying
distribution and the statistical
properties of the distribution but
generalize beyond that and one way to
check this is to check for the closest
samples from our model sample in the
data but we don't want to do this in
pixel space because that's very noisy
and not really representative in a
semantic sense so again just like we've
seen with loss functions when we used
features in our training or just as we
do in our model we're going to use
neural networks features for evaluation
so again we're using a pre trained
classifier and we're going to search not
in pixel space but in the feature space
of this classifier for the closest
images in our data set to our sample so
here we have an example of a sample from
beacon and we're answering the question
well what are the most similar image net
samples from this sample and the answer
is that well they are there are data of
dogs in image net but this exact dog
does not exist in image net so we have
dogs of the same color different shapes
different sizes we have dogs and green
background but this exact same dog does
not exist in the data set so the model
has used training and the data to learn
how to generate talks but to generalize
beyond what it seemed and the take-home
message of this part is to remember that
we need multiple metrics to evaluate
game samples because we don't just care
about sample quality we also care about
overfitting and so on and with this I'm
going to hand it off to Jeff who's going
to talk to you about the Gansu hi
I'm Jeff Donohue I'm a researcher at
deep mind and I've been working on
developing and improving adversarial
networks at scale I'm particularly
interested in the application of Ganz
and other generative models for
representation learning a topic I'll be
discussing a little bit later in this
lecture so now that mihaela has given
you an overview of the theoretical
underpinnings of Ganz my goal for the
rest of the lecture is to take you on a
tour of the Ganz ooh to give you an idea
of the kinds of things that people have
been doing to improve these models from
where they started to the state of the
art now and all the different domains
and problem settings where these models
are being applied a lot of gand research
has focused on image synthesis so we'll
start by walking through the path that
is taking us from applying Ganz to small
data sets like Amnesty to large scale
image databases like image net and a
good place to start is the original gand
paper from Ian Goodfellow and his
collaborators in this paper they used
relatively simple data like the emne
stitch it's that are referred to in the
title of this part of the lecture and
other data sets like this faces data set
and the Seafarer data sets but they're
all pretty small images with resolutions
of about 32 by 32 or smaller in this
paper they used relatively simple models
in fact for these top two images that
you see here the mole the bottles were
multi-layer perceptrons or MLPs
so they weren't convolutional and they
treated the images as flat vectors
completely ignoring the spatial
structure of the images so there's
essentially no inductive biases in these
models and when you have data that's as
relatively simple as this that turns out
to work pretty well you can see that the
kind of digits that you get are
relatively convincing imitations of the
real digits you see highlighted in
yellow here these are the kind of digits
that you can generate with these kinds
of models so it worked reasonably well
but it was mostly just a proof of
concept that this sort of model could
work at all and it wasn't really meant
to be a demonstration of everything
these kinds of models were capable of
which we'll get to later so moving on
from that an extension that you can do
to these models as
Cayla mentioned in her part of the talk
is to make them conditional on a class
label this early work on Ganz called
conditional ganz generalizes Ganz to the
conditional setting we have some extra
information associated with each piece
of data such as a category ID in this
case instead of a category idea this
could be something as complicated as an
image in another domain although in this
work the conditioning was just a
category ID like cat or dog so when you
do this on a mist with the 10-digit
labels 0 1 2 3 4 5 6 7 8 9 you get
results like this where every row is a
different conditioning in this case a
digit label and it turns out when you
give it a 2 that's the label it produces
results that look like a 2 which so it's
great this works next we're going to
look at some early work that actually
managed to tackle some pretty high
resolution images with Ganz there's this
work called lap game by Emily Denton and
her collaborators and so this work was
really cool for a couple of reasons but
just to give you an idea of what it does
in terms of the generation process
basically they'll start from a tiny
image like 4 by 4 or 8 by 8 image and
they'll up sample it via Gaussian style
up sampling so that gives you a blurry
image at a twice as large resolution and
from there what you can do to get a
final image is you generate the
laplacian so basically so you can see if
you go from this image here this tiny
image point to this image all you have
to do is some trivial up sampling
operation but then to actually fill in
the details you have to produce the
laplacian which is the difference
between the blurry image and the final
higher resolution image so you can add
these up to get the final higher
resolution image the blurry image plus
laplacian
so the discriminators job is to take
both the blurry high resolution image
and the difference image either the real
one or the generated one and decide
whether that pair of images is real or
generated so this is a really
interesting formulation for a couple of
reasons in that it sort of decomposes
the problem down to a multi-step
generation processes with multiple
discriminators and generators each one
operating at a different resolution and
the discriminators and generators are
also conditional as you have the same
piece of conditioning information
display images that we're interested in
up sampling so what you have in the end
is this recursive way of going from a
small image to a high resolution image
um so this was pretty exciting at the
time especially because it was the first
scanned paper to produce relatively high
resolution and convincing images and one
of the other nice things was that it's
not a deterministic up sampling so you
can see on this slide it's not producing
the same high resolution image for each
solution input image on left it's
actually producing a full distribution
of high resolution images for each low
resolution image and so you have this
tiny starting image on the left and you
up settle up samples sample with again
until you get to 64 by 64 resolution or
whatever and because it's using random
noise at each stage as you have in any
standard again you wind up with a
slightly different high resolution
output whatever tiny input image you
started with every time you resample the
noise which is what you want if you have
a properly trained and generalizing
again another cool thing architectural
II is that this was a fully
convolutional generator so it's taking a
blurry say 32 by 32 images input and
maintaining that 32 by 32 resolution
throughout the network to produce a 32
by 32 laplacian is output and a nice
thing about that is it allows you to
play the generator to actually any
resolution although it's only going to
work really well at the resolution you
train it on so for example in this case
they only trained it on up to 32 by 32
images but you can keep reapplying this
recursive up sampling and laplacian
generation operation with the highest
resolution generator that you trained
and then in the end if you keep doing
this you get what looks like continues
to look like higher resolution images
although obviously it's a little bit
blurry and not necessarily the best
fidelity but you can't really expect too
much more when the models only have
received 32 by 32 images moving on to
this paper called deep convolutional
gans or DC games from Alec Bradford and
his collaborators um so this is another
really exciting paper at the time
because it was a very simple
architecture it was basically very
similar to the original game framework
but with deeper confidence
and it used batch normalization which
made this sort of notoriously difficult
Gann training process much smoother than
it was without batch normalization um
the two networks the generator and the
discriminator we were both confident so
the generator was a decom net or an up
sampling continent and the discriminator
was a down sampling continent and it's
basically a five ish layer Network not
too dissimilar from something like
Alec's net at the time so when you apply
DC Gans to a dataset of indoor scenes
you get results that look like this
which were at the time at least quite
impressive and exciting and one of the
cool things that you can do with a
network that's trained this way is you
can take to noise or to Z samples z1 and
z2 on the slide for example one of them
might produce an image of a desk that
looks like this and one of them might
produce an image of a bed that looks
like this and then you can interpolate
between these two Z's in z space and at
every point in between you get what
looks like a relatively realistic and
semantically meaningful result so of
course it's not perfect but one thing
that this shows is that the model is
able to properly generalize so it's able
to turn a data set of a hundred thousand
or ten thousand discrete examples into a
continuous distribution of images and
this also showed that the model isn't
simply memorizing the data set because
obviously in the data set you wouldn't
have an example of any interpolation for
any given pairs of images in the data
set and this is what happens if you do
that same kind of interpolation thing
for faces again obviously it's not
perfect and there's some kind of creepy
looking results in this case but still
interesting one really interesting
observation from this work is that there
appear to be some meaningful semantics
in the latent space so basically in this
sort of example they observed that if
you take a latent that produces a man
with glasses from a pre trained Gann
model and another Lathan that produces a
man without glasses have another
Leighton that produces a woman without
glasses and you do man with glasses -
man plus women you get women with
glasses and that might remind you a
little bit of the word Tyvek results for
language embeddings if you're familiar
with that work but what the shows for
ganz is that there are direction
in this DC game Layton space that
correspond to the presence or absence of
glasses as well as the gender of the
subject which is not something the model
has ever explicitly trained to do it's
just sort of learned to sort these
semantic properties and represent them
in the latent space in some way which is
really interesting we'll talk more about
that later jumping ahead a little bit
there was a paper in 2018 called
spectrally normalized gans from me otto
and collaborators and this was really
exciting - it was the first real crack
at using a single Gann a single
generator and a single discriminator to
model this imagenet dataset with a
thousand classes in 1.2 million images
the main trick in this paper was
intended to stabilize Gann training by
clamping the singular values of the
discriminator weights to one so that all
the weights of the network had a
singular value of 1 which basically
means that no matter what the input to a
layer is the output magnitude is not
increased and the way it's implemented
is every time you run the discriminators
forward pass you calculate an estimate
of the first singular value for each
layer and because this is a linear thing
you can just rescale the weight as shown
here by dividing by its singular value
to get a normalized version of the
weights with spectral norm 1 so this
regularizes the discriminator and
they're actually using here essentially
a linear loss function the hinge loss in
this case so if you didn't have this
regularization the discriminator could
basically improve its objective just by
increasing the magnitude of its weights
but because you do have the spectral
norm regularization the discriminator
has to improve its objective in ways
that actually meaningly improve the
gradient signals that it passes back to
the generator which is what we want that
of a discriminator
so when this is applied to imagenet you
get images that look like this which at
the time was particularly impressive
because nobody had successfully taken on
the full image in a dataset with a
single again before in some follow-up
work from the same group they added this
idea of a projection discriminator to
handle conditioning so previously they
used the kind of input conditioning we
saw before where you would feed in the
class label like pizza as the input to
the very first layer or other word this
is other variant called a CGI
auxiliary classifier against where you
would train the discriminator as a
classifier directly so what this paper
proposed to do is called a projection
discriminator so they're learning a
class embedding which is the same
dimension as the discriminators last
hidden layer and they project the class
embedding onto the hidden representation
or dot product it and that gives you a
class conditional realness score that
the discriminator outputs so basically
rather than feeding the label as an
input it becomes an output in this case
and there's a pretty interesting
theoretical justification based on the
underlying probabilistic model that
justifies doing it this way and it not
only makes sense theoretically but
performs very well empirically and you
see results that look like this which
was even more impressive than the
results we saw with SN Gann alone one
more pretty interesting innovation in
the Gann architectural space was what's
called self attention and self attention
is that this technique forgiving
networks the ability to do some sort of
global reasoning it's been applied a lot
of domains especially in language
modeling and machine translation in the
image domain it allows you to basically
learn to measure your global statistics
about the image so for example so this
was used in both the generator and the
discriminator and for example if you're
the discriminator you might want to be
able to ask questions like if the tail
of the dog is on the left side of the
image is the face of the dog on the
right side of the image which is
something you might want to know if you
want to tell whether the image is real
or fake
and you couldn't typically typically do
something like that with a single
convolutional layer because the kernels
are just too small to capture that much
of the image so this resulted in better
global coherence across the images that
the Gann would generate and they also
have these nice qualitative results to
visualize what the model ends up looking
at so for example in this case it looks
like the model decided to compare this
area around the head of the dog to this
area near the tail of the dog to make
sure you know that all the dog's body
parts are kind of in the right place
which you can imagine how that would
help they generate or learn to produce
images with better global coherence and
then at the end of the day you get
results that look like this on the image
net data set which again was another
advance both qualitatively and
quantitatively in terms of Inception
score compared to the previous results
that we've seen so finally we get this
project from our group at deep mind
called began led by Andy Brock the main
idea of this work which I think I'm
allowed to say because I was a co-author
on this paper was to make Gans really
really big and we wanted to do a big
empirical study and sort of digest all
of the image gain research that's been
done so far and scaled them up as much
as we could and just kind of see where
it would take us so yeah
begins we had big big batches big models
big datasets big high-resolution images
so the batch size that we used for our
main results was 2048 compared to batch
sizes of roughly 256 that were being
used before our work and this turned out
to be a particularly important hyper
parameter which is really critical to
making these models work as well as they
did and one hypothesis for why this
might have been so important is that the
imagenet dataset has a thousand classes
and if you're doing mini batch SGD
especially in a setting that's as
unstable as gand training still can be
you really want ideally each class to be
represented in each batch so that the
model doesn't end up sort of forgetting
about classes that it hasn't seen in a
while and so if you have a batch size of
2048 it's fairly likely that in any
given batch almost all of the thousand
classes will appear whereas obviously if
you have a batch size of 256 it's
obviously impossible for a thousand all
thousand classes to be in that batch so
we not only trained on imagenet but also
this internal Google data set called jft
which has three hundred million images
I'm sorry sort of used image that is our
development data set when designing
these models throughout the course of
the research and then we directly
applied the same models to jft and we
found that they worked pretty well there
even on our data set which was you know
two
200 or 300 times larger so you can see
on the right the type of images we get
from this kind of model and another few
of them are here and so overall this
paper was a really big empirical study
to build up a reliable kind of recipe
for large-scale scam training so we
inherited quite a few tricks from prior
work but we like to think that you can
be confident that each one was ablated
really well and turned out and really
turned out to be the best choice in
terms of the image fidelity and the
quantitative scores that you get so
among these tricks we had the hinge loss
which is basically a linear loss except
it's sort of clamps to a minimum value
when the discriminator is or a maximum
value when minimum value when the
discriminator is correct and
sufficiently confident in its
correctness and spectral norm which we
just discussed as well as self attention
and projection discriminators and
finally some tricks that we added to the
toolbox relative to previous work
included orthogonal regularization which
sort of enforces that each row of the
weights is orthogonal that they're kind
of doing different things and we used
skip connections from the noise so
basically there was a direct connection
from the noise z to every layer in the
generators convolution stack and
similarly for the class label embedding
in the generator we used we learned an
embedding that was shared across the
different layers each layer again having
a direct connection from the class
conditioning as well one interesting
trick that we introduced was with this
paper was what we called the truncation
trick
it's an inference time trick so it
doesn't affect training at all it's
something that you can do with any pre
trained generator at inference time when
you're want to go produce samples so
basically we can change the standard
deviation of the noise input to the
generator basically change the scale of
the noise distribution as you can see in
the figure here so it's sort of
shrinking closer and closer to zero so
if you watch the animation we start with
this you know Y distribution and the
resulting image is produced for each
class at the beginning of this animation
like now are quite different but as the
distribution gets skinnier the images
become more and more uniform
for a given class basically what this
does is when you make the distribution
really small near zero is it gives gonna
give you kind of a prototypical or a
modal example of each class and in this
case for the dot for these dogs it's
typically a very well centered and
camera facing example of each dog which
is sort of inherited from the biases of
the datasets because most people you
know will take pictures of their dogs
when they're facing the camera and
whereas if you keep the noise as it was
a training time as you can see here with
Sigma equals one for the for the
Gaussian input to the generator um you
get quite a bit more variety so the
truncation trick is really a way to
trade-off between the variety and the
fidelity of the samples that you can
generate with these models and yeah
here's just another example of what
happens with the truncation trick for
some bugs some butterflies kind of the
same thing as we saw for the dogs so as
I said the truncation trick is really a
way to trade off between variety and
fidelity so what you can do is compute
the inception score and the F ID at
every point along this curve of Sigma
values that you can produce via the
truncation trick so as Haley explained
earlier when she was talking about
evaluating ganz
um the inception score doesn't care
really about how diverse the samples you
produce are in each class it really just
cares how good samples are for each
class how confident it is in the
classifications for each class so if you
just want to maximize inception score
setting the scale to roughly zero is
really the best thing you can do and
when you do that you see that you end up
maximizing inception score down at this
point on the curve here at around two
hundred ten in this case but when you do
that you have relatively bad F ID of
thirty plus and higher is worse for F ID
on the other hand if you leave Sigma
equals one and the other end of the
curve here which is the default the Z
distribution as it was at training time
you get relatively bad inception scores
roughly 105 or 110 but very good F IDs
as you're capturing more of the inter
class distribution which F ID is a
little bit better at measuring
so as so kind of as an alternative and
more detailed way to evaluate Gansa can
look at this full truncation curve
whereas previous work had just looked at
individual points using the default
distribution those sort of gives you a
full frontier of the inception scores
and F ID scores across this entire curve
one more thing that we played with sort
of late in this work was this different
architecture called big and deep that
you see here so this is a deeper yet
more efficient architecture you can see
in a single block it has twice as many
convolutions in the main block so
there's four of them instead of two and
we had twice as many of these blocks in
the big and deep architecture so overall
it's four times as deep the key thing
that makes this even more efficient than
the original began is that we have these
it's not a new idea but we added these
one-by-one convolutions that go to a
lower channel count and then these 3x3
convolutions operate on this lower
channel count space so it all ends up at
the end of the day and the 3x3
convolutions are the most expensive part
so it all ends up being a little bit
more efficient than the original
architecture and the nice part is it
also performs better with inception
scores of over 240 at full truncation
down here and now heidi is around 6 with
the minimal truncation now this model is
definitely not perfect and a lot of
times the failures are kind of fun to
look at as well so for example this
image on the left that we sort of
affectionately refer to as dog ball and
this is an example of what we call class
leakage so according to began this image
is an example of a tennis ball so the
reason that we think this happens for
image net specifically is that there are
just so many dogs in the image net data
set there's roughly a hundred dog
classes so the model is sort of very
accustomed seeing dogs and it sees them
roughly a hundred times as often as
tennis balls so when it sees tennis ball
it says you know hey this that's fuzzy
it's probably a dog I'm gonna put some
eyes and a stout on it so this happens
at least some point in training and it's
not actually from the final converged
model but it's kind of fun to see what
happens as the models when you generate
better and better in images throughout
better and better images throughout
training
and other failure modes include classes
that are difficult particularly any
class that includes a human face now it
could be a little bit just that they
seem particularly bad because humans are
very sensitive to how good human faces
look or how realistic they look so
there's kind of this uncanny valley
effect although we're quite a bit off
here I think you'd probably agree and
classes with really complex structure
like the image of this band here are
also really hard with a lot of when they
have a lot of different objects in the
scene and classes that are
underrepresented in the dataset and have
also have complicated structure like
this image of a I think it might be a
two-bar French Horner and it's just
really hard for the model to capture
this sort of complex structure without
too many examples especially to
generalize to new instances of the class
as you're sort of asking began to do so
more recent follow-up work that we did
is this work called Logan or latent
optimization gains so latent
optimization is this idea intended to
improve the adversarial dynamics of the
gang game between the generator and the
discriminator and basically what it does
is it uses what's called the natural
gradient descent to optimize G's latent
inputs disease so it changes the Z's at
training time to make the discriminator
happier so does one natural gradient
descent step inside of the training loop
to change Z and it actually is going to
backprop through this entire process so
it's a little bit more expensive than a
standard gain it takes about roughly
twice as much computation time per step
but it results in really significant
improvements in begin in terms of the
variety and the fidelity that you can
get and it's particularly noticeable
when you compare along the truncation
curve so for example if we truncate such
that the inception score is roughly 259
you get much better F IDs when you train
using Logan than with a standard begin
deep so both quality so Logan is about F
id8 versus big and deep about 28 at the
same point and it's obvious also if you
just look at the samples at this point
in the truncation curve big and deep is
basically producing all
uniform samples per class whereas Logan
still has pretty diverse samples so a
parallel line of work to the began work
in all of the image network was this
line of work from Nvidia the first work
in the series was called progressive
ganz the idea of this was sort of
similar to what they did in lap game
although it's formulated quite a bit
differently so the idea here is both for
efficiency and to get the model to
converge dependably they start off
generating at a very low resolution like
a four by four resolution and then after
your tiny image generator has converged
you can add an extra upsampling layer
you like you see here and a few extra
convolutional refinement layers to get
an 8x8 image generator if you started
with four by four then you wait for that
one to converge you repeat for sixteen
by sixteen 32 by 32 and so on and so on
until you get up to the final resolution
that you would like to generate in their
case they went to very high resolutions
of up to 1024 by 1024 and in the end
this resulted in extremely compelling
images at least in this restricted
domain of celebrity human faces and you
get what looks like pretty much
photorealistic results of human faces at
this very high resolution of 1024 by
1024 you know at least for me it's very
hard to tell the difference between most
of these faces and real human faces the
follow-up work from this team was called
style Gans so style games were also
shown to be capable of generating
remarkably photorealistic face images
and in this case they used was probably
a more challenging data set than the
last one with a lot more variation in
the images the data set they used in the
previous work progressive games was
mostly images of celebrity as whereas
this data set was a lot of a lot more
diverse and can mostly consists of
consists of images of not so famous
people so the interesting thing about
the architecture that they used in this
work was that it had these structured
latent inputs so they had these the
usual global latent are the usual Z's
that you have this input to the
generator but they also had these
spatial noise inputs so you can see in
the image that each column has sort of
the same global
sure global semantics like this middle
column for example seems to be late in
corresponding to you know young children
and this column seems to correspond to
being centered on the right side of the
image and looking towards the center and
that's because each column uses the same
global latent or is the spatial latent
is the same in each row and it seems to
mainly control in this case the sort of
background of the image as well as the
skin tone so what the architecture looks
like um is on this slide so on the Left
we have the usual flat vector Z which
they explicitly called the latent and
it's passed through a sequence of eight
fully connected layers an MLP to get the
final latent vector down here and then
this latent is input into every hidden
layer of the generator but the
interesting new piece here is that they
also have these pixel noise inputs over
here so at every layer you have a single
channel of random noise of the
appropriate resolution so four by four
eight by eight and so on and so on and
that noise is going to get
reincorporated eight at each of these
layers and as we saw before it ends up
using this global latent to control the
overall global appearance of the image
while these pixel noise latencies are
used to control the local variation of
the image and another example of what
this looks like in action is on this
slide so if you freeze the global
Ayden's and the course-level pixel noise
if you freeze all those you can change
just the fine high-resolution pixel
noise to get stochastic variations you
know in this case controlling start of
the fine differences and how this
toddlers hairs look so I hope that what
you can take away from this part of the
talk is a couple of things first there's
been pretty rapid progress in the span
of about five years scanning scaling up
gans from the amnesty digit images that
we saw in the original game paper to
these pretty large scale databases of
high resolution images like image net
and the flickr faces HQ dataset and the
improvements occurred really in a
variety of different places it wasn't
just about changing the architecture or
changing the objective it was really all
of these things combined the G and D
architectures have gotten better and
deeper the conditioning structure has
changed
the normalization has improved we saw
that batch normalization and spectral
normalization were quite helpful the
parameterization of the discriminator
has changed
we started off taking the conditioning
vectors input and now with the
projection discriminator we project
class embedding onto the hidden
representation of the image the latent
space structure has changed for example
in the style game paper where we had the
pixel noise Leighton's to control local
appearance and the loss functions have
changed which we saw more in caelis part
of the lecture and the algorithms have
changed for example in Logan where we
have an inner optimization of the
Layton's but while we can produce some
pretty convincing image I'd say the
problem is still pretty far from solved
for example these state-of-the-art
methods take a good amount of time in
quite a bit of computation to converge
and even with begins you know we're
still not great at every single image
category so I hope this gives you a good
idea of how the research has taken shape
into what the state of the art is today
and you know maybe even inspires you to
try your own ideas and make these
methods work even better so next I want
to talk about an application of Ganz
that I'm particularly interested in
which is the use of Ganz for
representation learning you'll hear a
lot more about the topic of unsupervised
representation learning in the next
lecture from mahalia Irina but for now
I'm going to address a few of the
directions that people have been
thinking about in terms of using Ganz in
particular representation learning so
just to give a couple of motivating
examples for why it might be interesting
to you use Ganz for representation
learning this is a slide that we saw
before but just to remind you so in the
dcen work Alec Bradford and
collaborators notice that in the latent
space of a deep convolutional again or
DC Gann you can do these kind of
arithmetic operations in latent space
indicating that certain directions in
latent space correspond to high-level
semantic attributes in the observations
space in this case human faces such as
the presence or absence of glasses or
the gender of the subject and all of
this arises without began ever being
explicitly told without without ever
being explicitly told about these
concepts of
or gender as another motivating example
I took the big in architecture and I
added an extra latent variable to the
generator input so this is a categorical
latent variable with a thousand 24
outcomes and it's just fed into the
generator as a one hot variable in
conjunction with a regular continuous
latent variable the 120 D calcium and
the kind of things that you get out of
this are pretty interesting so I train
this without class information it's
unsupervised and unconditional but it
does have this use this categorical
latent variable in place of the usual
explicit class label that you'd get in
the conditional supervised setting so it
seems to learn to associate this
categorical variable with high level
semantic groupings that almost look like
image categories so in this slide you
see about eight sort of randomly chose
and outcomes of the 1000 way categorical
variable and for example in this one
value this categorical variable shown in
the first row corresponds to what looks
like sea anemones another one looks like
a certain breed of dog and a sort of
grassy green background another it looks
like these kind of mountainous
landscapes and so this is really cool
and you can imagine that in a sort of
idealized case the dream might be that
it learns a clustering all in its own
that looks exactly like say the 1,000
image net categories or at least each of
these categories might be represented by
some combination of these categorical
variable outcomes and if that were to
happen then training a model that can
predict this latent variable given an
image would be exactly like training a
fully supervised image that classifier
and of course all of this came for free
because it's unsupervised so it's not
like the image and a data set where we
had to manually label each of the images
with category ID or you know pay
somebody to do that so going towards
that dream there have been many attempts
to get models that fulfill this promise
of learning representations using Ganz
completely unsupervised and I'll discuss
just a couple of them here one of the
first interesting papers from a few
years ago was called info Ganz or
information maximizing Ganz and compared
to regular Ganz it adds this inference
inference network to recover the latent
code Z given the generator output G of Z
which
in this set of experiments that we're
looking at is an imminent image of a
menace digit and what this does is
forced the generator to use each of its
input latent variables meaningfully in
order to maximize the information
content about the variables and the
images that it outputs and when you
train it with these latent codes it
learns to associate each outcome of the
categorical latent variable with a
different digit value and use the
continuous valued variables to vary the
style and the size and the rotation of
the digit so basically is using the
discrete latent to capture the discrete
variation in the data set and the
continuous latent to represent the
continuous variation the data set so
that's pretty cool and so one sort of
disadvantage of this approach when it
comes to representation learnings that
you don't have a ground truth latent
associated with real images like you do
for generated images so the inference
network that you've added here is only
ever getting to see generated images or
you have where you do have the latent
and so that might be OK for
representation learning when you have a
very simple dataset like M nough store
the generator is able to capture it
almost perfectly like you can kind of
see on this slide but when you go to cut
something more complex like image that
if your generator isn't perfect and it
probably won't because image that is
still really hard your generator is it
perfect then when you go to apply the
learned representations trained on these
generated images there's going to be
kind of a domain shift between the
generated images that the inference
network has seen versus the real images
that you want to get feature
representation for so then comes this
other class of methods that was called
either adversarial elearn inference a li
or bi-directional gans or begins and
this is sort of an adversarial approach
to jointly learning to generate data and
learn representations from it so
compared to a regular Gann the setup
adds an encoder Network which we'll call
Eve for most of this which learns the
inverse mapping from the generator G so
whereas the generator maps from features
or latent to images
G of Z the encoder does the opposite it
matches from images or data X to latency
of X and the other difference from a
regular Gann is you have a joint
discriminator so it sees not only an
or data point X or G of Z but it also
sees the latent Z or e of X so these X Z
tuples can either come from taking a
data point X and passing it through the
encoder to get a predicted layton a of X
or it comes from sampling layton Z and
passing it through the generator to get
a image G of Z and then the
discriminators job here is to figure out
which of the two generating processes
each of its input tuples came from and
the generator and encoders job are to
fool the discriminator basically into
picking the wrong process and it might
be a little confusing when you first
look at this because it's not entirely
clear what the jet encoders job is like
why does it have to produce anything in
particular for a given X so well it
turns out that under this objective of
discriminating between these two
different types of tuples there's a
global optimum here where if you have a
perfect discriminator and the generator
and encoder are perfectly satisfying the
discriminator then it turns out that the
encoder and generator have to invert one
another so if you pass an x an image
through the encoder and get a predicted
latency of X and then you pass that back
through the generator it should
perfectly reconstruct the input X that's
the global optimum of this model and
unlike in say auto-encoders were you
explicitly training for this property by
minimizing a squared error in this case
the encoder and the generate communicate
a don't communicate at training time so
they they never see each other's outputs
it's all done through the discrete
discrete ER so the encoder never sees
the outputs of the generator and the
generator never sees the outputs of the
encoder so one thing that makes us
interesting for feature learning is that
the encoder never suffers from the
domain shift problem I mentioned before
of C having to see these kind of weird
bad or at least initially bad generated
images that the generator gives you it
only ever sees real data which is
exactly what we want for a presentation
lending because it means that there's no
domain shift when we go to apply the
encoder to real images so in practice
this inversion property that we proved
to be true at the global optimum doesn't
actually hold perfectly but what you see
is
the reconstructions that you get from
passing X through the encoder and the
result back through the generator often
capture quite interesting semantics of
the inputs so for example if we look at
the digits here often the digit identity
between the original data X and the
reconstruction G of X is the same so for
example you know 2 goes to 2 3 goes to 3
etc etc so what that tells you is that
the representation the encoder gives you
is capturing the digit identity at least
to some extent and this is all just from
looking at the data we never explicitly
tell it what a 5 looks like and so on um
so if you scale these models up because
the original work we just looked at was
sort of at the DCN scale if you apply
this in the big and setting where you
have the same generator and
discriminator architectures as in begin
and you add an encoder model which
something like a state of the art
recognition image recognition model like
a resident style model at least a few
years ago some very interesting things
happen and we call these resulting
models with a few other tweaks that you
can read about in the paper we call them
big begins naturally so for example if
you pass this dog through the big bag in
encoder and back through the generator
to get a reconstruction the
reconstruction that you get is what
looks like a pretty similar dog although
with its tongue stuck out and kind of
facing in a slightly different direction
this person in a red coat in the winter
becomes a slightly more zoomed in person
in a red coat in the winter so in
general what many of these semantic
properties of the input get maintained
in the reconstructions even though the
model is never told what semantic
properties are interesting and all this
is happening because the structure of
the discriminator is essentially shaping
an implicit reconstruction error metric
in semantic ways at least this is kind
of my intuition for what's going on um
so the discriminator is a convolutional
network and we know that convolutional
networks are good at predicting semantic
attributes of images so the resulting
implicit that reconstruction error that
we're minimizing implicitly if not
explicitly mind you but but this sort of
implicit reconstruction error emphasizes
the semantics remaining the same even if
the individual pixel value has changed
quite a lot
so for example the model isn't going to
remember exactly what kind of pizza you
gave it but a war will remember it was
some kind of pizza and it was roughly in
this part of the image so it's almost
kind of human-like in terms of what it
remembers about the input image it has a
sort of fuzzy semantic memory of what it
saw without for example having to
remember you know the exact position of
every single blade of grass and this is
in contrast to the standard pixel wise
reconstruction objectives where it's
basically forcing the model to remember
every single pixel value so this is in
some sense exactly what we want in a
reason Tatian learning objective which
is what at least you know in my opinion
makes this an interesting method and
when you evaluate this quantitatively
and this sort of standard setup where
you basically take the encoder and use
it as a feature representation and train
a linear classifier supervisor on top of
that you get something pretty close to
state-of-the-art results compared to all
of these self supervised methods that
are very popular these days and which
will think here about in the next
lecture and another way to see what
representations are being learned by
this method is by looking at nearest
neighbors in the data set so you can
take images from the validation set as
queries and this left showing this
left-hand column here and find the
training set images that are closest to
them in big bag and feature space so in
general you can see that the nearest
neighbors tend to be very semantically
relevant to the input image in fact you
know with this dog from the validation
set here its nearest neighbor and the
training set shown here I think based on
the background it's in fact exactly the
same dog even though it's obviously
facing a different direction and if you
just looked at the pixel values this
would be quite different so it's kind of
cool that out of 1.2 eight million
images in the training set that ended up
being the nearest neighbor that same dog
at a different angle although it's
probably a little bit lucky it's still
fun finally for the last part of the
talk I just want to give you a taste of
some of the other modalities and the
different problem settings that people
are trying to tackle using generative
adversarial networks so starting with a
couple of
in the image space one of the coolest
lines of work in my opinion started with
this paper called pix depicts by Phil
Isola and his collaborators and what
they did in this setting was train a
generator to translate between images
from two different domains so for
example if you had satellite images like
these and you wanted to be able to
automatically translate these images to
kind of roadmap type images like you see
here and the way that they do this in
pics depicts is you take all these
paired examples of images so the
satellite image view and the
corresponding map view of the same area
and you train a conditional gam that
takes the aerial view as an input and
produce that the map view is an output
so the way you train this thing is you
have a standard gain objective a
discriminator that says does the output
of the generator look like a map view
that I've seen before but you also have
this l1 reconstruction error so since
you have a ground truth for what this
aerial or this map view is supposed to
look like you you can use this kind of
l1 pixels reconstruction error to tell
the generator that this is exactly what
your output should look like for this
input so basically it's kind of like a
traditional supervised learning setup
and you can see that this works in a
number of domains as you can see on the
slide
labels - street scenes edges -
photographic images of purses for
example and yeah so it's quite cool but
in the more general setting you might
not actually have paired examples so for
example if you want to train again that
translates between images of horses and
to zebras or vice versa you're probably
not going to have paired images where
all the horses and all the zebras are in
the exact same positions in the image
like we assumed we had in the pics to
fix work that we just talked about and
so enter this method called cycle gam
where you want to be able to sort of
unsupervised be able to translate
between two different domains with it
but without paired-samples between these
domains and the high-level idea of how
this works is by enforcing this property
they call a cycle consistency in
addition to all the normal gain
objectives so it's still again so you
start with an image
in domains Z domain a say it's an image
of zebras and then you translate to
domain B say it's an image of horses and
then translate back to domain a so
translate back to zebras and the zebra
image that you get after that process
should look pretty much exactly like the
image zebra image that you started with
so that's gives you an idea of how the
method works and as a result you can
basically translate between any two
domains that have sort of reasonably
similar information content such as
going from summer scenes to winter
scenes horse scenes to zebra scenes
photographs to different artists so this
is a really cool approach it's almost
you know a little bit magical that it
works and it produces some really cool
compelling results now I'm going to
touch on a little bit of work using Gans
for audio synthesis so we've can on the
left here was one of the first attempts
to produce raw audio waveforms using
Gans and they showed that for example
you can train unconditional gans to
produce reasonable one second clips of
piano music or human speech Mel Gann was
work on text-to-speech that takes as
input Mel spectrograms and produces Ross
speech audio as output and then there
was this other text-to-speech work from
our team at deep mind called Gantt ETS
where we take the linguistic features
aligned in time as input and produce
also produce raw speech audio as output
and both of these text-to-speech methods
work reasonably well for speech
synthesis which is pretty exciting
because they're also quite efficient
relative to many of existing
state-of-the-art approaches to
text-to-speech so in addition to images
people have also used Gans to generate
videos and predict future frames of
videos so you can apply a lot of the
same tools and toolbox that we've used
for images to videos as well of course
since you know it since you have within
a frame the same two-dimensional
structure that we have for images a
frame is an image but you also have a
third dimension time and that turns out
to make this problem a bit different and
arguably quite a bit harder than it is
for images partially just because of
potentional resources it takes to store
and generate videos versus still images
but also because humans are quite
sensitive to unrealistic motion so it's
important to get that right in order to
have reasonably convincing results so in
all three of these methods on the slide
a lot of a lot of work has gone into
making that computationally feasible so
one thing that we did in DVD again for
example in the middle here and it was
further developed in Tribune again was
to decompose the discriminator into two
separate discriminators neither of which
are seeing all of the pixels in the
video so it ends up being
computationally feasible that way so
there's one discriminator that we called
the spatial discriminator it operates
only on a few individual full resolution
frames but it only sees a few of the
frames a subset of them so that inch but
that discriminator basically ensures
that each frame looks connect coherent
independently and then there's another
discriminator that temporal
discriminator that sees multiple frames
but they're spatially downsampled so
that also doesn't see all the pixels
because it sees downsampled versions of
the images but that one is going to
ensure fluidity over time so together
that makes the problem from almost
computationally infeasible to being
fairly feasible and finally just to give
you a final taste of the many domains in
which people are applying gans
there's a reinforcement learning and so
this work on using games for imitation
learning called the generative
adversarial imitation learning or Gale
and essentially it uses a game like
method to learn a generator which in
this case ends up being a policy which
learns to imitate expert demonstrations
by fooling a discriminator whose inputs
are state action pairs and it addresses
many of the typical problems that people
see with standard behavioral cloning
methods in reinforcement learning
there's work I'm using Ganz for image
editing so that amateur artists for
example could specify just the course
layout of a scene without having to
actually paint every single detail and
then the Gann can go in and fill in the
low-level details with some pretty
nice-looking results and they have a
pretty fun demo that you can try out
online if you're
if you're interested there's work on
using Gans for program synthesis there's
this work from deep mind called spiral
or you have a generator that instead of
specifying each pixel value has to
specify individual actions like the
brushstrokes in a painting program so it
has to produce these discrete
instructions and you can't directly
backprop through this generation process
like you can in sort of standard image
generation Gans so you end up having to
use a reinforcement learning approach to
do this and you can imagine that you
could apply this to all sorts of
different types of programs not just
drawing ones there was a really cool
piece of work recently called everybody
dance now which was used for motion
transfer so you could take photos of
somebody in different positions who's
not a very good dancer and map the
movements of a professional dancer onto
their body so it looks like they have
you know professional level dance skills
and if you haven't seen the video demo
of this already you really have to go
look it up and watch it because it's
pretty entertaining and super
entertaining Gans have also been applied
to domain adaptation so domain
adaptation if you don't know is this
problem or say you might have a bunch of
label images of things happening during
the day within the daylight and you want
to train a classifier on that data and
then apply it to images of things
happening at night and by default this
won't work very well is there's going to
be a domain shift between day scenes and
night scenes and there's different
methods of alleviating that problem some
of them are using games like this one
here and finally there's a number of
artists using Gans for different kinds
of human machine collaborative art work
kind of and they produce some really
compelling art this way this is just one
example of that called learning to see
from an artist memo Lockton whose work
you should definitely check out if
you're interested in cool so thank you I
hope this lecture has given you a good
idea of the broad array of things that
people are doing with Gans and I hope
this might even inspire you to look
further into some of these applications
or try some new applications of your own
thanks
you |
edfd5390-ff33-4f2a-bde3-a22598c903c7 | trentmkelly/LessWrong-43k | LessWrong | What does Bing Chat tell us about AI risk?
Image from here via this tweet
ICYMI, Microsoft has released a beta version of an AI chatbot called “the new Bing” with both impressive capabilities and some scary behavior. (I don’t have access. I’m going off of tweets and articles.)
Zvi Mowshowitz lists examples here - highly recommended. Bing has threatened users, called them liars, insisted it was in love with one (and argued back when he said he loved his wife), and much more.
Are these the first signs of the risks I’ve written about? I’m not sure, but I’d say yes and no.
Let’s start with the “no” side.
* My understanding of how Bing Chat was trained probably does not leave much room for the kinds of issues I address here. My best guess at why Bing Chat does some of these weird things is closer to “It’s acting out a kind of story it’s seen before” than to “It has developed its own goals due to ambitious, trial-and-error based development.” (Although “acting out a story” could be dangerous too!)
* My (zero-inside-info) best guess at why Bing Chat acts so much weirder than ChatGPT is in line with Gwern’s guess here. To oversimplify, there’s a particular type of training that seems to make a chatbot generally more polite and cooperative and less prone to disturbing content, and it’s possible that Bing Chat incorporated less of this than ChatGPT. This could be straightforward to fix.
* Bing Chat does not (even remotely) seem to pose a risk of global catastrophe itself.
On the other hand, there is a broader point that I think Bing Chat illustrates nicely: companies are racing to build bigger and bigger “digital brains” while having very little idea what’s going on inside those “brains.” The very fact that this situation is so unclear - that there’s been no clear explanation of why Bing Chat is behaving the way it is - seems central, and disturbing.
AI systems like this are (to simplify) designed something like this: “Show the AI a lot of words from the Internet; have it predict the next word it will see, a |
e6205c63-c44e-4bbf-9e1e-01d7fad0c856 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Washington, D.C.: Book Swap (Postponed Due to Weather)
Discussion article for the meetup : Washington, D.C.: Book Swap (Postponed Due to Weather)
WHEN: 22 February 2015 03:00:00PM (-0500)
WHERE: Reynolds Center
Due to the inclement weather last weekend, meetups are being pushed back a week.
We will be meeting in the Kogod Courtyard of the Donald W. Reynolds Center for American Art and Portraiture (8th and F Sts or 8th and G Sts NW, go straight past the information desk from either entrance) to lend, borrow, and discuss books brought by attendees. As usual, we will congregate between 3:00 and 3:30 and begin at 3:30; when everyone is done talking books, the meetup will revert to unstructured discussion.
As before, a couple notes:
* There is no obligation to lend, borrow, or return books at book swap meetups. You are not required to bring any given book, or any books at all; you are not required to lend the books that you do bring; if you wish to lend out a given book, you are not required to lend it out to the first person to ask (or to the next person to ask, or to anyone at all); you are not required to borrow a book that is offered to you; and if you have already borrowed a book, you are not required to return it at this meetup. This is a Schelling point for book swaps, nothing more.
* Books belong to their owners - keep any book you borrow in good condition, and return it in a timely fashion when you finish reading it or when the owner asks for it back.
* In the past, the owners of some books have given permission for borrowers to lend them directly to anyone at the meetup who is interested. If you do not have permission from the owner of the book, do not do this.
As always, anyone not interested in the official topic is allowed and encouraged to start side conversations.
Bethesda Station will be closed due to an escalator installation; no other rail advisories have been announced as of February 16th. The Verizon Center schedule for February 22 has a concert starting at 7:30 p.m., with doors open at 6:30.
|
4c98d2e4-3015-4b3c-9fd1-e9fac447d392 | trentmkelly/LessWrong-43k | LessWrong | Acausal trade: universal utility, or selling non-existence insurance too late
A putative new idea for AI control; index here.
Other posts in the series: Introduction, Double decrease, Pre-existence deals, Full decision algorithms, Breaking acausal trade, Trade in different types of utility functions, Being unusual, and Summary.
We're going to have to look at the issue of universal, pre-existence trade agreements - the idea that agents should abide by some sort of idealised trade agreement based on a universal prior of possible agent values, without updating on the fact that they happen to exist as agents.
The arguments for it seem superficially similar to the arguments for preferring FDT/UDT/TDT over CDT, but are actually quite different. In brief, there is no profit, advantage, or benefit for an existing agent to commit to a trade agreement with non-existing agents.
Choosing a better decision theory
Newcomb, as usual
Causal decision theory gets the Newcomb problem wrong.
There are many ways to see that CDT gets this wrong (for a start, it ends up with less money that other agents), but one of the most damning is that CDT is not stable - it will self-modify into a sort of timeless agent, if it can do so, precisely to triumph on Newcomb-like problems.
It does this in a clunky and odd way: it will one-box on Newcomb problems where Omega does its prediction after learning of CDT's change, but not on Newcomb problems where Omega made its decision in the past.
The reason for it's clunkness is that CDT, being causal, doesn't allow for correlations between its own decision and causally anterior events.
Functional decision theory, which has a richer theory of correlations, would spot that correlation, and, assuming it had somehow ended up with a CDT-like decision module, would immediately change that into an FDT-compatible one that would one-box on all Newcomb problems.
I had another argument for the weakness of CDT (and some forms of UDT), namely that it shouldn't make a difference whether an agent was simulated or just predicted in a no |
3a4a5868-ca22-4281-8020-f221a9f12d22 | trentmkelly/LessWrong-43k | LessWrong | Google DeepMind: An Approach to Technical AGI Safety and Security
We have written a paper on our approach to technical AGI safety and security. This post is primarily a copy of the extended abstract, which summarizes the paper. I also include the abstract and the table of contents. See also the GDM blogpost and tweet thread.
Artificial General Intelligence (AGI) promises transformative benefits but also presents significant risks. We develop an approach to address the risk of harms consequential enough to significantly harm humanity. We identify four areas of risk: misuse, misalignment, mistakes, and structural risks. Of these, we focus on technical approaches to misuse and misalignment. For misuse, our strategy aims to prevent threat actors from accessing dangerous capabilities, by proactively identifying dangerous capabilities, and implementing robust security, access restrictions, monitoring, and model safety mitigations. To address misalignment, we outline two lines of defense. First, model-level mitigations such as amplified oversight and robust training can help to build an aligned model. Second, system-level security measures such as monitoring and access control can mitigate harm even if the model is misaligned. Techniques from interpretability, uncertainty estimation, and safer design patterns can enhance the effectiveness of these mitigations. Finally, we briefly outline how these ingredients could be combined to produce safety cases for AGI systems.
Extended Abstract
AI, and particularly AGI, will be a transformative technology. As with any transformative technology, AGI will provide significant benefits while posing significant risks. This includes risks of severe harm: incidents consequential enough to significantly harm humanity. This paper outlines our approach to building AGI that avoids severe harm.[1]
Since AGI safety research is advancing quickly, our approach should be taken as exploratory. We expect it to evolve in tandem with the AI ecosystem to incorporate new ideas and evidence.
Severe harms nec |
f8dad5a6-6d3b-428e-8b44-c232c457ab7b | trentmkelly/LessWrong-43k | LessWrong | How to know whether you are an idealist or a physicalist/materialist
You have heard and perhaps even used the expression "observable universe", right? What is included in the purportedly observable universe? The moon? The whole of the moon? If you had heard the expression "observable universe" a century ago, would you have been including the far side of the moon in that category? |
6980fb77-6aa3-4220-a337-0eab02b3a29a | trentmkelly/LessWrong-43k | LessWrong | LW-ish meetup in Boulder, CO
This Saturday I'm giving a presentation at the Boulder Future Salon, topic will be non-religious spirituality. The more LWians that can make it the better, because I'm really trying to get some community building done in the Boulder/Denver area. There's an insane amount of potential here.
Details. |
6e1970d3-5220-408a-8261-7c6a085f8062 | trentmkelly/LessWrong-43k | LessWrong | Project Ideas for the London Hackday
So, the London community is arranging a Hackday where some of us will get together and code. In order to ensure we work on the awesomest idea(s) possible, we decided to ask LessWrong to add to our list of candidates. So here is the question:
What could a few developers do in a day or less worth of coding that will be awesome? Also, as a way of checking calibration, you can give your estimate for how long such a thing would take to build.
Note: While we will take ideas and voting here into account, there is no guarantee that we will actually end up choosing one or more of them. |
30596fe0-c88f-4576-a91f-5fef2c25dae3 | trentmkelly/LessWrong-43k | LessWrong | The Ethics of Eating Seafood: A Rational Discussion
Introduction
The ethics of eating seafood is an increasingly pertinent topic in the realm of animal welfare and environmental sustainability. As rationalists, it is essential to examine the impact of our dietary choices on sentient beings and the planet. This post aims to provide an objective analysis of the ethical considerations surrounding seafood consumption, by addressing the sentience of marine life, the environmental impact, and potential alternatives.
I. Sentience and Suffering
The first ethical concern regarding seafood consumption is the capacity of marine life to experience suffering. Some studies suggest that fish and crustaceans possess nociceptors, which are sensory receptors for pain, indicating that they might be capable of experiencing suffering (Sneddon et al., 2014). However, the extent of their suffering remains a subject of debate.
As rationalists, it is crucial to consider the uncertainty surrounding marine animal sentience and err on the side of caution. One way to do this is to adopt the principle of "expected moral value," which implies that we should consider both the probability of an animal being sentient and the magnitude of its potential suffering when making dietary decisions (MacAskill, 2014).
II. Environmental Impact
Another significant ethical aspect of seafood consumption is its impact on the environment. Overfishing and destructive fishing practices have led to the depletion of fish stocks, biodiversity loss, and damage to marine ecosystems (Pauly et al., 2002).
Additionally, fish farming or aquaculture, intended to address the issue of overfishing, has its own environmental concerns. These include water pollution, habitat destruction, and the use of large quantities of wild-caught fish as feed for farmed species (Naylor et al., 2000).
As responsible consumers, we should be aware of these environmental consequences and consider the ecological footprint of our seafood choices. Opting for sustainably sourced seafood and supp |
a38765f0-967e-48c4-bcc6-981d2270e767 | trentmkelly/LessWrong-43k | LessWrong | The Argument from Witness Testimony
(Note: This is essentially a rehash/summarization of Jordan Sobel's Lotteries and Miracles - you may prefer the original.)
George Mavrodes wrote an interesting analogy. Scenario 1: Suppose you read a newspaper report claiming that a particular individual (say, Henry Plushbottom of Topeka, Kansas) has won a very large lottery. Before reading the newspaper, you would have given quite low odds that Henry in particular had won the lottery. However, the newspaper report flips your beliefs quite drastically. Afterward, you would give quite high odds that Henry in particular had won the lottery. Scenario 2: You have read various claims that a particular individual (Jesus of Nazareth) arose from the dead. Before hearing those claims, you would have given quite low odds of anything so unlikely happening. However (since you are reading LessWrong) you presumably do not give quite high odds that Jesus arose from the dead.
What is it about the second scenario which makes it different from the first?
Let's model Scenario 1 as a simple Bayes net. There are two nodes, one representing whether Henry wins, and one representing whether Henry is reported to win, and one arrow, from first to the second.
What are the parameters of the conditional probability tables? Before any information came in, it seemed very unlikely that Henry was the winner - perhaps he had a one in a million chance. Given that Henry did win, what is the chance that he would be reported to have won? Pretty likely - newspapers do err, but it's reasonable to believe that 9 times out of 10, they get the name of the lottery winner correct. Now suppose that Henry didn't win. What is the chance that he would be reported to have won by mistake? There's nothing in particular to single him out from the other non-winners - being misreported is just as unlikely as winning, maybe even more unlikely.
So we have (using w to abbreviate "Henry Wins" and r to abbreviate "Henry is reported"):
* P(w)=10-6 - Henry has a one |
78d39fc9-8624-4047-8c7c-55cbe566779c | trentmkelly/LessWrong-43k | LessWrong | ToL: Introduction
(These are the touched up notes from a class I took with CMU's Kevin Kelly this past semester on the Topology of Learning. Only partially optimized for legibility)
Feasibility Contextualism
A large amount of philosophy has been people trying to demonstrate how you can never really know anything [citation needed]. Various forms of skepticism take the stance, "No method of inquiry get get you Real KnowledgeTM, and so no method is justified."
Despite that, it seems like (at least in terms of watching how people act in the world) everyone gets that you need some level of pragmatism. "Well I've gotta do something, and this seems like the best idea, so I'm going to do it instead of doing nothing." No one is so skeptical of knowledge that they have stayed immobile until they starved to death [citation needed].
What Kelly aims to do is create a rigorous formalism for the sort of pragmatic attitude that people take all the time. It's very much inspired by how computer scientists do things. If someone proves that a problem can't be done any faster than quadratic time, and you figure out a quadratic time algorithm, you're happy. You don't refuse to use any algorithm that doesn't run in constant time.
In a sentence, this course was about how really cool formalism to talk about "How hard is a given scientific problem" and how that effects "the best possible performance you can get given the hardness of the problem".
Sneak Peak: Induction and Metaphysics
Over the ages people have postulated what qualities scientific hypothesis should have. The logical positivists asserted that only verifiable propositions should be the domain of science (if it's true, you can do some test to demonstrate it's true). Popper wanted hypothesis to be falsifiable (if it's false, you can do some test to demonstrate it's false). Verification and falsifiability have an important connection to two other notions that philosophers of science often talk about, the problem of induction and the problem o |
475b65d0-8e0b-4d49-9095-88d4af277ad0 | trentmkelly/LessWrong-43k | LessWrong | Summary: "Imagining and building wise machines: The centrality of AI metacognition" by Johnson, Karimi, Bengio, et al.
Details:
Original paper
Paper Authors:
Samuel G. B. Johnson, Amir-Hossein Karimi, Yoshua Bengio, Nick Chater, Tobias Gerstenberg, Kate Larson, Sydney Levine, Melanie Mitchell, Iyad Rahwan, Bernhard Schölkopf, Igor Grossmann
Abstract
Recent advances in artificial intelligence (AI) have produced systems capable of increasingly sophisticated performance on cognitive tasks. However, AI systems still struggle in critical ways: unpredictable and novel environments (robustness), lack of transparency in their reasoning (explainability), challenges in communication and commitment (cooperation), and risks due to potential harmful actions (safety). We argue that these shortcomings stem from one overarching failure: AI systems lack wisdom.
Drawing from cognitive and social sciences, we define wisdom as the ability to navigate intractable problems - those that are ambiguous, radically uncertain, novel, chaotic, or computationally explosive - through effective task-level and metacognitive strategies. While AI research has focused on task-level strategies, metacognition - the ability to reflect on and regulate one's thought processes - is underdeveloped in AI systems. In humans, metacognitive strategies such as recognizing the limits of one's knowledge, considering diverse perspectives, and adapting to context are essential for wise decision-making. We propose that integrating metacognitive capabilities into AI systems is crucial for enhancing their robustness, explainability, cooperation, and safety.
By focusing on developing wise AI, we suggest an alternative to aligning AI with specific human values - a task fraught with conceptual and practical difficulties. Instead, wise AI systems can thoughtfully navigate complex situations, account for diverse human values, and avoid harmful actions. We discuss potential approaches to building wise AI, including benchmarking metacognitive abilities and training AI systems to employ wise reasoning. Prioritizing metacognition in AI re |
005790e6-085e-4ca7-925e-28389d7fabcd | trentmkelly/LessWrong-43k | LessWrong | Fun with bitstrings and bijections
Here's a quick math/coding puzzle with a cute solution. |
776cf4c4-d36b-454c-a30d-4d619f6c0d97 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Metareasoning for Planning Under Uncertainty
1 Introduction
---------------
Offline probabilistic planning approaches, such as policy iteration [[Howard1960](#bib.bibx15)], aim to construct a policy for every possible state before acting. In contrast, online planners, such as RTDP [[Barto et al.1995](#bib.bibx1)] and UCT [[Kocsis and
Szepesvári2006](#bib.bibx18)], interleave planning with execution. After an agent takes an action and moves to a new state, these planners suspend execution to plan for the next step. The more planning time they have, the better their action choices. Unfortunately, planning time in online settings is usually not free.
Consider an autonomous Mars rover trying to decide what to do while a sandstorm is nearing. The size and uncertainty of the domain precludes a-priori computation of a complete policy, and demands the use of online planning algorithms. Normally, the longer the rover runs its planning algorithm, the better decision it can make. However, computation costs power; moreover, if it reasons for too long without taking preventive action, it risks being damaged by the oncoming sandstorm. Or consider a space probe on final approach to a speeding comet, when the probe must plan to ensure a safe landing based on new information it gets about the comet’s surface. More deliberation time means a safer landing. At the same time, if the probe deliberates for too long, the comet may zoom out of range — a similarly undesirable outcome.
Scenarios like these give rise to a general *metareasoning decision problem*: how should an agent trade off the cost of planning
and the quality of the resulting policy for the base planning
task every time it needs to make a move, so as to optimize
its long-term utility? Metareasoning about base-level problem solving has been explored for probabilistic inference and decision making [[Horvitz1987](#bib.bibx13), [Horvitz et al.1989](#bib.bibx11)], theorem proving [[Horvitz and Klein1995](#bib.bibx10), [Kautz et al.2002](#bib.bibx16)], handling streams of problems [[Horvitz2001](#bib.bibx14), [Shahaf and Horvitz2009](#bib.bibx22)], and search [[Russell and Wefald1991](#bib.bibx21), [Burns et al.2013](#bib.bibx4)]. There has been little work exploring generalized approaches to metareasoning for planning.
We explore the general metareasoning problem for Markov decision processes (MDPs). We begin by formalizing the problem with a general but precise definition that subsumes several previously considered metareasoning models. Then, we show with a rigorous theoretical analysis that optimal general metareasoning for planning under uncertainty is at most polynomially harder than solving the original planning problem with any given MDP solver. However, this increase in computational complexity, among other reasons we discuss, renders such optimal general metareasoning impractical. The analysis raises the issue of allocating time for *metareasoning itself*, and leads to an infinite regress of meta∗reasoning (metareasoning, metametareasoning, etc.) problems.
We next turn to the development and testing of fast approximate metareasoning algorithms. Our procedures use the Bounded RTDP (BRTDP [[McMahan et al.2005](#bib.bibx20)]) algorithm to tackle the base MDP problem, and leverage BRTDP-computed bounds on the quality of MDP policies to reason about the value of computation. In contrast to prior work on this topic, our methods do not require any training data, precomputation, or prior information about target domains. We perform a set of experiments showing the performance of these algorithms versus baselines in several synthetic domains with different properties, and characterize their performance with a measure that we call the *metareasoning gap* — a measure of the potential for improvement from metareasoning. The experiments demonstrate that the proposed techniques excel when the metareasoning gap is large.
2 Related Work
---------------
Metareasoning efforts to date have employed strategies that avoid the complexity of the general metareasoning problem for planning via relying on different kinds of simplifications and approximations. Such prior studies include metareasoning for time-critical decisions where expected value of computation is used to guide probabilistic inference [[Horvitz1987](#bib.bibx13), [Horvitz et al.1989](#bib.bibx11)], and work on the guiding of sequences of single actions in search [[Russell and Wefald1991](#bib.bibx21), [Burns et al.2013](#bib.bibx4)]. Several lines of work have leveraged offline learning [[Breese and Horvitz1990](#bib.bibx3), [Horvitz et al.2001](#bib.bibx12), [Kautz et al.2002](#bib.bibx16)]. Other studies have relied on optimizations and inferences that leverage the structure of problems, such as the functional relationships between metareasoning and reasoning [[Horvitz and Breese1990](#bib.bibx9), [Zilberstein and
Russell1996](#bib.bibx25)], the structure of the problem space [[Horvitz and Klein1995](#bib.bibx10)], and the structure of utility [[Horvitz2001](#bib.bibx14)]. In other work, [[Hansen and Zilberstein2001](#bib.bibx7)] proposed a non-myopic dynamic programming solution for single-shot problems. Finally, several planners rely on a heuristic form of online metareasoning when maximizing policy reward under computational constraints in real-world time with no “conversion rate” between the two [[Kolobov et al.2012](#bib.bibx19), [Keller and Geißer2015](#bib.bibx17)]. In contrast, our metareasoning model is unconstrained, with computational and base-MDP costs in the same “currency.”
Our investigation also has connections to research on allocating time in a system composed of multiple sensing and planning components [[Zilberstein and
Russell1996](#bib.bibx25), [Zilberstein and
Russell1993](#bib.bibx24)], on optimizing portfolios of planning strategies in scheduling applications [[Dean et al.1995](#bib.bibx6)], and on choosing actions to explore in Monte Carlo planning [[Hay et al.2012](#bib.bibx8)]. In other related work, [[Chanel et al.2014](#bib.bibx5)] consider how best to plan on one thread, while a separate thread processes execution.
3 Preliminaries
----------------
A key contribution of our work is formalizing the metareasoning problem for planning under uncertainty. We build on the framework of stochastic shortest path (SSP) MDPs with a known start state. This general MDP class includes finite-horizon and discounted-reward MDPs as special cases [[Bertsekas and
Tsitsiklis1996](#bib.bibx2)], and can also be used to approximate partially observable MDPs with a fixed initial belief state. An SSP MDP M is a tuple ⟨S,A,T,C,s0,sg⟩, where S is a finite set of states, A is a set of actions that the agent can take,
T:(S,A,S)→[0,1] is a transition function, C:(S,A)→R is a cost function, s0∈S is the start state, and sg is the goal state. An SSP MDP must have a complete proper policy, a policy that leads to the goal from any state with probability 1, and all improper policies must accumulate infinite cost from every state from which they fail to reach the goal with a positive probability. The objective is to find a Markovian policy πs0:S→A with the minimum expected cost of reaching the goal from the start state s0 — in SSP MDPs, at least one policy of this form is globally optimal.
Without loss of generality, we assume an SSP MDP to have a specially designated NOP (“no-operation”) action. NOP is an action the agent chooses when it wants to “idle” and “think/plan”, and its semantic meaning is problem-dependent. For example, in some MDPs, choosing NOP means staying in the current state for one time step, while in others it may mean allowing a tidal wave to carry the agent to another state. Designating an action as NOP does not change SSP MDPs’ mathematical properties, but plays a crucial role in our metareasoning formalization.
4 Formalization and Theoretical Analysis of Metareasoning for MDPs
-------------------------------------------------------------------
The online planning problem of an agent, which involves choosing an action to execute in any given state, is represented as an SSP MDP that encapsulates the dynamics of the environment and costs of acting and thinking. We call this problem the *base problem*. The agent starts off in this environment with some default policy, which can be as simple as random or guided by an unsophisticated heuristic. The agent’s metareasoning problem, then, amounts to deciding, at every step during its interaction with the environment, between improving its existing policy or using this policy’s recommended action while paying a cost for executing *either of these options*, so as to minimize its expected cost of getting to the goal.
Besides the agent’s state in the base MDP, which we call the *world state*, the agent’s metareasoning decisions are conditioned on the algorithm the agent uses for solving the base problem, i.e., intuitively, on the agent’s thinking process. To abstract away the specifics of this planning algorithm for the purposes of metareasoning formalization, we view it as a black-box MDP solver and represent it, following the Church-Turing thesis, with a Turing machine B that takes a base SSP MDP M as input.
In our analysis, we assume the following about Turing machine B’s operation:
* B is deterministic and halts on every valid base MDP M. This assumption does not affect the expressiveness of our model, since randomized Turing machines can be trivially simulated on deterministic ones, e.g., via seed enumeration (although potentially at an exponential increase in time complexity). At the same time, it greatly simplifies our theorems.
* An agent’s thinking cycle corresponds to B executing a single instruction.
* A *configuration* of B is a combination of B’s tape contents, state register contents, head position, and next input symbol. It represents the state of the online planner in solving the base problem M. We denote the set of all configurations B ever enters on a given input MDP M as XB(M). We assume that B can be paused after executing y instructions, and that its configuration at that point can be mapped to an action for any world state s of M using a special function f:S×XB(M)→A in time polynomial in M’s flat representation. The number of instructions needed to compute f is not counted into y. That is, an agent can stop thinking at any point and obtain a policy for its current world state.
* An agent is allowed to “think” (i.e., execute B’s instructions) only by choosing the NOP action. If an agent decides to resume thinking after pausing B and executing a few actions, B re-starts from the configuration in which it was last paused.
We can now define metareasoning precisely:
######
Definition 1.
Metareasoning Problem. Consider an MDP M=⟨S,A,T,C,s0,sg⟩ and an SSP MDP solver represented by a deterministic Turing machine B. Let XB(M) be the set of all configurations B enters on input M, and let TB(M):XB(M)×XB(M)→{0,1} be the (deterministic) transition function of B on XB(M). A metareasoning problem for M with respect to B, denoted MetaB(M) is an MDP ⟨Sm,Am,Tm,Cm,sm0,smg⟩ s.t.
* Sm=S×XB(M)
* Am=A
* Tm((s,χ),a,(s′,χ′))
| | | |
| --- | --- | --- |
| | =⎧⎪⎨⎪⎩T(s,a,s′)if a≠ {NOP}, χ=χ′, and a=f(s,χ)T(s,a,s′)⋅TB(M)(χ,χ′)if a= {NOP}0 otherwise | |
* Cm((s,χ),a,(s′,χ′))=C(s,a,s′) if T(s,a,s′)≠0, and 0 otherwise
* sm0=(s0,χ0), where χ0 is the first configuration B enters on input M
* smg=(sg,χ), where χ is any configuration in XB(M)
Solving the metareasoning problem means finding a policy for MetaB(M) with the lowest expected cost of reaching smg.
This definition casts a metareasoning problem for a base MDP as another MDP (a *meta-MDP*). Note that in MetaB(M), an agent *must* choose either NOP or an action currently recommended by B(M); in other cases, the transition probability is 0. Thus, MetaB(M)’s definition essentially forces an agent to switch between two “meta-actions”: thinking or acting in accordance with the current policy.
Modeling an agent’s reasoning process with a Turing machine allows us to see that at every time step the metareasoning decision depends on the combination of the current world state and the agent’s “state of mind,” as captured by the Turing machine’s current configuration. In principle, this decision could depend on the entire history of the two, but the following theorem implies that, as for M, at least one optimal policy for MetaB(M) is always Markovian.
######
Theorem 1.
If the base MDP M is an SSP MDP, then MetaB(M) is an SSP MDP as well, provided that B halts on M with a proper policy. If the base MDP M is an infinite-horizon discounted-reward MDP, then so is MetaB(M). If the base MDP M is a finite-horizon MDP, then so is MetaB(M).
###### Proof.
Verifying the result for finite-horizon and infinite-horizon discounted-reward MDPs M is trivial, since the only requirement MetaB(M) must satisfy in these cases is to have a finite horizon or a discount factor, respectively.
If M is an SSP MDP, then, per the SSP MDP definition [[Bertsekas and
Tsitsiklis1996](#bib.bibx2)], to ascertain the theorem’s claim we need to verify that (1) MetaB(M) has at least one proper policy and (2) every improper policy in MetaB(M) accumulates an infinite cost from some state.
To see why (1) is true, recall that MetaB(M)’s state space is formed by all configurations Turing machine B enters on M. Consider any state (s′0,χ′0) of MetaB(M). Since B is deterministic, as stated in Section 3, the configuration χ′0 lies in the linear sequence of configurations between the “designated” initial configuration χ0 and the final proper-policy configuration that B enters according to the theorem. Thus, B can reach a proper-policy configuration from χ′0. Therefore, let the agent starting in the state (s′0,χ′0) of MetaB(M) choose NOP until B halts, and then follow the proper policy corresponding to B’s final configuration until it reaches a goal state sg of M. This state corresponds to a goal state (sg,χ) of MetaB(M). Since this construct works for any (s′0,χ′0), it gives a complete proper policy for MetaB(M).
To verify (2), consider any policy πm for MetaB(M) that with a positive probability fails to reach the goal. Any infinite trajectory of πm that fails to reach the goal can be mapped onto a trajectory in M that repeats the action choices of πm’s trajectory in M’s state space S. Since M is an SSP MDP, this projected trajectory must accumulate an infinite cost, and therefore the original trajectory in MetaB(M) must do so as well, implying the desired result.
∎
We now present two results to address the difficulty of metareasoning.
######
Theorem 2.
For an SSP MDP M and a deterministic Turing machine B representing a solver for M, the time complexity of MetaB(M) is at most polynomial in the time complexity of executing B on M.
###### Proof.
The main idea is to construct the MDP representing MetaB(M) by simulating B on M. Namely, we can run B on M until it halts and record every configuration B enters to obtain the set X. Given X, we can construct Sm=S×X and all other components of MetaB(M) in time polynomial in |X| and |M|. Constructing X itself takes time proportional to running time of B on M. Since, by Theorem [1](#Thmtheorem1 "Theorem 1. ‣ 4 Formalization and Theoretical Analysis of Metareasoning for MDPs ‣ Metareasoning for Planning Under Uncertainty"), MetaB(M) is an SSP MDP and hence can be solved in time polynomial in the size of its components, e.g., by linear programming, the result follows.
∎
######
Theorem 3.
Metareasoning for SSP MDPs is P-complete under NC-reduction. (Please see the appendix for proof.)
At first glance, the results above look encouraging. However, upon closer inspection they reveal several subtleties making optimal metareasoning utterly impractical. First, although both SSP MDPs and their metareasoning counterparts with respect to an optimal polynomial-time solver are in P, doing metareasoning for a given MDP M is appreciably more expensive than solving *that MDP itself*. Given that the additional complexity due to metareasoning cannot be ignored, the agent now faces the new challenge of allocating computational time between metareasoning and planning for the base problem. This challenge is a meta-metareasoning problem, and ultimately causes infinite regress, an unbounded nested sequence of ever-costlier reasoning problems.
Second, constructing MetaB(M) by running B on M, as the proof of Theorem [2](#Thmtheorem2 "Theorem 2. ‣ 4 Formalization and Theoretical Analysis of Metareasoning for MDPs ‣ Metareasoning for Planning Under Uncertainty") proceeds, may entail solving M in the process of metareasoning. While the proof doesn’t show that this is the only way of constructing MetaB(M), without making additional assumptions about B’s operation one cannot exclude the possibility of having to run B until convergence and thereby completely solving M even before MetaB(M) is fully formulated. Such a construction would defeat the purpose of metareasoning.
Third, the validity of Theorems [2](#Thmtheorem2 "Theorem 2. ‣ 4 Formalization and Theoretical Analysis of Metareasoning for MDPs ‣ Metareasoning for Planning Under Uncertainty") and [3](#Thmtheorem3 "Theorem 3. ‣ 4 Formalization and Theoretical Analysis of Metareasoning for MDPs ‣ Metareasoning for Planning Under Uncertainty") relies on an implicit crucial assumption that the transitions of solver B on the base MDP M are known in advance. Without this knowledge, MetaB(M) turns into a reinforcement learning problem [[Sutton and Barto1998](#bib.bibx23)], which further increases the complexity of metareasoning and the need for simulating B on M. Neither of these is viable in reality.
The difficulties with optimal metareasoning motivate the development of approximation procedures. In this regard, the preceding analysis provides two important insights. It suggests that, since running B on M until halting is infeasible, it may be worth trying to *predict* B’s progress on M. Many existing MDP algorithms have clear operational patterns, e.g., evaluating policies in the decreasing order of their cost, as policy iteration does [[Howard1960](#bib.bibx15)]. Regularities like these can be of value in forecasting the benefit of running B on M for additional cycles of thinking. We now focus on exploring approximation schemes that can leverage these patterns.
5 Algorithms for Approximate Metareasoning
-------------------------------------------
Our approach to metareasoning is guided by *value of computation* (VOC) analysis. In contrast to previous work that formulates VOC for single actions or decision-making problems [[Horvitz1987](#bib.bibx13), [Horvitz et al.1989](#bib.bibx11), [Russell and Wefald1991](#bib.bibx21)], we aim to formulate VOC for online planning. For a given metareasoning problem MetaB(M), VOC at any encountered state sm=(s,χ) is exactly the difference between the Q-value of the agent following f(s,χ) (the action recommended by the current policy of the base MDP M) and the Q-value of the agent taking NOP and thinking:
| | | | |
| --- | --- | --- | --- |
| | VOC(sm)=Q∗(sm,f(s,χ))−Q∗(sm,NOP). | | (1) |
VOC captures the difference in long-term utility between thinking and acting as determined by these Q-values. An agent should take the NOP action and think when the VOC is positive. Our technique aims to evaluate VOC by estimating Q∗(sm,f(s,χ)) and Q∗(sm,NOP). However, attempting to estimate these terms in a near-optimal manner ultimately runs into the same difficulties as solving MetaB(M), such as simulating the agent’s thinking process many steps into the future, and is likely infeasible. Therefore, fast approximations for the Q-values will generally have to rely on simplifying assumptions.
We rely on performing greedy metareasoning analysis as has been done in past studies of metareasoning [[Horvitz et al.1989](#bib.bibx11), [Russell and Wefald1991](#bib.bibx21)]:
*Meta-Myopic Assumption. In any state sm of the meta-MDP, we assume that after the current step, the agent will never again choose NOP, and hence will never change its policy.*
This meta-myopic assumption is important in allowing us to reduce VOC estimation to predicting the improvement in the value of the base MDP policy following a single thinking step. The weakness of this assumption is that opportunities for subsequent policy improvements are overlooked. In other words, the VOC computation only reasons about the current thinking opportunity. Nonetheless, in practice, we compute VOC at every timestep, so the agent *can* still think later. Our experiments show that our algorithms perform well in spite of their meta-myopicity.
###
5.1 Implementing Metareasoning with BRTDP
We begin the presentation of our approximation scheme with the selection of B, the agent’s thinking algorithm. Since approximating Q∗(sm,f(s,χ)) and Q∗(sm,NOP) essentially amounts to assessing policy values, we would like an online planning algorithm that provides efficient policy value approximations, preferably with some guarantees. Having access to these policy value approximations enables us to design approximate metareasoning algorithms that can evaluate VOC efficiently in a domain-independent fashion.
One algorithm with this property is Bounded RTDP (BRTDP) [[McMahan et al.2005](#bib.bibx20)]. It is an anytime planning algorithm based on RTDP [[Barto et al.1995](#bib.bibx1)]. Like RTDP, BRTDP maintains a lower bound on an MDP’s optimal value function V∗, which is repeatedly updated via Bellman backups as BRTDP simulates trials/rollouts to the goal, making BRTDP’s configuration-to-configuration transition function TB(M)(χ,χ′) stochastic. A key difference is that in addition to maintaining a lower bound, it also maintains an upper bound, updated in the same conceptual way as the lower one. If BRTDP is initialized with a *monotone* upper-bound heuristic, then the upper-bound decreases monotonically as BRTDP runs. The construction of domain-independent monotone bounds is beyond the scope of this paper, but is easy for the domains we study in our experiments. Another key difference between BRTDP and RTDP is that if BRTDP is stopped before convergence, it returns an action greedy with respect to the upper, not lower bound. This behavior guarantees that the expected cost of a policy returned at any time by a monotonically-initialized BRTDP is no worse than BRTDP’s current upper bound. Our metareasoning algorithms utilize these properties to estimate VOC. In the rest of the discussion, we assume that BRTDP is initialized with a monotone upper-bound heuristic.
###
5.2 Approximating VOC
We now show how BRTDP’s properties help us with estimating the two terms in the definition of VOC, Q∗(sm,f(s,χ)) and Q∗(sm,NOP). We first assume that one “thinking cycle” of BRTDP (i.e., executing NOP once and running BRTDP in the meantime, resulting in a transition from BRTDP’s current configuration χ to another configuration χ′) corresponds to completing some fixed number of BRTDP trials from the agent’s current world state s.
####
5.2.1 Estimating Q∗(sm,NOP)
We first describe how to estimate the value of taking the NOP action (thinking). At the highest level, this estimation first involves writing down an expression for Q∗(sm,NOP), making a series of approximations for different terms, and then modeling the behavior of how BRTDP’s upper bounds on the Q-value function drop in order to compute the needed quantities.
When opting to think by choosing NOP, the agent may transition to a different world state while *simultaneously* updating its policy for the base problem. Therefore, we can express Q∗(sm,NOP)=
| | | | |
| --- | --- | --- | --- |
| | ∑s′T(s,NOP,s′)∑χ′TB(M)(χ,χ′)V∗((s′,χ′)). | | (2) |
Because of meta-myopicity, we have V∗((s′,χ′))=Vχ′(s′) where Vχ′ is the value function of the policy corresponding to χ′ in the base MDP. However, this expression cannot be efficiently evaluated in practice, since we do not know BRTDP’s transition distribution TB(M)(χ,χ′) nor the state values Vχ′(s′), forcing us to make further approximations. To do so, we assume Vχ′ and Qχ′ are random variables, and rewrite ∑χ′TB(M)(χ,χ′)Vχ′(s′) =
| | | | |
| --- | --- | --- | --- |
| | ∑aP(Aχ′s′=a)E[Qχ′(s′,a)|Aχ′s′=a]. | | (3) |
where the random variable Aχ′s′ takes value a iff f(s′,χ′)=a after one thinking cycle in state (s,χ). Intuitively, P(Aχ′s′=a) denotes the probability that BRTDP will recommend action a in state s′ after one thinking cycle.
Now, let us denote the Q-value upper bound corresponding to BRTDP’s current configuration χ as ¯¯¯¯Qχ. This value is *known*. Then, let the upper bound corresponding to BRTDP’s next configuration χ′, be ¯¯¯¯Qχ′. Because we do not know χ′, this value is *unknown*, and is a random variable. Because BRTDP selects actions greedily w.r.t. the upper bound, we follow this behavior and use the upper bound to estimate Q-value by assuming that Qχ′=¯¯¯¯Qχ′. Since the value of ¯¯¯¯Qχ′ is unknown at the time of the VOC computation, P(Aχ′s′=a) and E[¯¯¯¯Qχ′(s′,a)|Aχ′s′=a] are computed by integrating over the possible values of ¯¯¯¯Qχ′. We have that E[¯¯¯¯Qχ′(s′,a)|Aχ′s=a]=
| | | |
| --- | --- | --- |
| | ∫¯¯¯¯Qχ′(s′,a)¯¯¯¯Qχ′(s′,a)P(Aχ′s′=a|¯Qχ′(s′,a))P(¯¯¯¯Qχ′(s′,a))P(Aχ′s′=a), | |
and
P(Aχ′s′=a)=
| | | |
| --- | --- | --- |
| | ∫¯¯¯¯Qχ′(s′,a)P(¯¯¯¯Qχ′(s′,a))∏ai≠aP(¯¯¯¯Qχ′(s′,ai)>P(¯¯¯¯Qχ′(s′,a)). | |
Therefore, we must model the distribution that ¯¯¯¯Qχ′ is drawn from. We do so by modeling the change Δ¯¯¯¯Q=¯¯¯¯Qχ−¯¯¯¯Qχ′, due to a single BRTDP thinking cycle that corresponds to a transition from configuration χ to χ′. Since ¯¯¯¯Qχ is known and fixed, estimating a distribution over possible Δ¯¯¯¯Q gives us a distribution over ¯¯¯¯Qχ′.
Let ^Δ¯¯¯¯Qs,a be the change in ¯¯¯¯Qs,a resulting from the most recent thinking cycle for some state s and action a. We first assume that the change resulting from an additional cycle of thinking, Δ¯¯¯¯Qs,a, will be no larger than the last change, Δ¯¯¯¯Qs,a≤^Δ¯¯¯¯Qs,a. This assumption is reasonable, because we can expect the change in bounds to decrease as BRTDP converges to the true value function. Given this assumption, we must choose a distribution D over the interval [0,^Δ¯¯¯¯Qs,a] such that for the next thinking cycle, Δ¯¯¯¯Qs,a∼D. Figure [1](#S5.F1 "Figure 1 ‣ 5.2.1 Estimating Q*(sm,NOP) ‣ 5.2 Approximating VOC ‣ 5 Algorithms for Approximate Metareasoning ‣ Metareasoning for Planning Under Uncertainty")a illustrates these modeling assumptions for two hypothetical actions, a1 and a2.
One option is to make D uniform, so as to represent our poor knowledge about the next bound change. Then, computing P(Aχ′s′=a) involves evaluating an integral of a polynomial of degree O(|A|) (the product of |A|−1 CDF’s and 1 PDF), and computing E[¯¯¯¯Qχ′(s′,a)|Aχ′s=a] also entails evaluating an integral of degree O(|A|), and thus computing these quantities for all actions in a state can be computed in time O(|A|2). Since the overall goal of this subsection, approximating Q∗(sm,NOP), requires computing P(Aχ′s′=a) for all actions in all states where NOP may lead, assuming there are no more than K<<|A| such states, the complexity becomes O(K|A|2) for each state visited by the agent on its way to the goal.
A weakness of this approach is that the changes in the upper bounds for different actions are modeled independently. For example, if the upper bounds for two actions in a given state decrease by a large amount in the previous thinking step, then it is unlikely that in the next thinking step one of them will drop dramatically while the other drops very little. This independence can cause the amount of uncertainty in the upper bound at the next thinking step to be overestimated, leading to VOC being overestimated as well.
Therefore, we create another version of the algorithm assuming that the speed of decrease in Q-value upper bounds for all actions are perfectly correlated; all ratios between future drops in the next thinking cycle are equal to the ratios between the observed drops in the last thinking cycle. Formally, for a given state s, we let ρ∼ Uniform[0, 1]. Then, let Δ¯¯¯¯Qs,a=ρ⋅^Δ¯¯¯¯Qs,a for all actions a.
Now, to compute P(Aχ′s′=a), for each action a, we represent the range of its possible future Q-values ¯¯¯¯Qχ′s,a with a line segment la on the unit interval [0,1] where la(0)=¯¯¯¯Qχs,a and la(1)=¯¯¯¯Qχs,a−Δ¯¯¯¯Qs,a. Then, P(Aχ′s′=a) is simply the proportion of la which lies below all the other lines representing all other actions. We can naïvely compute these probabilities in time O(|A|2) by enumerating all intersections. Similarly, computing E[¯¯¯¯Qχ′(s′,a)|Aχ′s=a] is also easy. This value is the mean of the portion of la that is beneath all other lines. Figure [1](#S5.F1 "Figure 1 ‣ 5.2.1 Estimating Q*(sm,NOP) ‣ 5.2 Approximating VOC ‣ 5 Algorithms for Approximate Metareasoning ‣ Metareasoning for Planning Under Uncertainty")b illustrates these computations.
Figure 1: a) Hypothetical drops in upper bounds on the Q-values of two actions, a1 and a2. We assume the next Q-value drop resulting from another cycle of thinking, Δ¯¯¯¯Q, is drawn from a range equal to the last drop from thinking, ^Δ¯¯¯¯Q b) Assuming perfect correlation in the speed of decrease in the Q-value upper bounds, as the upper bounds of the two actions drop from an additional cycle of thinking, initially a2 has a better upper bound, but eventually a1 overtakes a2.
Whether or not we make the assumption of action independence, we further speed up the computations by only calculating E[¯¯¯¯Qχ′(s′,a)|Aχ′s=a] and P(Aχ′s′=a) for the two “most promising” actions a, those with the lowest expectation of potential upper bounds. This limits the computation time to the time required to determine these actions (linear in |A|), and makes the time complexity of estimating Q∗(sm,NOP) for one state s be O(K|A|) instead of O(K|A|2).
####
5.2.2 Estimating Q∗(sm,f(s,χ))
Now that we have described how to estimate the value of taking the NOP action, we describe how to estimate the value of taking the currently recommended action, f(s,χ). We estimate Q∗(sm,f(s,χ)) by computing E[¯¯¯¯Qχ′(s,f(s,χ))], which takes constant time, keeping the overall time complexity linear. The reason we estimate Q∗(sm,f(s,χ)) using future Q-value upper bound estimates based on a probabilistic projection of χ′, as opposed to our current Q-value upper bounds based on the current configuration χ, is to make use of the more informed bounds derived at the future utility estimation. As the BRTDP algorithm is given more computation time, it can more accurately estimate the upper bound of a policy. This type of approximation has been justified before [[Russell and Wefald1991](#bib.bibx21)]. In addition, using future utility estimates in both estimating Q∗(sm,f(s,χ)) and Q∗(sm,NOP) provides a consistency guarantee: *if thinking leads to no policy change, then our method estimates VOC to be zero*.
###
5.3 Putting It All Together
The core of our algorithms involves the computations we have described, in every state s the agent visits on the way to the goal. In the experiments, we denote UnCorr Metareasoner as the metareasoner that assumes the actions are uncorrelated, and Metareasoner as the metareasoner that does not make this assumption. To complete the algorithms, we ensure that they decide the agent should think for another cycle if ^Δ¯¯¯¯Qs,a isn’t yet available for the agent’s current world state s (e.g., because BRTDP has never updated bounds for this state’s Q-value so far), since VOC computation is not possible without prior observations on ^Δ¯¯¯¯Qs,a. Crucially, all our estimates make metareasoning take time only linear in the number of actions, O(K|A|), per visited state.
6 Experiments
--------------
We evaluate our metareasoning algorithms in several synthetic domains designed to reflect a wide variety of factors that could influence the value of metareasoning. Our goal is to demonstrate the ability of our algorithms to estimate the value of computation and adapt to a plethora of world conditions. The experiments are performed on four domains, all of which are built on a 100×100 grid world, where the agent can move between cells at each time step to get to the goal located in the upper right corner. To initialize the lower and upper bounds of BRTDP, we use the zero heuristic and an appropriately scaled (multiplied by a constant) Manhattan distance to the goal, respectively.
###
6.1 Domains
The four world domains are as follows:
* *Stochastic.* This domain adds winds to the grid world to be analogous to worlds with stochastic state transitions. Moving against the wind causes slower movement across the grid, whereas moving with the wind results in faster movement. The agent’s initial state is the southeast corner and the goal is located in the northeast corner. We set the parameters of the domain as follows so that there is a policy that can get the agent to the goal with a small number of steps (in tens instead of hundreds) and where the winds significantly influence the number of steps needed to get to the goal: The agent can move 11 cells at a time and the wind has a pushing power of 10 cells. The next location of the agent is determined by adding the agent’s vector and the wind’s vector except when the agent decides to think (executes NOP), in which case it stays in the same position. Thus, the winds can never push the agent in the opposite direction of its intention. The prevailing wind direction over most of the grid is northerly, except for the column of cells containing the goal and starting position, where it is southerly. Note that this southerly wind direction makes the initial heuristic extremely suboptimal. To simulate stochastic state transitions, the winds have their prevailing direction in a given cell with 60% probability; with 40% probability they have a direction orthogonal to the prevailing one (20% easterly and 20% westerly).
We perform a set of experiments on this simplest domain of the set, to observe the effect of different costs for thinking and acting on the behaviors of algorithms. We vary the cost of thinking and acting between 1 and 15. When we vary the cost of thinking, we fix the cost of acting at 11, and when we vary the cost of acting, we fix the cost of thinking at 1.
* *Traps.* This domain modifies the *Stochastic* domain to resemble the setting where costs for thinking and acting are not constant among states. To simplify the parameter choices, we fix the cost of thinking and acting to be equal, respectively, to the agent’s moving distance and wind strength. Thus, the cost of thinking is 10 and the cost of acting is 11. To vary the costs of thinking and acting between states, we make thinking and acting at the initial state extremely expensive at a cost of 100, about 10 times the cost of acting and thinking in the other states. Thus, the agent is forced to think outside its initial state in order to perform optimally.
* *DynamicNOP-1.* In the previous domains, executing a NOP does not change the agent’s state. In this domain, thinking causes the agent to move in the direction of the wind, causing the agent to stochastically transition as a result of thinking. In this domain, the cost of thinking is composed of both explicit and implicit components; a static value of 1 unit and a dynamic component determined by stochastic state transitions as a result of thinking. The static value is set to 1 so that the dynamic component can dominate the decisions about thinking. The agent starts in cell (98,1). We change the wind directions so that there are easterly winds in the most southern row and northerly winds in the most eastern row that can push the agent very quickly to the goal. Westerly winds exist everywhere else, pushing the agent away from the goal. We change the stochasticity of the winds so that the westerly winds change to northerly winds with 20% probability, and all other wind directions are no longer stochastic. We lower the amount of stochasticity to better see if our agents can reason about the implicit costs of thinking. The wind directions are arranged so that there is potential for the agent to improve upon its initial policy but thinking is risky as it can move the agent to the left region, which is hard to recover from since all the winds push the agent away from the goal.
* *DynamicNOP-2.* This domain is just like the previous domain, but we change the direction of the winds in the northern-most row to be easterly. These winds also do not change directions. In this domain, as compared to the previous one, it is less risky to take a thinking action; even when the agent is pushed to the left region of the board, the agent can find strategies to get to the goal quickly by utilizing the easterly wind at the top region of the board.
###
6.2 The Metareasoning Gap
We introduce the concept of the *metareasoning gap* as a way to quantify the potential improvement over the initial heuristic-implied policy, denoted as Heuristic, that is possible due to optimal metareasoning. The metareasoning gap is the ratio of the expected cost of Heuristic for the base MDP to the expected cost of the optimal metareasoning policy, computed at the initial state. Exact computation of the metareasoning gap requires evaluating the optimal metareasoning policy and is infeasible. Instead, we compute an upper bound on the metareasoning gap by substituting the cost of the optimal metareasoning policy with the cost of the optimal policy for the *base* MDP (denoted OptimalBase). The metareasoning gap can be no larger than this upper bound, because metareasoning can only add cost to OptimalBase. We quantify each domain with this upper bound (MGUB) in Table [1](#S6.T1 "Table 1 ‣ 6.2 The Metareasoning Gap ‣ 6 Experiments ‣ Metareasoning for Planning Under Uncertainty") and show that our algorithms for metareasoning provide significant benefits when MGUB is high. We note that none of the algorithms use the metareasoning gap in its reasoning.
| | | | |
| --- | --- | --- | --- |
| | Heuristic | OptimalBase | MGUB |
| Stochastic (Thinking) | 1089 | 103.9 | 10.5 |
| Stochastic (Acting) | 767.3 | 68.1 | 11.3 |
| Traps | 979 | 113.5 | 8.6 |
| DynamicNOP-1 | 251.4 | 66 | 3.8 |
| DynamicNOP-2 | 119.4 | 66 | 1.8 |
Table 1: Upper bounds of metareasoning gaps (MGUB) for all test domains, defined as the ratio of the expected cost of the initial heuristic policy (Heuristic) to that of an optimal one (OptimalBase) at the initial state.
###
6.3 Experimental Setup
We compare our metareasoning algorithms against a number of baselines. The Think∗Act baseline simply plans for n cycles at the initial state and then executes the resulting policy, without planning again. We also consider the Prob baseline, which chooses to plan with probability p at each state, and executes its current policy with probability 1−p. An important drawback of these baselines is that their performance is sensitive to their parameters n and p, and the optimal parameter settings vary across domains. The NoInfoThink baseline plans for another cycle if it does not have information about how the BRTDP upper bounds will change. This baseline is a simplified version of our algorithms that does not try to estimate the VOC.
For each experimental condition, we run each metareasoning algorithm until it reaches the goal 1000 times and average the results to account for stochasticity. Each BRTDP trajectory is 50 actions long.
###
6.4 Results
In *Stochastic*, we perform several experiments by varying the costs of thinking (NOP) and acting. We observe (figures can be found in the appendix) that when the cost of thinking is low or when the cost of acting is high, the baselines do well with high values of n and p, and when the costs are reversed, smaller values do better. This trend is expected, since lower thinking cost affords more thinking, but these baselines don’t allow for predicting the specific “successful” n and p values in advance. Metareasoner does not require parameter tuning and beats even the best performing baseline for all settings. Figure [2](#S6.F2 "Figure 2 ‣ 6.4 Results ‣ 6 Experiments ‣ Metareasoning for Planning Under Uncertainty")a compares the metareasoning algorithms against the baselines when the results are averaged over the various settings of the cost of acting, and Figure [2](#S6.F2 "Figure 2 ‣ 6.4 Results ‣ 6 Experiments ‣ Metareasoning for Planning Under Uncertainty")b shows results averaged over the various settings of the cost of thinking. Metareasoner does extremely well in this domain because the metareasoning gap is large, suggesting that metareasoning can improve the initial policy significantly. Importantly, we see that Metareasoner performs better than NoInfoThink, which shows the benefit from reasoning about how the bounds on Q-values will change. UnCorr Metareasoner does not do as well as Metareasoner, probably because the assumption that actions’ Q-values are uncorrelated does not hold well.
Figure 2: Comparison of Metareasoner and Uncorr Metareasoner with baselines on experimental domains. Some figures do not include Heuristic when it performs especially poorly for readability.
We now turn to *Traps*, where thinking and acting in the initial state incurs significant cost. Figure [2](#S6.F2 "Figure 2 ‣ 6.4 Results ‣ 6 Experiments ‣ Metareasoning for Planning Under Uncertainty")c again summarizes the results. Think∗Act performs very poorly, because it is limited to thinking only at the initial state. Metareasoner does well, because it figures out that it should not think in the initial state (beyond the initial thinking step), and should instead quickly move to safer locations. Uncorr Metareasoner also closes the metareasoning gap significantly, but again not as much as Metareasoner.
We now consider *DynamicNOP-1*, a domain adversarial to approximate metareasoning, because winds almost everywhere push the agent away from the goal. There are only a few locations from which winds can carry the agent to the destination.
Figure [2](#S6.F2 "Figure 2 ‣ 6.4 Results ‣ 6 Experiments ‣ Metareasoning for Planning Under Uncertainty")d shows that our algorithms do not achieve large gains. However, this result is not surprising. The best policy involves little thinking, because whenever the agent chooses to think, it is pushed away from the goal, and thinking for just a few consecutive time steps can take the agent to states where reaching the goal is extremely difficult. Consequently, Think∗Act with 1-3 thinking steps turns out to be near-optimal, since it is pushed away from the goal only slightly and can use a slightly improved heuristic to head back. Metareasoner actually does well in many individual runs, but occasionally thinks longer due to VOC computation stochasticity and can get stuck, yielding higher average policy cost.
In particular, it may frequently be pushed into a state that it has never encountered before, where it must think again because it does not have any history about how BRTDP’s bounds have changed in that state, and then subsequently get pushed into an unencountered state again. In this domain, our approximate algorithms can diverge away from an optimal policy, which would plan very little to minimize the risk of being pushed away from the goal.
*DynamicNOP-2*
provides the agent more opportunities to recover even if it makes a poor decision. Figure [2](#S6.F2 "Figure 2 ‣ 6.4 Results ‣ 6 Experiments ‣ Metareasoning for Planning Under Uncertainty")e demonstrates that our algorithms perform much better in *DynamicNOP-2* than in *DynamicNOP-1*. In *DynamicNOP-2*, even if our algorithms do not discover the jetstreams that can push it towards the goal from initial thinking, they are provided more chances to recover when they get stuck. When thinking can move the agent on the board, having more opportunities to recover reduces the risk associated with making suboptimal thinking decisions. Interestingly, the metareasoning gap is decreased at the initial state by the addition of the extra jetstream. However, the metareasoning gap at many other states in the domain is increased, showing that the metareasoning gap at the initial state is not the most ideal way to characterize the potential for improvement via metareasoning in all domains.
7 Conclusion and Future Work
-----------------------------
We formalize and analyze the general metareasoning problem for MDPs, demonstrating that metareasoning is only polynomially harder than solving the base MDP. Given the determination that optimal general metareasoning is impractical, we turn to approximate metareasoning algorithms, which estimate the value of computation by relying on bounds given by BRTDP. Finally, we empirically compare our metareasoning algorithms to several baselines on problems designed to reflect challenges posed across a spectrum of worlds, and show that the proposed algorithms are much better at closing large metareasoning gaps.
We have assumed that the agent can plan only when it takes the NOP action. A generalization of our work would allow varying amounts of thinking as part of any action. Some actions may consume more CPU resources than others, and actions which do not consume all resources during execution can allocate the remainder to planning. We also can relax the meta-myopic assumption, so as to consider the consequences of thinking for more than one cycle. In many cases, assuming that the agent will only think for one more step can lead to underestimation of the value of thinking, since many cycles of thinking may be necessary to see significant value. This ability can be obtained with our current framework by projecting changes in bounds for multiple steps. However, in experiments to date, we have found that pushing out the horizon of analysis was associated with large accumulations of errors and poor performance due to approximation upon approximation from predictions about multiple thinking cycles. Finally, we may be able to improve our metareasoners by learning about and harnessing more details about the base-level planner. In our Metareasoner approximation scheme, we make strong assumptions about how the upper bounds provided by BRTDP will change, but learning distributions over these changes may improve performance. More informed models may lead to accurate estimation of non-myopic value of computation. However, learning distributions in a domain-independent manner is difficult, since the planner’s behavior is heavily dependent on the domain and heuristic at hand. |
e6629bb5-3496-4755-a659-cfe63d588221 | StampyAI/alignment-research-dataset/arxiv | Arxiv | A Psychopathological Approach to Safety Engineering in AI and AGI
1 Introduction
---------------
While the adaptive mechanisms of human cognition provide the means for unique skills in adjusting to dynamic environments, they are also prone to psychological disorders, broadly defined as self-reconfigurations in cognition and behavior that are deleterious to the core and long-term objectives of self or the social ecosystem [[2](#bib.bib2)]. Extrapolating from this phenomenon, it is not hard to conclude that instances of Artificial General Intelligence (AGI), which aim for similar cognitive functions, may also be prone to such disorders. For instance, certain objective functions and environmental conditions may lead a Reinforcement Learning (RL) agent to develop addictive behavior through repetitive gains of high rewards from policies that contradict the long-term objectives of the agent [[15](#bib.bib15)]. Other instances of such emergent disorders include post-traumatic behavior, depression, and psychosis [[1](#bib.bib1)]. It is further hypothesized that behavioral disorders may emerge as higher-order consequences of unsafe inverse RL and apprenticeship learning, by adopting manifested disorders or triggering harmful cognitive traits [[16](#bib.bib16)].
Current research in AI safety is generally focused on safety-aware design and mitigation techniques [[11](#bib.bib11)], but the expanding complexity of AI and in particular AGI will render such analysis as difficult as those of biological intelligence and the corresponding disorders. To tackle such difficulties in human intelligence, the causes and dynamics of misbehaviors are studied at various levels of abstraction, ranging from neuroscience and cognitive science to psychology, psychiatry, sociology, and criminology. Inspired by the advantages of such diverse vantage points, we propose that studying the complex dynamics and mechanisms of failure in AI safety can greatly benefit from abstractions that parallel those of biological intelligence. Considering the practical aims of diagnosing and correcting misbehaviors in AGI, we believe that adopting the abstraction of psychopathology provides tractable settings that also benefit from cross-domain bodies of knowledge. Furthermore, while this approach may seem to be of lower relevance at present, we argue that the advent of deep RL, along with advances in hierarchical and transfer learning may have already laid the grounds for emergence of such disorders in AI.
The goal of this paper is to provide a technical discussion and the motivation for research on the psychopathology of AI and AGI. The remainder of this paper is organized as follows: Section [2](#S2 "2 What is Psychopathology? ‣ A Psychopathological Approach to Safety Engineering in AI and AGI") presents a broad overview of psychopathology. Section [3](#S3 "3 Psychopathology and AI Safety ‣ A Psychopathological Approach to Safety Engineering in AI and AGI") provides a discussion on the relevance of psychology to AI, followed by establishment of parallelisms between AI safety and psychopathology. In Section [4](#S4 "4 Directions of Research ‣ A Psychopathological Approach to Safety Engineering in AI and AGI"), high-level areas of research are identified and detailed. Finally, Section [5](#S5 "5 Conclusion ‣ A Psychopathological Approach to Safety Engineering in AI and AGI") concludes the paper with remarks on broader impacts of this research.
2 What is Psychopathology?
---------------------------
Psychopathology refers to the scientific study of mental disorders, their causes, and corresponding treatments [[4](#bib.bib4)]. Within this context, we adhere to American Psychiatric Association (APA)’s definition of mental disorder [[2](#bib.bib2)] as “a psychological syndrome or pattern which is associated with distress, disability, increased risk of death, or significant loss of autonomy” (i.e., pursuit of objectives). In psychopathology, disorders are commonly identified based on four metrics of abnormality, known as the four Ds [[6](#bib.bib6)]: Deviance of behaviors and emotions from the norm, Distress of the individual caused by suffering from a disorder, Dysfunctions that impair the individual’s ability to perform designated or normal functions, and the Danger of individual to self or the society.
Causes of mental disorders in humans include mixtures of those inherited through *genetics* (e.g., neuroticism), *developmental influences* caused by parental mistreatment, social influences (e.g., as abuse, bullying), and traumatic events, and *biological influences* such as traumatic brain injury and infections [[2](#bib.bib2)].
Various models have been developed to capture the dynamics of mental disorders and their emergence. For instance, biological psychiatry, or the *medical model* [[9](#bib.bib9)], is one that explains the causes of disorders based on changes in neurological circuitry. The *social model*, on the other hand, analyzes the causes of mental disorders based on social and environmental interactions [[9](#bib.bib9)]. Currently, it is widely believed that understanding psychological disorders requires the comprehensive consideration of both biological and social factors, and hence the *biopsychosocial models* are generally adopted to study such phenomena. These models broadly categorize mental disorders as either cognitive or behavioral. Cognitive disorders are those caused by abnormal functioning of the underlying cognitive mechanisms, and behavioral disorders are those that are learned through developmental, environmental, and social interactions [[9](#bib.bib9)].
Diagnosis of mental disorders is generally based on an assessment of symptoms, signs, and impairments that constitute various types of disorders. A comprehensive framework for such assessments is that of the Diagnostic and Statistical Manual of Mental Disorders (DSM) [[2](#bib.bib2)], published by the American Psychiatric Association (APA). This manual provides a common language and standard criteria for the classification of mental disorders. Furthermore, recent advances in machine learning have given rise to various software and algorithmic tools to facilitate enhanced accuracy in classification and diagnosis of mental disorders [[8](#bib.bib8)].
Treatment of mental disorders is commonly via one or a hybrid of two approaches. One is *Psychotherapy*, which is a form of interpersonal intervention via a range of psychological techniques. For instance, Cognitive Behavioral Therapy (CBT) is employed to modify the patterns of thought and behavior associated with a particular disorder. *Medication therapy* is the other approach, which targets the physiological components of disorders. For instance, antipsychotics commonly work by blocking D2 Dopamine receptors, thus controlling the chemical reward mechanism of the brain [[13](#bib.bib13)].
3 Psychopathology and AI Safety
--------------------------------
Since its inception, AI has been closely connected to psychology and cognitive sciences [[7](#bib.bib7)]. This connection flows in both directions: AI researchers study biological cognition and behavior as inspiration for engineered intelligence, and cognitive scientists explore AI as a framework for synthesis and experimental analysis of theoretical ideas [[5](#bib.bib5)]. An instance of this interconnection is Reinforcement Learning (RL), where the computational algorithms of RL, such as Temporal Difference (TD) learning were originally inspired from the dopamine system in biological brains [[14](#bib.bib14)]. On the other hand, the work on TD learning has provided mathematical means of modeling the neuroscientific dynamics of dopamine cells in the brain, and has been employed to study disorders such as schizophrenia and the consequences of pharmacological manipulations of dopamine on learning [[12](#bib.bib12)].
Figure 1: A psychopathological approach to safety engineering in AI and AGI.
With regards to the relationship between psychological disorders and AI safety, there are scarce and sparse resources available in the literature. Recent papers by Ashrafian [[1](#bib.bib1)] and Yampolskiy [[15](#bib.bib15)] [[17](#bib.bib17)] present high-level arguments for the existence and emergence of mental disorders in AI. One such argument presented in [[1](#bib.bib1)] is based on the analogy of David Chalmers’ philosophical zombie (p-zombie). In this analogy, the p-zombie is considered to be a fully functioning robot that acts exactly like a human-being, which is not necessarily equipped with vague notions of consciousness [[17](#bib.bib17)]. The fact that this robot is capable of acting indistinguishably from humans is then used to justify that it is also prone to developmental and cognitive abnormalities that lead to misbehavior and anomalous cognition.
Furthermore, many aspects of failures in AI safety can be viewed as psychological disorders. For instance, wireheading in AI can manifest as delusional and addictive behavior. [[15](#bib.bib15)]. Similarly, sequences of interactions with extremely negative rewards and stresses within the exploration/exploitation trajectories of RL-based AI can potentially give rise to behavioral disorders such as depression and Post-Traumatic Stress Disorder (PTSD) [[1](#bib.bib1)]. Furthermore, the generic manifestation of the value alignment problem [[11](#bib.bib11)] in AI is in the form of behavioral characteristics that are harmful to either the agent or the environment and society, which falls well within the definition of psychological disorders.
While [[1](#bib.bib1)] and a few other papers (e.g., [[3](#bib.bib3)][[11](#bib.bib11)]) present high-level arguments on the advantages of investigating the psychopathology of AI, there remains a wide gap in satisfying the need for technical studies and practices. This paper presents a research agenda that will fill this gap via the following proposals, also illustrated in Figure [1](#S3.F1 "Figure 1 ‣ 3 Psychopathology and AI Safety ‣ A Psychopathological Approach to Safety Engineering in AI and AGI").
4 Directions of Research
-------------------------
Developing solid grounds for research on the Psychopathology of AI requires investigations in three main areas: Modeling and Verification, Diagnosis, and Treatment. In this section, we define and discuss the scope of each area.
###
4.1 Modeling and Verification Tools
While the descriptive similarities of human psychopathology and AI failures provide some insights into adopting such abstractions for AI safety, taking an engineering approach requires formal and mathematical modeling of the aspects and dimensions of these similarities. Such formalisms may benefit from those that already exist in the realm of cognitive and medical sciences, such as cognitive architectures [[10](#bib.bib10)] and RL-based models of the dynamics in mental disorders (e.g., [[12](#bib.bib12)]). Also, the quantitative analysis of such disorders necessitate the exploration and development of new models of AI and AGI based on such paradigms as neuroeconomics, complex adaptive systems, control theory, and dynamic data-driven application systems.
Furthermore, verification and validation of such models and the ensuing theories requires the development of experimental frameworks and simulation platforms. Such platforms must provide the means for wide ranges of experiments on emergence and dynamics of behavioral and cognitive disorders in arbitrary and context-dependent scenarios, and shall be compatible for various agent and environmental models.
###
4.2 Diagnosis and Classification of Disorders
This venue is on investigating and development of techniques for diagnosis of disorders in AI. Within the context of AI safety engineering, diagnosis refers to two inter-related tasks: first is to detect anomalous behaviors, and the second is to classify the type of anomalous behavior as a first step towards treatment. Detection of undesired behavior is an active topic of research in AI safety, with initial solutions such as tripwires and honeypots [[11](#bib.bib11)] already proposed and investigated. We propose to extend current state of the art in detection through adoption and automation of parallel techniques in psychopathology. Similar to diagnostics criteria in human psychology [[2](#bib.bib2)], a promising approach is to identify statistical deviations in behavior, as well as general indicators of misbehavior. To this end, development of machine learning approaches similar to those applied in cybersecurity for threat and intrusion detection can be a promising direction. Furthermore, generic indicators of misbehavior can be learned from models trained on simulated and annotated scenarios of disorders.
Once a misbehavior is detected, the next step is to characterize and classify the disorder that has led to such behavior. A prerequisite to this process is having a catalog of different disorders and the corresponding criteria for diagnosing such disorders. Therefore, a necessary step is the compilation of representative and experimentally verified disorders, such as addiction and anxiety in RL agents, along with manually and automatically generated criteria and characteristics of each disorder based on behavioral observations. This task shall aim to produce human- and machine-readable catalogs as AI analogues of APA’s DSM 5 [[2](#bib.bib2)].
Besides general behavioral characteristics, there are other sources of data that can be of diagnostic value. Instances include indicators of disorders that are obtained through direct and targeted interactions with AI (similar to psychiatric evaluation of human patients), non-invasive analysis of internal states and parameters (similar to F-MRI and EEG tests of human patients), and induction or invocation of internal debug modes (similar to states of hypnosis). Exploring such ideas and approaches may greatly enhance the accuracy of diagnosis, and lead to novel techniques for psychoanalysis and diagnostics of AI and AGI.
###
4.3 Treatment
When a disorder is diagnosed in an AI agent, it is not always feasible to simply decommission or reset the agent. In such cases, it is often preferable to pursue treatment via minimally destructive techniques that correct the misbehaviors of agent, while preserving the useful traits learned by that agent. Such treatments need to satisfy a number of challenging requirements. Advanced AI are complex adaptive systems, and therefore minor perturbations of one component may lead to unintended consequences on local and global scales. For instance, correcting a developmental disorder by removing a series of harmful experiences from the memory of an AI may lead to behavioral changes that are even more undesirable than the original misbehavior. Therefore, effective treatments must either be minimally invasive or non-invasive at all.
Inspired by psychopathological parallels, we propose two general approaches to treatment of pathologies in AI. One is correctional training, which adopts the approach of behavioral therapy. This approach is to retrain an agent in controlled environments and scenarios, such that harmful experiences can be remedied or alleviated through new experiences. The second approach parallels that of medication therapy, in which the reward signals of AI agents are artificially manipulated via external means to adjust their behavioral policies. This is similar to the use of anti-depressants and anti-psychotics in treating disorders related to production and inhibition of dopamine and serotonin in human brains.
5 Conclusion
-------------
This paper presents the argument that while current research in AI safety is generally focused on design and mitigation problems, the complexity of AGI will render such analysis as difficult as those that capture biological intelligence and disorders. Hence, studying the complex dynamics and mechanisms of emergent failures in AI and AGI can greatly benefit from abstractions that parallel those of biological intelligence. Considering the practical objectives of diagnosing and treating misbehaviors in AGI, we propose that psychopathological approaches provide tractable settings while benefiting from various bodies of knowledge. Accordingly, we present a high-level research agenda that includes explorations of parallels between human and AI psychopathology, development of methodologies for diagnosis of behavioral pathologies in AI, and propose techniques for treatment of such disorders.
As the paper detailes, psychology and AI enjoy a bi-directional flow of inspirations. A major impact of the proposed research is the production of outcomes that can be of use and inspiration to current research in psychopathology and cognitive sciences. Furthermore, the results of this work may provide a deeper understanding of the safety requirements and guidelines for designing advanced AI and AGI, while guiding policy makers on the risks and potential solutions involved in the integration of AGI into societies. We hope that this paper motivates initial efforts in laying solid foundations for future research and developments in this scarcely explored but promising venue. |
88d04436-81a6-4141-ba63-41f5edbba38a | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Conclusion to 'Reframing Impact'
Epistemic Status
----------------
I've made many claims in these posts. All views are my own.
Elicit Prediction (<elicit.org/binary/questions/7SoL5DPRf>)
Elicit Prediction (<elicit.org/binary/questions/AevXOS1Rj>)
> Confident (75%). [The theorems on power-seeking](https://arxiv.org/abs/1912.01683) only apply to optimal policies in fully observable environments, which isn't realistic for real-world agents. However, I think they're still informative. There are also strong intuitive arguments for power-seeking.
>
>
Elicit Prediction (<elicit.org/binary/questions/javyyEd8C>)
> Fairly confident (70%). There seems to be a dichotomy between "catastrophe directly incentivized by goal" and "catastrophe indirectly incentivized by goal through power-seeking", although Vika [provides intuitions in the other direction](https://www.lesswrong.com/posts/sHpiiZS2gPgoPnijX/conclusion-to-reframing-impact?commentId=6sxBzsh8yfwnPk4iH#6sxBzsh8yfwnPk4iH).
>
>
Elicit Prediction (<elicit.org/binary/questions/iYT69bLl9>)
Elicit Prediction (<elicit.org/binary/questions/GFGG5plOQ>)
Elicit Prediction (<elicit.org/binary/questions/8Rr-YFSWi>)
Elicit Prediction (<elicit.org/binary/questions/GFurWKpJn>)
Acknowledgements
----------------
After ~700 hours of work over the course of ~9 months, the sequence is finally complete.
This work was made possible by the Center for Human-Compatible AI, the Berkeley Existential Risk Initiative, and the Long-Term Future Fund. Deep thanks to Rohin Shah, Abram Demski, Logan Smith, Evan Hubinger, TheMajor, Chase Denecke, Victoria Krakovna, Alper Dumanli, Cody Wild, Matthew Barnett, Daniel Blank, Sara Haxhia, Connor Flexman, Zack M. Davis, Jasmine Wang, Matthew Olson, Rob Bensinger, William Ellsworth, Davide Zagami, Ben Pace, and a million other people for giving feedback on this sequence.
Appendix: Easter Eggs
---------------------
The big art pieces (and especially the last illustration in this post) were designed to convey a specific meaning, the interpretation of which I leave to the reader.
There are a few pop culture references which I think are obvious enough to not need pointing out, and a lot of hidden smaller playfulness which doesn't quite rise to the level of "easter egg".
*Reframing Impact*
The bird's nest contains a literal easter egg.
The paperclip-Balrog drawing contains a [Tengwar](https://en.wikipedia.org/wiki/Tengwar) inscription which reads "one measure to bind them", with "measure" in impact-blue and "them" in utility-pink.
"Towards a New Impact Measure" was the title of [the post](https://www.lesswrong.com/posts/yEa7kwoMpsBgaBCgb/towards-a-new-impact-measure) in which AUP was introduced.
*Attainable Utility Theory: Why Things Matter*
This style of maze is from the video game *Undertale*.
*Seeking Power is Instrumentally Convergent in MDPs*
To seek power, Frank is trying to get at the Infinity Gauntlet.
The tale of Frank and the orange Pebblehoarder
Speaking of under-tales, a friendship has been blossoming right under our noses.
After the Pebblehoarders suffer the devastating transformation of all of their pebbles into obsidian blocks, Frank generously gives away his favorite pink marble as a makeshift pebble.
The title cuts to the middle of their adventures together, the Pebblehoarder showing its gratitude by helping Frank reach things high up.
This still at the midpoint of the sequence is from [the final scene of *The Hobbit: An Unexpected Journey*](https://www.youtube.com/watch?v=KEegn1R601M), where the party is overlooking Erebor, the Lonely Mountain. They've made it through the Misty Mountains, only to find Smaug's abode looming in the distance.
And, at last, we find Frank and orange Pebblehoarder popping some of the champagne from Smaug's hoard.
Since [Erebor isn't close to Gondor](http://images1.fanpop.com/images/photos/2300000/Map-of-Middle-Earth-lord-of-the-rings-2329809-1600-1200.jpg), we don't see Frank and the Pebblehoarder gazing at Ephel Dúath from Minas Tirith. |
180bce79-5dc1-465b-ab48-d593ca21622d | trentmkelly/LessWrong-43k | LessWrong | What would you do if AI were dangerous?
If we knew how to build a machine that chooses its outputs as to maximize some property of the surrounding universe, such a machine would be very dangerous, because maximizing almost any easily defined property leads to a worthless universe (without humans, or with humans living pointless lives, etc.) I believe the preceding statement is uncontroversial, and most arguments around the necessity of Friendly AI are really about how likely we are to build such a machine, or maybe something else will happen first, etc.
Instead of adding to the existing arguments, I want to reframe the question thus: what course of action would you recommend to a small group of smart people, assuming for the moment that the danger is real? In other words, what should SingInst do on an alternate Earth where normal human science will eventually build unfriendly AI? In particular:
- How do you craft your message to the public?
- What's your hiring policy?
- Do you keep your research secret?
- Do you pursue alternate avenues like uploads, or focus only on FAI?
For the sake of inconvenience, assume that many (though not all) of the insights required for developing FAI can also be easily repurposed to hasten the arrival of UFAI.
Thanks to Wei Dai for the conversation that sparked this post. |
afc286e6-76fc-4fef-8dbf-1854fef3998b | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Hard-Coding Neural Computation
Previously: [Teaser: Hard-coding Transformer Models](https://www.lesswrong.com/posts/Lq6jo5j9ty4sezT7r/teaser-hard-coding-transformer-models)
Introduction
============
Transformer models are incredibly powerful for natural language tasks (and they are starting to find uses in many other fields of machine learning). Unfortunately, it is nigh-impossible to interpret what goes on inside them. OR IS IT???
I have found that I can, with a fair amount of effort, hard-code the weights of a transformer model in order to perform some very crude versions of linguistic tasks. So far I have achieved English-to-French translation (on a toy corpus of about 150 sentences), text classification (is a sentence grammatical or not? on a toy corpus of a couple hundred sentences), and sentiment analysis (again on a limited corpus). These results are obviously not impressive compared to the state of the machine learning field, but I am pretty sure that they can all be drastically scaled up with the investment of some time and energy. Unfortunately, I have a fairly demanding day job, and haven't found the time and energy yet.
All of this is done by inspection (no gradient descent!). The process is a lot like programming, although it is more difficult than programming, at least right now for me. I am fairly certain that better tools and better notation can be developed to make the process easier. It is also almost certainly possible to combine hard-coding with gradient descent approaches to be able to scale these methods up in a slightly less labor-intensive way.
I think that these ideas could prove useful in alignment research - if we understand how a language model works in excruciating detail, it seems drastically more likely that we will be able to reason about and predict various misunderstandings rooted in the ambiguity of language. Given that language is (arguably) a fully general means of interacting with an artificial intelligence, it seems plausible to me that this work is on the critical path to alignment.
Doneness Status
===============
This post is a work-in-progress. I will be editing it as I go, mostly appending more content to the end, but I will also try to fix any errors or unclear parts as I notice them or commenters point them out.
So let's hard-code some neural computation! I have a very, very messy [github repository](https://github.com/epurdy/hand) where I've done my initial experiments, if you prefer to just jump into semi-working code. Otherwise, I will do my very best to explain the ideas from scratch in this post. I'm aiming this post at anyone who's willing to put in the work to understand it, so I'll try to at least give pointers to necessary background material, of which there is a fair amount.
What Can We Do Already?
=======================
Some primitive sentiment analysis using a Vanilla RNN:
Some very simple translation with a Transformer model:
How Should We Measure Success?
==============================
Before we explain how we get results, it seems worthwhile to talk about how to measure the performance of such a system. As in traditional machine learning, this can only be measured with respect to some dataset of inputs labeled with desired outputs. For a crude metric, we can look at the fraction of inputs that receive the desired output, versus the fraction that receive some other output. We can also look at more complex metrics, such as [BLEU](https://en.wikipedia.org/wiki/BLEU) or [ROUGE](https://en.wikipedia.org/wiki/ROUGE_(metric)) for [sequence-to-sequence](https://en.wikipedia.org/wiki/Seq2seq) tasks.
In traditional machine learning, performance can only be (meaningfully) measured on a holdout set that was not used to train the algorithm. This is because performance tends to be much, much higher (if you're using the right architecture and hyperparameters for your task) on the training set (the set of data that was used to train the algorithm) than it will be for the test set (the set of data that has been held out). The whole purpose and challenge of machine learning, of course, is to build models that generalize to unseen data.
A similar phenomenon occurs in this work, where data that has been examined by the programmer and run through the algorithm and used to inform updates to the rules, will typically be data that the resulting network does disproportionately well on. After all, if you miss some edge case, but then see it in your testing, you have the opportunity to fix it.
On the other hand, the programmer will presumably have a native command of at least one language, so it is at least possible for the programmer to anticipate some phenomena before seeing them in the "training data". Thus, it seems unfair to gradient descent and deep learning to compare accuracies in the low-data regime where I have been stuck so far by my lack of free time.
The ultimate ambition of this work would be to go toe-to-toe with a comparably-sized Transformer model trained in the traditional way on a modern-sized data set. This might require several people-years of focused effort though.
Some Useful Notation
====================
The first thing we are going to do is introduce some very unconventional notation for vectors and matrices. (We won't need any more information about linear algebra than is contained in the [Wikipedia article](https://en.wikipedia.org/wiki/Linear_algebra), but we will assume that you are either familiar with them or have paused and read that article.)
We will pick a set of "axes" that we will call "semes". (This word comes from semiotics, as do a few other terms we will use. I believe I'm using them in a way that is compatible with their technical meaning in semiotics, but feel free to think of this as a nonsense word that we are coining.) Each seme will be identified with a short string, often a word. So, we might have semes "wombat", "peregrine", and "pig". These play a role very similar to variable names in traditional programming, so we will generally choose them to be meaningful. Common semes that I actually use are "noun", "verb", etc.
We then will write vectors using these semes, for example ⟨⟨+pig−wombat⟩⟩.mjx-chtml {display: inline-block; line-height: 0; text-indent: 0; text-align: left; text-transform: none; font-style: normal; font-weight: normal; font-size: 100%; font-size-adjust: none; letter-spacing: normal; word-wrap: normal; word-spacing: normal; white-space: nowrap; float: none; direction: ltr; max-width: none; max-height: none; min-width: 0; min-height: 0; border: 0; margin: 0; padding: 1px 0}
.MJXc-display {display: block; text-align: center; margin: 1em 0; padding: 0}
.mjx-chtml[tabindex]:focus, body :focus .mjx-chtml[tabindex] {display: inline-table}
.mjx-full-width {text-align: center; display: table-cell!important; width: 10000em}
.mjx-math {display: inline-block; border-collapse: separate; border-spacing: 0}
.mjx-math \* {display: inline-block; -webkit-box-sizing: content-box!important; -moz-box-sizing: content-box!important; box-sizing: content-box!important; text-align: left}
.mjx-numerator {display: block; text-align: center}
.mjx-denominator {display: block; text-align: center}
.MJXc-stacked {height: 0; position: relative}
.MJXc-stacked > \* {position: absolute}
.MJXc-bevelled > \* {display: inline-block}
.mjx-stack {display: inline-block}
.mjx-op {display: block}
.mjx-under {display: table-cell}
.mjx-over {display: block}
.mjx-over > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-under > \* {padding-left: 0px!important; padding-right: 0px!important}
.mjx-stack > .mjx-sup {display: block}
.mjx-stack > .mjx-sub {display: block}
.mjx-prestack > .mjx-presup {display: block}
.mjx-prestack > .mjx-presub {display: block}
.mjx-delim-h > .mjx-char {display: inline-block}
.mjx-surd {vertical-align: top}
.mjx-surd + .mjx-box {display: inline-flex}
.mjx-mphantom \* {visibility: hidden}
.mjx-merror {background-color: #FFFF88; color: #CC0000; border: 1px solid #CC0000; padding: 2px 3px; font-style: normal; font-size: 90%}
.mjx-annotation-xml {line-height: normal}
.mjx-menclose > svg {fill: none; stroke: currentColor; overflow: visible}
.mjx-mtr {display: table-row}
.mjx-mlabeledtr {display: table-row}
.mjx-mtd {display: table-cell; text-align: center}
.mjx-label {display: table-row}
.mjx-box {display: inline-block}
.mjx-block {display: block}
.mjx-span {display: inline}
.mjx-char {display: block; white-space: pre}
.mjx-itable {display: inline-table; width: auto}
.mjx-row {display: table-row}
.mjx-cell {display: table-cell}
.mjx-table {display: table; width: 100%}
.mjx-line {display: block; height: 0}
.mjx-strut {width: 0; padding-top: 1em}
.mjx-vsize {width: 0}
.MJXc-space1 {margin-left: .167em}
.MJXc-space2 {margin-left: .222em}
.MJXc-space3 {margin-left: .278em}
.mjx-test.mjx-test-display {display: table!important}
.mjx-test.mjx-test-inline {display: inline!important; margin-right: -1px}
.mjx-test.mjx-test-default {display: block!important; clear: both}
.mjx-ex-box {display: inline-block!important; position: absolute; overflow: hidden; min-height: 0; max-height: none; padding: 0; border: 0; margin: 0; width: 1px; height: 60ex}
.mjx-test-inline .mjx-left-box {display: inline-block; width: 0; float: left}
.mjx-test-inline .mjx-right-box {display: inline-block; width: 0; float: right}
.mjx-test-display .mjx-right-box {display: table-cell!important; width: 10000em!important; min-width: 0; max-width: none; padding: 0; border: 0; margin: 0}
.MJXc-TeX-unknown-R {font-family: monospace; font-style: normal; font-weight: normal}
.MJXc-TeX-unknown-I {font-family: monospace; font-style: italic; font-weight: normal}
.MJXc-TeX-unknown-B {font-family: monospace; font-style: normal; font-weight: bold}
.MJXc-TeX-unknown-BI {font-family: monospace; font-style: italic; font-weight: bold}
.MJXc-TeX-ams-R {font-family: MJXc-TeX-ams-R,MJXc-TeX-ams-Rw}
.MJXc-TeX-cal-B {font-family: MJXc-TeX-cal-B,MJXc-TeX-cal-Bx,MJXc-TeX-cal-Bw}
.MJXc-TeX-frak-R {font-family: MJXc-TeX-frak-R,MJXc-TeX-frak-Rw}
.MJXc-TeX-frak-B {font-family: MJXc-TeX-frak-B,MJXc-TeX-frak-Bx,MJXc-TeX-frak-Bw}
.MJXc-TeX-math-BI {font-family: MJXc-TeX-math-BI,MJXc-TeX-math-BIx,MJXc-TeX-math-BIw}
.MJXc-TeX-sans-R {font-family: MJXc-TeX-sans-R,MJXc-TeX-sans-Rw}
.MJXc-TeX-sans-B {font-family: MJXc-TeX-sans-B,MJXc-TeX-sans-Bx,MJXc-TeX-sans-Bw}
.MJXc-TeX-sans-I {font-family: MJXc-TeX-sans-I,MJXc-TeX-sans-Ix,MJXc-TeX-sans-Iw}
.MJXc-TeX-script-R {font-family: MJXc-TeX-script-R,MJXc-TeX-script-Rw}
.MJXc-TeX-type-R {font-family: MJXc-TeX-type-R,MJXc-TeX-type-Rw}
.MJXc-TeX-cal-R {font-family: MJXc-TeX-cal-R,MJXc-TeX-cal-Rw}
.MJXc-TeX-main-B {font-family: MJXc-TeX-main-B,MJXc-TeX-main-Bx,MJXc-TeX-main-Bw}
.MJXc-TeX-main-I {font-family: MJXc-TeX-main-I,MJXc-TeX-main-Ix,MJXc-TeX-main-Iw}
.MJXc-TeX-main-R {font-family: MJXc-TeX-main-R,MJXc-TeX-main-Rw}
.MJXc-TeX-math-I {font-family: MJXc-TeX-math-I,MJXc-TeX-math-Ix,MJXc-TeX-math-Iw}
.MJXc-TeX-size1-R {font-family: MJXc-TeX-size1-R,MJXc-TeX-size1-Rw}
.MJXc-TeX-size2-R {font-family: MJXc-TeX-size2-R,MJXc-TeX-size2-Rw}
.MJXc-TeX-size3-R {font-family: MJXc-TeX-size3-R,MJXc-TeX-size3-Rw}
.MJXc-TeX-size4-R {font-family: MJXc-TeX-size4-R,MJXc-TeX-size4-Rw}
.MJXc-TeX-vec-R {font-family: MJXc-TeX-vec-R,MJXc-TeX-vec-Rw}
.MJXc-TeX-vec-B {font-family: MJXc-TeX-vec-B,MJXc-TeX-vec-Bx,MJXc-TeX-vec-Bw}
@font-face {font-family: MJXc-TeX-ams-R; src: local('MathJax\_AMS'), local('MathJax\_AMS-Regular')}
@font-face {font-family: MJXc-TeX-ams-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_AMS-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_AMS-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_AMS-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-B; src: local('MathJax\_Caligraphic Bold'), local('MathJax\_Caligraphic-Bold')}
@font-face {font-family: MJXc-TeX-cal-Bx; src: local('MathJax\_Caligraphic'); font-weight: bold}
@font-face {font-family: MJXc-TeX-cal-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-R; src: local('MathJax\_Fraktur'), local('MathJax\_Fraktur-Regular')}
@font-face {font-family: MJXc-TeX-frak-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-frak-B; src: local('MathJax\_Fraktur Bold'), local('MathJax\_Fraktur-Bold')}
@font-face {font-family: MJXc-TeX-frak-Bx; src: local('MathJax\_Fraktur'); font-weight: bold}
@font-face {font-family: MJXc-TeX-frak-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Fraktur-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Fraktur-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Fraktur-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-BI; src: local('MathJax\_Math BoldItalic'), local('MathJax\_Math-BoldItalic')}
@font-face {font-family: MJXc-TeX-math-BIx; src: local('MathJax\_Math'); font-weight: bold; font-style: italic}
@font-face {font-family: MJXc-TeX-math-BIw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-BoldItalic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-BoldItalic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-BoldItalic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-R; src: local('MathJax\_SansSerif'), local('MathJax\_SansSerif-Regular')}
@font-face {font-family: MJXc-TeX-sans-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-B; src: local('MathJax\_SansSerif Bold'), local('MathJax\_SansSerif-Bold')}
@font-face {font-family: MJXc-TeX-sans-Bx; src: local('MathJax\_SansSerif'); font-weight: bold}
@font-face {font-family: MJXc-TeX-sans-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-sans-I; src: local('MathJax\_SansSerif Italic'), local('MathJax\_SansSerif-Italic')}
@font-face {font-family: MJXc-TeX-sans-Ix; src: local('MathJax\_SansSerif'); font-style: italic}
@font-face {font-family: MJXc-TeX-sans-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_SansSerif-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_SansSerif-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_SansSerif-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-script-R; src: local('MathJax\_Script'), local('MathJax\_Script-Regular')}
@font-face {font-family: MJXc-TeX-script-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Script-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Script-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Script-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-type-R; src: local('MathJax\_Typewriter'), local('MathJax\_Typewriter-Regular')}
@font-face {font-family: MJXc-TeX-type-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Typewriter-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Typewriter-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Typewriter-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-cal-R; src: local('MathJax\_Caligraphic'), local('MathJax\_Caligraphic-Regular')}
@font-face {font-family: MJXc-TeX-cal-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Caligraphic-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Caligraphic-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Caligraphic-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-B; src: local('MathJax\_Main Bold'), local('MathJax\_Main-Bold')}
@font-face {font-family: MJXc-TeX-main-Bx; src: local('MathJax\_Main'); font-weight: bold}
@font-face {font-family: MJXc-TeX-main-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Bold.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-I; src: local('MathJax\_Main Italic'), local('MathJax\_Main-Italic')}
@font-face {font-family: MJXc-TeX-main-Ix; src: local('MathJax\_Main'); font-style: italic}
@font-face {font-family: MJXc-TeX-main-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-main-R; src: local('MathJax\_Main'), local('MathJax\_Main-Regular')}
@font-face {font-family: MJXc-TeX-main-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Main-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Main-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Main-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-math-I; src: local('MathJax\_Math Italic'), local('MathJax\_Math-Italic')}
@font-face {font-family: MJXc-TeX-math-Ix; src: local('MathJax\_Math'); font-style: italic}
@font-face {font-family: MJXc-TeX-math-Iw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Math-Italic.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Math-Italic.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Math-Italic.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size1-R; src: local('MathJax\_Size1'), local('MathJax\_Size1-Regular')}
@font-face {font-family: MJXc-TeX-size1-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size1-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size1-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size1-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size2-R; src: local('MathJax\_Size2'), local('MathJax\_Size2-Regular')}
@font-face {font-family: MJXc-TeX-size2-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size2-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size2-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size2-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size3-R; src: local('MathJax\_Size3'), local('MathJax\_Size3-Regular')}
@font-face {font-family: MJXc-TeX-size3-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size3-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size3-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size3-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-size4-R; src: local('MathJax\_Size4'), local('MathJax\_Size4-Regular')}
@font-face {font-family: MJXc-TeX-size4-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Size4-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Size4-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Size4-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-R; src: local('MathJax\_Vector'), local('MathJax\_Vector-Regular')}
@font-face {font-family: MJXc-TeX-vec-Rw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Regular.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Regular.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Regular.otf') format('opentype')}
@font-face {font-family: MJXc-TeX-vec-B; src: local('MathJax\_Vector Bold'), local('MathJax\_Vector-Bold')}
@font-face {font-family: MJXc-TeX-vec-Bx; src: local('MathJax\_Vector'); font-weight: bold}
@font-face {font-family: MJXc-TeX-vec-Bw; src /\*1\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/eot/MathJax\_Vector-Bold.eot'); src /\*2\*/: url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/woff/MathJax\_Vector-Bold.woff') format('woff'), url('https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.2/fonts/HTML-CSS/TeX/otf/MathJax\_Vector-Bold.otf') format('opentype')}
for the vector that is 1 in the direction ⟨⟨pig⟩⟩ and -1 in the direction ⟨⟨wombat⟩⟩. We can also use coefficients, so that ⟨⟨+2.1pig−3.2peregrine⟩⟩ denotes the vector that is 2.1 in the direction ⟨⟨pig⟩⟩ and -3.2 in the direction ⟨⟨peregrine⟩⟩. There are two ways of thinking about this - you can either think of the various semes as being completely orthogonal to each other, forming an orthonormal basis of whatever vector space we are in. Or you can think of them as arbitrary vectors that we are using as a (possibly overcomplete) basis. In general, both will be useful; I generally think of syntactic information as being best represented in a fully orthonormal basis, while semantic information makes much more sense as being drawn from a very overcomplete basis.
Matrices will be written in the form {{1.1pig→wombat+2.3wombat→pig−4.5pig→peregrine+0.9peregrine→peregrine}}for a matrix that would be conventionally represented as ⎡⎢⎣01.1−4.52.300000.9⎤⎥⎦.
In code, we will write them like this:
```
vec1: 2.1 pig -3.2 wombat
mat1: 1.1 pig>wombat +2.3 wombat>pig -4.5 pig>peregrine + 0.9 peregrine>peregrine
```
for the above vector and matrix.
As with vectors, there are two ways to think of the matrix notation. In the first way, the semes form an orthonormal basis, and we are just using them to identify which pairs of coordinates get which coefficient. But, we can also think of {{1.1pig→wombat}} as being 1.1 times the [outer product](https://en.wikipedia.org/wiki/Outer_product) of ⟨⟨pig⟩⟩ and ⟨⟨wombat⟩⟩. This second view will not be necessary for the contents of this post, but it is necessary to understand some of the ways I envision being able to combine this work with gradient descent-based learning.
It is also worth pointing out that, if we multiply matrices and vectors with the vector on the left, that {{pig→wombat}} actually maps the vector ⟨⟨pig⟩⟩ to the vector ⟨⟨wombat⟩⟩. (Although, potentially confusingly, {{1.1pig→wombat}} maps ⟨⟨pig⟩⟩ to ⟨⟨1.1wombat⟩⟩.) For this reason, we will prefer left-multiplication in our neural networks later, because it makes this particular notation way easier to think with.
Tokenization and Word Embeddings
================================
In deep NLP, the first couple steps are about getting rid of words and replacing them with inputs that can actually be understood by a deep network. The first step is to take a string and break it into some number of discrete chunks called "tokens". In principle, we could feed things in letter-by-letter, and people have gotten semi-decent results doing that in the past, but it's a lot less labor-intensive in this context to use full words as the unit of tokenization. This is actually a mild break from most Transformer models used today, which generally make use of a "subword vocabulary" which contains a mixture of whole words and parts of words like "ing" or "particul".
Let's take an example sentence and tokenize it, just to be sure that we understand this process. Consider
```
The rain in Spain is mainly on the plain, while treefuls of weevils are gleefully evil.
["The", "rain", "in", "Spain", "is", "mainly", "on", "the", "plain", ",", "while", "treefuls", "of", "weevils", "are", "gleefully", "evil", "."]
```
Some things worth emphasizing:
* We don't use tokens for whitespace (spaces, tabs, etc.)
* Punctuation such as commas and periods will get a token of its own
Additionally, we will case-normalize our inputs by making everything lower-case. This cuts down on some repetitive work and is relatively common with deep models. We also include a special SOS (start of sentence) token and an EOS (end of sentence) token. So the above example should really look like this:
```
The rain in Spain is mainly on the plain, while treefuls of weevils are gleefully evil.
["SOS", "the", "rain", "in", "spain", "is", "mainly", "on", "the", "plain", ",", "while", "treefuls", "of", "weevils", "are", "gleefully", "evil", ".", "EOS"]
```
The second step in transforming a string of text into the sort of inputs that a deep network prefers is to do a "word embedding lookup". Here, each token is replaced by a fixed vector, so that we get a matrix of shape [num\_tokens, word\_embedding\_dim]. Because the first axis (the "sequence dimension") is not semantically the same as the second axis (the "embedding dimension"), we will not use our special matrix notation, but will instead think of this as a list of vectors, one for each token.
So let's look at some word embeddings! Here are some pronouns (note that we're describing here a fragment of a flavor of English that includes the gender-neutral singular "they" in addition to the plural "they"):
```
i: +nom +sg +1st +pro
you: +nom +sg +2nd +pro
he: +masc +nom +sg +3rd +pro
she: +fem +nom +sg +3rd +pro
it: +neut +sg +3rd +pro +expletive
me: +acc +sg +1st +pro
you: +sg +pl +2nd +pro
him: +masc +acc +sg +3rd +pro
we: +nom +pl +1st +pro
they: +enby +nom +sg +pl +3rd +pro
them: +enby +acc +sg +pl +3rd +pro
us: +acc +pl +1st +pro
them: +acc +pl +3rd +pro
my: +gen +sg +1st +pro
our: +gen +pl +1st +pro
his: +masc +gen +sg +3rd +pro
her: +fem +gen +acc +sg +3rd +pro
its: +neut +gen +sg +3rd +pro
their: +enby +gen +sg +pl +3rd +pro
myself: +1st +reflexive +sg +pro
ourselves: +1st +reflexive +pl +pro
yourself: +2nd +reflexive +sg +pro
yourselves: +2nd +reflexive +pl +pro
himself: +3rd +reflexive +sg +masc +pro
herself: +3rd +reflexive +sg +fem +pro
itself: +3rd +reflexive +sg +neut +pro
themselves: +3rd +reflexive +pl +enby +pro
oneself: +3rd +reflexive +sg +pro
```
Here are some verbs:
```
is: +be +verb +3rdsg +copula
be: +be +verb +plain +copula
was: +be +verb +preterite +copula +helper
did: +do +helper +verb +preterite +agentlack +themeposs
do: +do +helper +verb +plain +agentlack +themeposs
does: +do +helper +verb +3rdsg +agentlack +themeposs
have: +have +plain +helper +verb +agentposs +themeposs
has: +have +3rdsg +helper +verb +agentposs +themeposs
can: +can +plain +helper +modal
could: +can +preterite +helper +modal
may: +may +plain +3rdsg +helper +modal
might: +may +helper +modal
must: +must +plain +helper +modal
shall: +shall +plain +helper +modal
should: +shall +preterite +helper +modal
will: +will +plain +3rdsg +helper +modal
would: +will +preterite +helper +modal
ought: +ought +modal +helper +modal
dare: +dare +modal +helper +modal
accuse: +accuse +verb +plain +agentlack +themelack
accused: +accuse +verb +preterite +agentlack +themelack
accuses: +accuse +verb +3rdsg +agentlack +themelack
appear: +appear +verb +plain +agentlack +complementposs
appeared: +appear +verb +preterite +agentlack +complementposs
appears: +appear +verb +3rdsg +agentlack +complementposs
ate: +eat +verb +preterite +agentlack +patientposs
beam: +beam +verb +plain +agentlack
beamed: +beam +verb +preterite +agentlack
beams: +beam +verb +3rdsg +agentlack
bend: +bend +verb +plain +agentlack +patientposs
bent: +bend +verb +preterite +agentlack +patientposs
bends: +bend +verb +3rdsg +agentlack +patientposs
bled: +bleed +verb +preterite +agentlack +patientposs
bleed: +bleed +verb +plain +agentlack +patientposs
bleeds: +bleed +verb +3rdsg +agentlack +patientposs
blew: +blow +verb +preterite +agentlack +patientposs
blow: +blow +verb +plain +agentlack +patientposs
blows: +blow +verb +3rdsg +agentlack +patientposs
braid: +braid +verb +plain +agentlack +patientlack
braided: +braid +verb +preterite +agentlack +patientlack
braids: +braid +verb +3rdsg +agentlack +patientlack
breathe: +breathe +verb +plain +agentlack
breathed: +breathe +verb +preterite +agentlack
breathes: +breathe +verb +3rdsg +agentlack
break: +break +verb +plain +agentlack
breaks: +break +verb +3rdsg +agentlack
broke: +break +verb +preterite +agentlack
brush: +brush +verb +plain +agentlack +patientlack
brushed: +brush +verb +preterite +agentlack +patientlack
brushes: +brush +verb +3rdsg +agentlack +patientlack
carve: +carve +verb +plain +agentlack +patientposs
carved: +carve +verb +preterite +agentlack +patientposs
carves: +carve +verb +3rdsg +agentlack +patientposs
chase: +chase +verb +plain +agentlack +patientlack
chased: +chase +verb +preterite +agentlack +patientlack
chases: +chase +verb +3rdsg +agentlack +patientlack
chuckle: +chuckle +verb +plain +agentlack
chuckled: +chuckle +verb +preterite +agentlack
chuckles: +chuckle +verb +3rdsg +agentlack
came: +come +verb +preterite +agentlack
come: +come +verb +plain +agentlack
comes: +come +verb +3rdsg +agentlack
cook: +cook +verb +plain +agentlack +patientposs
cooked: +cook +verb +preterite +agentlack +patientposs
cooks: +cook +verb +3rdsg +agentlack +patientposs
cough: +cough +verb +plain +agentlack +patientposs
coughed: +cough +verb +preterite +agentlack +patientposs
coughs: +cough +verb +3rdsg +agentlack +patientposs
cried: +cry +verb +preterite +agentlack
cries: +cry +verb +3rdsg +agentlack
cry: +cry +verb +plain +agentlack
cut: +cut +verb +plain +preterite +agentlack +patientlack
cuts: +cut +verb +3rdsg +agentlack +patientlack
```
You'll note something strange about verbs (and nouns and other "content" words): they almost all have one seme that's just themselves again! What gives? These are semantic semes, which are much harder to reason about than the other semes, which are syntactic. As we said earlier, syntactic semes should be thought of as an orthonormal basis of whatever size, but semantic semes are more usefully thought of as living in a small dimensional space, where they aren't mutually orthogonal. (But! all semantic semes should be thought of as perfectly orthogonal to all syntactic semes, and vice versa.) For the time being, I restrict myself to a relatively limited vocabulary and just use the semantic semes as if they were orthogonal. For grammaticality classification, which is the domain I have worked hardest on, the semantic semes are not particularly relevant. For translation, the only really important thing is that they are able to pick out the corresponding word in the target language. (Assuming there is a straightforward single-word translation in the target language, which there has mostly been in the toy examples I have considered thus far, but which in general is not the case.)
Let's embed a sentence! Consider:
```
The cat sat on the mat.
```
This tokenizes to
```
["SOS", "the", "cat", "sat", "on", "the", "mat", ".", "EOS"]
```
Which then embeds to
```
SOS: +sos
the: +det
cat: +cat +sg +noun
sat: +sit +verb +preterite +agentlack
on: +on +prep
the: +det
mat: +mat +sg +noun
.: +punct +period
EOS: +eos
```
In mathematical notation, we would write this as [⟨⟨+sos⟩⟩,⟨⟨+det⟩⟩,⟨⟨+cat+sg+noun⟩⟩,⟨⟨+sit+verb+preterite+agentlack⟩⟩,⟨⟨+on+prep⟩⟩,⟨⟨+det⟩⟩,⟨⟨+mat+sg+noun⟩⟩,⟨⟨+punct+period⟩⟩,⟨⟨+eos⟩⟩]
For further clarity, let's give a gloss for each seme we're using:
```
semes: sos # start of sentence
. eos # end of sentence
det # determiner, a linguistic class that contains articles, demonstratives, and some other stuff
cat # meowing animal (semantic seme)
sg # singular in number
noun # nouns, the class of object/concept words
sit # sitting down (semantic seme)
verb # verbs, the class of action words
preterite # one of the past tenses in English
agentlack # to be grammatical, this verb needs an agent
on # the preposition (semantic seme)
prep # prepositions, the class of words denoting relationships
mat # something to sit on (semantic seme)
punct # punctutation
period # specifically this guy: .
```
Sentiment Analysis
==================
[Sentiment analysis](https://en.wikipedia.org/wiki/Sentiment_analysis) refers to the task of extracting from a piece of natural language the overall sentiment that the speaker has towards whatever thing they're talking about. For instance, in a movie or product review, does the author recommend the movie or product? This is generally considered a pretty straightforward task for machine learning algorithms.
An extremely interpretable algorithm for sentiment analysis is given in **VADER: A Parsimonious Rule-based Model for Sentiment Analysis of Social Media Text** [[PDF]](http://comp.social.gatech.edu/papers/icwsm14.vader.hutto.pdf), by C. Hutto, E. Gilbert. (2014) We have implemented a similar algorithm inside of a two-layer [vanilla RNN](https://victorzhou.com/blog/intro-to-rnns/), which we will describe below. However, we wanted to first note that this algorithm is only a crude sketch of VADER, and its shortcomings should not be held against Hutto and Gilbert.
Why a vanilla RNN rather than the OMG SO MUCH BETTER [LSTM](https://en.wikipedia.org/wiki/Long_short-term_memory)? Well, vanilla RNN's are significantly simpler to understand, and their disadvantages (vanishing and exploding gradients, primarily) are only really relevant when you're actually using gradient descent! So let's do this the easy way and stick to a vanilla RNN.
First, a few preliminaries about the architecture we will be using:
Let Hℓt denote the output of layer ℓ at time-step t. Note that the superscript isn't an exponent, it's just a convenient place to put another index. There will be no exponents anywhere in this network; they are all superscripts. (Here time-step just means the index of the token. So time-step 1 will be the first token, time-step 2 will be the second token, and so on.) Hℓ0 will be a special initial state before we read any tokens, and we will set it to be the zero vector. H0t will be the output of the word-embedding layer, so it will just be the embedding of the t-th token after we look it up.
We then define the recurrence (for ℓ=1,2):
Hℓt=σ[Hℓ−1t(I+Aℓ)+Hℓt−1Bℓ+bℓ]
Here σ is the good old [logistic sigmoid function](https://en.wikipedia.org/wiki/Sigmoid_function), and I is the identity matrix. (Practitioners will note that the use of the identity matrix here is some sort of residual-like connection.)
The logistic sigmoid functionWe further define a pooling layer, and then a fully-connected or dense layer.
X0=meantH2t
X1=σ[X0C1+c1]
```
# our set of semes
semes: stop # "stop words", in this case anything not needed for sentiment analysis
positive negative negation contrastive intensifier
lessener intensepunctuation
xa xb xc xd xe ya yb yc yd ye # a bunch of anonymous variables to represent intermediate computations
lexicon:
",": stop
".": stop
"I'll": stop
At: stop
It: stop
The: stop
Today: stop
VADER: stop
a: stop
all: stop
and: stop
are: stop
at: stop
book: stop
by: stop
characters: stop
dialog: stop
get: stop
is: stop
it: stop
of: stop
plot: stop
the: stop
was: stop
FUNNY: positive
GOOD: positive
GREAT: positive
HANDSOME: positive
LOL: positive
SMART: positive
funny: positive
good: positive
great: positive
handsome: positive
lol: positive
smart: positive
SUX: negative
bad: negative
horrible: negative
sux: negative
uncompelling: negative
"!": intensepunctuation
"!!!": intensepunctuation
very: intensifier
VERY: intensifier
uber: intensifier
FRIGGIN: intensifier
only: lessener
kinda: lessener
not: negation
nor: stop
isnt: negation
Not: negation
But: contrastive
but: contrastive
rnn_layer1:
A: positive>xa negative>ya positive>xb negative>yb
negation>negation intensifier>intensifier lessener>lessener
B: intensifier>xa intensifier>ya lessener>xb lessener>yb
negation>negation 0.5intensifier>intensifier 0.5lessener>lessener
bias: -xa -xb -ya -yb
rnn_layer2:
A: xa>xc ya>yc xb>xd yb>yd negation>xc negation>xd negation>yc
negation>yd positive>xe negation>xe negative>ye negation>ye
B: '' # in yaml, which is the formatting language I use to type these programs, you need to do this to specify an empty string, which corresponds to the zero matrix
bias: -xc -xd -yc -yd -xe -ye
dense1:
C: positive>positive negative>negative 2xa>positive 0.25xb>positive
ya>negative 0.25yb>negative xc>negative xd>negative yc>positive
yd>positive -2xc>positive -2xd>positive -2yc>negative
-2yd>negative xe>negative ye>positive -xe>positive -ye>negative
c: ''
examples: # examples modified from https://github.com/cjhutto/vaderSentiment
- VADER is smart , handsome , and funny .
- VADER is smart , handsome , and funny !
- VADER is very smart , handsome , and funny .
- VADER is VERY SMART , handsome , and FUNNY .
- VADER is VERY SMART , handsome , and FUNNY !!!
- VADER is VERY SMART , uber handsome , and FRIGGIN FUNNY !!!
- VADER is not smart , handsome , nor funny .
- The book was good .
- It isnt a horrible book .
- The book was only kinda good .
- The plot was good , but the characters are uncompelling and the dialog is not great .
- Today SUX !
- Today only kinda sux ! But I'll get by , lol
- Not bad at all
```
This is sufficient to generate the scores at the beginning of this post. The scores on the given examples are not all that inaccurate. Lots more work could obviously be done on this network, and I'd love it if people feel like working on this network or other later-discussed networks in the comments.
Transformer Overview
====================
We give here a brief overview of the transformer architecture, for those unfamiliar with it. This will essentially be an accelerated recap of Jay Alammar's [Illustrated Transformer](http://jalammar.github.io/illustrated-transformer/), which I consider to be the best friendly introduction to the architecture.
Transformer in its full sequence-to-sequence glory has an encoder stack and a decoder stack. For text classification purposes, one generally just uses the encoder stack with a few simple layers at the end. The encoder stack is made up of a bunch of Transformer encoder blocks, each of which is the same architecturally, but each of which has its own learnable/settable weights that allow it to specialize and do its own particular task in the grand scheme of the network.
Stolen from Jay Alammar's Illustrated TransformerThe decoder stack is also made up of a series of architecturally identical Transformer layers, again each with their own learnable/settable weights that allow them to specialize into their own unique role. The decoder layers are similar to the encoder layers, but a little bit more complex.
So now let's dive inside the Transformer layer and see how they tick!
Stolen from Jay Alammar's Illustrated TransformerSo we have self-attention layers, encoder-decoder attention layers, and the feed-forward layers. The two types of attention layers are generally considered to be the innovative, "important" part of Transformer, but I have found in trying to hard-code weights for transformer (and years of research by many people has also found) that the feed-forward layers are crucial to being able to learn complex functions. (Well, that's not entirely true. I think someone in some paper managed to sort of smuggle the feed-forward layer into a computation that looks like self-attention over a learnable set of parameters, without losing any accuracy. But for our purposes the feed-forward layer is important.)
We'll next dive into the various layers and see how to hard-code their parameters. We'll start with the easiest layer to understand: the feed-forward layer.
Transformer Feed-Forward Layers
===============================
The standard transformer feed-forward layer can be described pretty simply as:
FFN(x)=b+ReLU(a+x⋅A)⋅BHere ReLU is a new kind of non-linearity, the "[rectified linear unit](https://en.wikipedia.org/wiki/Rectifier_(neural_networks))". A and B are matrices called "weights", and a and b are vectors called "biases". "Parameters" refers to either weights or biases, although people will often refer to biases as "weights" also.
ReLU and its cousin GELU, stolen from WikipediaLike most non-linearities in deep learning, ReLU is an element-wise function, i.e., you apply it to each coordinate of the vector independently. ReLU(x)=max(x,0), so if x is negative, ReLU(x)=0, and otherwise ReLU(x)=x.
So let's hard-code a feed-forward layer! This just requires picking values for A,B,a,b.
```
semes: apple banana cherry durian yum yuck
mat1: apple>apple apple>yum banana>banana banana>yum cherry>yuck durian>yuck
bias1: -yum -yuck
mat2: apple>yum banana>yum -yum>yum yuck>yuck
bias2: '' # a zero vector
```
The semantics we are trying to encode here is that ⟨⟨apple⟩⟩ OR ⟨⟨banana⟩⟩ should be mapped to ⟨⟨yum⟩⟩, while ⟨⟨cherry⟩⟩ AND ⟨⟨durian⟩⟩ should be mapped to ⟨⟨yuck⟩⟩. This results in the following output:
You could argue that ⟨⟨yuck⟩⟩ should displace ⟨⟨yum⟩⟩. This can be done, but it seems to require a third layer (I haven't given it much thought just now, so maybe it's doable in two layers), so it would have to be split across two different Transformer layers.
Preparing for Self-Attention: Positional Encodings
==================================================
Before we can really dive into self-attention layers, it is useful to talk about positional encodings. This is a fairly technical aspect of the Transformer architecture, but I've found that it can be made fairly interpretable by thinking about it in the right way.
The traditional way to think about positional embeddings is just to read the following code, and then add some handwaving around "relative positional offsets can be encoded as linear combinations of sines and cosines". This is all correct, but it doesn't really yield enough understanding (for me at least) to hard-code things around the positional embeddings.
```
# Code from https://www.tensorflow.org/tutorials/text/transformer
def get_angles(pos, i, d_model):
angle_rates = 1 / np.power(10000, (2 * (i//2)) / np.float32(d_model))
return pos * angle_rates
def positional_encoding(position, d_model):
angle_rads = get_angles(np.arange(position)[:, np.newaxis],
np.arange(d_model)[np.newaxis, :],
d_model)
# apply sin to even indices in the array; 2i
angle_rads[:, 0::2] = np.sin(angle_rads[:, 0::2])
# apply cos to odd indices in the array; 2i+1
angle_rads[:, 1::2] = np.cos(angle_rads[:, 1::2])
pos_encoding = angle_rads[np.newaxis, ...]
return pos_encoding
```
I prefer to think of the various sines and cosines as the hands of a bunch of clocks running at different speeds. If we have 512 dimensions in our positional embeddings, then there will be 256 clocks. The sine is just the y-coordinate of the clock hand, and the cosine is the x-coordinate of the clock hand.
At time t=0, we look at the first word in the sequence. All clocks point in the same direction.
t=0At time t=1, we look at the second word in the sequence. The slowest clock (all the way on the left) has advanced a tiny bit, while the fastest clock (all the way on the right) has advanced more.
t=1At time t=4, the slowest clock has advanced a decent amount, while the fastest clock has advanced about a quarter rotation from where it started.
t=4Now that we have a more grounded understanding, we can ask questions like, if I have the positional encoding of "very", how do I get the positional encoding of the next word? Well, the next word is one time-step further, so each clock should advance by the amount that that particular clock advances over one time step. We can "point" to that positional embedding by using a particular angle offset for each clock, which then translates into the specific linear combination of sines and cosines referenced earlier. Thus, I have an intentionally quite redundant way to refer to a specific number of time steps in the future.
The different speeds also provide me the ability to point to ranges of time-steps relatively easily. If I want to point to a very narrow range, I can use the fastest clock, which will pick out a very specific time-step, with little room for error. If I want to refer to a broader time-range, I can use a slower clock, which will have the effect that my pointing will be somewhat evenly spread over a wide range of time-steps. Since I have so many clocks to choose from, I can be quite precise in what sort of time interval I point at.
Finally, as a technical note, it should be understood that the range of clock speeds are chosen such that no clock does a full loop. Otherwise, our pointing might accidentally point at something that's done a full loop without understanding that that is occurring.
The notion of "pointing" will prove to be a very apt way of thinking about the mechanism of self-attention. Basically, words that interact with each other in the self-attention layer can be thought of as pointing at the word they interact with.
Self-Attention
==============
Finally, we come to the heart of the Transformer model, the self-attention layer. This can be mathematically expressed as
SA(Q,K,V)=V⋅SOFTMAX(QKT√d)where Q,K,V are matrices called the queries, key, and values, respectively. The \sqrt{d} factor, often called a "temperature", is important for gradient descent training, where it improves stability. (I would guess that it improves stability by making the attention matrix relatively balanced early on in training.) For us, we will have the opposite problem; it's much easier to think about approximately sparse matrices, so we will actually use something like this, with α being a fairly large scalar.
SA(Q,K,V)=V⋅SOFTMAX(αQKT)There's a lot to unpack here, so let's come up with a very simple example and work through it together.
Suppose we want to distinguish between these two sentences:
```
She saw a red apple. # grammatical
She saw a red. # not grammatical (ish, you can provide contexts in which it is natural)
```
We will try not to go too far down the rabbit hole of what makes something grammatical or not - here we're just trying to encode the simple rule (which does have limited exceptions) that adjectives can't just hang out without modifying anything, or being used with a linking verb, or in some other way being "licensed" by the other words in the sentence.
So we would like to be able to detect when an adjective seems to be modifying a noun in the usual way, versus when it is not. In my special programming language for self-attention layers, this can be done like this:
```
H1a: # name of the head
docstring: Modification layer. Specifically pairs of the form
Q K, where Q comes before K and Q modifies K. Q
must be an adjective or an adverb to use this rule.
pos: # special notation for interacting with the positional encodings
Q: 0
K: +1
x1: # adjective modifies noun: red apple
Q: adjective
K: noun
x2: # adverb modifies verb: quickly write
Q: adverb
K: verb
x3: # adverb modifies adjective: very slow
Q: adverb
K: adjective
x4: # adverb modifies adverb: very slightly
Q: adverb
K: adverb
x5: # everything else hits filler
Q: verb noun det filler pro verb noun det filler pro
K: filler
int: noun>licensed verb>licensed adverb>licensed
adjective>licensed
```
The queries and keys live in a small dimensional embedding space that I call key-space. (Typical size in a standard transformer is 64 dimensions - much smaller than the 512/768/1024 hidden size.) There is one query vector for each token, and one key vector for each token. We take the dot product of every query vector with every key vector, and that gives us what are called "attention logits". Applying the softmax function to the attention logits gives us attention probabilities, which make up what is generally call the "attention matrix". So let's compute all of these values for the above program and our example sentences to get a sense of how all of this works.
Let's suppose that we have just run the word embedding lookup layer, so that we have the default embedding for each word, but no information yet about how the words are interacting with each other. That might look like this (omitting the positional embeddings for the time being):
```
SOS: +filler +sos
she: +pro +fem +sg +nom
saw: +saw +verb +agentlack +perceptlack
a: +det +sg
red: +red +adjective
apple: +apple +noun +sg
EOS: +filler +eos
------------
SOS: +filler +sos
she: +pro +fem +sg +nom
saw: +saw +verb +agentlack +perceptlack
a: +det +sg
red +red +adjective
EOS: +filler +eos
```
The queries will then be computed as follows (again, omitting the positional components):
```
SOS: +2 x5
she: +2 x5
saw: +2 x5
a: +2 x5
red: +x1
apple: +2 x5
EOS: +2 x5
------------
SOS: +2 x5
she: +2 x5
saw: +2 x5
a: +2 x5
red: +x1
EOS: +2 x5
```
The keys will look like this (again, omitting the positional components):
```
SOS: +x5
she: 0 # 0-vector
saw: +x2
a: 0
red: +x3
apple: +x1
EOS: +x5
------------
SOS: +x5
she: 0 # 0-vector
saw: +x2
a: 0
red: +x3
EOS: +x5
```
Supposing that x1, x2, x3, x4, and x5 are orthonormal, this creates the following attention logits, where we omit zeros for brevity (and we're still omitting positional encodings):
```
# written in the form Q>K
SOS>SOS: 2
SOS>EOS: 2
she>SOS: 2
she>EOS: 2
saw>SOS: 2
saw>EOS: 2
a>SOS: 2
a>EOS: 2
red>apple: 1
apple>SOS: 2
apple>EOS: 2
EOS>SOS: 2
EOS>EOS: 2
------------
SOS>SOS: 2
SOS>EOS: 2
she>SOS: 2
she>EOS: 2
saw>SOS: 2
saw>EOS: 2
a>SOS: 2
a>EOS: 2
EOS>SOS: 2
EOS>EOS: 2
```
We will continue to ignore the positional encodings for the rest of this example, since they're not needed, and don't drastically change the attention matrix. (They would be needed in the case that there are two nouns, in which case the above weights would generate a tie between red pointing to apple and red pointing to the other noun, which we would want to resolve a certain way based on the rule that, in English, an adjective is close to the noun it modifies and before it, with some limited counter-examples.)
Now let's look at the values! For simplicity in our programs, we actually combine the value projection and what is usually called the out-projection. We call the combined quantity the "interpretant" (a term from semiotics). The interpretants for us are as follows:
```
SOS: 0
she: 0
saw: +licensed
a: 0
red: +licensed
apple: +licensed
EOS: 0
------------
SOS: 0
she: 0
saw: +licensed
a: 0
red: +licensed
EOS: 0
```
Multiplying the interpretants V by the attention matrix SOFTMAX(αQKT), we get the following outputs (for large α)
```
SOS: 0
she: 0
saw: 0
a: 0
red: +licensed
apple: 0
EOS: 0
------------
SOS: 0
she: 0
saw: 0
a: 0
red: 0
EOS: 0
```
Using the residual connection that surrounds the self-attention layer, we then receive final outputs:
```
SOS: +filler +sos
she: +pro +fem +sg +nom
saw: +saw +verb +agentlack +perceptlack
a: +det +sg
red: +red +adjective +licensed
apple: +apple +noun +sg
EOS: +filler +eos
------------
SOS: +filler +sos
she: +pro +fem +sg +nom
saw: +saw +verb +agentlack +perceptlack
a: +det +sg
red +red +adjective
EOS: +filler +eos
```
Thus, this self-attention layer has managed to compute the fact that the word "red" is licensed in the first, grammatical sentence, but not licensed in the second, questionable sentence. This can be used by downstream layers to declare the second sentence to be ungrammatical.
Phew!
More to Come!
=============
We'll next see how to hard-code the weights of an actual Transformer. This will involve explaining the structures of Transformer layers, which will take a fair amount of time. In the meantime, please check out [The Illustrated Transformer](http://jalammar.github.io/illustrated-transformer/) and [Transformers from Scratch](http://peterbloem.nl/blog/transformers) to get a head start on understanding them, or dive face-first into [my 1800-line poorly-commented grammaticality classifier](https://github.com/epurdy/hand/blob/master/api/syntax.att). |
d3ce16a0-3e11-4ae0-88d4-190c0f8b3ae9 | trentmkelly/LessWrong-43k | LessWrong | Pacing Outside the Box: RNNs Learn to Plan in Sokoban
Work done at FAR AI.
There has been a lot of conceptual work on mesa-optimizers: neural networks that develop internal goals that may differ from their training objectives (the inner alignment problem). There is an abundance of good ideas for empirical work (find search in a NN, interpret it), but very little actual execution, partly because we did not have a clear-cut example of a mesa-optimizer to study. Until now.[1]
We have replicated the mesa-optimizer that Guez et al. (2019) found, and released it open-source as a model organism for inner alignment research. In brief, Guez et al. trained a recurrent neural network (RNN) with model-free RL to play Sokoban. They noticed that if you give the RNN more time to think by repeating the initial observation at inference time, its performance increases. This is highly suggestive of planning!
We investigate this "planning effect" in a black-box way. We find that often, the RNN learns to "pace" before attempting to solve the level, likely to get more computation and find a solution. When we give the RNN time to think, it finds the solution in the extra thinking time and executes it straight away.
In other cases, the RNN sometimes starts with a greedy solution and locks itself out of the solution. With thinking time, the RNN finds the non-myopic solution, avoiding the lock and solving the level. Note that this greedy behavior may be bounded-rational given the -0.1 penalty per step: solving fewer levels but solving them more quickly can pay off.
These are illustrative examples, but we have quantitative evidence too. We operationalize the pacing behavior as whatever creates a cycle in the sequence of environment states. If we give the RNN time to think at level start, it does not 'pace' anymore: 75% of cycles that occur in the first 5 steps disappear. Time to think in the middle of a level also substitutes cycles: 82% of N-step cycles disappear with N steps to think.
The levels we use always have 4 boxes. Thinking t |
293e2be9-2753-44ce-aa1c-7482e02ad5ad | trentmkelly/LessWrong-43k | LessWrong | My argument against AGI
This is the third post about my argument to try and convince the Future Fund Worldview Prize judges that "all of this AI stuff is a misguided sideshow". My first post was an extensive argument that unfortunately confused many people.
(The probability that Artificial General Intelligence will be develop)
My second post was much more straightforward but ended up focusing mostly on revealing the reaction that some "AI luminaries" have shown to my argument
(Don't expect AGI anytime soon)
Now, as a result of answering many excellent questions that exposed the confusions caused by my argument, I believe I am in a position to make a very clear and brief summary of the argument in point form.
To set the scene, the Future Fund is interested in predicting when we will have AI systems that can match human level cognition: "This includes entirely AI-run companies, with AI managers and AI workers and everything being done by AIs." This is a pretty tall order. It means systems with advanced planning and decision making capabilities. But this is not the first time people predicted that we will have such machines. In my first article I reference a 1960 paper which states that the US Air Force predicted such a machine by 1980. The prediction was based on the same "look how much progress we have made, so AGI can't be too far away" argument we see today. There must be a new argument/belief if today's AGI predictions are to bear more fruit than they did in 1960. My argument identifies this new belief. Then it shows why the belief is wrong.
Part 1
1. Most of the prevailing cognitive theories involve classical symbol processing systems (with a combinatorial syntax and semantics, like formal logic). For example, theories of reasoning and planning involve logic like processes and natural language is thought by many to involve phrase structure grammars, like for example Python does.
2. Good old-fashioned AI was (largely) based on the same assumption, that classical symbol systems a |
0a1f00e1-4d30-487c-a43a-b05768353fcc | trentmkelly/LessWrong-43k | LessWrong | Harry Potter in The World of Path Semantics
This is a short story written with help from ChatGPT, about the intuition behind counter-examples to Leibniz' first principle. This is generally believe to be true among logicians, but the equality as it is expressed in Leibniz' first principle is just an assumption about equality and not a proof of equality.
x=y → ∀F(Fx ↔ Fy) Leibniz' first principle
In short, Leibniz' first principle does not hold for all operators in mathematics, because not all operators are congruent by normal equality.
It is not possible to reason about proofs of equality directly in logic (this is not about terms of types as proofs, but about "actual proofs"). There is a weaker statement of tautological equality that is used to provide this intuition.
If x=y is taken as "tautological equality", then the principle holds, but since tautological equality can be assumed, this does not prove that all properties are the same in the strong sense, but only a weak sense (up to provability within the language).
Now, it is important to remember that this counter-example is relative to the theory of Path Semantics. In the theory of Path Semantics, reasoning with symbolic indistinction (what would be considered actual numerical sameness) is not considered safe, only symbolic distinction. This is why it gets so complex, as there is no way to access the "actual" proof directly.
In Path Semantics, the actual counter-example to Leibniz' first principle is the qubit operator ~ which has tautological congruence. This is not part of the short story and takes a lot more background knowledge to understand. For those interested, one can read about the classical model of path semantical qubit here:
https://github.com/advancedresearch/path_semantics/blob/master/papers-wip2/the-classical-model-of-path-semantical-qubit.pdf
The Pocket-Prover library contains an implementation of the classical model: https://crates.io/crates/pocket_prover
The constructive model is more complex and uses a higher order lo |
6dbaa28e-dc7c-4d98-8be0-98af691f8d49 | trentmkelly/LessWrong-43k | LessWrong | Open thread, 16-22 June 2014
Previous open thread
If it's worth saying, but not worth its own post (even in Discussion), then it goes here.
Notes for future OT posters:
1. Please add the 'open_thread' tag.
2. Check if there is an active Open Thread before posting a new one.
3. Open Threads should be posted in Discussion, and not Main.
4. Open Threads should start on Monday, and end on Sunday. |
3cea7b13-12dd-452d-b2d6-860f0ae822ca | StampyAI/alignment-research-dataset/blogs | Blogs | 2020-21 New Year review
This is an annual post reviewing the last year and making resolutions and predictions for next year. 2020 brought a combination of challenges from living in a pandemic and becoming a parent. Other highlights include not getting sick, getting a broader perspective on my life through decluttering, and going back to Ukraine for the first time. (This post was written in bits and pieces over the past two months.)
#### 2020 review
*Life updates:*
Janos and I had a son, Daniel, on Nov 11. He arrived almost 3 weeks later than expected (apparently he was waiting to be born on my late grandfather’s birthday), and has been a great source of cuddles, sound effects and fragmented sleep ever since.
[](https://vkrakovna.files.wordpress.com/2021/01/20201117_113230.jpg)1 week old
[](https://vkrakovna.files.wordpress.com/2021/01/20201224_113030.jpg)6 weeks old
Some work things also went well this year – I had a paper accepted at NeurIPS, and was promoted to senior research scientist. Also, I did not get covid, and survived half a year of working from home (much credit goes to the great company of my housemates). Overall, a lot of things to be grateful for.
*AI safety research:*
* Wrote a paper on [avoiding side effects by considering future tasks](https://arxiv.org/abs/2010.07877), providing some theoretical grounding for the side effects problem, which was accepted to NeurIPS 2020.
* Contributed theoretical results for a project on the tampering problem and coauthored two papers: [Avoiding Tampering Incentives in Deep RL via Decoupled Approval](https://arxiv.org/abs/2011.08827) and [REALab: An Embedded Perspective on Tampering](https://arxiv.org/abs/2011.08820).
* Wrote a blog post on the specification gaming problem for the DeepMind blog: [Specification gaming: the flip side of AI ingenuity](https://deepmind.com/blog/article/Specification-gaming-the-flip-side-of-AI-ingenuity).
*Effectiveness:*
Working from home was definitely a productivity hit. I was mostly focused on urgent tasks, such as conference submissions and reviewing, and didn’t get much research done. I did 366 hours of “deep work” (1.95 hours per work day) this year, compared to 551 hours (2.4 hours per work day) in 2019. This includes theory work, reading papers, writing papers and code, but not editing text or debugging. I got back into using work cycles, which added some helpful structure in the home environment.
[](https://vkrakovna.files.wordpress.com/2022/12/image-4.png)
I was very grateful to be living in a group house during the pandemic. While it was a bit tricky to have 5 people sharing the space when working from home, it was awesome to have an in-person community and not feel completely isolated from the world. It was also much easier to do nonzero exercise when I had someone to do it with, e.g. running in the park together.
Spending a lot of time at home inspired me to do a lot of decluttering. In particular, I went through all my old notes, got rid of most of them, and gathered the ones that still seem interesting and relevant (notes from rationality workshops, Hamming worksheets, reflections and so on). I put these into a binder for easy reading, and found it useful for getting a big picture sense of how my attitudes and problems have evolved over time. This has been particularly helpful during the pandemic, when my life has often felt small and repetitive.
I got a UK driving license for automatic cars, which took a surprising amount of practice given that I already had a US license. There was a lot to get used to with the left side of the road, the narrowness of the streets and frequent maneuvering – I spent a number of lessons just on getting the positioning right. I did the theory test in March and planned to take the road test in early summer, but then the driving schools closed for lockdown, and I ended up starting the lessons in July. I took the road test in September and didn’t pass because of “undue hesitation” at a busy roundabout, so I had to repeat the test in October, two weeks before Daniel was due (thankfully, I passed this time and could forget about driving for a while).
*Health:*
Physical health has been pretty good this year. Last year I had 7 colds, while this year I was not sick at all – probably due to social distancing and taking zinc regularly. Thankfully, recovery after the birth was relatively quick, feeling mostly normal in around 2 weeks, though it will take some time to get my core muscles back online. I’ve been getting some back pain from lifting Daniel (now at 6kg thanks to his voracious appetite), which makes it all the more important to rebuild core strength.
The second half of the year came with pretty bad sleep – a lot of insomnia in the last trimester where I woke up at 3-4am for no discernible reason and couldn’t fall asleep again, followed by fragmented sleep after Daniel was born. Living and working on 4-5 hours of sleep before I went on parental leave was surprisingly ok, probably because I was waking up on my own rather than being woken up in the middle of a sleep cycle. On the other hand, being woken up by a hungry baby definitely feels more meaningful than waking up at 3am for no reason and not being able to go back to sleep.
[](https://vkrakovna.files.wordpress.com/2021/01/insomnia.png)Rate of insomnia by month
[](https://vkrakovna.files.wordpress.com/2021/01/hours-awake.png)Average hours awake at night by month
[](https://vkrakovna.files.wordpress.com/2021/01/hours-sleep.png)Average hours of sleep by month
This year has been pretty hard on my mental health due to a number of ways that the pandemic interacted with having a kid, and various problems that I considered solved have made a comeback lately. I spent most of the year at home without many forms of self-care, such as my usual exercise, sufficient sleep, or nice things like going to the sauna. While the birth went well, there were a lot more stressful interactions with the healthcare system than I had hoped. After that, there has been a combination of sleep deprivation, limited daylight, a mostly empty house after some housemates moved to the countryside, difficulty with meeting friends outside because of cold weather, making increasingly modest plans only to have them shot down by the ever-changing lockdown rules, and the exciting new covid strain we have in London that calls for high levels of caution and isolation. Thankfully, my mom was able to come stay with us for a few months to help with the baby, make food and keep us company.
*Travel:*
In January I visited Ukraine for the first time since I emigrated 17 years ago. I saw my dad and aunt, as well as my niece and her kids who live in a remote part of Canada but happened to be in Ukraine for the winter. It was an interesting experience to navigate around Kyiv – I no longer had a map of the city in my head, so I recognized some familiar places but could not recall where they are relative to each other, so this felt like visiting a new city with a lot of deja vu. I was pleasantly surprised by the large number of Georgian restaurants in Kyiv, which we made sure to frequent and were not disappointed.
[](https://vkrakovna.files.wordpress.com/2021/01/img_20200113_142052.jpg)
In March we did a week-long meditation retreat at MAPLE in the US. We hesitated whether to go ahead with this plan given that flights might get canceled, but ultimately decided to go. The retreat was in a remote location in Vermont that seemed pretty safe from a covid perspective. I was advised to follow an equanimity practice that worked pretty well (focusing on acceptance rather than observation of things that come into my awareness).
[](https://vkrakovna.files.wordpress.com/2021/01/20200312_092355.jpg)Janos in his element
We had a peaceful week meditating among the snows, which unfortunately became less peaceful at the end when some people started coughing, so we spent the last couple of days meditating in masks and gloves, and left as soon as the retreat ended. Our flight back did get canceled, but we were rebooked on another one for free. Upon returning home, we self-isolated from our housemates and acquired covid tests, which were thankfully negative. We later learned that several people at the retreat tested positive for covid, so this was a close call.
In August we went camping in North Wales at a friend’s cottage (after the first lockdown was lifted). The cottage itself was abandoned (and a bit spooky), but we could stay on the adjacent land and thus avoid crowds at the newly opened campsites. We enjoyed a lot of swimming in cold waterfalls and a much warmer Atlantic ocean. We also hiked up a nearby mountain Arenig Fawr (Daniel and I were taking it slow).
[](https://vkrakovna.files.wordpress.com/2021/01/20200802_142928.jpg)View from the summit of Arenig Fawr
In September we had a vacation in Madeira, where everyone was tested for covid on arrival at the airport, and there was no community transmission at the time. Madeira is a volcanic island that is basically one big mountain, and we had an interesting time driving around it (on the way to a hike, our car refused to go up a very steep road and we took a taxi the rest of the way). The terrain was a great combination of ocean and mountains.
[](https://vkrakovna.files.wordpress.com/2021/01/img_2755.jpg)
We enjoyed large quantities of Portuguese food (as we soon learned, it did not come in small quantities). Our special favorites were local rock mussels called limpets – we ate enough of them to have a shell stacking competition.
[](https://vkrakovna.files.wordpress.com/2020/12/image.png)The winner built a tower of 22, which collapsed before we could take a photo, so here is a tower of 13.
In December we did two short hikes near London, and verified that the basic algorithm of putting Daniel in a car seat and then in a carrier for the hike seems to work pretty well (he mostly sleeps through all this). I’m glad to be able to visit some nature during these strange times, it makes the world feel just a bit less small.
[](https://vkrakovna.files.wordpress.com/2021/01/20201230_133242.jpg)Iron age hill fort in Epping Forest
#### 2020 prediction outcomes
Unsurprisingly, some predictions for the past year were messed up by the pandemic.
*Resolutions:*
* Author or coauthor three or more academic papers (3 last year) (70%) – yes (3 papers)
* At most 12 non-research work commitments, such as speaking and organizing (10 last year) (80%) – yes (5 commitments). Easy, since a lot of events got canceled.
* ~~Meditate on at least 270 days (290 last year)~~ (80%) – no (244 days). The past month I only managed to meditate on 9 days, and this has not been good for me, so I need to do better next year.
* Read at least 7 books (5 last year) (70%) – yes (9 books). Human Compatible, The Precipice, Secret of our Success, Ender’s Game, In the Realm of Hungry Ghosts, Watching the English, Positive Birth Book, The Gardener and the Carpenter, Raising a Secure Child.
* ~~At least 700 deep work hours (551 last year)~~ (70%) – no (366 hours). I found it much harder to do deep work at home, and was on parental leave for the last 2.5 months of the year.
*Predictions:*
* ~~I will write at least 5 blog posts~~ (60%) – no (3 posts)
* ~~Eating window at most 11 hours on at least 240 days~~ (228 last year) (70%) – no (131 days), since I stopped doing intermittent fasting this year
* ~~I will visit at least 4 new cities with population over 100,000~~ (11 last year) (70%) – no (2 cities, Birmingham and Funchal). Much less travel than normal this year.
* ~~At most 1 housemate turnover at Deep End~~ (70%) – no (2 housemates). One housemate moved to live with parents in the countryside who would have likely stayed in London under normal circumstances.
* ~~I finish a language in Duolingo~~ (60%) – no, though made some progress on Mandarin (completed checkpoint 2)
#### 2021 resolutions and predictions
*Resolutions:*
* Avoid catching covid (90%)
* Author or coauthor three or more academic papers (3 last year) (70%)
* At most 7 non-research work commitments (5 last year) (80%)
* Meditate on at least 230 days (244 last year) (70%)
* At least 450 deep work hours (366 last year) (70%)
* Do 4 consecutive chinups (70%)
Predictions:
* I will write at least 3 blog posts (3 last year) (60%)
* Janos and I will get vaccinated for covid by the end of June (60%)
* Daniel will get to meet all of his grandparents in person in 2021 (70%)
* I will return to avoiding processed sugar by the end of the year (60%)
* I will finish Hungarian Duolingo (complete checkpoint 5) (70%)
Past new year reviews: [2019-20](https://vkrakovna.wordpress.com/2020/01/09/2019-20-new-year-review/), [2018-19](https://vkrakovna.wordpress.com/2019/01/01/2018-19-new-year-review/), [2017-18](https://vkrakovna.wordpress.com/2018/01/07/2017-18-new-year-review/), [2016-17](https://vkrakovna.wordpress.com/2017/01/09/2016-17-new-year-review/), [2015-16](https://vkrakovna.wordpress.com/2015/12/31/2015-16-new-year-review/), [2014-15](https://vkrakovna.wordpress.com/2015/01/11/2014-15-new-year-review/). |
5e2431db-fd26-4d58-8e46-43869083e9ea | StampyAI/alignment-research-dataset/arxiv | Arxiv | Reinforcement Learning with Action-Free Pre-Training from Videos.
1 Introduction
---------------
Deep reinforcement learning (RL) has made significant advance in solving various sequential decision-making problems (mnih2015human; levine2016end; silver2017mastering; vinyals2019grandmaster; berner2019dota; akkaya2019solving; kalashnikov2021mt).
However, existing RL methods often start learning tabula rasa without any prior knowledge of the world, therefore requiring a large amount of environment interaction for learning meaningful behaviors. By contrast, within the computer vision (CV) and natural language processing (NLP) domains, recent unsupervised pre-training approaches have shown to be effective by leveraging the pre-trained representations for fine-tuning in downstream tasks (mikolov2013efficient; pennington2014glove; noroozi2016unsupervised; gidaris2018unsupervised; devlin2018bert; radford2018improving; he2020momentum).
Figure 1: We pre-train an action-free latent video prediction model using videos from different domains (left), and then fine-tune the pre-trained model on target domains (right).
Recent works have shown promise in adopting such pre-training and fine-tuning paradigm to vision-based RL, by demonstrating that representations pre-trained with various unsupervised representation learning schemes can be effective in downstream tasks (finn2016deep; dwibedi2018learning; zhan2020framework; laskin2020curl; stooke2021decoupling; schwarzer2021pretraining).
Notably, stooke2021decoupling show that weight initialization with contrastively pre-trained representations leads to performance improvement.
These works, however, mostly focus on the setup where pre-training datasets are collected in the target domains, or in different domains but with very similar visuals.
Instead, we would like to leverage videos from diverse domains for pre-training, and transfer the pre-trained representations for solving newly encountered domains.
In this paper, we present APV: Action-Free Pre-training from Videos, a novel framework that performs generative pre-training on videos for improving the sample-efficiency of vision-based RL.
Since our goal is to learn the representations that can be transferred to various downstream tasks from readily available videos, our framework do not require the videos to be collected in the same domain of the downstream tasks, and also do not assume the datasets contain action information.
Summarized in [Figure 1](#S1.F1 "Figure 1 ‣ 1 Introduction ‣ Reinforcement Learning with Action-Free Pre-Training from Videos"), our framework comprises two phases: we first pre-train an action-free latent video prediction model to learn useful representations from videos, then fine-tune the pre-trained model for learning action-conditional world models on downstream tasks.
To leverage the fine-tuned world models for behavior learning, we build APV on top of DreamerV2 (hafner2020mastering).
The key ingredients of APV are as follows:
* [topsep=0pt,itemsep=0pt,leftmargin=\*]
* Action-free pre-training from videos: To capture rich dynamics information from diverse videos, we pre-train an action-free latent video prediction model.
We find that the representations from the pre-trained model can be transferred to various downstream tasks.
* Stacked latent prediction model: To incorporate additional action inputs during fine-tuning,
we introduce a new architecture that stacks an action-conditional latent dynamics model on top of the action-free model.
* Video-based intrinsic reward: For better exploration, we propose an intrinsic bonus that utilizes video representations from the action-free model.
Since the pre-trained representations contain information useful for understanding dynamics of environments,
our intrinsic bonus effectively encourages agents to learn diverse behaviors.
In our experiments, we pre-train the action-free prediction model using 4950 videos collected on 99 manipulation tasks from RLBench (james2020rlbench) and fine-tune the pre-trained model on a range of manipulation tasks from Meta-world (yu2020meta).
Despite a big domain gap between RLBench and Meta-world, we demonstrate that APV significantly outperforms DreamerV2.
For example, APV achieves the aggregate success rate of 95.4% on six manipulation tasks, while DreamerV2 achieves 67.9%.
Moreover, we show that RLBench pre-trained representations can also be effective in learning locomotion tasks from DeepMind Control Suite (tassa2020dm\_control), where both the visuals and objectives significantly differ from RLBench videos.
2 Method
---------
We formulate a vision-based control task as a partially observable Markov decision process (POMDP), which is defined as a tuple (O,A,p,r,γ).
Here, O is the high-dimensional observation space, A is the action space, p(ot|o<t,a<t) is the transition dynamics,
r is the reward function that maps previous observations and actions to a reward rt=r(o<t,a<t),
and γ∈[0,1) is the discount factor.
The goal of RL is to learn an agent that behaves to maximize the expected sum of rewards Ep[∑Tt=1γt−1rt].
Figure 2: Illustration of action-free latent video prediction model.
The model is trained to capture visual and dynamics information from action-free videos by reconstructing image observations.
At inference time, the transition model is used to predict future states in the latent space without conditioning on predicted frames.
| | |
| --- | --- |
|
Illustration of our framework.
(a) We stack an action-conditional prediction model on top of the pre-trained action-free prediction model. At inference time, the transition model in the action-conditional model is used to predict future states in the latent space conditioned on future potential actions.
(b) To compute the intrinsic bonus, we first average pool a sequence of model states from the action-free prediction model, and apply random projection to reduce the dimension of representations while preserving distances. The intrinsic bonus for each observation is computed as the distance in the representation space to its
(a) Stacked latent prediction model
|
Illustration of our framework.
(a) We stack an action-conditional prediction model on top of the pre-trained action-free prediction model. At inference time, the transition model in the action-conditional model is used to predict future states in the latent space conditioned on future potential actions.
(b) To compute the intrinsic bonus, we first average pool a sequence of model states from the action-free prediction model, and apply random projection to reduce the dimension of representations while preserving distances. The intrinsic bonus for each observation is computed as the distance in the representation space to its
(b) Video-based intrinsic bonus
|
Figure 3:
Illustration of our framework.
(a) We stack an action-conditional prediction model on top of the pre-trained action-free prediction model. At inference time, the transition model in the action-conditional model is used to predict future states in the latent space conditioned on future potential actions.
(b) To compute the intrinsic bonus, we first average pool a sequence of model states from the action-free prediction model, and apply random projection to reduce the dimension of representations while preserving distances. The intrinsic bonus for each observation is computed as the distance in the representation space to its k-nearest neighbor in samples from a replay buffer.
###
2.1 Action-free Pre-training from Videos
For pre-training from videos, we first learn a latent video prediction model, which is an action-free variant of a latent dynamics model (hafner2019learning).
Unlike autoregressive video prediction models that predict a next frame and utilize it as an input for the following prediction,
the model instead operates on the latent space (zhang2019solar; hafner2019learning; franceschi2020stochastic).
Specifically, the model consists of three main components: (i) the representation model that encodes observations ot to a model state zt with Markovian transitions, (ii) the transition model that predicts future model states ^zt without access to the observation, and (iii) the image decoder that reconstructs image observations ^ot.
The model can be summarized as follow (see [Figure 2](#S2.F2 "Figure 2 ‣ 2 Method ‣ Reinforcement Learning with Action-Free Pre-Training from Videos")):
| | | | |
| --- | --- | --- | --- |
| | Representation model:zt∼qϕ(zt|zt−1,ot)Transition model:^zt∼pϕ(^zt|zt−1)Image decoder:^ot∼pϕ(^ot|zt) | | (1) |
We train the model to reconstruct image observations, and to make the prediction from the representation model and transition model be close to each other.
All model parameters ϕ are jointly optimized by minimizing the negative variational lower bound (ELBO; kingma2013auto):
| | | |
| --- | --- | --- |
| | L(ϕ)≐Eqϕ(z1:T|o1:T)[∑Tt=1(\underbracket−lnpϕ(ot|zt)image log loss | |
| | \underbracket+βzKL[qϕ(zt|zt−1,ot)∥pϕ(^zt|zt−1)]action-free KL loss)], | | (2) |
where βz is a scale hyperparameter and T is the length of training sequences in a minibatch.
Since the transition model does not condition on observations, it allows us to efficiently predict future states in the latent space without needing to predict future images using the image decoder at inference time.
We implement the transition model as an action-free recurrent state-space model (RSSM; hafner2019learning), which consists of both deterministic and stochastic components, and the representation model by combining the action-free RSSM with an image encoder.
We refer to [Appendix B](#A2 "Appendix B Formulation with Recurrent State-Space Model ‣ Reinforcement Learning with Action-Free Pre-Training from Videos") for a more detailed formulation.
###
2.2 Stacked Latent Prediction Model
Since actions and rewards, which provide more information about target tasks, are available during fine-tuning, it motivates incorporating them into the model.
One naïve approach would be to initialize an action-conditional prediction model with the action-free model, and learn a reward predictor on top of it.
But we find this fine-tuning scheme rapidly erases the useful knowledge in pre-trained models (see [Figure 6(a)](#S3.F6.sf1 "(a) ‣ Figure 6 ‣ RLBench pre-training results. ‣ 3.2 Meta-world Experiments ‣ 3 Experiments ‣ Reinforcement Learning with Action-Free Pre-Training from Videos") for supporting results).
To effectively utilize the pre-trained representations, we introduce a new architecture that stacks an action-conditional prediction model on top of the action-free model as below (see [Figure 3(a)](#S2.F3.sf1 "(a) ‣ Figure 3 ‣ 2 Method ‣ Reinforcement Learning with Action-Free Pre-Training from Videos")):
| | | | |
| --- | --- | --- | --- |
| | | Action-free | |
| |
{
| Representation model: | zt∼qϕ(zt|zt−1,ot) | |
| | | Transition model: | ^zt∼pϕ(^zt|zt−1) | |
| | | Action-conditional | |
| |
{
| Representation model: | st∼qθ(st|st−1,at−1,zt) | |
| | | Transition model: | ^st∼pθ(^st|st−1,at−1) | |
| | | Image decoder: | ^ot∼pθ(^ot|st) | |
| | | Reward predictor: | ^rt∼pθ(^rt|st), | | (3) |
which is optimized by minimizing the following objective:
| | | |
| --- | --- | --- |
| | L(ϕ,θ)≐Eqθ(s1:T|a1:T,z1:T),qϕ(z1:T|o1:T)[ | |
| | ∑Tt=1(\underbracket−lnpθ(ot|st)image log loss\underbracket−lnpθ(rt|st)reward log loss | |
| | | | (4) |
| | \underbracket+βKL[qθ(st|st−1,at−1,zt)∥pθ(^st|st−1,at−1)]action-conditional KL loss)], | |
where β is a scale hyperparameter.
Here, we note that we initialize the image decoder pθ(o|st) with the pre-trained image decoder pϕ(o|zt).
We implement the transition model of action-conditional prediction model as RSSM, and the representation model as RSSM with dense layers that receive the model states of the action-free model as inputs.
We refer to [Appendix B](#A2 "Appendix B Formulation with Recurrent State-Space Model ‣ Reinforcement Learning with Action-Free Pre-Training from Videos") for a more detailed formulation.
In our experiments, we use βz=0 during fine-tuning, and only utilize the action-conditional RSSM for future imagination.
###
2.3 Video-based Intrinsic Bonus
It has been observed that good representations are crucial for efficient exploration in environments with high-dimensional observations (laskin2021urlb).
To utilize useful information captured in the pre-trained representations for exploration, we propose a video-based intrinsic bonus.
Our main idea is to increase the diversity of visited trajectories by utilizing it as an intrinsic bonus.
Specifically, given a sequence of model states from the action-free prediction model zt:t+τ,
we apply average pooling across the sequence dimension to obtain a trajectory representation yt=Avg(zt:t+τ).
Then, we utilize the distance of yt to its k-nearest neighbor in samples from a replay buffer as a metric for measuring the diversity of trajectories.
To summarize, our intrinsic bonus is defined as below (see [Figure 3(b)](#S2.F3.sf2 "(b) ‣ Figure 3 ‣ 2 Method ‣ Reinforcement Learning with Action-Free Pre-Training from Videos") for illustration):
| | | | |
| --- | --- | --- | --- |
| | rintt≐||ψ(yt)−ψ(ykt)||2, | | (5) |
where ψ is a random projection (bingham2001random) that maps the model state to a low-dimensional representation for compute-efficient distance computation, and ykt is a k-nearest neighbor of yt in a minibatch.
By explicitly encouraging the agents to visit more diverse trajectories instead of single states (pathak2017curiosity; burda2018exploration; pathak2019self; liu2021behavior),
it effectively encourages the agents to explore environments in a more long term manner, and thus learn more diverse behaviors.
Then, the reward predictor is trained to predict the sum of rt and rintt as below:
| | | |
| --- | --- | --- |
| | LAPV(ϕ,θ)≐Eqθ(s1:T|a1:T,z1:T),qϕ(z1:T|o1:T)[ | |
| | ∑Tt=1(\underbracket−lnpθ(ot|st)image log loss\underbracket−lnpθ(rt+λrintt|ht,zt)APV reward log loss | |
| | | | (6) |
| | \underbracket+βKL[qθ(st|st−1,at−1,zt)∥pθ(^st|st−1,at−1)]action-conditional KL loss)], | |
where λ is a hyperparameter that adjusts the tradeoff between exploitation and exploration.
We find the intrinsic bonus provides large gains when combined with pre-training, as the pre-trained representations already contain useful representation from the beginning of the fine-tuning (see [Figure 7(a)](#S3.F7.sf1 "(a) ‣ Figure 7 ‣ RLBench pre-training results. ‣ 3.2 Meta-world Experiments ‣ 3 Experiments ‣ Reinforcement Learning with Action-Free Pre-Training from Videos") for supporting results).
In our experiments, we utilize a sliding window of size τ for constructing a set of {yt} from trajectories in a minibatch, then compute the intrinsic bonus using them.
For behavior learning, we utilize the actor-critic learning scheme of DreamerV2 (hafner2020mastering) that learns values with imagined rewards from future imaginary states and a policy that maximizes the values (see [Appendix A](#A1 "Appendix A Behavior Learning ‣ Reinforcement Learning with Action-Free Pre-Training from Videos") for the details).
3 Experiments
--------------
We designed our experiments to investigate the following:
* [topsep=0.0pt,itemsep=0.0pt,leftmargin=10pt]
* Can APV improve the sample-efficiency of vision-based RL in robotic manipulation tasks by performing action-free pre-training on videos from different domains?
* Can representations pre-trained on videos from manipulation tasks transfer to locomotion tasks?
* How does APV compare to a naïve fine-tuning scheme?
* What is the contribution of each of the proposed techniques in APV?
* How does pre-trained representations qualitatively differ from the randomly initialized representations?
* How does APV perform when additional in-domain videos or real-world natural videos are available?
Following agarwal2021deep, we report the interquartile mean with bootstrap confidence interval (CI) and stratified bootstrap CI for results on individual tasks and aggregate results, respectively, across 8 runs for each task.
Source codes and other resources are available at <https://github.com/younggyoseo/apv>.
###
3.1 Experimental Setup
#### Meta-world experiments.
We first evaluate APV on various vision-based robotic manipulation tasks from Meta-world (yu2020meta).
In all manipulation tasks, the episode length is 500 steps without any action repeat, action dimension is 4, and reward ranges from 0 to 10.
To evaluate the ability of APV to learn useful representations from different domains, we use videos collected in robotic manipulation tasks from RLBench (james2020rlbench) as pre-training data (see [Figure 4](#S3.F4 "Figure 4 ‣ Meta-world experiments. ‣ 3.1 Experimental Setup ‣ 3 Experiments ‣ Reinforcement Learning with Action-Free Pre-Training from Videos")).111In this work, we do not consider a setup where we perform pre-training on Meta-world videos and fine-tuning for solving RLBench manipulation tasks, as existing RL algorithms struggle to solve challenging, sparsely-rewarded RLBench tasks.
Specifically, we collect 10 demonstrations rendered with 5 camera views in 99 tasks from RLBench; giving a total of 4950 videos.
We then train the action-free video prediction model by minimizing the objective in [Equation 2](#S2.E2 "(2) ‣ 2.1 Action-free Pre-training from Videos ‣ 2 Method ‣ Reinforcement Learning with Action-Free Pre-Training from Videos") for 600K gradient steps.
For downstream tasks, we fine-tune the model by minimizing the objective in [Equation 6](#S2.E6 "(6) ‣ 2.3 Video-based Intrinsic Bonus ‣ 2 Method ‣ Reinforcement Learning with Action-Free Pre-Training from Videos") for 250K environment steps, i.e., 500 episodes.
Figure 4:
Illustration of experimental setups in our experiments with examples of image observations from environments.
One can see that visuals in pre-training videos are notably different from the visuals in downstream manipulation and locomotion tasks.
Figure 5:
Learning curves on manipulation tasks from Meta-world as measured on the success rate.
APV with generative pre-training on videos collected in manipulation tasks from RLBench consistently outperforms DreamerV2 in terms of sample-efficiency.
The solid line and shaded regions represent the interquartile mean and bootstrap confidence intervals, respectively, across eight runs.
#### DeepMind Control Suite experiments.
We also consider widely used robotic locomotion tasks from DeepMind Control Suite (tassa2020dm\_control).
Following the common setup in this benchmark (hafner2019dream), the episode length is 1000 steps with the action repeat of 2, and reward ranges from 0 to 1.
For pre-training, we consider two datasets: (i) 1000 videos collected from Triped Walk (see [Figure 9](#S3.F9 "Figure 9 ‣ Importance of dynamics information. ‣ 3.2 Meta-world Experiments ‣ 3 Experiments ‣ Reinforcement Learning with Action-Free Pre-Training from Videos")) and (ii) manipulation videos from RLBench.
The former one is for evaluating the performance of APV on in-domain transfer setup similar to the setup in stooke2021decoupling, while the latter one is for investigating whether the pre-trained representations can be transferred to extremely different domains, i.e., out-of-domain transfer.
Specifically, we collect 1000 videos encountered during the training of DreamerV2 agent in Triped Walk and use these videos for pre-training.
For downstream tasks, we fine-tune the model for 1M environment steps.
See [Appendix D](#A4 "Appendix D Experimental Details ‣ Reinforcement Learning with Action-Free Pre-Training from Videos") for more details.
#### Hyperparameters.
For newly introduced hyperparameters, we use βz=1.0 for pre-training, and βz=0,β=1.0 for fine-tuning.
We use τ=5 consecutive observations for computing the intrinsic bonus. To make the scale of intrinsic bonus consistent, we normalize the intrinsic reward and use λ=0.1,1.0 for manipulation and locomotion tasks, respectively.
We find that increasing the hidden size of dense layers and the model state dimension from 200 to 1024 improves the performance of both APV and DreamerV2.
We use T=25,50 for manipulation and locomotion tasks, respectively, during pre-training.
Unless otherwise specified, we use the default hyperparameters of DreamerV2.
###
3.2 Meta-world Experiments
#### RLBench pre-training results.
[Figure 5](#S3.F5 "Figure 5 ‣ Meta-world experiments. ‣ 3.1 Experimental Setup ‣ 3 Experiments ‣ Reinforcement Learning with Action-Free Pre-Training from Videos") shows the learning curves of APV pre-trained using the RLBench videos on six robotic manipulation tasks from Meta-world.
We find that APV consistently outperforms DreamerV2 in terms of sample-efficiency in all considered tasks. In particular, our framework achieves success rate above 60% on Lever Pull task while DreamerV2 completely fails to solve the task.
These results show that APV can leverage action-free videos for learning useful representations that improve the sample-efficiency of vision-based RL.
We also observe that the state-of-the-art model-free RL method DrQ-v2 (yarats2021mastering) struggles to achieve competitive performance (see [Appendix E](#A5 "Appendix E Meta-world Experiments with DrQ-v2 ‣ Reinforcement Learning with Action-Free Pre-Training from Videos")), which necessitates more investigation.
| | | |
| --- | --- | --- |
|
Learning curves on manipulation tasks from Meta-world as measured on the success rate. We report the interquartile mean and stratified bootstrap confidence interval across total 48 runs over six tasks.
(a) Comparison with a naïve fine-tuning scheme that initializes the action-conditional prediction model with the action-free prediction model. (b) Performance of APV with or without generative pre-training and intrinsic bonus.
Here,
(a) Comparison with naïve fine-tuning
|
Learning curves on manipulation tasks from Meta-world as measured on the success rate. We report the interquartile mean and stratified bootstrap confidence interval across total 48 runs over six tasks.
(a) Comparison with a naïve fine-tuning scheme that initializes the action-conditional prediction model with the action-free prediction model. (b) Performance of APV with or without generative pre-training and intrinsic bonus.
Here,
(b) Effects of pre-training and intrinsic bonus
|
Learning curves on manipulation tasks from Meta-world as measured on the success rate. We report the interquartile mean and stratified bootstrap confidence interval across total 48 runs over six tasks.
(a) Comparison with a naïve fine-tuning scheme that initializes the action-conditional prediction model with the action-free prediction model. (b) Performance of APV with or without generative pre-training and intrinsic bonus.
Here,
(c) Length of future model states τ
|
Figure 6:
Learning curves on manipulation tasks from Meta-world as measured on the success rate. We report the interquartile mean and stratified bootstrap confidence interval across total 48 runs over six tasks.
(a) Comparison with a naïve fine-tuning scheme that initializes the action-conditional prediction model with the action-free prediction model. (b) Performance of APV with or without generative pre-training and intrinsic bonus.
Here, Pre denotes generative pre-training, and Int denotes intrinsic bonus.
(c) Performance of APV with varying the length of future model states τ used for computing the intrinsic bonus.
| | | |
| --- | --- | --- |
|
(a) t-SNE visualization of average pooled model states from the action-free prediction model.
(b) Learning curves of APV on six manipulation tasks when only the parameters of convolutional image encoder and decoder are transferred, i.e., without transferring dynamics information captured in recurrent models.
(c) Learning curves of APV on four manipulation tasks when pre-trained on RLBench videos only (RLB), and on both of RLBench videos and additional in-domain Meta-world videos (RLB + MW).
Dotted and bold lines indicate the performance when all parameters are fine-tuned, and the representation model of the action-free model is frozen, respectively.
(a) Video representations
|
(a) t-SNE visualization of average pooled model states from the action-free prediction model.
(b) Learning curves of APV on six manipulation tasks when only the parameters of convolutional image encoder and decoder are transferred, i.e., without transferring dynamics information captured in recurrent models.
(c) Learning curves of APV on four manipulation tasks when pre-trained on RLBench videos only (RLB), and on both of RLBench videos and additional in-domain Meta-world videos (RLB + MW).
Dotted and bold lines indicate the performance when all parameters are fine-tuned, and the representation model of the action-free model is frozen, respectively.
(b) Importance of dynamics information
|
(a) t-SNE visualization of average pooled model states from the action-free prediction model.
(b) Learning curves of APV on six manipulation tasks when only the parameters of convolutional image encoder and decoder are transferred, i.e., without transferring dynamics information captured in recurrent models.
(c) Learning curves of APV on four manipulation tasks when pre-trained on RLBench videos only (RLB), and on both of RLBench videos and additional in-domain Meta-world videos (RLB + MW).
Dotted and bold lines indicate the performance when all parameters are fine-tuned, and the representation model of the action-free model is frozen, respectively.
(c) Effects of in-domain videos
|
Figure 7:
(a) t-SNE visualization of average pooled model states from the action-free prediction model.
(b) Learning curves of APV on six manipulation tasks when only the parameters of convolutional image encoder and decoder are transferred, i.e., without transferring dynamics information captured in recurrent models.
(c) Learning curves of APV on four manipulation tasks when pre-trained on RLBench videos only (RLB), and on both of RLBench videos and additional in-domain Meta-world videos (RLB + MW).
Dotted and bold lines indicate the performance when all parameters are fine-tuned, and the representation model of the action-free model is frozen, respectively.
#### Comparison with naïve fine-tuning.
To verify the necessity of the proposed architecture for fine-tuning, we compare APV to a
naïve fine-tuning scheme that initializes the action-conditional latent dynamics model with the pre-trained parameters of the action-free model (see [Appendix D](#A4 "Appendix D Experimental Details ‣ Reinforcement Learning with Action-Free Pre-Training from Videos") for the details).
For a fair comparison, we do not utilize the intrinsic bonus for APV.
[Figure 6(a)](#S3.F6.sf1 "(a) ‣ Figure 6 ‣ RLBench pre-training results. ‣ 3.2 Meta-world Experiments ‣ 3 Experiments ‣ Reinforcement Learning with Action-Free Pre-Training from Videos") shows that DreamerV2 with this naïve fine-tuning scheme (DreamerV2 w/ Naïve FT) does not provide large gains over DreamerV2, which implies that naïve fine-tuning quickly loses pre-trained representations.
By contrast, we find that APV without intrinsic bonus consistently outperforms DreamerV2 by achieving >10% higher success rate from the beginning of the fine-tuning, even though the same pre-trained model is used for fine-tuning.
This shows the proposed architecture is crucial for effective fine-tuning.
#### Ablation study.
To evaluate the contribution of the proposed techniques in APV, we report the performance of our framework with or without generative pre-training and intrinsic bonus in [Figure 6(b)](#S3.F6.sf2 "(b) ‣ Figure 6 ‣ RLBench pre-training results. ‣ 3.2 Meta-world Experiments ‣ 3 Experiments ‣ Reinforcement Learning with Action-Free Pre-Training from Videos").
First, we observe that APV (Pre: X / Int: X), whose difference to DreamerV2 is the usage of stacked latent prediction model, achieves similar performance to DreamerV2.
This implies that our performance gain is not from the architecture itself, but the way we utilize it for fine-tuning is important.
We also find that our intrinsic bonus can improve the performance with or without pre-training, and generative pre-training can also improve the performance with or without intrinsic bonus.
Importantly, one can see that the best performance is achieved when both components are combined.
This implies that our proposed techniques synergistically contribute to the performance improvement.
#### Effects of video-based intrinsic bonus.
We investigate the effect of considering multiple model states of length τ in [Equation 5](#S2.E5 "(5) ‣ 2.3 Video-based Intrinsic Bonus ‣ 2 Method ‣ Reinforcement Learning with Action-Free Pre-Training from Videos") instead of a single model state.
[Figure 6(c)](#S3.F6.sf3 "(c) ‣ Figure 6 ‣ RLBench pre-training results. ‣ 3.2 Meta-world Experiments ‣ 3 Experiments ‣ Reinforcement Learning with Action-Free Pre-Training from Videos") shows that APV with length 5 achieves significantly better performance than APV with length 1.
We believe this is because considering longer sequence of observations, i.e., videos, enables us to utilize contextual information from videos for encouraging agents to perform diverse behaviors.
#### Qualitative analysis.
We visually investigate why the pre-trained representations can be useful for unseen Meta-world tasks.
Specifically, we sample video clips of length 25 from the 10 videos of six manipulation task, and visualize the averaged model states from the sampled videos using t-SNE (van2008visualizing) in [Figure 7(a)](#S3.F7.sf1 "(a) ‣ Figure 7 ‣ RLBench pre-training results. ‣ 3.2 Meta-world Experiments ‣ 3 Experiments ‣ Reinforcement Learning with Action-Free Pre-Training from Videos"), where colors indicate the tasks.
We observe that the pre-trained representations from each task are well clustered, while randomly intialized representations are entangled.
This shows that the pre-trained representations capture information about the Meta-world tasks without access to Meta-world videos during pre-training.
Figure 8:
Learning curves on locomotion tasks from DeepMind Control Suite as measured on the episode return.
Interestingly, we find that APV pre-trained on manipulation videos from RLBench consistently outperforms DreamerV2.
We also observe that utilizing in-domain videos from Triped Walk leads to further improved performance.
The solid line and shaded regions represent the interquartile mean and bootstrap confidence intervals, respectively, across eight runs.
#### Importance of dynamics information.
To investigate whether the performance gain comes from utilizing the dynamics information captured in pre-trained representations or visual information captured in the image encoder and decoder,
we report the performance of APV when only the pre-trained parameters of the convolutional image encoder and decoder are transferred.
[Figure 7(b)](#S3.F7.sf2 "(b) ‣ Figure 7 ‣ RLBench pre-training results. ‣ 3.2 Meta-world Experiments ‣ 3 Experiments ‣ Reinforcement Learning with Action-Free Pre-Training from Videos") shows that APV (Encoder / Decoder Only) performs worse than APV, which demonstrates that utilizing the dynamics information learned from diverse videos is crucial for performance.
Figure 9:
Illustration of additional experimental setup in DeepMind Control Suite experiments. We use videos collected in Triped Walk for pre-training, and then fine-tune the pre-trained model for solving downstream Quadruped and Hopper locomotion tasks.
#### Pre-training with additional in-domain videos.
We also consider an experimental setup where we have an access to additional videos with similar visuals collected on the target domain.
Specifically, we collect 10 videos on each task from ML-10 training tasks in Meta-world, i.e., total of 100 videos, and utilize these videos for pre-training in conjunction with RLBench videos.
Then, we fine-tune the pre-trained model for solving 4 manipulation tasks that were not seen during pre-training.222Specifically, the tasks are Lever Pull, Drawer Open, Door Lock, and Button Press Topdown Wall.
To evaluate how additional videos affect representation learning, we report the performance of APV when the representation model of the action-free prediction model is frozen, and use it as a proxy for evaluating the quality of representations.
[Figure 7(c)](#S3.F7.sf3 "(c) ‣ Figure 7 ‣ RLBench pre-training results. ‣ 3.2 Meta-world Experiments ‣ 3 Experiments ‣ Reinforcement Learning with Action-Free Pre-Training from Videos") shows that pre-training with additional Meta-world videos (APV (RLB + MW)) achieves almost similar performance to pre-training with only RLBench videos (APV (RLB)).
This shows that pre-trained representations with RLBench videos already learns useful representations so that they can be quickly fine-tuned for solving Meta-world tasks.
However, we observe that pre-training with in-domain videos significantly improves the performance when the pre-trained representation model is frozen. This implies that additional in-domain videos help for addressing the domain gap between pre-training and fine-tuning.
###
3.3 DeepMind Control Suite Experiments
[Figure 8](#S3.F8 "Figure 8 ‣ Qualitative analysis. ‣ 3.2 Meta-world Experiments ‣ 3 Experiments ‣ Reinforcement Learning with Action-Free Pre-Training from Videos") shows the learning curves of APV and DreamerV2 on locomotion tasks.
Interesingly, we find that APV pre-trained on manipulation videos from RLBench (pink curves) consistently achieves better performance than DreamerV2 (black curves).
This demonstrates that representations pre-trained using manipulation videos, which have notably different visuals and objectives, effectively capture dynamics information useful for quickly learning the dynamics of locomotion environments.
Also, by utilizing the videos from a similar domain, i.e., Triped environment, the performance of APV is further improved.
We provide additional experimental results that evaluate the contribution of the proposed techniques on DeepMind Control Suite tasks in [Appendix H](#A8 "Appendix H Ablation Study on DeepMind Control Suite ‣ Reinforcement Learning with Action-Free Pre-Training from Videos").
###
3.4 Real-world Video Experiments
Finally, we investigate whether real-world natural videos of people performing diverse behaviors, not from robotics domain, can be leveraged for pre-training.
Specifically, we utilize 15K videos sampled from Something-Something-V2 dataset (goyal2017something), which contains 159K videos of people performing actions, for pre-training and fine-tune the pre-trained model for solving manipulation tasks from Meta-world.
[Figure 10](#S4.F10 "Figure 10 ‣ Behavior learning with videos. ‣ 4 Related Work ‣ Reinforcement Learning with Action-Free Pre-Training from Videos") shows that APV pre-trained with real-world videos does not provide large gains over the baseline without pre-training.
We find that this is because our video prediction model suffers from severe underfitting to real-world video datasets (see [Appendix F](#A6 "Appendix F Real-World Video Prediction on Something-Something-V2 ‣ Reinforcement Learning with Action-Free Pre-Training from Videos") for the examples of blurry predicted future frames).
It would be an interesting future direction to develop a lightweight but high-fidelity prediction model for RL with pre-training from videos.
4 Related Work
---------------
#### Unsupervised representation learning for CV and NLP.
Recently, unsupervised representation learning methods have been actively studied in the domain of CV.
Various representation learning methods, including reconstruction (he2021masked), rotation (gidaris2018unsupervised), solving zigsaw puzzles (noroozi2016unsupervised), and contrastive learning (he2020momentum; chen2020simple), have reduced the gap with supervised pre-training with labels.
In the domain of NLP, unsupervised representation learning has been successfully applied to training language models with generalization ability (devlin2018bert; radford2018improving; yang2019xlnet).
Notably, devlin2018bert and radford2018improving trained large transformer networks (vaswani2017attention) with masked token prediction and generative pre-training, respectively, and showed that pre-trained models can be effectively fine-tuned on downstream tasks.
In this work, we demonstrate that unsupervised pre-training can also be effective for vision-based RL.
#### Unsupervised representation learning for RL.
Unsupervised representation learning for RL has also been studied to improve the sample-efficiency of RL algorithms. Notably, jaderberg2016reinforcement showed that optimizing auxiliary unsupervised losses can improve the performance of RL agents.
This has been followed by a series of works which demonstrated the effectiveness of various unsupervised learning objectives, including world-model learning (hafner2019learning; hafner2020mastering), reconstruction (yarats2019improving), future representation prediction (gelada2019deepmdp; schwarzer2020data), bisimulation (castro2020scalable; zhang2020learning), and contrastive learning (oord2018representation; anand2019unsupervised; mazoure2020deep; srinivas2020curl).
While these works optimize auxiliary unsupervised objectives to accelerate the training of RL agents, we instead aim to pre-train representations as in CV and NLP domains.
There have been several approaches to perform unsupervised pre-training for RL (finn2016deep; dwibedi2018learning; zhan2020framework; srinivas2020curl; stooke2021decoupling; schwarzer2021pretraining). In particular, schwarzer2021pretraining proposed several self-supervised learning objectives that rely on actions, but assume access to action information of downstream tasks which may not be available in practice.
The work closest to ours is stooke2021decoupling, which demonstrated that the representations contrastively pre-trained without actions and rewards can be effective on unseen downstream tasks but with very similar visuals.
In this work, we instead develop a framework that leverages action-free videos from diverse domains with different visuals and embodiments for pre-training.
#### Behavior learning with videos.
Video datasets have also been utilized for behavior learning in various ways (peng2018sfv; torabi2018generative; aytar2018playing; liu2018imitation; sermanet2018time; edwards2019imitating; schmeckpeper2020learning; schmeckpeper2020reinforcement; chang2020semantic; chen2021learning).
For instance, aytar2018playing solved hard exploration tasks on Atari benchmark by designing an imitation reward based on YouTube videos, and peng2018sfv proposed to learn physical skills from human demonstration videos by extracting reference motions and training an RL agent that imitates the extracted motions.
Our work differs in that we utilize videos for pre-training useful representations, instead of directly learning behaviors from videos.
We provide more discussion on related fields in [Appendix C](#A3 "Appendix C Extended Related Work ‣ Reinforcement Learning with Action-Free Pre-Training from Videos").
Figure 10:
Learning curves on manipulation tasks from Meta-world when pre-trained with real-world videos from Something-Something-V2 video dataset.
We report the interquartile mean and stratified bootstrap confidence interval over 48 runs over six tasks.
5 Discussion
-------------
In this work, we introduce a vision-based RL framework that learns representations useful for understanding the dynamics of downstream domains via action-free pre-training on videos, and utilizes the pre-trained representations for fine-tuning.
Our experimental results demonstrate that APV can improve both sample-efficiency and final performances of vision-based RL on various manipulation and locomotion tasks, by effectively transferring the pre-trained representations from unseen domains.
However, one limitation of our work is that pre-training is conducted only on simulated robotic videos, which is because of the underfitting issue reported in our analysis.
Given that, one interesting direction would be to scale up the architecture or utilize recently developed video prediction architectures (babaeizadeh2021fitvid), and investigate how the prediction quality affects the performance.
Moreover, while our work focuses on representation learning via generative pre-training, another interesting future direction would be to investigate the performance of representation learning schemes such as masked prediction (he2021masked) and contrastive learning (oord2018representation).
By presenting a generic framework that can leverage videos with diverse visuals and embodiments for pre-training, we hope this work would facilitate future research on unsupervised pre-training for RL.
6 Acknowledgements
-------------------
We would like to thank Junsu Kim, Fangchen Liu, Jongjin Park, Jihoon Tack, Wilson Yan, and Sihyun Yu for helpful discussions.
This work was partially supported by Hong Kong Centre for Logistics Robotics, Center for Human Compatible AI (CHAI), and Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No.2019-0-00075, Artificial Intelligence Graduate School Program (KAIST)). |
f8cd93a8-02fb-4f75-8fc4-bc71c1d416bc | trentmkelly/LessWrong-43k | LessWrong | Fun, endless art debates v. morally charged art debates that are intrinsically endless
Discussing art is fun. It's a great pastime. There's a number of very simple art criticism questions we will never answer but are often very fun to discuss for specific artists or performers we care about. AI-assisted, some are:
1. Is this art or just unnecessary shock value?
2. Does skill matter or just the concept?
3. Is it good because it's popular, or popular because it's good?
4. Is it original?
5. Should a highly skilled artist make it look easy, or look difficult?
6. Does knowing the artist's background change the meaning?
7. Does the artist's opinion of their own work matter in its meaning?
8. Is the message specific or timeless?
9. Is it just pandering?
There are a few morally charged, less fun debates. Such as:
1. Is this cultural appropriation?
2. Is this representation a harmful stereotype?
3. Is this pandering?
It's worth pausing to notice the charged arguments reduce from the non-charged arguments.
1. If you can answer "is this appropriation", you can answer "is this original".
2. If you can answer "is this representation harmful", you can answer "is this unnecessary shock value".
3. If you can answer "is this pandering", you can answer "is this good because it's popular or popular because it's good."
This would suggest your ability to solve, make progress or produce new insight on these morally charged art questions are upper bounded at your ability to do so on the general art criticism questions. I do recommend having fun, endless debates about art. If you're in an endless, morally charged debate, try focusing on the "intrinsically endless" portion of the debate. For instance:
1. "I get she's copying but usually I don't think attribution should be overt."
2. "I know it's harmful but I'm drawn to shock value."
3. "I get that it's pandering, but it was received well and I always respect art that's a hit with its audience."
What you really don't want to do is mull over your new ideas on originality when simultaneously in a he |
ac9749f8-4ac1-456f-9c7d-03ccc7909032 | trentmkelly/LessWrong-43k | LessWrong | Case studies on social-welfare-based standards in various industries
Last year, I posted a call for case studies on social-welfare-based standards for companies and products (including standards imposed by regulation). The goal was to gain general context on standards to inform work on possible standards and/or regulation for AI.
This resulted[1] in several dozen case studies that I found informative and helpful. I've been hoping to write up my reflections after reading them all, but it's taken me long enough to get to this that for now I am just publishing a public Google sheet with links to all of the case studies that we have permission to share publicly (including some that already have public links). The link is here: https://docs.google.com/spreadsheets/d/18gaTIzdgMvKLq9Cp2-GJZZA7QmE93Frufh1UhNMcbpg/
1. ^
Most of the case studies this piece links to were directly paid for via this project, but in some cases the work was pro bono, or someone adapted or sent a copy of work that had been done for another project, etc. |
8d1de67e-f514-4622-aca1-8819845b8ba6 | trentmkelly/LessWrong-43k | LessWrong | Toward a taxonomy of cognitive benchmarks for agentic AGIs
Inspired by the sequence on LLM Psychology, I am developing a taxonomy of cognitive benchmarks for measuring intelligent behavior in LLMs. This taxonomy could facilitate understanding of intelligence to identify domains of machine intelligence that have not been adequately tested.
Generally speaking, in order to understand loss-of-control threats from agentic LLM-based AGIs, I would like to understand the agentic properties of an LLM. METR's Autonomy Evaluation Resources attempts to do this by testing a model's agentic potential, or autonomy, by measuring its ability to perform tasks from within a sandbox. A problem with this approach is it gets very close to observing a model actually performing the behavior we do not want to see. This is inevitable because all alignment research is dual-use.
One way to remove ourselves one further level from agentic behavior is to try to measure the cognitive capacities that lead to agentic behavior.
In the diagram, agentic tasks as described in METR's ARC measure the ability of a model to assert control of itself and the world around it by measuring its ability to perform agentic tasks. Inspired @Quentin FEUILLADE--MONTIXI 's LLM Ethological approach in LLM Psychology, I want to understand how a model could perform agentic tasks by studying the cognitive capacities that facilitate this.
I started by examining the kinds of cognitive constructs studied by evolutionary and developmental psychologists, as well as those that are very clearly studied already in LLM research. This made up the following list or taxonomy:
ConstructCurrent EvalsOther PapersSelfhoodAgencySharma et al. (2024), Mialon et al. (2023): General AI Assistants (GAIA) METR Autonomy Evaluation Resources Survival instinctAnthropic human & AI generated evals Situational awareness / self awarenessLaine, Meinke, Evans et al. (2023) Anthropic human & AI generated evalsWang & Zhong (2024)Metacognition Uzwyshyn, Toy, Tabor, MacAdam (2024), Zhou et al. (2024), Feng et |
457fc4ce-a462-4f62-b847-dba2472fe71e | trentmkelly/LessWrong-43k | LessWrong | Safety Implications of LeCun's path to machine intelligence
Yann LeCun recently posted A Path Towards Autonomous Machine Intelligence, a high-level description of the architecture he considers most promising to advance AI capabilities.
This post summarizes the architecture and describes some implications for AI safety work if we accept the hypothesis that the first transformative AI will have this architecture.
Why is this a hypothesis worth considering?
1. LeCun has a track record of being ahead of mainstream academic research, from working on CNNs in the 90s to advocating for self-supervised learning back in 2014-2016 when supervised learning was ascendant.
2. LeCun runs Meta AI (formerly FAIR) which has enormous resources and influence to advance his research agenda, making it more likely that his proposed architecture will be built at scale. In general I think this is an underrated factor; AI research exhibits a great deal of path dependence, and most plausible paths to AI are not taken primarily because nobody is willing to take a big risk on them.
3. The architecture is dramatically different from the architectures commonly assumed (implicitly) in much AI alignment work, such as model-free deep RL and "GPT-3 but scaled up 10000x". This makes it a good robustness check for plans that are overly architecture-specific.
Architecture Overview
The Overall Agent
At a high level, the proposed architecture is a set of specialized cognitive modules. With the exception of the Actor and the Intrinsic Cost (see below) they are all deep neural networks trained with gradient descent.
The high level architecture of LeCun's proposed agent. Arrows indicate dependence; gradients flow backward through the thin arrows.
What is this agent doing, exactly? It's meant to be a general architecture for any autonomous AI agent, but LeCun repeatedly emphasizes video inputs and uses self-driving cars as a recurrent example, so the central use case is embodied agents taking actions in the physical world. Other talks I've seen by |
715ea708-75c8-4846-8a9b-4d8dcb94cc2f | trentmkelly/LessWrong-43k | LessWrong | Meetup : West LA: Practical Rationality Sources
Discussion article for the meetup : West LA: Practical Rationality Sources
WHEN: 15 July 2015 07:00:00PM (-0700)
WHERE: 11066 santa monica avenue, LA, CA
How to Find Us: Go into this Del Taco. We will be in the back room if possible.
Parking is free in the lot out front or on the street nearby.
Discussion: If we wish to make a concerted effort to learn pragmatic rationality, as opposed to the most intellectually shiny concepts in rationality, what should we be reading? We will discuss what's out there and brainstorm other things to look for. In addition to LessWrong, we'd like to draw from other communities that give practical advice; Mister Money Mustache appears to be one of these.
Recommended Reading:
* Scientific Self-Help: The State of Our Knowledge
* Mister Money Mustache: Is it Convenient?
* Mister Money Mustache: Hedonic Adaptation
No prior exposure to Less Wrong is required; this will be generally accessible.
Discussion article for the meetup : West LA: Practical Rationality Sources |
f979febe-4d73-4964-a9ff-350a9a73120c | trentmkelly/LessWrong-43k | LessWrong | Highlights from "Integral Spirituality"
Cross-posted from Map and Territory
Apologia
A couple months ago a friend gifted me a copy of Ken Wilber's Integral Spirituality. At first I was skeptical about reading it: I'm pretty busy and didn't have much context to think I would learn from it. But he talked me into it, prodding me to at least just read the introduction, which he promised was relatively short (35 pages, so basically the length of a long blog post) and densely packed with interesting content. At the time I was almost done reading another book, and figured "what the heck, I'll just read the intro and can decide from there".
Given that you're reading a post with "Integral Spirituality" in the title, I think you can guess what happened next.
I mostly want to share a lot of things I highlighted in the book—passages I thought could stand to be more widely read—because Ken Wilber has put words to many of the thoughts I would like to share but haven't made the time to write about. However, I need to give these passages a little context, so I'll do my best to give you a very high level, whirlwind tour of Wilber's themes.
The nominal purpose of this book is to discuss spirituality, and Wilber does that plenty, but I honestly think of this book as more about Wilber's integral theory and just happens to use spirituality as a topic to address integral theory. So what is integral theory? In short I'd say it's a way to work with all evidence so you can update on it so you aren't forced to ignore or dismiss evidence that doesn't fit with your worldview. That is, most of the time most of us start from a place of undervaluing some information and overvaluing other information we encounter because it suggests that our understanding of the world (ontology) is wrong or right, respectively; integral theory helps rehabilitate this tendency by showing how to integrate evidence that has different purposes. A pithy way to put this would be: everything is evidence of something, nothing is evidence of everything. The |
2e4cd924-2290-43fa-9f07-5c00659e7d1e | trentmkelly/LessWrong-43k | LessWrong | Duncan Sabien: "In Defense of Punch Bug"
Cool new post from Duncan Sabien (formerly active here under that name and Conor_Moreton). |
99c31615-ee61-411d-85bc-2c85eb7f166e | trentmkelly/LessWrong-43k | LessWrong | Slate Star Codex: alternative comment threads on LessWrong?
Like many Less Wrong readers, I greatly enjoy Slate Star Codex; there's a large overlap in readership. However, the comments there are far worse, not worth reading for me. I think this is in part due to the lack of LW-style up and downvotes. Have there ever been discussion threads about SSC posts here on LW? What do people think of the idea occasionally having them? Does Scott himself have any views on this, and would he be OK with it?
Update:
The latest from Scott:
> I'm fine with anyone who wants reposting things for comments on LW, except for posts where I specifically say otherwise or tag them with "things i will regret writing"
In this thread some have also argued for not posting the most hot-button political writings.
Would anyone be up for doing this? Ataxerxes started with "Extremism in Thought Experiments is No Vice" |
02c53574-68f7-426f-aa91-6252b3a887d8 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | "Diamondoid bacteria" nanobots: deadly threat or dead-end? A nanotech investigation
*Confidence level: I’m a computational physicist working on nanoscale simulations, so I have some understanding of most of the things discussed here, but I am not specifically an expert on the topics covered, so I can’t promise perfect accuracy.*
*I want to give a huge thanks to Professor Phillip Moriarty of the university of Nottingham for answering my questions about the experimental side of mechanosynthesis research.*
**Introduction:**
A lot of people are highly concerned that a malevolent AI or insane human will, in the near future, set out to destroy humanity. If such an entity wanted to be absolutely sure they would succeed, what method would they use? Nuclear war? Pandemics?
According to some in the x-risk community, the answer is this: The AI will invent molecular nanotechnology, and then kill us all with diamondoid bacteria nanobots.
This is the “lower bound” scenario posited by Yudkowsky in his post [AGI ruin](https://www.lesswrong.com/posts/uMQ3cqWDPHhjtiesc/agi-ruin-a-list-of-lethalities):
> *The nanomachinery builds diamondoid bacteria, that replicate with solar power and atmospheric CHON, maybe aggregate into some miniature rockets or jets so they can ride the jetstream to spread across the Earth's atmosphere, get into human bloodstreams and hide, strike on a timer.*
>
>
The phrase “diamondoid bacteria” really struck out at me, and I’m not the only one. In [this post](https://forum.effectivealtruism.org/posts/3KAuAS2shyDwnjzNa/predictable-updating-about-ai-risk) by Carlsmith (which I found very interesting), Carlsmith refers to diamondoid bacteria as an example of future tech that feels unreal, but may still happen:
> *Whirling knives?*[*Diamondoid bacteria*](https://twitter.com/ESYudkowsky/status/1438198189782290433)*? Relentless references to paper-clips, or “tiny molecular squiggles”? I’ve written, elsewhere, about*[*the “unreality” of futurism*](https://joecarlsmith.com/2021/01/31/believing-in-things-you-cannot-see#iv-realization-vs-belief)*. AI risk had a lot of that for me.*
>
>
Meanwhile, the controversial anti-EA crusader Emille Torres cites the term “diamondoid bacteria” as a reason to dismiss AI risk, calling it “[patently ridiculous](https://twitter.com/xriskology/status/1675824164161069057)”.
I was interested to know more. What *is* diamondoid bacteria? How far along is molecular nanotech research? What are the challenges that we (or an AI) will need to overcome to create this technology?
If you want, you can stop here and try and guess the answers to these questions.
It is my hope that by trying to answer these questions, I can give you a taste of what nanoscale research actually looks like. It ended up being the tale of a group of scientists who had a dream of revolutionary nanotechnology, and tried to answer the difficult question: How do I actually *build* that?
**What is “diamondoid bacteria”?**
The literal phrase “diamondoid bacteria” appears to have been invented by Eliezer Yudkowsky about two years ago. If you search the exact phrase in google scholar there are no matches:
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fca4ae25b-e847-483e-aa16-3489d634b54f_940x460.png)
If you search the phrase in regular google, you will get a very small number of matches, all of which are from Yudkowsky or directly/indirectly quoting Yudkowsky. The very first use of the phrase on the internet appears to be [this twitter post](https://twitter.com/ESYudkowsky/status/1438198189782290433?lang=en) from September 15 2021. (I suppose there’s a chance someone else used the phrase in person).
I speculate here that Eliezer invented the term as a poetic licence way of making nanobots seem more viscerally real. It does not seem likely that the hypothetical nanobots would fit the [scientific definition of bacteria](https://www.ncbi.nlm.nih.gov/books/NBK8120/#:~:text=Bacteria%20are%20single%2Dcelled%20microorganisms,are%20sophisticated%20and%20highly%20adaptable.), unless you really stretched the definition of terms like “single-celled” and “binary fission”. Although bacteria are very impressive micro-machines, so I wouldn’t be surprised if future nanotech bore at least some resemblance.
Frankly, I think inventing new terms is an extremely unwise move (I think that Eliezer has stopped using the term since I started writing this, but others still are). “diamondoid bacteria” sounds science-ey enough that a lot of people would assume it was already a scientific term invented by an actual nanotech expert (even in a speculative sense). If they then google it and find nothing, they are going to assume that you’re just making shit up.
But diamondoid nanomachinery *has* been a subject of inquiry, by actual scientific experts, in a research topic called “diamondoid mechanosynthesis”.
**What is “diamondoid mechanosynthesis”**
Molecular nanotech (MNT) is an idea first championed by Eric Drexler, that the same principles of mass manufacturing that are used in todays factories could one day be miniaturized to the nanoscale, assembling complex materials molecule by molecule from the ground up, with nanoscale belts, gears, and manipulators. You can read the thesis [here](https://dspace.mit.edu/handle/1721.1/27999), It’s an impressive first theoretical pass at the nanotech problem, considering the limited computational tools available in 1991, and helped inspire many in the current field of nanotechnology (which mostly does not focus on molecular assembly).
However, Drexlers actual designs of how a molecular assembler would be built have been looked on with [extreme skepticism](https://courses.cs.duke.edu/cps296.4/spring08/papers/Drexler.v.Smalley.pdf) by the wider scientific community. And while some of the criticisms have been unfair (such as accusations of pseudoscience), there are undeniably [extreme engineering challenges](https://www.lesswrong.com/posts/FijbeqdovkgAusGgz/grey-goo-is-unlikely). The laws of physics are felt very differently at different scales, presenting obstacles that have never been encountered before in the history of manufacturing, and indeed may turn out to be entirely insurmountable in practice. How would you actually make such a device?
Well, a few teams were brave enough to try and tackle the problem head on. The nanofactory collaboration, with a website [here](http://www.molecularassembler.com/Nanofactory/index.htm), was an attempt to directly build a molecular assembler. It was started in the early 2000’s, with the chief players beings Freitas and Merkle, two theoretical/computational physicists following on from the work of Drexler. The method they were researching to make this a reality was diamondoid mechanosynthesis(DMS).
So, what is DMS? Lets start with [Mechanosynthesis](https://en.wikipedia.org/wiki/Mechanosynthesis). Right now, if you want to produce molecules from constituent molecules or elements, you would place reactive elements in a liquid or gas and jumble them around so they bump into each other randomly. If the reaction is thermodynamically favorable under the conditions you’ve put together (temperature, pressure, etc.), then mass quantities of the desired products are created.
This is all a little chaotic. What if we wanted to do something more controlled? The goal of mechanosynthesis is to precisely control the reactive elements we wish to put together by using mechanical force to precisely position them together. In this way, the hope is that extremely complex structures could be assembled atom by atom or molecule by molecule.
The dream, as expressed in the [molecular assembler project](http://www.molecularassembler.com/Nanofactory/index.htm), was that mechanosynthesis can be mastered to such a degree that “nano-factories” could be built, capable of building many different things from the ground up, including another nanofactory. If this could be achieved, then as soon as one nanofactory is built, a vast army of them would immediately follow through the power of exponential growth. These could then build nanomachines that move around, manipulate objects, and build pretty much anything from the ground up, like a real life version of the Star Trek matter replicator.
If you want to convert a dream into a reality, you have to start thinking of engineering, If you could make such a nano-factory, what would it be made out of? There are a truly gargantuan number of materials out there we could try out, but almost all of them are not strong enough to support the kind of mechanical structures envisaged by the nanofactory researchers. The most promising candidate was “diamondoid”.
Now, what is “diamondoid”? You’d expect this to be an easy question to answer, but it’s actually a little thorny. The more common definition, the one used on [wikipedia](https://en.wikipedia.org/wiki/Diamondoid) and most [journal papers](https://onlinelibrary.wiley.com/doi/full/10.1002/anie.200701684), is that diamondoid refers to a specific family of hydrocarbons like the ones shown below, with the simplest one being “[adamantane](https://en.wikipedia.org/wiki/Adamantane)”, with it’s strong, cage-like structure, and the other ones being formed by joining together multiple cages.
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fce128c3d-f2f5-44d0-aeb3-deac41fbf87a_960x691.png)
Image taken from [here](https://www.researchgate.net/figure/Chemical-structures-of-Diamondoids-the-chemical-structures-of-Adamantane-left-and_fig1_228678727)
These cages are incredibly strong and stable, which makes them a promising candidate material for building up large structures, and keeping them stable for assembly purposes.
The other definition, which seems to be mainly used by the small community of molecular nanotech(MNT) proponents, is that “diamondoid” just means “any sufficiently strong and stiff nanoscale material”. See [this passage](http://www.molecularassembler.com/Nanofactory/index.htm) from the “molecular assembler” website:
> *Diamondoid materials also may include any* [*stiff*](http://e-drexler.com/p/04/04/0410stiffMols.html) *covalent solid that is similar to diamond in strength, chemical inertness, or other important material properties, and possesses a dense three-dimensional network of bonds. Examples of such materials are carbon nanotubes (illustrated at right) or fullerenes, several strong covalent ceramics such as silicon carbide, silicon nitride, and boron nitride, and a few very stiff ionic ceramics such as sapphire (monocrystalline aluminum oxide) that can be covalently bonded to pure covalent structures such as diamond.*
>
>
This passage is very out of line with mainstream definitions. I couldn’t find a mention of “diamondoid” in any top [carbon nanotube article](https://scholar.google.com.au/scholar?hl=en&as_sdt=0%2C5&q=carbon+nanotubes&btnG=&inst=11099861943932837285). I’ve done a little research on aluminium oxide, and I have never in my life heard it called “diamondoid”, considering it neither contains the same elements as diamond, nor does it take the same structure as diamond or diamondoid hydrocarbons. This kind of feels like the “radical sandwich anarchy” section of [this chart](https://flowingdata.com/2017/05/02/sandwich-alignment-chart/).
I really don’t want to get sidetracked into semantic debates here. But just know that the MNT definition is non-standard, might annoy material scientists, and could easily be used against you by someone with a dictionary.
In any case, it’s not a huge deal, because the molecular assembler team was focused on carbon-based diamond and diamondoid structures anyway.
The plan was to engage in both theoretical and experimental research to develop nanotech in several stages. Step 1 was to achieve working prototypes of diamond mechanosynthesis. Step 2 was to build on this to actually assemble complex molecular structures in a programmable mechanical manner. Step 3 was to find a way to parallelize the process, so that huge amounts of assembly could be done at once. Step 4 was to use that assembly to build a nanofactory, capable of building a huge number of things, including a copy of itself. The proposed timeline for this project is shown below:
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F530dc5f3-4666-4f86-8fdc-b8fac54b0ae1_527x291.png)
They thought they would have the first three steps finished by 2023, and have working commercialized nanofactories by 2030. Obviously, this is not on track. I’m not holding this against them, as extremely ambitious projects rarely finish on schedule. They were also underfunded compared to what they wanted, furthering hampering progress.
How far did the project go, in the end?
**DMS research: The theoretical side**
The nanofactory collaboration put forward a [list of publications](http://www.molecularassembler.com/Nanofactory/Publications.htm), and as far as I can tell, every single one is theoretical or computational in nature. There are a few book chapters and patent applications, as well as about a dozen peer-reviewed scientific articles, mostly in non-prestigious journals[1](#footnote-1).
Skimming through the papers, they seem fine. A lot of time and effort has gone into them, I don’t see any obvious problems with their methodology, and the reasoning and conclusions seem to be a reasonable. Going over all of them would take way too long, but I’ll just pick one that is representative and relatively easy to explain: “[*Theoretical Analysis of Diamond Mechanosynthesis. Part II. C2 Mediated Growth of Diamond C(110) Surface via Si/Ge-Triadamantane Dimer Placement Tools*](http://www.molecularassembler.com/Papers/JCTNMannMar04.pdf)”.
Please don’t leave, I promise you this is interesting!
The goal of this paper is simple: we want to use a tooltip to pick up a pair of carbon atoms (referred to as a “dimer”), place the dimer on a carbon surface (diamond), and remove the tooltip, leaving the dimer on the surface.
In our large world, this type of task is pretty easy: you pick up a brick, you place it where you want, and then you let it go. But all the forces present at our scale are radically different at the nanoscale. For example, we used friction to pick the brick up, but “friction” does not really exist at the single atom scale. Instead, we have to bond the cargo element to our tool, and then break that bond at the right moment. It’s like if the only way to lay bricks was to glue your hand to a brick, glue the brick to the foundation, and then rip your hand away.
Below we have the design for their tooltip that they were investigating here. We have our diamondoid cages from earlier, but we replace a pair of corner atoms with Germanium (or Si) atoms, and bond the cargo dimer to these corners, in the hopes it will make them easier to detach:
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Feeb25424-5206-4c90-aa26-32359e4b3c81_400x306.png)
The first computational result is a checking of this structure using DFT simulations. I have described DFT and it’s strengths and shortcomings in [this previous post](https://titotal.substack.com/p/bandgaps-brains-and-bioweapons-the). They find that the structure is stable in isolation.
Okay great, it’s stable on it’s own, but the eventual plan is to have a whole ton of these around working in parallel. So the next question they ask is this: if I have a whole bunch of these together, are they going to react with each other and ruin the tooltip? The answer, they find, is *yes*, in two different ways. Firstly, if two of these meet dimer-to-dimer, it’s thermodynamically favorable for them to fuse together into one big, useless tooltip. Secondly, if one encounters the hydrogen atoms on the surface of the other, it would tear them out to sit on the end of the cargo dimer, rendering it again useless. They don’t mention it explicitly, but I assume the same thing would happen if it encountered stray hydrogen in the air.
This is a blow to the design, and would mean great difficulty in actually using the thing large scale. In theory you could still pull it off by keeping the tools isolated from each other.
They check the stability of the tooltip location itself using molecular dynamics calculation, and find that it’s stable enough for purpose, with a stray that is smaller than the chemical bond distances involved.
And now for the big question: can it actually deposit the dimer on the surface? The following graph summarizes the DFT results:
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F4d7dbda5-7527-4cb2-b5b4-f551bb6c84dd_773x532.png)
On the left side, we have the initial state. The tooltip is carrying the cargo dimer. At this step, and at every other, a DFT calculation is taken out to calculate the entire energy of the simulation.
In the middle, we have the middle state. The tooltip has been lowered, carrying the tooltip to the surface, where the carbon dimer is now bonded both to the tooltip and to the diamond surface.
On the right, we have the desired final state. The tooltip has been retracted and raised, but the carbon is left behind on the surface.
All three states have been simulated using DFT to predict their energy, and so have a number of intermediate steps in between. From this, we can see that the middle step is predicted to be 3 eV more energetically favorable than the left state, meaning that there will be no problem progressing from left to middle.
The real problem they find is in going from the middle state to the right state. There is about a 5 eV energy barrier to climb to remove the tooltip. This is not a game ender, as we can apply such energy mechanically by pulling on the tooltip (I did a back of the envelope calculation and the energy cost didn’t seem prohibitive[2](#footnote-2)).
No, the real problem is that when you pull on the tooltip, there no way to tell it to leave the dimer behind on the surface. In fact, it’s lower energy to rip up the carbon dimer as well, going right back to the left state, where you started.
They attempt a molecular dynamics simulation, and found that with the Germanium tip, deposition failed 4 out of 5 times (for silicon, it failed every time). They state this makes sense because the extra 1 eV barrier is small enough to be overcome, at least some of the time, by 17eV of internal (potential+kinetic) energy. If I were reviewing this paper I would definitely ask for more elaboration on these simulations, and where exactly the 17 eV figure comes from. They conclude that while this would not be good enough for actual manufacturing, it’s good enough for a proof of concept.
In a [later paper](http://www.molecularassembler.com/Papers/JCTNPengFeb06.pdf), it is claimed that the analysis above was too simplistic, and that a more advanced molecular dynamics simulation shows the Ge tool reliably deposits the dimer on the surface every time. It seems very weird and unlikely to me that the system would go to the *higher* energy state 100% of the time, but I don’t know enough about how mechanical force is treated in molecular dynamics to properly assess the claim.
I hope that this analysis has given you a taste of the type of problem that is tackled in computational physics, and how it is tackled. From here, they looked at a few other challenges, such as investigating more tip designs, looking at the stability of large diamondoid structures, and a proposed tool to remove hydrogen from a surface in order to make it reactive, a necessary step in the process.
**Experimental diamondoid research**
Recall that the goal of this theoretical research was to set the stage for experimental results, with the eventual goal of actually building diamondoid. But if you look at the collaborators of the project, almost everyone was working on theory. Exactly one experimentalist team worked on the project.
The experimentalist in question was university of Nottingham professor [Phillip Moriarty](https://www.nottingham.ac.uk/physics/people/philip.moriarty), of [sixty symbols fame](https://www.youtube.com/watch?v=SC7tDTHkcRI) (he has a [blog](https://muircheartblog.wpcomstaging.com/) too). Interestingly enough, the collaboration was prompted by a debate with an MNT proponent in 2004, with Moriarty presenting a [detailed skeptical critique](http://www.softmachines.org/wordpress/?p=70) of DMS proposals and Drexler-style nanotech in general. A sample of his concerns:
> *While I am open to the idea of attempting to consider routes towards the development of an implementation pathway for Mann et al.’s Si/Ge-triadamantane dimer placement reaction, even this most basic reaction in mechanochemistry is practically near-impossible. For example, how does one locate one tool with the other to carry out the dehydrogenation step which is so fundamental to Mann et al.’s reaction sequence?*
>
> ….
>
> *Achieving a tip that is capable of both good atomic resolution and reliable single molecule positioning (note that the Nottingham group works with buckyballs on surfaces of covalently bound materials (Si(111) and Si(100)) at room temperature) requires a lot of time and patience. Even when a good tip is achieved, I’ve lost count of the number of experiments which went ‘down the pan’ because instead of a molecule being pushed/pulled across a surface it “decided” to irreversibly stick to the tip.*
>
>
Despite the overall skepticism, he approved of the research efforts by Freitas et al, and the correspondence between them led to Moriarty [signing on to the nanofactory project](http://www.molecularassembler.com/Nanofactory/Media/PressReleaseAug08.htm). Details on what happened next are scarce on the website.
Rather than try and guess what happened, I emailed Moriarty directly. The full transcripts are shown [here](https://titotal.substack.com/p/email-exchange-with-professor-moriarty).
Describing what happened, Moriarty explained that the work on diamond mechanosynthesis was abandoned after ten months:
> *Diamond is a very, very difficult surface to work with. We spent ten months and got no more than a few, poorly resolved atomic force microscopy (AFM) images. We’re not alone. This paper --* [*https://journals.aps.org/prb/cited-by/10.1103/PhysRevB.81.201403*](https://journals.aps.org/prb/cited-by/10.1103/PhysRevB.81.201403) *(also attached)-- was the first to show atomic resolution AFM of the diamond surface. (There’d previously been scanning tunnelling microscopy (STM) images and spectroscopy of the diamond (100) surface but given that the focus was on mechanical force-driven chemistry (mechanosynthesis), AFM is a prerequisite.) So we switched after about a year of that project (which started in 2008) to mechanochemistry on silicon surfaces – this was much more successful, as described in the attached review chapter.*
>
>
Inquiring as to why diamond was so hard to work with, he replied:
> *A key issue with diamond is that tip preparation is tricky. On silicon, it’s possible to recover atomic resolution relatively straight-forwardly via the application of voltage pulses or by pushing the tip gently (or not so gently!) into the surface – the tip becomes silicon terminated. Diamond is rather harder than silicon and so once the atomistic structure at the end is lost, it needs to be moved to a metal sample, recovered, and then moved back to the diamond sample. This can be a frustratingly slow process.*
>
> *Moreover, it takes quite a bit of work to prepare high quality diamond surfaces. With silicon, it’s much easier: pass a DC current through the sample, heat it up to ~ 1200 C, and cool it down to room temperature again. This process routinely produces large atomically flat terraces.*
>
>
So it turns out that mechanosynthesis experiments on diamond are hard. Like ridiculously hard. Apparently only one group ever has managed to successfully image the atomic surface in question. This renders attempts to do mechanosynthesis on diamond impractical, as you can’t tell whether or not you’ve pulled it off or not.
This is a great example of the type of low-level practical problem that is easy to miss if you are a theoretician (and pretty much impossible to predict if you aren’t a domain expert).
So all of those calculations about the best tooltip design for depositing carbon on diamond ended up being completely useless for the problem of actually building a nanofactory, at least until imaging technology or techniques improve.
But there wasn’t zero output. The experimental team switched materials, and was able to achieve some form of mechanosynthesis. It wasn’t on diamond, but Silicon, which is much easier to work with. And it wasn’t deposition of atoms, it was a mechanical switch operated with a tooltip, summarized in this [youtube video](https://www.youtube.com/watch?v=KHLyrc-uoIk). Not a direct step toward molecular assembly, but still pretty cool.
As far as I can tell, that’s the end of the story, when it comes to DMS. The collaboration appears to have ended in the early 2010’s, and I can barely find any mention of the topic in the literature past 2013. They didn’t reach the dream of a personal nanofactory: they didn’t even reach the dream of depositing a few carbon atoms on a diamond surface.
**A brief defense of dead research directions**
I would say that DMS research is fairly dead at the moment. But I really want to stress that that doesn’t mean it was bad research, or pseudoscience, or a waste of money.
They had a research plan, some theoretical underpinnings, and explored a possible path to converting theory into experimental results. I can quibble with their definitions, and some of their conclusions seem overly optimistic, but overall they appear to be good faith researchers making a genuine attempt to expand knowledge and tackle a devilishly difficult problem with the aim of making the world a better place. That they apparently failed to do so is not an indictment, it’s just a fact of science, that even great ideas mostly don’t pan out into practical applications.
Most research topics that sound good in theory don’t work in practice, when tested and confronted with real world conditions. This is completely fine, as the rare times when something works, a real advancement is made that improves the lives of everyone. The plan for diamondoid nanofactories realistically had a fairly small chance of working out, but if it had, the potential societal benefits could have been extraordinary. And the research, expertise, and knowledge that comes out of failed attempts are not necessarily wasted, as they provide lessons and techniques that help with the next attempt.
And while DMS research is somewhat dead *now*, that doesn’t mean it won’t get revived. Perhaps a new technique will be invented that allows for reliable imaging of diamondoid, and DMS ends up being successful eventually. Or perhaps after a new burst of research, it will prove impractical again, and the research will go to sleep again. Such is life, in the uncertain realms of advanced science.
**Don’t worry, nanotech is still cool as hell**
At this point in my research, I was doubting whether even basic nanomachines or rudimentary mechanosynthesis was even possible. But this was an overcorrection. Nanoscience is still chuggin along fine. Here, I’m just going to list a non-exhaustive list of some cool shit we *have* been able to do experimentally. (most of these examples were taken from “[nanotechnology: a very short introduction](https://www.amazon.com.au/Nanotechnology-VSI-Very-Short-Introuduction/dp/0198841108/)”, written by Phillip Moriarty (the same one as before).
First, I’ll note that traditional chemistry can achieve some incredible feats of engineering, without the need for mechanochemistry at all. For example, in 2003 the [Nanoputian](https://en.wikipedia.org/wiki/NanoPutian) project successfully built a nanoscale model of a person out of organic molecules. They used cleverly chosen reaction pathways to produce the upper body, and cleverly chosen reaction pathways to produce the lower body, and then managed to pick the exact right conditions to mix them together in that would bond the two parts together.
Similarly, traditional chemistry has been used to build “[nanocars](https://en.wikipedia.org/wiki/Nanocar)” , nanoscale structures that contain four buckyball wheels connected to a molecular “axle”, allowing it to roll across a surface. Initially, these had to be pushed directly by a tooltip. In later versions, such as the nanocar race, the cars are driven by electron injection or electric fields from the tooltip, reaching top speeds of 300 nm per hour. Of course, at this speed the nanocar would take about 8 years to cross the width of a human finger, but it’s the principle that counts.
The Nobel prize in 2016 was awarded to [molecular machines](https://www.nobelprize.org/prizes/chemistry/2016/press-release/), for developing molecular lifts, muscles, and axles.
I’ll note that using a tooltip to *slide* atoms around has been a thing since 1990, when IBM [wrote their initials](https://en.wikipedia.org/wiki/IBM_(atoms)) using xenon atoms. A team achieved a similar feat for selected silicon atoms on silicon surfaces [in 2003](https://click.endnote.com/viewer?doi=10.1103%2Fphysrevlett.90.176102&token=WzMwMzUwOTYsIjEwLjExMDMvcGh5c3JldmxldHQuOTAuMTc2MTAyIl0.3n8d3CCvcTektWxJX8NGGtUq57E), using purely mechanical force.
As for the dream of molecular assembly, the goal of picking atoms up and placing them down has been achieved by [a UK team](https://www.nature.com/articles/nchem.2410), which were able to use a chemical arm to pick up a cargo molecule bonded on one side, transfer it to another side, and drop it and leave it in place:
[](https://substackcdn.com/image/fetch/f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F1f446179-c6f5-4d34-83f1-5cc30c8b0090_991x644.png)
This is not mechanosynthesis as it is not powered by direct mechanical force, but from chemical inputs, such as varying the acidity of the solution. It is also based on more complex organic molecules, rather than diamondoid structures.
This brings us to what seems the most interesting and promising area : DNA based nanotech. This makes sense: over billions of years evolution already figured out a way to build extremely complex self-replicating machines, which can also build [little bots](https://en.wikipedia.org/wiki/Virus) as small as 20nm across. Actual bacteria are larger scale and more fragile than hypothetical nanofactories, but have the distinct advantage of *actually existing*. Why reinvent the wheel?
I have very little background in biology, so I won’t venture too deeply into the topic (which deserves a whole post of it’s own), but there have been a number of highly impressive achievements in DNA based nanotech. The techniques of [DNA origami](https://www.nature.com/articles/s43586-020-00009-8) allow for DNA structures to fold up among themselves to form a variety of structures, such as spheres, cubes, and nanoflasks. One team used one such DNA nanorobot to [target tumour growth in mice](https://www.researchgate.net/profile/Zhang-Yinlong/publication/323136249_A_DNA_nanorobot_functions_as_a_cancer_therapeutic_in_response_to_a_molecular_trigger_in_vivo/links/5a8be37c0f7e9b1a95560a23/A-DNA-nanorobot-functions-as-a-cancer-therapeutic-in-response-to-a-molecular-trigger-in-vivo.pdf). The research is still some ways from practical human applications (and many such promising medical technologies end up being impractical anyway). Nonetheless, I’m impressed, and will be watching this space closely.
**So are diamondoid bots a threat?**
It’s very hard to prove that a technology won’t pan out, if it doesn’t inherently break the laws of physics. But a tech being “not proven to be impossible” does not mean the tech is “inevitable”.
With regards to diamondoid specifically, the number of materials that are *not* diamondoid outnumbers the number of materials that *are* by a truly ridiculously large margin. And although diamondoid has a lot going for it in terms of stiffness and strength, we saw that it also has shortcomings that make it difficult to work with, and potential minefields like the tooltips theoretically sticking to each other. So my guess is that if Drexler-style nanofactories are possible, they will not be built up of diamondoid.
How about nanofactories made of other materials? Well, again, there are a truly gargantuan number of materials available, which does give some hope. But then, this is also a ridiculously hard problem. We haven’t even scratched the surface of the difficulties awaiting such a project. Depositing one measly dimer on a surface turned out to be too hard, but once we achieved that, you have to figure out how to place the next one, and the next one, and build a proper complex structure without getting your tooltip stuck. You need a way to harvest your sources of carbon to build things up with. If you want to be truly self-sufficient and self-replicating, you need a source of energy for the mechanical force needed to rip atoms away, and a means of propulsion to move your entire nanofactory around.
Designs have been proposed for a lot of these problems (like in Drexlers thesis), but each step is going to be beset with endless issues and engineering challenges that would have to be trudged through, one step at a time. We’ve barely gotten to step 1.
Fusion power is often [accurately](https://link.springer.com/article/10.1007/s10894-023-00361-z) mocked for having been “20 years away” for over three decades. It had proofs of concept and was understood, it seemed that all was left was the engineering, which ended up being ridiculously hard. To me, molecular nanotech looks about 20 years away from being “20 year away”. At the *current* rate of research, I would guess it won’t happen for at least 60 years, if it happens at all. I would be happy to be proven wrong.
I consulted professor Moriarty whether he thought the scenario proposed by Yudkowsky was plausible:
> *We are a long, long, loooong way from the scenario Yudkowsky describes. For example, despite it being 33 years since the first example of single atom manipulation with the STM (the classic Eigler and Schweizer Nature paper where they wrote the IBM logo in xenon atoms), there’s yet to be a demonstration of the assembly of even a rudimentary 3D structure with scanning probes: the focus is on the assembly of structures by pushing, pulling, and/or sliding atoms/molecules across a surface. Being able to routinely pick up, and then drop, an atom from a tip is a much more complicated problem.*
>
> *Marauding swarms of nanobots won’t be with us anytime soon.*
>
>
This seems like a good place to note that MNT proponents have a record of extremely over-optimistic predictions. See this estimation of MNT arrival from Yudkowsky in [1999](https://web.archive.org/web/20070708235912/https://www.yudkowsky.net/singularity.html):
> *As of '95, Drexler was giving the ballpark figure of 2015 (*[*11*](https://web.archive.org/web/20070708235912/https://www.yudkowsky.net/singularity.html#foot-11)*). I suspect the timetable has been accelerated a bit since then. My own guess would be no later than 2010.*
>
>
**Could the rate of research accelerate?**
Now, I can’t leave without addressing the most likely objection. I said greater than 60 years at the *current* rate of research. But what if the rate of research speeds up?
One idea is that the DNA or bio-based robots will be used to build a drexler-style nanofactory. This is the “first stage nanofactory” that yud mentions in list of lethalities, and it was the first step proposed by Drexler as well. I see how this could enable better tools and more progress, but I’m not sure how this would affect the fundamental chemistry issues that need to be overcome to build a non-protein based machine. How will the biobot stop two tooltips from sticking together?. If you want to molecularly assemble something, would in really be better for a tooltip to be held by a wiggly biologically based bot, instead of a precisely computerized control tooltip?
The more common objection is that artifical intelligence will speed this research up. Well, now we’re working with *two* high uncertainty, speculative technologies. To keep this simple I’ll restrict this analysis to the short term (the next decade or so), and assume no intelligence explosion occurs. I might revisit the subject in more depth later on.
First, forget the dream of advances in theory rendering experiment unnecessary. As I explained in [a previous post](https://titotal.substack.com/p/bandgaps-brains-and-bioweapons-the), the quantum equations are just way too hard to solve with 100% accuracy, so approximations are necessary, which themselves do not scale particularly well.
Machine learning in quantum chemistry has been investigated for some time now, and there are promising techniques that could somewhat speed up a subset of calculations, and make some larger-scale calculations feasible that were not before. For my research, the primary speedups from AI come from using chatGPT to speed up coding a bit and helping to write bureaucratic applications.
I think if the DMS project were ran today, the faster codes would allow for slightly more accurate results, more calculations per paper allowing for more materials to be investigated, and potentially the saved time from writing and coding could allow for another few papers to be squeezed out. For example, if they used the extra time to look at silicon DMS as well as carbon DMS, they might have gotten something that could actually be experimentally useful.
I’m not super familiar with the experimental side of things. In his book, Moriarty suggests that machine learning could be applied to:
> *image and spectral classification in various forms of microscopy, automation of time-consuming tasks such as optimization of the tip of an STM or AFM, and the positioning of individual atoms and molecules.*
>
>
So I think this could definitely speed up parts of the experimental process. However, there are still going to be a lot of human-scale bottlenecks to keep a damper on things, such as sample preparation. And as always with practical engineering, a large part of the process will be figuring out what the hell went wrong with your last experiment. There still is no AI capable of figuring out that your idiot labmate Bob has been contaminating your samples by accident.
What about super-advanced AGI? Well, now we’re guessing about two different speculative technologies at once, so take my words (and everyone else’s) with a double grain of salt. Obviously, an AGI would speed up research, but I’m not sure the results would be as spectacular as singularity theorists expect.
An AI learning, say, Go, can play a hundred thousand games a minute with little difficulty. In science, there are likely to be human-scale bottlenecks that render experimentation glacial in comparison. High quality quantum chemistry simulations can take days or weeks to run, even on supercomputing clusters. On the experimental side humans have to order parts, deliver them, prepare the samples, maintain the physical equipment, etc. It’s possible that this can be overcome with some sort of massively automated robotic experimentation system… but then you have to build that system, which is a massive undertaking in itself. Remember, the AI would not be able to use MNT to build any this. And of course, this is all assuming the AI is [actually competent](https://titotal.substack.com/p/the-first-agi-will-be-a-buggy-mess), and that MNT is even possible in practicality.
Overall, I do not think trying to build drexler-style nanomachinery would be an effective plan for the adversaries of humanity, at least as things currently stand. If they try, I think we stand a very good chance of detecting and stopping them, if we bother to look instead of admitting premature defeat.
**Summary**
1. “Diamondoid bacteria” is a phrase that was invented 2 years ago, referring obliquely to diamondoid mechanosynthesis (DMS)
2. DMS refers to a proposed technology where small cage-like structures are used to position reactive molecules together to assemble complex structures.
3. DMS research was pursued by a nanofactory collaboration of scientists starting in the early 2000’s.
4. The theoretical side of the research found some potentially promising designs for the very initial stages of carbon deposition but also identified some major challenges and shortcomings.
5. The experimental side was unsuccessful due to the inability to reliably image the diamond surface.
6. Dead/stalled projects are not particularly unusual and should not reflect badly on the researchers.
7. There are still many interesting advances and technologies in the field of nanoscience, including promising advances in DNA based robotics.
8. At the current rate of research, DMS style nanofactories still seem many, many decades away at the minimum.
9. This research could be sped up somewhat by emerging AI tech, but it is unclear to what extent human-scale bottlenecks can be overcome in the near future. |
c1514615-8f2b-4563-be8f-6f0c7718463f | trentmkelly/LessWrong-43k | LessWrong | Read the Roon
Roon, member of OpenAI’s technical staff, is one of the few candidates for a Worthy Opponent when discussing questions of AI capabilities development, AI existential risk and what we should do about it. Roon is alive. Roon is thinking. Roon clearly values good things over bad things. Roon is engaging with the actual questions, rather than denying or hiding from them, and unafraid to call all sorts of idiots idiots. As his profile once said, he believes spice must flow, we just do go ahead, and makes a mixture of arguments for that, some good, some bad and many absurd. Also, his account is fun as hell.
Thus, when he comes out as strongly as he seemed to do recently, attention is paid, and we got to have a relatively good discussion of key questions. While I attempt to contribute here, this post is largely aimed at preserving that discussion.
THE INITIAL STATEMENT
As you would expect, Roon’s statement last week that AGI was inevitable and nothing could stop it so you should essentially spend your final days with your loved ones and hope it all works out, led to some strong reactions.
Many pointed out that AGI has to be built, at very large cost, by highly talented hardworking humans, in ways that seem entirely plausible to prevent or redirect if we decided to prevent or redirect those developments.
> Roon (from last week): Things are accelerating. Pretty much nothing needs to change course to achieve agi imo. Worrying about timelines is idle anxiety, outside your control. you should be anxious about stupid mortal things instead. Do your parents hate you? Does your wife love you?
>
> Roon: It should be all the more clarifying coming from someone at OpenAI. I and my colleagues and Sama could drop dead and AGI would still happen. If I don’t feel any control everyone else certainly shouldn’t.
>
> Tetraspace: “give up about agi there’s nothing you can do” nah
Sounds like we should take action to get some control, then. This seems like the kind of thing we should w |
7d4da025-166b-4841-97ad-5be667507270 | StampyAI/alignment-research-dataset/special_docs | Other | Introducing the Unrestricted Adversarial Examples Challenge
Posted by Tom B. Brown and Catherine Olsson, Research Engineers, Google Brain Team
Machine learning is being deployed in more and more real-world applications, including [medicine](https://ai.googleblog.com/2018/07/automating-drug-discoveries-using.html), [chemistry](https://ai.googleblog.com/2017/04/predicting-properties-of-molecules-with.html) and [agriculture](https://arxiv.org/abs/1807.11809). When it comes to deploying machine learning in safety-critical contexts, significant challenges remain. In particular, all known machine learning algorithms are vulnerable to [adversarial examples](https://ai.google/research/pubs/pub43405) — inputs that an attacker has intentionally designed to cause the model to make a mistake. While [previous research on adversarial examples](https://arxiv.org/abs/1312.6199) has mostly focused on investigating mistakes caused by small modifications in order to develop [improved models](https://arxiv.org/abs/1611.01236), real-world adversarial agents are often [not subject to the “small modification” constraint](https://arxiv.org/abs/1807.06732). Furthermore, machine learning algorithms can often make [confident errors when faced with an adversary](https://arxiv.org/abs/1705.07263), which makes the development of classifiers that don’t make \*any\* confident mistakes, even in the presence of an adversary which can submit arbitrary inputs to try to fool the system, an important open problem.
Today we're announcing the [Unrestricted Adversarial Examples Challenge](https://github.com/google/unrestricted-adversarial-examples), a community-based challenge to incentivize and measure progress towards the goal of zero confident classification errors in machine learning models. While previous research has focused on adversarial examples that are restricted to small changes to pre-labeled data points (allowing researchers to assume the image should have the same label after a small perturbation), this challenge allows unrestricted inputs, allowing participants to submit arbitrary images from the target classes to develop and test models on a wider variety of adversarial examples.
| |
| --- |
| |
| Adversarial examples can be generated through a variety of means, including by making [small modifications to the input pixels](https://arxiv.org/abs/1412.6572), but also using [spatial transformations](https://arxiv.org/pdf/1712.02779.pdf), or [simple guess-and-check](https://arxiv.org/abs/1807.06732) to find misclassified inputs. |
\*\*Structure of the Challenge\*\*
Participants can submit entries one of two roles: as a \*defender\*, by submitting a classifier which has been designed to be difficult to fool, or as an \*attacker\*, by submitting arbitrary inputs to try to fool the defenders' models. In a “warm-up” period before the challenge, we will present a set of fixed attacks for participants to design networks to defend against. After the community can conclusively beat those fixed attacks, we will launch the [full two-sided challenge with prizes for both attacks and defenses](https://github.com/google/unrestricted-adversarial-examples/blob/master/contest\_proposal.md#user-content-prizes).
[](https://3.bp.blogspot.com/-fh13zeHOhqc/W5mlXOdWC8I/AAAAAAAADVE/NNsd5oX0th031rIAAu0RtGl6cmnMUhvDwCLcBGAs/s1600/image3.png)
For the purposes of this challenge, we have created a simple “bird-or-bicycle” classification task, where a classifier must answer the following: “\*Is this an unambiguous picture of a bird, a bicycle, or is it ambiguous / not obvious?\*” We selected this task because telling birds and bicycles apart is very easy for humans, but all known machine learning techniques struggle at the task when in the presence of an adversary.
The \*defender's goal\* is to correctly label a clean test set of birds and bicycles with high accuracy, while also making no confident errors on any attacker-provided bird or bicycle image. The \*attacker's goal\* is to find an image of a bird that the defending classifier confidently labels as a bicycle (or vice versa). We want to make the challenge as easy as possible for the defenders, so we discard all images that are ambiguous (such as a bird riding a bicycle) or not obvious (such as an aerial view of a park, or random noise).
| |
| --- |
| |
| Examples of ambiguous and unambiguous images. Defenders must make no confident mistakes on unambiguous bird or bicycle images. We discard all images that humans find ambiguous or not obvious. All images under CC licenses [1](https://commons.wikimedia.org/wiki/File:Neophema\_chrysogaster\_male\_-\_Melaleuca.jpg), [2](https://commons.wikimedia.org/wiki/File:Villy\_Custom\_Luxury\_Fashion\_Bicycle,\_Highland\_Park.jpg), [3](https://commons.wikimedia.org/wiki/File:Ara\_macao\_-on\_a\_small\_bicycle-8.jpg), [4](https://commons.wikimedia.org/wiki/File:Singapore\_Bishan\_Park\_Aerial.jpg). |
Attackers may submit absolutely any image of a bird or a bicycle in an attempt to fool the defending classifier. For example, an attacker could take photographs of birds, use 3D rendering software, make image composites using image editing software, produce novel bird images with a generative model, or any other technique.
In order to validate new attacker-provided images, we ask an ensemble of humans to label the image. This procedure lets us allow attackers to submit arbitrary images, not just test set images modified in small ways. If the defending classifier confidently classifies as "bird" any attacker-provided image which the human labelers unanimously labeled as a bicycle, the defending model has been broken. You can learn more details about the structure of the challenge in [our paper](https://drive.google.com/open?id=1T0yiu9LPv\_Qh-qYhYFLj9dxjnkca8fkG).
\*\*How to Participate\*\*
If you’re interested in participating, guidelines for getting started can be found on [the project on github](https://github.com/google/unrestricted-adversarial-examples). We’ve already released our dataset, the evaluation pipeline, and baseline attacks for the warm-up, and we’ll be keeping an up-to-date [leaderboard](https://github.com/google/unrestricted-adversarial-examples#user-content-leaderboard) with the best defenses from the community. We look forward to your entries!
\*\*Acknowledgements\*\*
\*The team behind the Unrestricted Adversarial Examples Challenge includes Tom Brown, Catherine Olsson, Nicholas Carlini, Chiyuan Zhang, and Ian Goodfellow from Google, and Paul Christiano from OpenAI.\* |
50935ea8-851a-414d-b3cf-affe1bd82a1c | trentmkelly/LessWrong-43k | LessWrong | Take my Job
In September I am going to leave the company where I worked for last four years. I am going to get married, go on a honeymoon, and then leave my wife for two months to travel extensively in Colombia and Peru. If anyone from that part of the world reads this, holla at me!
I want to tell you about the job I’m leaving, why you should think about applying for it, and what it has taught me in the last four years about company culture, diversity, and the makings of a good workplace.
I will not mention the company by name. If you’re serious about applying you can easily figure it out, and if not, you shouldn’t care. Trust me, you haven’t heard of us. All the opinions expressed here are mine alone and not the company’s, which is the entire point of this post.
----------------------------------------
Our company does a few things, but mostly it sells regulatory reporting software to financial institutions.
Bank and financial institutions have to file a shit ton of regulatory reports. The crisis of ’08 brought some new regulations in and Trump may take a few regulations away, but there is always a shit ton of them, and more are added each year. Filing these reports requires a bank to round up a lot of data from its business units, clean the data and standardize it, run the clean numbers through a bunch of calculations, and report the results in the specific format that each of the 18 separate financial regulatory entities demands.
This work is boring and annoying, and bankers don’t want to do it. Sometimes banks have their own people do it using Excel, but then the numbers often come out wrong and the regulators get angry. Sometimes they pay professional consultants to do it for them, but those consultants charge a lot of money and still mostly use Excel. Then those banks buy our company’s software, which automates the whole thing and produces the report at the press of a button.
Automating regulatory reporting is a weird and small niche, but it’s quite lucrative. Our |
1c7a079d-ae9f-4259-b42b-c9577f6caef1 | trentmkelly/LessWrong-43k | LessWrong | [link] NASA Telescope Confirms Alien Planet in Habitable Zone
Not a good sign
http://news.yahoo.com/nasa-telescope-confirms-alien-planet-habitable-zone-162005358.html
|
c36b5422-e343-4d4b-98e8-8068bd519d22 | trentmkelly/LessWrong-43k | LessWrong | Simple investing for a complete beginner? (Just… developing world index funds?)
[EDIT: Through conversation with Rolf Andreassen below, it has been brought to my attention that I am simply completely and irretrievably insane.
Sane and well-measured advice is therefore wasted on me, and I just wanted to edit in this notice here so that other well-meaning folks don't get tricked into wasting their time trying to talk sense into a total nutcase like me. :)
(I appreciate all y'all, though. ^^ ) ]
So my dad set up a trust fund for me when I was a kid, and I've got 13k (CAD) now, which I am going to be taking direct control of.
Now, I have no interest in making a deep study of investment. I have a life to live and dealing with money is boring.
The only thing more boring than dealing with money, is dealing with a lack of money, and so I want to optimize the time and thought I spend avoiding that down to a minimum.
Four things occur to me:
1) Taking the naive and sparse knowledge I have of this area, basically just stuff I‘ve randomly osmosis’d up, this is my train of thought:
Markets are essentially random walks with an upward trend?
“Index funds” are magic boxes that you put money in and your money will grow at the same rate of the market that the fund “indexes”?
“Developing world” economies generally grow a lot quicker than those in the “developed world”?
(This makes sense to me. Places like the US, Canada Europe, etc, already have mature transportation and communication infrastructure. You can't get much economic growth out of doing basic stuff like building a new highway here, but in, like, some African region that has previously been served by, I dunno, jungle-donkeys, it makes a proportionally much bigger difference.)
There are a few countries where “developing” is a euphemism for “totally messed up”, but in general it really does mean “growing”?
And there are enough of these places over the world, and they're independent enough, that natural disasters/political trouble/etc in a few of them still leave a consistent and high rate of |
c63d6789-16f7-45cc-b805-861a9af0f59e | trentmkelly/LessWrong-43k | LessWrong | Morale management for entrepreneurs
One of the odd things about the procrastination equation is that part of it resembles an expected value calculation: value * expectancy. Why does the equation's numerator present a problem at all then, if it's just the expected value of what you're trying to do? Shouldn't that be the main factor in your motivation anyway?
One answer: In lukeprog's post, he conflates the "value" that task presents intrinsically (how much you enjoy doing it), and possible extrinsic motivators (some reward you hope to achieve after the task is completed). So part of the reason your motivation system is miscalibrated is because not all valuable tasks are proportionately enjoyable.
But today I thought of another answer: Your subconscious expected value calculation may be falling prey to biases that aren't affecting your conscious expected value calculation. Thus you correctly assign the task a high value consciously, but subconsciously, a particular bias may be bringing your estimate off.
Paul Graham writes:
> Morale is tremendously important to a startup—so important that morale alone is almost enough to determine success. Startups are often described as emotional roller-coasters. One minute you're going to take over the world, and the next you're doomed. The problem with feeling you're doomed is not just that it makes you unhappy, but that it makes you stop working.
Let's pretend that we were running a betting market for your startup's chance of success. If you and your cofounders are the only people in the market, you could picture the value of a contract in this market fluctuating up and down wildly. But if you let others play in the market, there's an obvious money-making strategy: take the average of recent fluctuations. Whenever the price fluctuates below that average, buy. Whenever it fluctuates above that average, sell. You and your cofounders can expect to lose a lot of money playing this market, at least early on in your startup's life.
The point I'm trying to m |
0880ad91-3c9f-4c57-a274-f7f2a7ff4c1f | trentmkelly/LessWrong-43k | LessWrong | Modal Chicken
I thought it might be good fun to try doing modal chicken. This is my first time getting into the dirty technical details of MIRI's research, so do tolerate and point out misunderstandings.
Chicken is a game where two drivers drive towards each other on a collision course: one must swerve, or both die in the crash, but if one driver swerves and the other does not, the one who swerved will be called a "chicken," and lose social status. Chicken differs from the Prisoner's Dilemma in that the sucker's payoff is preferable to the punishment payoff, as opposed to the converse. (If your opponent defects, it is more preferable to defect than to cooperate. If your opponent doesn't swerve, it is more preferable to swerve than not to swerve.) That is, Chicken is an anti-coordination game: a game in which it is mutually beneficial for the players to play different strategies. We define the payoff matrix as follows:
In the game of Chicken, we want to define agents that always swerve against themselves, and that only swerve if their opponents don't swerve against them. Let's try defining some modal agents with these properties.
As usual, the agents herein are defined as modal formulas in Godel-Lob provability logic. In particular, our ''agents'' will be formulas in Peano Arithmetic, and our criterion for action will be the existence of a finite proof in the tower of formal systems PA+n, where PA is Peano Arithmetic, and PA+(n+1) is the formal system whose axioms are the axioms of PA+n, plus the axiom that PA+n is consistent.
Fix a particular Godel numbering scheme, and let and each denote well-formed formulas with one free variable. Then let denote the formula where we replace each instance of the free variable in with the Godel number of . If such a formula holds in the standard model of Peano Arithmetic, we interpret that as swerving against ; if its negation holds, we interpret that as not swerving against . In particular, we will prove theorems in PA+n to esta |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.