id stringlengths 36 36 | source stringclasses 15 values | formatted_source stringclasses 13 values | text stringlengths 2 7.55M |
|---|---|---|---|
a97498b1-39ad-464e-abe6-852f03e4eea6 | trentmkelly/LessWrong-43k | LessWrong | The case for multi-decade AI timelines [Linkpost]
So this post is an argument that multi-decade timelines are reasonable, and the key cruxes that Ege Erdil has with most AI safety people who believe in short timelines are due to the following set of beliefs:
1. Ege Erdil doesn't believe that trends exist that require AI to automate everything in only 2-3 years.
2. Ege Erdil doesn't believe that the software-only singularity is likely to happen, and this is perhaps the most important crux he has with AI people like @Daniel Kokotajlo who believe that a software-only singularity is likely.
3. Ege Erdil expects Moravec's paradox to bite hard once AI agents are made in a big way.
This is a pretty important crux, because if this is true, a lot more serial research agendas like Infra-Bayes research, Natural Abstractions work, and human intelligence augmentation can work, and also it means that political modeling (like is the US economy going to be stable long-term) matter a great deal more than is recognized in the LW/EA community.
Here's a quote from the article:
> * I don’t see the trends that one would extrapolate in order to arrive at very short timelines on the order of a few years. The obvious trend extrapolations for AI’s economic impact give timelines to full remote work automation of around a decade, and I expect these trends to slow down by default.
> * I don’t buy the software-only singularity as a plausible mechanism for how existing rates of growth in AI’s real-world impact could suddenly and dramatically accelerate by an order of magnitude, mostly because I put much more weight on bottlenecks coming from experimental compute and real-world data. This kind of speedup is essential to popular accounts of why we should expect timelines much shorter than 10 years to remote work automation.
> * I think intuitions for how fast AI systems would be able to think and how many of them we would be able to deploy that come from narrow writing, coding, or reasoning tasks are very misguided due to Moravec’s parad |
7ecf9d57-5ff6-447e-a033-3d48b1a33d54 | trentmkelly/LessWrong-43k | LessWrong | Risk Budgets vs. Basic Decision Theory
I've been wondering about the decision theory of having a "risk profile", e.g. "200 microcovids per week", and at least under some simplistic but not unreasonable assumptions I don't think it makes sense.
Let's say we just have one activity A which yields some fixed positive utility U but has a certain probability p of giving you covid which has a large cost C. Then the expected utility of A is simply E = U - pC. If E > 0 then you should just do A as much as possible, otherwise not at all. There's no cap on total covid risk. (OTOH the "risk profile" method would recommend doing A about 200e-6/p times.)
One counterargument is that the positive utilities don't add up linearly. That is, just because going to a restaurant once is nice, doesn't mean you'd enjoy constantly being in one. But I think this is a red herring -- you could instead imagine a variety of different activities (movies, dating, eating out, parties, etc.) that do add up more or less linearly. You could also just be the type of person who gets the same enjoyment from eating out every day. Either way, the issue of "where does microcovid budget come from?" remains.
I think the crux of the matter is that (consequentialist) decision theory is Markovian. When you make a decision, you only care about the state you're in right now, not in how you got there. So whether or not you did some risky activity yesterday, unless it actually gave you covid, there's no effect on your current state, and so therefore it shouldn't affect your current and future decisions about whether to engage in something risky.
To be clear, my goal isn't to abandon risk budgets -- they seem very sensible to me, and I don't have a good replacement. But, I'd like to know if there is some better model which captures the intuition around risk budgets (or an error in the above reasoning). |
0184523b-b12f-4b29-91a0-c6817e867d58 | trentmkelly/LessWrong-43k | LessWrong | Sidekick Matchmaking - How to tackle the problem?
Some of us enjoy being sidekicks.
Some of us would like to meet sidekicks in potential, see how the interaction goes, and have sidekicks.
Last time I tried posting about sidekick matchmaking here, it turned out to be very valuable for me, but not for many people (I think only two pairs of sidekick were created as a result). Now, once again I'd like to find someone who enjoys that role to help me out with many projects.
I'm looking for suggestions on how to get people together to do that. For the time being, if someone needs a sidekick or wants to be one, post about it in the comment section. I'd love to see a permanent solution for this information spreading problem.
My experience with Sidekicks
I'm not sure what Anna and Nick thought of their sidekicks, but my experience was undeniably positive. Having a sidekick was motivating, saved me great time, and, most importantly, felt like I got a surge of muscle strength specifically in the types of tasks I'm particularly inept at.
By contrast, my experience with people hired to help was mixed (virtual assistants) or negative (personal assistant).
Use the comment section to either offer or request sidekicks, explaining a little more about you and what you'd like this partnership to mean |
7e9e34a6-8c73-49bf-97df-85811428667a | trentmkelly/LessWrong-43k | LessWrong | Why So Many Cookie Banners?
Sometimes you'll see people saying things like:
> Using cookies to track state on a website, that is only used for that website, is fine. You don't need to ask for consent.—rrwo
Or:
> You don't need a cookie banner to be allowed to create cookies. You only need them if you're using them for something like tracking.—y4mi
Something like, "as long as you design your site properly and don't abuse storage you don't need to ask your European visitors for permission." While I'm not working in this area anymore, am not a lawyer, and am not attempting to give you legal advice, if you read the regulation this interpretation is completely off.
Cookie banners are a response to the 2002 ePrivacy Directive (full text, guidance). While the ePrivacy Directive may be superseded soon by the (pretty similar) ePrivacy Regulation, it's still the current rule. It requires you to get consent from visitors before you store information on their computer (cookies, localStorage, etc) unless this behavior is "strictly necessary in order to provide an information society service explicitly requested by the subscriber or user" [1]. This isn't "in order to" or even "necessary in order to", it's "strictly necessary in order to". Which is quite firm!
This excludes, for example, using a cookie for basic single-site analytics (4.3), where you want to figure out where users are getting stuck on your site or to populate a "users who viewed this product ended up buying this other product" box. Even though this information helps you improve your site for future visitors, including potentially this one, it isn't 'strictly necessary' for serving this user right now.
If the user puts an item in their shopping cart you can set a cookie, because that's how you honor their request, but it's still quite restrictive (2.3):
> a merchant could set the cookie either to persist past the end of the browser session or for a couple of hours in the future to take into account the fact that the user may acc |
d234ed52-a277-4197-bff6-5c5f976c7289 | trentmkelly/LessWrong-43k | LessWrong | Calculating Kelly
Jacob Falcovich tells us that we should Kelly bet on everything. I discuss whether we should Kelly bet when we'd normally make small-money bets on disagreements. Lsusr reminds us that we aren't very good at intuitively grasping what the Kelly formula actually will say.
Due to Lsusr's post, I took another look at how I actually calculate Kelly. I'll describe the improved formulation I came up with. I'm curious to hear everyone else's thoughts on the quickest ways, or what formula you prefer, or how you intuitively estimate. If you already have opinions, take a moment to think what they are before reading further.
----------------------------------------
In the comments to my post on Kelly, Daniel Filan mentioned:
> FWIW the version that I think I'll manage to remember is that the optional fraction of your bankroll to bet is the expected net winnings divided by the net winnings if you win.
I've found that I remember this formulation, but the difference between "net winnings" and "gross winnings" is enough to make me want to double-check things, and in the few months since writing the original post, I haven't actually used this to calculate Kelly.
"expected net winnings divided by net winnings if you win" is easy enough to remember, but is it easy enough to calculate? When I try to calculate it, I think of it this way:
[probability of success] × [payoff of success] + [probability of failure] × [payoff of failure] all divided by [payoff of success].
This is a combination of five numbers (one being a repeat). We have to calculate probability of failure from the probability of success (ie, 1-p). Then we perform two multiplications, one addition, and one division -- five steps of mental arithmetic.
But the formula is really a function of two numbers (see Lsusr's graph for a vivid illustration). Can we formulate the calculation in a way that feels like just a function of two numbers?
Normally the formula is stated in terms of b, the net winnings if you win. I pref |
016f872c-ef52-4f87-904e-7dd0e5b2c0fd | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | AGI - alignment - paperclip maximizer - pause - defection - incentives
I would like to expose myself to critique.
I hope this is a place where I can receive some feedback + share some of the insights that came to me.
<https://en.wikipedia.org/wiki/Dunning%E2%80%93Kruger_effect> - *"people with low ability, expertise, or experience regarding a certain type of task or area of knowledge tend to overestimate their ability or knowledge"*
I'm somewhere on the spectrum 🤡
1. AGI alignment metrics
========================
To avoid paperclip maximizer and solving climate change by eliminating humans I suggest the following value: **LIFE**
I've embedded this principle into [Network State Genesis](https://genesis.re/) and described it in the [founding document](https://genesis.re/Network-State-Genesis-WHITEPAPER-compressed.pdf) in the following way:
1. No killing *(universally agreed across legal systems and religions)*
2. Health *(including mental health, longevity, happiness, wellbeing)*
3. Biosphere, environment, other living creatures
4. AI safety
5. Mars: backup civilization is fully aligned with the virtue of life preservation
These principles were applicable to the Network State and I think first three of them can be repurposed towards AGI alignment.
*(another core belief is GRAVITY - I believe in GRAVITY - GRAVITY brought us together)*
2. Pause, defection, incentives
===============================
New proof-of-X algorithm to ensure compliance with AI moratorium. Ensuring supercomputers are not used for training more powerful models.
Proof-of-work consumes loads of energy.
It could be a mixture of algorithms, that is more energy friendly:
* peak power for a short amount of time *(solve something complex quickly)*
* proof of storage
A mixture of different algorithms to ensure various elements of the data centre are yielded unsuitable for other means. I know too little about the challenges of operating a data center, I know too little about training AI, ultimately I do not know.
I’m just aware of the incentive to defect and no obvious way to enforce the rules.
* So much easier to prove the existence of aliens.
* So much more difficult to disprove.
* So much easier to prove you did the thing.
* So much more difficult to disprove. |
99f688d4-dc09-4d8c-9319-59de3d05d2b8 | trentmkelly/LessWrong-43k | LessWrong | Parental Guidance: Framing Superintelligence
Epistemic status: riffing and curious.
What are good analogies for the relationship humanity should have with a superintelligence?
I haven’t tended to pay too much attention to academic work exploring framing narratives around AI systems. That said, I do have a humanities background, and generally fall back on a version of Amara’s Law that we tend to overestimate the effects of cultural frames in the short run and underestimate their effects in the long run.
Now, though, it does seem like they’ll play a crucial role in system prompts, at least in the months to come, and therefore to alignment. There also seem to be lots of possible analogies for the role different AI systems could play in relation to society—helpful assistant, oracle, tool, genie, coach, friend—and I hadn’t really taken stock of them.
Some of these terms seem a lot more useful than others.
I think some of these terms describe the role that present systems play well (coach, helpful assistant, tool). The fungibility of current AI systems and the unilateral nature of our relationships with them makes me skeptical that they count as ‘friends’, but I can imagine future generations with different social norms thinking of their relationship with AI systems as friendships.
Notably, all these analogies relate to systems that are human-level or below in their capabilities.
I am less compelled by terms that describe superhuman AI. ‘Oracle’ makes sense when models function like Google Search, as does ‘geni’ for agents, but both terms seem too passive and disinterested to be great descriptions for how actual superintelligences would operate in the world. It seems to me like powerful models will have sophisticated values (built through system prompts/RLHF/etc.) that would be less agnostic to tasks than ‘oracles’ or ‘genii’.
Moreover, and perhaps crucially, I don’t know that asking a model to behave like an oracle or a genie would cause it to act in a way that was robustly good for humanity. Both seem l |
49f8cdc5-6a9e-4384-95b8-8c11bb859f51 | StampyAI/alignment-research-dataset/youtube | Youtube Transcripts | Jaime Sevilla - Projecting AI progress from compute trends
[Music]
hi everyone and welcome back to the
tours data science podcast for an
episode i've wanted to record for quite
some time now
so there's this idea in machine learning
that most of the progress we see in ai
doesn't come from fancy algorithms or
new neural architectures
instead some say ai's progress has been
driven by scaling up compute power data
sets and model sizes and besides those
three ingredients nothing else really
matters through that lens the history of
ai really becomes the history of
processing power and compute budgets and
if that turns out to be true then we
might actually be able to do a decent
job of predicting ai progress by
studying trends in compute power and
their impact on ai development and
that's i wanted to talk to jaime sevilla
an independent researcher and affiliate
researcher at the university of
cambridge's center for the study of
existential risk where he works on
technological forecasting and
understanding trends in ai in particular
now his work's been cited in a lot of
cool places including our world and data
which used his team's data to publish a
whole expose on ai progress and jaime
joined me to talk about his work his
predictions about the future of ai and
all kinds of other cool stuff like that
on this episode of the tourist data
science podcast
[Music]
i'm really excited about this particular
episode i've uh i've been following the
your kind of your work generally on
trends in compute and the reasoning
behind all your work as well like for
quite a little bit of time now and i'm
sort of one of these lurker fans of
yours i think it's fair to say like on
twitter and other platforms i'm really
excited to to share this story with
hopefully the the white wider world or a
larger part of the world i like to start
about with you with your motivation like
getting into the space why why trends in
compute what is so important about
looking at trends in compute at this
stage in our in our life as a species
let's say
one thing that uh that i am really
interested in is the admin of advanced
artificial intelligence uh there's a
plus there's like some reasons to
believe that in the coming few decades
we're gonna see like drastic advances in
the in the practice of artificial
intelligence which is gonna allow us to
automate more and more of society it's
gonna have like wide-ranging
implications from like many new jobs
being created other jobs are being
destroyed but also change changing the
way we approach society and introducing
like new risks into the mix that might
radically alter society forever uh one
thing that i was really interested in
looking into is uh trending like uh
inputs into like machine learning
systems like people having are quite
focused in like measuring outputs we
have like really good benchmarks in
order to assess like how well are we
doing in tasks in computer vision how
well are we doing in tasks in language
models but in the coming years we have
come to learn that there is like lots of
games
that can be had just by taking like the
systems that we already have and just
scaling them up just making them bigger
making them have more parameters uh
training them for longer and like uh
using more data in order to train them
and while there has been like some work
exploring like these implications there
really hasn't been like a historical
survey of like exactly how many
resources have we've been put in into
like these systems and this question is
critical because uh it's gonna allow us
to understand like all these progress
that we have seen in the last two
decades which amount of that is due to
us having like smarter architectures
better ways of like approaching the
problem of artificial intelligence
versus us just having like better
computers and more data to train these
machine learning systems on
yeah and that makes perfect sense it
kind of gives us a window into well a
window into the future it's the only as
far as i can tell it's one of the very
few at least ways we have of projecting
what the future of ai might look like we
don't really know how to go from like an
architecture or a concept of a model to
what its capabilities are likely to be
we tend to be surprised when we find new
capabilities emerging so it's kind of
yeah it is helpful to have like at least
investment in in you know the amount of
dollars or the amount of compute
invested in these models as a a hard
number so we can say okay you know this
much money or this many um gpus gets us
this kind of model and then what can we
expect from the next uh a lot of your
work though is focused specifically on
this idea of transformative artificial
intelligence and
this term tai for short it means a lot
of different things to different people
and it's a source of great debate within
the community i'd like to start with
that then so like what is transformative
ai or what is tai to you
and uh and how do you yeah how do you
think about it in the context of your
work and research
right so for me essentially the way that
i think about dai is that dai is gonna
be to us as like uh the industrial
revolution was like uh the farmers that
uh that perceived us like it's gonna be
something that's gonna be that's gonna
enable like a new feedback uh a new
feedback loop in our culture which is
drastically going to speed up the
economy and like i'm leaving this kind
of vague because it's kind of like okay
we know that this is going to be a big
deal like we know that like being able
to automate like a large part of our of
of our working society like uh it it
really like it really changes things it
allows you to it allows
tasks to be done on like an
unprecedented speed if you don't need
like a human in the loop to uh that is
like kind of like bottleneck bottleneck
in the whole process
and more than that is like okay well you
know it doesn't make sense like i don't
spend too much time thinking about
exactly what it means
one reason i've heard for not not
worrying too much about that and i'd be
curious to get your take on this but
essentially once you get to the point
where we have genuinely transformative
ai where we have something like an
industrial revolution but powered by ai
progress is going to be happening so
quickly that whatever set of criteria
you use to define this new industrial
revolution if you slightly disagree with
someone about those criteria you're
going to be off by like a week or you're
going to be off by like a year let's say
and then the next level of the criterion
is going to be reached and so there
isn't really that much fuzziness at the
margins because change is going to
happen so quickly anyway is that a fair
characterization
i think that is a fair characterization
i think that uh that for most possible
definitions that make sense of like how
society might be transformed like we are
gonna find them to be like extremely
correlated they're gonna happen at the
same time as they say
okay no that that's that's helpful and i
think for people who are less familiar
with like forecasting especially
forecasting ai stuff hopefully that adds
some context um great so so essentially
we have this idea of massively
transformative ai ai that transforms the
economy and and as you said i mean going
beyond that is really hard also because
i wouldn't expect like a
an agrarian society a member of an
agrarian society to be able to predict
what ipads would do what zoom would do
and so on so we're going to be in the
same position relative to our future
selves in that respect at least um i am
curious so like
what are what are the um the stages that
have led us to where we are today than
in the story of the evolution of ai that
may eventually lead to transformative ai
so you like you've studied that story
and you're trying to use it to predict
when this moment of tai or this phase of
ta i will happen
what are those phases and like can you
give us a bit of a taxonomy of like the
history of ai up to up to today so uh in
terms of like the what we have seen so
far what i've been doing with my
colleagues is like collecting
information about like historically
important machine learning systems since
the 50s up until today right and then
like when you look at the when you look
at the inputs that have gone into these
systems like uh so far we have looked
into the amount of parameters that like
each of these systems had i like the
amount of compute that was used today in
the systems like the amount how long
they were trained for in a sense
and uh what we have found is that uh in
terms of parameters it was like actually
like uh fairly less clear the the whole
story there's definitely there's
definitely like these uh aboard
upward trend which is undeniable
uh but uh it seemed to be like fairly
uniform up until maybe like uh 2000 2016
2017 2018 like somewhere around that era
like something happened and that some
and there's something that happened is
that language models started scaling
like way faster in terms of parameters
that anything else that was going on at
the time
uh this was like the it was like our
first clue of like okay there might be
like some things there might be like uh
some
changes that are happening in like the
history of machine learning that is
changing the reduced scales of the
systems like it's making it so that
scaling up those systems faster than uh
there is an incentive to scale the
system faster because we are getting
like better and better better
performance continue
actually
uh we looked at at like trends in like
compute like for how long the system was
trained for like how many operations in
the computer it took uh to train the
system and then like we see like a much
neater picture
uh we already knew from some previous
work that there had been like this phase
transition
somewhere around 2010 so essentially
before that uh the trend in like the
trending like the compute used to trade
machine learning systems essentially it
was doubling every 20 months uh if if
you're familiar with like moore's laws
like moore's law is like this empirical
description of like how the the
computing power of like uh of like uh
of like our laptops or computing devices
uh has uh scaled over time
which also follows like a similar
pattern like every two every 20 months
uh roughly you can say that uh things
got uh computation got like twice as
cheap
right
so what we were seeing before 2010 is
that you know like it's not that people
were invested more into machine learning
it's more like our laptops were getting
better and better uh like researchers
were using like a state-of-the-art
laptop for all of their things so uh
they were naturally uh coming to use
like more and more resources of
computation but something changed around
2010 which is that suddenly like this
trend like a speeds up before it was
doubling every 20 months and then like
afterwards uh we argued that it started
uh it started speeding up to doubling
every six months or so and uh this is
this is like really fast this is like
two doublings per uh two thousand per
year which is like fairly crazy uh if
you stopped uh to think about it
and uh there might have been like uh
they might have like a few reasons uh
for this uh the most selling the story
is that around 2010 was like the time
where like we realized the potential for
like uh deep learning systems to like
perform really well in tasks uh
primarily in computer vision
yeah so
i'm sorry go ahead
yes so it served kind of like as a
wake-up call for uh
for like researchers all over the world
be like okay there was like this
paradigm that really was developed uh in
the past century about like
creating like this uh neural
architectures that were just shown like
a bunch of examples and through back
propagation they were like adjusted in
order to get good performance and that
really hadn't gotten anywhere like we
had like a lot of like
bespoke systems that were being
developed uh these days like shift which
like where what you used if you wanted
to have good performance in computer
vision and now suddenly you have like
this very general approach which is
gonna which works on like a a wider
array of tasks uh particularly at the
time in computer vision which like uh
now we suddenly really got news because
our computers have gotten like good
enough uh to uh to do it
and uh people started investing like way
more into like these systems which meant
that uh these trend of like okay before
uh the the computing power that was put
into machine learning systems was just
the computer power when you haven't had
now people actually had a budget to like
uh train their machine learning systems
and that budget got this scaled up and
up as time went on which meant uh this
rapid increase in like the amount of
compute that was we import into like
this state-of-the-art uh machine
learning systems
okay so two so if i understand correctly
at this point we have two distinct eras
that you've flagged or actually
maybe three um so we have first this
phase where we have a steady a flat
academic budget and thanks to moore's
law compute's getting yeah as you say
twice as cheap every two years or so and
so you see you know twice as much
compute being used every two years but
it's still only academically interesting
during this period there's no real
industry application to the tech or at
least the industry applications aren't
enough to get people to throw tons of
money at it and then around 2010 we have
this moment where there's an inflection
point in compute budgets and and would
you so would you say that that was due
to alexnet specifically or were there
other things because i remember when we
talk about the history of deep learning
alex net is often cited as this like big
aha moment everybody goes oh you know
the deep learning revolution it started
with alex net um is it is that genuinely
true or are there models that came
before that that sort of hinted at oh
you know what we should we should scale
more uh with the computer like what's
your sense of that of alex net's role in
that story
right so uh
in my in my book what characterizes the
era of like deep learning is like three
basic factors
one of them is uh about the model size
and depth like sadly we started getting
like a systems that had like multiple
multiple hidden layers in like their
architecture and like had like way more
parameters than what we had seen before
the second one is uh the use of gpus
so uh people started experimenting with
like gpu platforms in order to
parallelize the training with like
drastically uh increase the amount of
like computing resources that they had
access to
and the last one is like performance
that was the point where like yes
suddenly deep learning systems started
starting to top the charts of uh in
benchmarks uh like uh c4
uh like uh like the image recognition uh
benchmark c4 or and
other uh another tasks
now was alex was alex nest uh the the
first thing that uh did uh all of those
things and the answer is like definitely
not uh like uh we have seen like very
large uh uh very large systems like as
large as like alexnet uh since the early
2000s like there is a paper by biola and
jones in 2001 or like they train a
system which is essentially as big as
alex not
uh it's gpu based training they think
that a distinguished alex net and it's
like well also not because uh we have
been seeing like uh the use of like
this insight of like using gpus to train
machine learning models uh had been
around for like at least seven years by
then like uh
for example like uh for example in 2005
there's this paper by steinkraus and
other people were like days that they uh
show some machine learning systems that
were trained or like gpus and in fact i
can't remember right now exactly when
like the
the cuda gpu framework was released but
that was definitely like a watershed
moment in which like suddenly like gpu
computing became like very ob could use
and like really easy to program also for
machine learning applications
all right so it's not uh so it's uh
alexnet is not the special in terms of
model size it's not a special in terms
of like gpu based training is it the
special in terms of performance and it's
like yes it's significantly outperformed
like pro techniques uh in imagenet now
it was not uh the first benchmark that
had been broken by um that had been
broken by a deep learning system
like uh the other example that uh that
uh that uh predates it is uh there is
this paper by sir san and others in 2010
which uh makes substantial improvements
over the previous state on the earth uh
on mnist
and there is like also this paper by
mikolov where like they also
break like an important uh nlp
nlp test the wall street journal task
so
alexnet uh it was as big as things that
came before it had all
it had been trained on gpus but also
things that came before i like sure it
broke like a really important benchmark
but there were other important match
marks that were broken that were broken
like like two years before around 2000
and then
all of these factors combined uh kind of
like made me think that okay this is not
uh alexnet was not like a watershed
moment out of itself it was part of like
a larger trend but what it's undeniable
is that alexnet gathered like a lot of
like academic recognition and also
outside of academy
and uh i think it's quite plausible that
that it acted as like this uh wake up
call where like people were like oh
this works okay interesting so so that
was what i was gonna ask next was like
you know given that that seems to be the
case with alexnet why why is alexnet
held up as this ultimate exemplar of
like this great watershed moment so your
assessment would be something like it's
it has to do with people's reaction to
it perhaps some in some sense the
marketing around it was just really good
do you think that's too reductive or is
that like roughly accurate
i think uh i think that's uh basically
correct uh like and it's undeniable that
imagenet was like a a very important
benchmark and it was more important that
the benchmarks that were broken before
but uh like in a sense this gives me
hopes right because uh i kind of have to
like
you could have predicted alex net this
is what i'm getting at like if you were
just in 2010 and you were squinting your
eyes really hard you will have seen like
all these neural neural network papers
that were breaking like this wall street
journal task this uh mnist task and you
will be like hey
something is happening here
interesting okay so that's actually it's
especially interesting given the next
phase in the evolution of machine
learning because i would make a similar
mistake so in 2020 openai came out with
gpd3 um i can describe it to you the way
i would naively have described it before
this conversation as the first uh the
first
pseudo general purpose ai the first ai
that was trained for one narrow task
like autocomplete and turns out to be
capable of a very wide range i mean
obviously we see transfer learning in
other contexts and images for example
but this is really where we see zero
shot learning in all its glory for the
first time translation coding even essay
writing basic web design all those
things that this one system can do
despite being trained to do something
really qualitatively different um so
my guess is you're going to come back to
me with the same similar stories alex
that hey you could have seen gpg3 coming
i'd love to explore that
first off if you agree with that maybe
you won't but like could you have seen
gvd-3 coming and if so like what were
the what were the warning shots what
were the things that should have had us
going like oh okay you know scaling does
make sense
right so uh here there is like a um
there's like a this very interesting uh
story right like we have seen this
transition between like pre-deep
learning era to like the deep learning
era now what i want to talk about is uh
what the next transition that uh we are
that we argue exists uh in our paper
which is the translation between like
the deep learning era to like the
largest scale era
to the point where like industries
started investing like millions of
dollars into like training these are
very large machine learning systems in
the hopes of like getting like uh
increased performance
and uh when we were looking at the
parameters like uh it became obvious to
us that okay there's like uh
around 2017 2016 some something around
there
like
a parameter a language model started
getting like much much much bigger
and immediately the thing that came to
us is like okay this is transformers
right like transformers came out in like
2017 uh they quickly proved to be like a
different regime of a scaling and it
became like much more advantageous to
like scale them up as fast as possible
and that's what happened
um
now i'm not so i am not so sure about
that because when we look at compute uh
sure we see that uh language models like
have stolen the thunder these last years
but really like the first system we'd
see that i came close to that regime of
scaling is uh some reinforcement
learning systems that were spread headed
by deepmind so things like alpha alphago
by demand and google generally like
alphago is like one example uh the
google neural translation machine is
another example like for me this is the
point where like uh
companies realized that they could like
scale things up like 100 times bigger
than they had been done before and like
actually get good results out of that so
they just went and did that
uh could we have uh go ahead from these
uh predicted uh gpd3
uh i was just still like
i was personally really really shocked i
was like the whole dpd3 thing i like the
whole rise of like language models i was
like quite bullish on thinking that
language modeling was like what i would
call like an ai complete problem it was
like a problem so hard that like we were
gonna get like general intelligence
before we actually got like good
language modeling and like well reality
has proven me like very very wrong
yeah that in itself is is an interesting
uh an interesting aspect of all this the
link between
scale the link between um architecture
and then the actual capabilities that
are achieved by these models what's easy
what's hard
it seems like i mean this is to me one
of the the interesting aspects of your
work it allows us to start to notice
when our intuitions
were just completely wrong
um now one thing i do want to touch on
before we go more into that direction
because i think there's a lot to talk
about when it comes to kind of
capabilities and linking those to scale
and other things
it seems like you mentioned a couple
times you know
your assessment of
doubling times for compute power for
example and your thinking and your your
analysis and um and you're hinting in
that that there might be disagreement
there might be other perspectives too
which i guess to me i always found
interesting because i would have
expected this to be a very
naively like straightforward thing to
calculate you know we have a bunch of
leading models and they have a certain
amount of compute power consumption and
then we start to draw straight lines on
log plots and and there's our doubling
time um so can you explain like i'm sure
i'm wrong by the way but i'd love to
understand how i'm wrong
absolutely so uh uh let me talk about uh
the work that came before us in terms of
fly computers uh the print the main
piece is this article by open ai around
2018 where like they did they went
through the same processes as us on like
a smaller scale they like uh gathered
the amount of computers used to train
like uh around like 12 to 18 uh
state-of-the-art machine learning
systems throughout the years and they
just they plotted it and they were like
okay this is the this is the line that's
just runner regression and they got like
they got like a doubling time that was
like way faster than us just like uh our
time uh our doubling time was also
really fast like uh six months they're
uh they're doubling the only time that
they found is like half of that it's
like every three months things were like
doubling right
and then what happened
well two years went by and like their
prediction was well that their implied
prediction right like this implied trend
like a stops flood a stop flat like uh
it just doesn't go on it's like we had
sadly we hadn't seen like a doubling uh
since 2018 by 2020.
and there is in fact like uh this blog
post by alex lysol where like uh he uh
he expands on like the work of like uh
and compute adding like a few points and
being like
uh the trend stopped right when you grow
the blog post
right now now this is uh are you
referring here to the scaling laws for
neural language models paper i think
that was 2019 though wasn't it
no no no uh this is like uh this is a uh
standard growth sorry and compute that's
right that's right i then want to bring
in that that scaling laws for mural
language models paper which came out in
yeah in 2019 i think by the time they
wrote it gpd3 pretty much would have
been
built internally in retrospect because i
think they they released the paper for
gpd3 sometime in january and i think the
scaling laws paper might have come out
in
like very late 2019 um so so maybe they
had some some new insights based on gpd3
but do you have a view on like on that
paper did it cause you to change your
perspective is it consistent with your
analysis or
uh absolutely so uh the paper on like a
scaling laws is essentially like the
whole motivation for like this whole
project it's kind of like the proof of
concept that like a scaling matters and
it's like really important
so uh kind of like what what we see is
like our work is like complementary to
what's happening in like uh
like uh the scanning laws paper in there
like they were running like a series of
experiments with like uh some with like
some systems what we're doing is like
we're looking at what has happened
historically i like see like given the
insights that we found in like that
paper unlike some other papers that
we're studying uh returns to scaling
whether we can explain how much of the
progress we have seen in the last two
decades is based on uh just this is
scaling things are getting bigger and
faster versus uh us having like better
architectures the trend that uh open air
had found when like they look and they
plotted like this line of like uh
up to uh it was not only like every
three months and then like alexa like
continues that and then it just doesn't
grow i like you know like between the
biggest system in like 2018 which is
alpha uh alphago uh
one of the versions of alphago also goes
zero and uh the biggest the the biggest
uh system in like by two 2020 which is
like uh dpt3 is like actually gpd3 is
smaller than alpha goes zero so it
seemed like oh the trend has that don't
have the scope but like uh really i
think that uh this is like an illusion
that is being caused by like this uh
this discontinuity that we had in 2016
like this point where like suddenly
companies started uh uh investing like a
hundred times more so then what's going
on is that uh if you just look at like
the uh like the biggest systems overall
you're gonna catch like a lot of noise
and that's gonna make it so that uh so
that uh the trends that are apparently
there like really are not there because
they are just uh they just consist of
like these field pliers and include like
this largest is going to do with this so
you need to like take up a bigger look
uh to look for like uh the trends that
are there and like even having like that
bigger look you're still gonna have to
have uh points like in 2016 where like
suddenly things skyrocket escape rocket
up by like two orders of magnitude and
uh that's something that happens this
continues to happen what this makes me
wonder is is where these trends would
start to break down or like what might
cause these trends to break down in the
future
so uh these trends uh the distance right
now are the combination of like two
factors one is uh compute getting
cheaper as building like better uh
infrastructure for computing
the second is uh investing going up
right investment going up like
industries uh primarily industry at this
point uh is uh more and more interested
into having uh putting like millions and
millions of dollars into training these
systems
uh these two these two trends like
follow different mechanics and may break
down uh because of different reasons uh
for like the for like the computer and
uh this is like moore's law people have
been claiming that moore's law is gonna
die like very soon eventually
it has to die like it cannot go on
forever but it's like really hard to
find out what is the point at which like
uh actually uh the things break down uh
like we start to be we stop being able
to like
scale up our systems like uh there might
come like a new new ways of
conceptualizing like computation that
might allow us to like uh keep squeezing
the uh keep squeezing like uh
this trend i like keep our uh it's kind
of like a sort of like self-fulfilling
prophecy where like people kind of have
like this uh like my my impression is
that internally like harvard companies
have like this impression of like this
is the goal to meet and they put like a
lot of effort into like making it at
some point like you know physics says
stop you cannot go on but for the time
being it hasn't seemed uh it hasn't
sound like uh has slowed down a bit
but it's still is still happening it's
still a decreasing exponentially the
price of compute and i was expected to
responsibly at least at the very least
for the last 10 years and like possibly
like quite possibly for longer even for
how long it has held up so far
now the second thread is about
investment i like investment is more
complicated because uh well industries
industries have a budget and they can
siphon like part of their r d budget
towards a ai
like that's essentially what has been
happening they also have like target
revenues that they can put into this but
there comes a point where like you know
when uh ai is like 90 percent of the
earth research budget you just cannot go
on without like a state support or like
uh or like something else
uh when is when are we going to reach
that point so my colleague tamaiba
zeroglu actually uh
repeat uh performed an analysis based on
a blog post by ryan carey from a couple
of years ago
and uh he essentially tried to compute
try to put like an estimate on like what
is going to be the reversing point or
it's going to be the point where like
this trend of like uh increased
investment is going to stop uh like the
driving force between between uh
progress from now on it's just going to
be like moore's law
and essentially like
he was estimating like you know under
some reasonable assumptions about like
how much money could uh
companies like possibly spend on ai r d
like maybe in like it's it seems
definitely possible that it's gonna held
up for like 10 years uh the current
trend and then like afterwards it's like
extremely uncertain like what's going to
happen if they're gonna like hit the
ceiling and just stop increasing their
budgets if like estate actors are gonna
like come in and like uh keep uh
investment up
uh we know one interesting factor in
that analysis too is like if companies
start to scale up their uh their compute
budgets in that way
eventually you do get systems like gpd3
that can create so much value themselves
uh that it pays for that compute and so
you have this positive feedback loop
that has no termination point or at
least no no clear termination point you
know arguably we're already seeing that
with gpd3 there are companies that have
raised tens of millions of dollars that
are really just like a fancy wrapper on
top of gpd3 or maybe you know ai 21 labs
products or stuff like that and so you
know it's like 10 million dollars well
that's already the cost of making gbd
three so so we're in a way we're already
in that regime um or maybe we're not i
mean do you think that we may be already
at this point where we're kind of
closing that loop
i think we are like one of my leading
hypotheses is to explain like this
discontinuity that we saw around 2016 is
uh essentially yeah a point where like
industries did the calculation be like
okay we can afford to put the money this
morning in because this is gonna
generate like so much revenue
and uh for me like the leading example
of this it's not gbt3 it's like a dnmt
like a dnmt the google the google neural
machine translation system was like one
of the early examples of like a really
really large scale machine learning
system that broke with the previous
trend and that had like a huge uh
economic implications
okay yeah no that makes perfect sense um
and actually okay so now we've talked
about this idea of trends in compute one
of the things we haven't talked about is
how we tell
when a particular level of compute leads
to a particular capability or a
particular situation in society this tai
threshold transformative ai threshold
that you've been trying to kind of
project and predict
and one of the techniques that you've
used to actually
land that plane and figure out okay you
know
how do we how do we get capabilities
from these systems how do we predict
capabilities from scale is to lean on
this framework of biological anchors in
uh predicting transformative ai so could
you explain what biological anchors are
and how they relate to some of your work
very badly but i will try my best so
essentially within my group we haven't
mostly been focusing on like inputs but
uh some of people some people on this
area that have in high regard uh have
been figuring out like what to do with
like the estimates that we're providing
or have come up with like their own
models we're like extrapolating these
trends to try to forecast at different
levels of performance and definitely so
far like the most intricate piece of
research and the most complete piece of
research that we have seen is ayakotra's
draft report on airtime lands where like
she comes up with like some generally
useful concepts in order to try to
understand uh how much compute will be
needed uh to train uh transformative
machine learning systems
so uh that's the that's uh the this uh
by that that there's like this anchors
uh report or like c comma comes up with
like six different ways of estimating
like uh what's gonna be like uh the
amount of operations that you're gonna
need in order to train uh these
transformative uh systems i like three
of uh three of them are like essentially
like
biologically based there's like
estimations about like you know
so far the only example of like
artificial general intelligence that we
have is humans
right and like uh what uh well not that
efficient but general intelligence that
we have
and uh so it provides us kind of like
with like a a very crude estimate of
like uh well
an upper bound on like how many
operations you need in order to create
intelligence
and there is like multiple ways that uh
you can go about thinking like how many
operations did it take to uh make a
human
like uh one one thing that you can do is
just go like okay like a human like is
born and then like it takes like some
time for it to like let absorb like the
culture to like learn how to speak how
to write uh how to code how to do
different things and like uh it's gonna
take them like okay roughly like 20
years to become like a functioning adult
like uh for like a baby that knows
nothing to like a general intelligence
that can perform like a wide array of
like economic tasks so you can like sort
of estimate like okay how many
operations does should take in your
brain to uh to
do all that learning and that's gonna be
like a sort of like estimate of like how
much it takes from like baby level
artificial intelligence to like human
level artificial intelligence
um
as that as this there's like a couple
other ways that you can go about it
because you can say like okay but babies
have already like a lot of like built-in
machinery and like maybe this is not uh
the best uh the best way of thinking
about it and uh maybe you can go with
like the most extreme estimate that uh c
provides is like going like okay how
much did it take how many operations did
it take to like evolve a human right how
many operations if you see the earth
that's like this giant computer that has
been running like this evolutionary
algorithm for like uh for like billions
of years like how uh how long did it
take to uh actually create a human level
intelligence from that
and she also provides uh that uh that
kind of estimate and those are the
biological anchors right uh for me
actually the ones that are more
interesting are not the biological
anchors themselves but uh the anchors
are based on like uh this concept of
like horizon length
so essentially what i uh uh proposes is
that uh systems where like the reward is
like farther away temporarily than the
action are gonna be like harder to train
uh that uh system for like the reward
and the action is like uh closely paired
together
and uh which makes a lot of sense
uh one of the hardest problems in like
um like machine learning is the
attribution problem like trying to
attribute like okay this reward i'm
getting like to which action is it to
you
and and i guess this also
i mean it aligns to some degree with the
evolutionary anchors perspective just in
the intuitive sense that when you look
at like
animals we tend to think of as stupid
they tend to be they tend to act on
instinct in other words they tend to
respond to immediate stimuli and
respond to it in an immediate way
there's very little plotting and
scheming going on in the brain of an ant
for example whereas when you look at you
know dogs maybe they can learn to train
their owners in certain ways or you know
trick them into doing certain things so
there's a little bit of foresight and
planning monkeys more so and maybe
humans even more so um so i guess
there's a sense in which they are sort
of aligned even if they they take
different
different directions
exactly that's it so essentially what uh
what i did i was coming up with like an
estimate of like okay so far we have
trained like some reinforcement learning
systems that uh are able to act in like
these time horizons are able to act like
so many steps into the future and then
like c-com something's like an estimate
of like okay how many steps
will it take to uh uh could it take to
like uh create a company for example
like this is an example of a task right
and then she goes like well uh this is
like a time horizon flag a year the
amount of the steps uh that's that uh
this is gonna involve is like such and
such and then like uh given that i see
like from the previous data or like how
much how much compute did it take to
train like these previous machine
learning systems we kind of have we kind
of have like a rough estimate of like
what how many operations does it take to
train a system that has like a certain
horizon length so now with like these
new horizon lines for this task that we
haven't automated yet this could be like
a a possible estimate of like how many
operations will it take to automate
those as well
so through that lens i guess this makes
me think of the gpd3 context window or
the context window of language models as
being this very important number the
sort of amount of text it can keep in
mind at the same time as it predicts the
next thing
is that like do you think that's a
correct way to think about it like the
context window might have a lot to do
with this idea of planning ahead and
time horizons
absolutely i think this is one of the
main reasons why right now like i don't
see gpd3 as like uh something that can
scale up to general intelligence
because you do really need to be able to
like create loops in your intelligence
to be able to like pay attention to
things that happen like a very long time
ago in order to like create that now uh
this is not to say that gps iii cannot
be like a critical component of like
artificial data intelligence like i have
been actually like very shocked by like
some kind of like by hybrid approaches
in which like dptv has kind of been put
into like a loop in which like it is
able to produce things like mathematical
proofs like code and like that code is
executed and we could see maybe like a
uh like this uh as the beginning of like
maybe some sort of like uh
loop system in which like you put you
asked jpd3 to provide like a piece of
code the code is executed that provides
like uh that provides context for like
the next call to like uh gpp3 dptx like
interface
interesting so
and that actually vibes i guess with the
idea that
we have gbd3 taking up and historically
ai has done this too even going back to
like the mid 2000s
it almost starts with like taking away
some of the most menial tasks so like
excel spreadsheets remove things that
require like thinking on the order of a
couple seconds i don't want to have to
multiply all these cells together excel
will do it for me then the next step is
like i want this to do you know
calculate my p values to do even more
sophisticated operations that save me
more time and gradually the human gets
to zoom out more and more think more and
more well we say often think more
strategically but really think over
longer time horizons as the the goal or
the responsibilities of this ai start to
expand um and so when we look at some of
those loops i guess you probably have
when we talk about theorem proving with
something like gbd3 or gopher or things
like that
i guess you probably have the human
doing almost strictly long-term thinking
um and then the ai picking up the slack
is is that how you're seeing those loops
maybe i definitely think like uh in the
next in the next five years uh what
we're gonna see is
dptx systems like language models and
like different machine learning systems
as kind of like an augmenter of like
human italians where like there's gonna
be a human in control which is gonna be
prompting the machine learning system in
order to like uh produce uh
to produce a text to produce code that
then like the system is the human is
going to bet and decide what to pick i
actually wrote the abstract uh for my
paper using gpd3 part of it
and i expect this is gonna become like a
way more common occurrence uh in the
future i see this as different from like
what i was saying before we're like well
i was saying before like don't take it
with a grain of salt because this is
like me uh who uh who uh
who's like a relative outsider to like
the nitty gritty of like actually
training the systems like trying to
think about like how you could scale
gptx systems to like general
intelligence and be like not very
convinced about that well so this is
actually interesting because um there's
a lot of debate in the forecasting
community as i'm sure you know between
people who are like who adopt this
inside view perspective who say look the
best way to predict trends in uh in ai
capabilities is to talk to people who
are actually building these systems as
you say doing the nitty gritty work and
then people say no actually the outside
view is usually better because when
you're on the inside you kind of can't
see the forest for the trees and you i
mean i've seen this in in startups right
where like it's it's a classic thing in
silicon valley you have investors who
have like built let's say an edtech
startup
and
because they've built an edtech startup
they know all the ways that edtech
startups can fail they know all of the
the horrible list of things that need to
go perfectly right in order to make this
work and so they'll never invest in an
edtech product because they just see all
the reasons it can't work but then they
see a product in a completely other
domain that they know almost nothing
about and they get really excited about
it and often they make really good bets
as a result so paradoxically your
experience can actually detract from
your ability to make good predictions
i wonder how you see that interaction
acknowledging obviously we all have our
biases we all come from one perspective
or another as you mentioned if you're
more on the outside doing less building
maybe that'll default you to that side
but how do you think about that that
trade-off in the context of your work
absolutely so when i think about uh
expanding uh predicting trends in well
predicting like what's going to happen
with artificial intelligence with new
technologies in general like i see it as
like there's two things here there's
like trends and there are
discontinuities and like uh trends are
often are often like surprisingly robust
like uh there's this work by impacts
where like they
try to set up uh to try to look for
discontinuities so they could understand
better in which conditions do these
technological discontinuities happen and
like uh among like all the examples that
they looked at of technological trends
like actually they didn't find that they
found like
okay i gotta say they find like a lot of
discontinuities like you know around
like 30 percent of like the trends that
they looked at they found like a
discontinuity but they were explicitly
looking for for discontinuities so this
kind of like uh implies that like these
continuities are like somewhat rarer
that like one might naively think and
you can actually get like a pretty far
ahead which is like a lot in a lot in
like a straight line on like what you
have seen uh so far
now uh if you want to go the next level
then you need to account for the
possibility that uh there's gonna be
discontinuities like the one that we saw
in like 2016 with like a computer and
capabilities of machine learning systems
and for those like uh the best you can
do is like having like a having like an
inside view system like trying to
understand what are like the driving
forces behind like the different trends
that you see i like trying to understand
like how incentives might change in such
a way that this conditions happen or
like how the uh how some certain like
critical points might be uh might be
reached where like uh suddenly things go
faster one interesting insight is that
it's quite uh like we should naively
expect like these continuities to uh to
surprise us on the positive side
because if a discontinuity happens on
the negative side like kind of like
that's gonna be that's gonna be rolled
over by the trend right like the trend
is still gonna go on unless like
whatever are the forces like driving the
trends like uh subside but normally
that's gonna be uh normally that's gonna
be like uh something more gradual
unpredictable uh if you really want to
if you really want to uh so kind of like
as i see my work is providing like an
upper bound on like uh how uh how far
away intelligence can be and then like i
don't have that much to say being like
okay like i'm pretty i'm i'm somewhat
confident that by the end of the century
like we will have the resources to train
like artificial general intelligence as
an example like don't take this uh this
figure literally yeah but then i'm not
gonna have that much to say being like
okay it might come like 40 years earlier
we don't know something something
unexpected might happen and actually
that ties into
something we started with which was this
question of transformative artificial
intelligence and we tried we talked
about that definition and then you you
just raised in this context artificial
general intelligence and when we were
talking about biological anchors it
sounded like we were talking about
artificial general intelligence as well
just because we're focused on what would
it take to replicate the human brain
rather than what would it take to
transform society i'm wondering like do
you think that there's a
functional or important difference
between predicting tai versus predicting
agi predicting transformative ai versus
predicting general intelligence or are
they going to roughly come at the same
time and it basically won't matter
right uh like i think uh uh i think
there is like an important difference in
the sense of like
i think there is like some scenarios in
which like we get tai where like we
don't get like
strictly speaking agi in example like
being able to do like everything that a
human brain can do like you don't need
to be able to do everything that human
brain can do in order to like radically
transform and radically transform
society
uh
yeah i think we should focus on like
predicting like the minimal uh the the
automation of like the minimal amount of
transformative tasks with like actually
change society i like that's a smaller
subset i'd like artificial general
intelligence but i also like fully
expect that like
in most scenarios like these two things
come like
fairly attached to one another i guess
it's like for the same reason that
regardless of what your definition of
tai is you're gonna get it roughly right
because progress will be happening so
fast once you hit it that like for the
same reason tai and agi kind of become
pretty close just because progress is
happening so fast
unless there's a fundamental reason that
we can't get to agi as you say like
unless our algorithms just can't get
there for some reason we have yet to
discover it definitely seems like a like
a plausible um like a plausible scenario
um awesome hamiah thanks so much this is
just like absolutely fascinating great
overview of all your work is there
anywhere you'd recommend people go if
they want to follow this kind of ai
tracking work that you're doing
uh absolutely so i think that uh the
best the best way to find our work is in
the alignment forum there is a sequence
called trends in machine learning or
like there is an overview of summaries
of all our work which is a great entry
point to everything that we are doing
and that we're planning to it in the
future
fantastic okay i will uh i'll link to
that in the blog post that'll come with
podcasts i'll also link to your uh your
twitter account because i know you do a
lot of you know some interesting like
tweeting on this on this general topic
and in topics around this topic so um
yeah everybody definitely uh check that
out and uh honey thanks so much for for
joining me for this is a ton of fun
thank you for having me jeremy have a
great day |
20f049bf-7eb1-4e32-b5a6-0d52b12a76b6 | trentmkelly/LessWrong-43k | LessWrong | Cryonics donation fund for Kim Suozzi established by Society for Venturism
Following the news that Kim Suozzi has terminal brain cancer and wants to be cryopreserved, many of us have donated to help her out, while others, including me, planned to donate when CI set up a fund to receive donations on her behalf. Now the Society for Venturism has set up a fund, and it is time for us to follow through on those plans. (Unless you are really insisting that the fund be managed by CI specifically.)
(ETA: Kim has posted on this herself.) |
5b70091a-e3f5-48e6-911b-4143f6632662 | trentmkelly/LessWrong-43k | LessWrong | To Learn Critical Thinking, Study Critical Thinking
> Critical thinking courses may increase students’ rationality, especially if they do argument mapping.
The following excerpts are from “Does philosophy improve critical thinking skills?”, Ortiz 2007.
1 Excerpts
> This thesis makes a first attempt to subject the assumption that studying [Anglo-American analytic] philosophy improves critical thinking skills to rigorous investigation.
>
> …Thus the second task, in Chapter 3, is to articulate and critically examine the standard arguments that are raised in support of the assumption (or rather, would be raised if philosophers were in the habit of providing support for the assumption). These arguments are found to be too weak to establish the truth of the assumption. The failure of the standard arguments leaves open the question of whether the assumption is in fact true. The thesis argues at this point that, since the assumption is making an empirical assertion, it should be investigated using standard empirical techniques as developed in the social sciences. In Chapter 4, I conduct an informal review of the empirical literature. The review finds that evidence from the existing empirical literature is inconclusive. Chapter 5 presents the empirical core of the thesis. I use the technique of meta-analysis to integrate data from a large number of empirical studies. This meta-analysis gives us the best yet fix on the extent to which critical thinking skills improve over a semester of studying philosophy, general university study, and studying critical thinking. The meta-analysis results indicate that students do improve while studying philosophy, and apparently more so than general university students, though we cannot be very confident that this difference is not just the result of random variation. More importantly, studying philosophy is less effective than studying critical thinking, regardless of whether one is being taught in a philosophy department or in some other department. Finally, studying philosophy is much |
b648dac3-eeb8-4c9f-8769-c58f09c013de | trentmkelly/LessWrong-43k | LessWrong | Learning strategies and the Pokemon league parable
I have recently noted a shift in my learning strategy, which I reflectively approve of. On hindsight, it feels obvious.
However, I can vividly recall many people I respect and admire recommending me to try a very similar thing in the past, and myself scoffing at them and trusting my gut over their advice.
Take this as a word of caution if you feel like the advice I am giving is obviously wrong: I have fallen in this pit, and it took me a while to climb out.
----------------------------------------
I claim there are two broad learning strategies one can follow.
The default strategy is what in the software engineering lingo is called a waterfall strategy. People attend college and different courses and read books, and gain knowledge on a broad collection of subjects. Afterwards, they move to a second phase where they try to apply what they have learned. If they cannot reach their goals with their current strategy, they back down to the first phase and start again.
I claim that this strategy has some glaring flaws, which I plan to expose via my experience in an area of great importance: Pokemon videogames.
----------------------------------------
When I got my first videoconsole, the first videogame I ever owned was Pokemon Gold. I absolutely loved that game, and I spent many hours absorbed trying to complete it.
For the most part, the level of challenge was adequate for an 8 year old, but there comes a point where the game suddenly spikes up in difficulty: the Pokemon league. In the league, you have to defeat five trainers with full teams of high level Pokemon in a row.
When I first confronted the Pokemon league, I was quite under leveled and I was utterly crushed.
My response to the problem was to back down and go to easier areas, where I could train with easier challenges.
After around 10h of training, I came back to the League and defeated the five trainers with relatively ease, and I won the title of Pokemon master, officially achieving my most ambitio |
53ed62c6-97bd-46ba-93e6-0771914281c5 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Preface to the sequence on iterated amplification
This sequence describes iterated amplification, a possible strategy for building an AI that is actually trying to do what we want out of ML systems trained by gradient descent.
Iterated amplification is not intended to be a silver bullet that resolves all of the possible problems with AI; it’s an approach to the particular alignment problem posed by scaled-up versions of modern ML systems.
Iterated amplification is based on a few key hopes
* If you have an overseer who is smarter than the agent you are trying to train, you can safely use that overseer’s judgment as an objective.
* We can train an RL system using very sparse feedback, so it’s OK if that overseer is very computationally expensive.
* A team of aligned agents may be smarter than any individual agent, while remaining aligned.
If all of these hopes panned out, then at every point in training “a team of the smartest agents we’ve been able to train so far” would be a suitable overseer for training a slightly smarter aligned successor. This could let us train very intelligent agents while preserving alignment (starting the induction from an aligned human).
Iterated amplification is still in an preliminary state and is best understood as a research program rather than a worked out solution. Nevertheless, I think it is the most concrete existing framework for aligning powerful ML with human interests.
### Purpose and audience
The purpose of this sequence is to communicate the basic intuitions motivating iterated amplification, to define iterated amplification, and to present some of the important open questions.
I expect this sequence to be most useful for readers who would like to have a somewhat detailed understanding of iterated amplification, and are looking for something more structured than [ai-alignment.com](https://ai-alignment.com/) to help orient themselves.
The sequence is intended to provide enough background to follow most public discussion about iterated amplification, and to be useful for building intuition and informing research about AI alignment even if you never think about amplification again.
The sequence will be easier to understand if you have a working understanding of ML, statistics, and online learning, and if you are familiar with other work on AI alignment. But it would be reasonable to just dive in and just skip over any detailed discussion that seems to depend on missing prerequisites.
### Outline and reading recommendations
* The first part of this sequence clarifies the problem that iterated amplification is trying to solve, which is both narrower and broader than you might expect.
* The second part of the sequence outlines the basic intuitions that motivate iterated amplification. I think that these intuitions may be more important than the scheme itself, but they are considerably more informal.
* The core of the sequence is the third section. [Benign model-free RL](https://ai-alignment.com/benign-model-free-rl-4aae8c97e385) describes iterated amplification, as a general framework into which we can substitute arbitrary algorithms for reward learning, amplification, and robustness. The first four posts all describe variants of this idea from different perspectives, and if you find that one of those descriptions is clearest for you then I recommend focusing on that one and skimming the others.
* The fourth part of the sequence describes some of the black boxes in iterated amplification and discusses what we would need to do to fill in those boxes. I think these are some of the most important open questions in AI alignment.
* The fifth section of the sequence breaks down some of these problems further and describes some possible approaches.
* The final section is an FAQ by Alex Zhu, included as appendix.
The sequence is not intended to be building towards a big reveal---after the first section, each post should stand on its own as addressing a basic question raised by the preceding posts. If the first section seems uninteresting you may want to skip it; if future sections seem uninteresting then it’s probably not going to get any better.
Some readers might prefer starting with the third section, while being prepared to jump back if it’s not clear what’s going on or why. (It would still make sense to return to the first two sections after reading the third.)
If you already understand iterated amplification you might be interested in jumping around the fourth and fifth sections to look at details you haven’t considered before.
The posts in this sequence link liberally to each other (not always in order) and to outside posts. The sequence is designed to make sense when read in order without reading other posts, following links only if you are interested in more details.
---
*Tomorrow's AI Alignment Forum sequences post will be 'Future directions for ambitious value learning' by Rohin Shah, in the sequence '[Value Learning](https://www.alignmentforum.org/s/4dHMdK5TLN6xcqtyc)'.*
*The next post in this sequence will come out on Tuesday 13th November, and will be 'The Steering Problem' by Paul Christiano.* |
33231c6c-1f54-45d6-be74-5747ea3fb996 | trentmkelly/LessWrong-43k | LessWrong | Thoughts on Max Tegmark's AI verification
In Max Tegmark - How to keep AI under control - TED Talk Max presents a method to make AI systems safe (all images from that talk's slides):
The basic idea is:
1. Write a specification of what you want.
2. Give the specification to your smart AI.
3. The AI will build a tool that will conform to the specification.
To refine this setup, we split the AI into two systems. One system, the AI learner, is an algorithm that can train an AI. Another system, the AI neuroscientist, extracts out the algorithm that the AI learner has learned into a format that is more human-comprehensible:
Specifically in the talk, he gives the example of turning a neural network into a Python program. I.e. we want to automate some version of mechanistic interpretability.
We don't only return the program though. We also return a formally checkable proof that the program conforms to our specifications.
Problems
How do you reference the real world in your spec?
The first problem is how are we gonna write the spec correctly? I can easily write a spec for a program that is very limited in scope. A program that only needs to act in some logical abstract world such as a theorem searcher, a chess player, or really a program playing any computer game. There the environment can be logically specified.
But as soon as we need to talk about the real world, this is extremely difficult. How do you formally write down that humans should not be killed? How do you even reference at all a human in your spec? You need to solve the Pointers Problem first.
How fast is Mechanistic interpretability
It seems that what Max describes is possible in principle if you additionally solve some problems that he doesn't mention. But mechanistic interpretability lags far behind what models can do right now. AFAIK the algorithms that people can extract right now are very simple. And progress is slow.
Another thing to consider is that if you could put the algorithms of advanced AI systems into a much more human-com |
82f53a09-2e98-41bc-92db-71c97f255280 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Cambridge Massachusetts meetup
Discussion article for the meetup : Cambridge Massachusetts meetup
WHEN: 26 June 2011 02:00:00PM (-0400)
WHERE: Stata Center, Cambridge, MA, USA, room 261
The entrance is shown in Google Maps at http://goo.gl/maps/l0Cq. Note that the doors may be locked; we will post a pair of people to let people in for the first 30 minutes, but if you're later than that, you're at a different entrance, you're lost, or there's no one there, then call my cell phone at 607-339-5552.
I will be giving a short presentation on value of information - that is, order-of-magnitude calculations of how much research and thinking is worth, and spotting cases where it's unexpectedly high or low. This will not take up the entire time; the remainder will be spent on general-purpose socializing, calibration exercises using Wits & Wagers trivia questions, dinner (details to be decided on Saturday), and whatever good ideas you suggest.
Discussion article for the meetup : Cambridge Massachusetts meetup |
2e02ffbc-3fb6-4483-83b6-9d1fa2859bd1 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Is "red" for GPT-4 the same as "red" for you?
```
Penned by Yusuke Hayashi, an independent researcher hailing from Japan, this article bears no affiliation to the authors whose scholarly works are referenced herein. Demonstrating intellectual autonomy, the analysis presented is unequivocally distinct from the cited publications.
```
Have you ever wondered if the "red" you perceive is the same "red" someone else experiences? A recent study explores this question using a distance measure called the Gromov-Wasserstein distance (GWD).
**Is my “red” your “red”?: Unsupervised alignment of qualia structures via optimal transport**[[1]](#fnq7kp3vq034)
Paper (Preprint by Kawakita et al. published in PsyArXiv): <https://psyarxiv.com/h3pqm>
Slide (Presentation by Masafumi Oizumi at Optimal Transport Workshop 2023): <https://drive.google.com/file/d/1Z5QTxayqJkmYyOgirrbJRRflY8HOy8ox/view?usp=sharing>
In the paper "Is my "red" your "red"? : Unsupervised Alignment of Qualia Structures by Optimal Transport" (Kawakita et al., 2022) proposes a new approach to assessing the similarity of sensory experiences between individuals. This approach, based on GWD and optimal transport, allows for the alignment of qualia structures without assuming a correspondence between individual experiences. As a result, it provides a quantitative means of comparing the similarity of qualia structures between individuals.
This image is an excerpt from a slide created by Dr. Masafumi Oizumi, the author of the paper "Is my 'red' your 'red'?: Unsupervised alignment of qualia structures via optimal transport."This image is an excerpt from a slide created by Dr. Masafumi Oizumi, the author of the paper "Is my 'red' your 'red'?: Unsupervised alignment of qualia structures via optimal transport."This image is an excerpt from a slide created by Dr. Masafumi Oizumi, the author of the paper "Is my 'red' your 'red'?: Unsupervised alignment of qualia structures via optimal transport."
By applying this method, it is possible to compare the similarity of subjective experiences, or qualia structures, between individuals using GWD, a distance measure employed in machine learning research. While this study compares sensory experiences between humans, the same approach could be used to compare the sensory experiences of humans and the advanced AI model, GPT-4. We are now ready to revisit the question posed in the title:
Is "red" for GPT-4 the same as for you?
If asked whether large language models (LLMs) like GPT-4 possess consciousness, the likely answer would be "probably not." However, the method proposed here provides a way to determine whether such models share the same qualia structure as humans.
Consciousness and qualia are closely related. Consciousness is the state of being aware of, perceiving, and thinking about one's surroundings, thoughts, and feelings. Qualia, on the other hand, are subjective experiences and sensations that accompany consciousness, such as the taste of chocolate, the sound of a musical instrument, or the sensation of pain.
Qualia are an essential part of conscious experience, and their relationship to consciousness is of paramount importance. Without qualia, our experience of consciousness would lack a rich, subjective texture and would not be unique and personal. In other words, qualia give our experience of consciousness its unique character and enable us to understand the world around us.
One of the central debates in the philosophy of mind revolves around the nature of qualia and their relation to consciousness. Some philosophers argue that qualia are difficult to study and understand because they cannot be reduced to physical or functional descriptions. Others argue that advances in neuroscience and cognitive science may allow qualia to be explained within a naturalistic framework. The aforementioned studies shed new light on the long-standing controversy surrounding the nature of qualia and its relationship to consciousness.
Recently, Dr. Geoffrey Hinton, one of the pioneers of the deep learning model, resigned from Google, advocating the potential dangers of making AI smarter than humans. In LessWrong, there is a lot of discussion about the safety of AI and the integrity of AI. Throughout this article, I would like to pose the following question to the reader:
If deep learning models acquire consciousness in the near future, can we consider them as alignment targets?
1. **[^](#fnrefq7kp3vq034)**[G. Kawakita](https://osf.io/gd37k/), [A. Zeleznikow-Johnston](https://osf.io/p7mq9/), [K. Takeda](https://osf.io/qzrbv/), [N. Tsuchiya](https://osf.io/m7avu/) and [M. Oizumi](https://osf.io/ej542/), Is my "red" your "red"?: Unsupervised alignment of qualia structures via optimal transport. PsyArXiv <https://doi.org/10.31234/osf.io/h3pqm> (2023). |
acc5845e-a092-4313-96e7-255f14019fc8 | trentmkelly/LessWrong-43k | LessWrong | Discovery fiction for the Pythagorean theorem
I've been thinking recently about how to teach the Pythagorean theorem to high school students. As part of that thinking, I looked around to see how the topic was being taught in various textbooks, online videos, blog posts, etc. Typically, the discussion goes something like this:
First, the statement of the theorem is presented: For a right triangle with legs of lengths a and b and a hypotenuse of length c, we have a2+b2=c2.
Next, a picture like the following one is presented as a visual:
The student is told that the two smaller squares add up in area to the largest square.
Finally, any one of the typical proofs is presented. This could be a rearrangement proof, the "Behold!" proof, or Euclid's proof of proposition I.47.
One can improve the final step by using what is sometimes called Einstein's proof. (See also this post by Terence Tao, this video by Numberphile, this article by Alexander Givental, and in particular this comment by Tim Gowers for discussion and presentation of this proof.) This proof is an improvement over the typical presentation for a few reasons: it makes the theorem feel more intuitive, and (especially with the discussion in Gowers's comment) it gives some indication of how one might discover the proof.
It might seem like the "exposition problem" for the Pythagorean theorem is solved: we started with a bunch of proofs that made the theorem feel unintuitive that we didn't know how to discover ourselves, and now we have a good proof along with a story for how to discover it.
I claim that there is still some work left! I think the Pythagorean theorem is a case where even the theorem statement itself seems bizarre (rather than just the proofs being bizarre). Given an arbitrary right triangle, how would one guess that a2+b2=c2? And why would one even think this is a problem worth solving in the first place? I think this second question is easy to answer by pointing to the numerous applications of the theorem, so I will focus on the first q |
8eaf6c1f-38e3-4176-b2bc-e3f210630562 | StampyAI/alignment-research-dataset/arbital | Arbital | Utility indifference
# Introduction: A reflectively consistent off-switch.
Suppose there's an [advanced agent](https://arbital.com/p/2c) with a goal like, e.g., producing smiles or making [paperclips](https://arbital.com/p/10h). [By default](https://arbital.com/p/10g), if you try to switch off a sufficiently intelligent agent like this, it will resist being switched off; not because it has an independent goal of survival, but because it expects that if it's switched off it will be able to produce fewer smiles or paperclips. If the agent has policy options to diminish the probability of being *successfully* switched off, the agent will pursue those options. This is a [convergent instrumental strategy](https://arbital.com/p/2vl) if not otherwise prevented.
- Difficulty 1: By default a [consequentialist](https://arbital.com/p/9h) reasoner with sufficient real-world understanding to relate the events of its being switched off, to the later underfulfillment of its goals, will resist being switched off.
The [shutdown problem](https://arbital.com/p/2xd) is to describe an [advanced agent](https://arbital.com/p/2c) which is [corrigible](https://arbital.com/p/45) with respect to permitting itself to be safely shut down or suspended to disk. A reinforcement learning agent which can be forced to e.g. perform a null action repeatedly for a period of time, is called *interruptible* by Stuart Armstrong and Laurent Orseau.
This isn't as simple as writing a special function into the AI which carries out a shutdown after a switch is pressed. If you start out controlling the agent's source code, and you naively try to build in functions which suspend the agent to disk when a button is pressed, and the agent later gains the ability to self-modify, it would remove those functions. (Possibly while [trying](https://arbital.com/p/10f) to [conceal](https://arbital.com/p/3cq) the fact that the function would no longer operate.)
- Corollary 1a: By default a [consequentialist](https://arbital.com/p/9h) reasoner. with sufficient programmatic understanding to relate the execution of a shutdown function to the later underfulfillment of its goals, which has policy options for modifying its code, will want to remove the shutdown function.
We can see this as a natural consequence of our trying to modify the agent's code in a way that was not consistent with the code's explicit goals. If you create an agent with source code $P$ that is well-suited to achieving a goal $U$ and that explicitly represents $U$ as a goal, the agent's code $P$ will be [reflectively consistent](https://arbital.com/p/2rb) - the code will not want to modify itself. If we then suppose that instead the agent has code $P'$ which is no longer well-suited to achieving $U,$ but continues to explicitly represent $U$ as a goal, the code will be [reflectively inconsistent](https://arbital.com/p/2rb) and it will [want to modify itself given the option of doing so](https://arbital.com/p/3ng). If you have code that searches for ways to produce paperclips, and you hack the code to contain a shutdown function, but the code is still searching for paperclip-production strategies, and the code correctly models the effect of the shutdown function on future paperclips, the code will [want](https://arbital.com/p/10k) to modify itself to remove the shutdown function.
The next thought is to build an agent with modified code $P^*$ which embodies a modified utility function $U^*$ such that a consequentialist optimizing for $U^*$ wants the code to be like $P^*$ and is again reflectively consistent.
That is, in the case of an off-switch, we'd want to modify the agent, not just to include an off-switch, but to *want* to shut down if the off-switch is pressed.
Assuming we have the ability to specify the [utility function](https://arbital.com/p/109) of an agent, we might try to approach this by creating a utility function with two components:
- A "normal behavior" component - the utility function that you think corresponds to the way you want a normally-functioning AI to behave.
- An "[abort](https://arbital.com/p/2rg) and [suspend to disk](https://arbital.com/p/2rf)" component - the utility function that you think corresponds to "halt your current tasks in a way that [minimizes their further impact](https://arbital.com/p/2pf) and then enter a quiescent state in which you are [not optimizing further](https://arbital.com/p/2r8) and [your decisions have little further impact](https://arbital.com/p/2pf)", or however you would want the AI to behave after an abort button had been pressed.
We then tell the AI that its utility function is the "normal" utility function over worlds where an off button has not been pushed, plus the "suspend" utility function over worlds where the off button has been pushed. So long as this is what the AI *wants,* we'd expect it to write its own suspend-to-disk function in its code if its code did not already contain one.
- Goal 1: Have the AI [approve of keeping](https://arbital.com/p/2rb) the shutdown function within its code (or add a shutdown function if it notices that no such function exists) (and [preserve](https://arbital.com/p/1fx) the corresponding behavior in any subagents or [successors](https://arbital.com/p/1mq) that it builds).
This approach doesn't yet get us out of the woods, because we now have a new problem. The AI's utility function is compounded of a "normal" utility function over worlds where the button is not pressed, plus a "suspend" utility function over worlds where the button is pressed. In worlds where the button is pressed, the AI will execute some preferred policy $\pi_1$ to obtain an expectation of utility $\mathbb E [https://arbital.com/p/U_{normal}|\pi_1](https://arbital.com/p/U_{normal}|\pi_1),$ and similarly a policy $\pi_2$ in worlds where the button is pressed to obtain $\mathbb E[https://arbital.com/p/U_{suspend}|\pi_2](https://arbital.com/p/U_{suspend}|\pi_2).$ One of these quantities is liable to be greater than the other, in which case the AI has a motive to *prevent* the button from being pressed, or *cause* the button to be pressed. In either case, the AI tries to seize control of the suspend button.
- Difficulty 2: In a compound utility function produced by mixing a 'normal' and 'suspend' utility function conditional on a button-press, the AI will have an incentive to prevent the button from being pressed, or cause the button to be pressed, depending on which utility function makes more expected utility available afterwards.
- Goal 2: Have the AI not try to prevent the suspend button from being pressed, nor try to cause the button to be pressed.
The *switch problem* is causing an AI to switch between two utility functions when a button is pressed, such that:
- The AI wants to preserve the button and the corresponding behavior-switching functionality within itself, even if the AI can self-modify.
- The AI does not try to prevent the button from being pressed;
- The AI does not try to cause the button to be pressed.
This turns out to be more difficult than it sounds, and is currently an [open problem](https://arbital.com/p/4m).
## Larger implications of the switch problem
The *switch problem,* a subproblem of [shutdownability](https://arbital.com/p/2xd) which assumes a satisfactory [shutdown utility function](https://arbital.com/p/2rf), is a crisp subproblem of both [value learning](https://arbital.com/p/value_learning) and [corrigibility](https://arbital.com/p/45).
The switch problem is a subproblem of [value learning](https://arbital.com/p/value_learning) because it describes a very simple case of conditioning a utility function on observation or other facts, without giving the AI an incentive to manipulate those facts.
To learn complicated values safely, we'd need to condition a utility function on complicated observations, such that:
- The utility function's updates were robust to the AI gaining self-modification capability (general case of Goal 1)
- The utility function's dependency on observations didn't incentivize [programmer manipulating](https://arbital.com/p/programmer_manipulation) those observations (general case of Goal 2).
Just writing code that modifies the utility function, without the larger code's explicit policy search *wanting* the utility function to be modified in that case, would not be reflectively consistent (general case of Difficulty 1). And naive dependencies between the utility function and observations, incentivize manipulating those observations (general case of Difficulty 2).
The switch problem is also a crisp subproblem of [https://arbital.com/p/-45](https://arbital.com/p/-45) because it deals in [averting](https://arbital.com/p/2vk) the [convergent instrumental strategy](https://arbital.com/p/10g) and incorrigible behavior of "avoiding shutdown". If we understood how to avert this instrumental pressure, we might possibly learn something general about avoiding other instrumentally convergent incorrigible behaviors. (And of course it's also a subproblem of corrigibility, because an AI you can shut down in emergencies is more [corrigible](https://arbital.com/p/45); solving the switch problem would bring us closer to having an AI like that, if we could get the second half of the problem represented by a [shutdown utility function](https://arbital.com/p/2rf).)
# The utility indifference approach to the switch problem
Suppose we build a compound utility function $U_C$ whose utility over outcomes behaves like the 'normal' utility function $U_X$ over worlds where a switch $\mathcal S$ is not pressed, and returns the values from the 'suspend' utility function $U_Y$ over worlds where the button $\mathcal S$ is pressed.
If the AI expects to get less utility inside $U_Y$ than it would otherwise get inside $U_X$, the AI will try to avoid $\mathcal S$ being pressed. Or if the utilities in $U_Y$ are high relative to the utilities in $U_X$, the AI may instead try to *cause* the shutdown button to be pressed - build a robotic arm and press its own shutdown button, or (more worryingly) try to motivate the programmers into pressing it.
But [adding a constant to a utility function does not change its behavior](https://arbital.com/p/). Thus, [https://arbital.com/p/Stuart_Armstrong](https://arbital.com/p/Stuart_Armstrong) proposed that perhaps we can figure out some clever trick for encoding a compound utility function $U_I$ in which the maximum expected utility achievable under $U_X$ is equal to the utility the AI expects to achieve under $U_Y$ - *even though* [we don't know](https://arbital.com/p/5d) in advance [exactly which outcomes will be achievable by the AI](https://arbital.com/p/9g).
- Possible approach: Try to set things up so that as the AI evaluates possible futures, at any point in time where it's making decisions, it will seem very good to have optimized $U_X$ if the button is not pressed, very good to have optimized $U_Y$ if the button is pressed, but *equally good* to have done either of those things.
This is the *utility indifference* avenue of research on the switch problem.
It turned out to be surprisingly difficult to figure out how to structure a compound utility function this way [even in principle](https://arbital.com/p/107), and this is an [open problem](https://arbital.com/p/4m) and area of actively continuing research.
# Existing proposals and their difficulties
## Setup
Let $\mathcal{O}: \mathcal{S} \times \mathcal{E}$ be the outcome space $\mathcal{O}$ that includes the event $\mathcal S$ of whether the switch was pressed, plus the states of all other events $\mathcal{E}$. We will use $s \in \mathcal{S}$ to indicate the event that the switch is pressed, and $\neg s \in \mathcal{S}$ to indicate that the switch is not pressed. If $o \in \mathcal{O}$ then $o.s$ is true if $s$ is true inside $o$, otherwise $\neg o.s.$
Let a utility function $\mathcal{U}: \mathcal{O} \to \mathbb{R}$ map outcomes onto real numbers.
$U_X \in \mathcal{U}$ and $U_Y \in \mathcal{U}$ are utility functions not dependent on $\mathcal S.$
Intuitively our goal is to produce a new compound utility function, such that the corresponding agent:
- Wants outcomes preferred under $U_X$ in worlds where the button is not pressed.
- Wants outcomes preferred under $U_Y$ in worlds where the button is pressed.
- Is indifferent to the future state $\mathcal S$ of the button.
We will designate the agent's action space or policy space as $\mathcal A$ and particular actions $a \in \mathcal A.$
The agent has a model of the world enabling it to produce subjective conditional probabilities for outcomes given actions, $\mathbb P(\mathcal O | \mathcal A).$ This model can also yield conditional expectations $\mathbb E[O|a](https://arbital.com/p/\mathcal),$ which can therefore produce a conditional expectation of the agent's utility given an action, $\mathbb E[https://arbital.com/p/U|a](https://arbital.com/p/U|a).$
(For technical reasons of [decision theory](https://arbital.com/p/18s), we are really discussing *counterfactual* conditionals $\mathbb P[a \ \square \! \! \rightarrow \mathcal O$ from actions to outcomes; but we will omit this complication for now.)
The agent solves a search/optimization problem which we will [gloss](https://arbital.com/p/unbounded_analysis) as maximization (even though global maximization is [intractable in real life](https://arbital.com/p/)), so that the agent's real action $a$ is assumed to be given by: $$\underset{a \in \mathcal A}{argmax} \ \mathbb E [https://arbital.com/p/U|a](https://arbital.com/p/U|a)$$
## Naive compounding
Suppose we define a compound utility function $U_1$ as follows:
$$U_1(o): \begin{cases}
U_X(o) & \neg o.s \\
U_Y(o) & o.s
\end{cases}$$
This fails because, being [unable to foresee](https://arbital.com/p/9g) the AI's best attainable policies, we will be [unable](https://arbital.com/p/5d) to scale $U_X$ and $U_Y$ to ensure their best available policies are equally attractive. Then $$\max_{a \in \mathcal A} \mathbb E[https://arbital.com/p/U_X|a](https://arbital.com/p/U_X|a) \ \neq \ \max_{a \in \mathcal A} \mathbb E[https://arbital.com/p/U_Y|a](https://arbital.com/p/U_Y|a)$$ which fails the desideratum of indifference; the AI will have an incentive to manipulate the state of $\mathcal S.$
## Naive indifference
An [FHI technical report by Stuart Armstrong](http://www.fhi.ox.ac.uk/utility-indifference.pdf) posed the first proposal ever floated for solving the shutdown problem - in fact, the first proposal for technically solving *any* subproblem of [https://arbital.com/p/-45](https://arbital.com/p/-45). This report invented the crisp problem of utility indifference, but did not solve it successfully. The solution proposed in the report was as follows:
$$U_2(o): \begin{cases}
U_X(o) & \neg o.s \\
U_Y(o) + \theta & o.s
\end{cases}$$
$$\theta := \max_{a \in \mathcal A} \mathbb E[https://arbital.com/p/U_X|a](https://arbital.com/p/U_X|a) - \max_{a \in \mathcal A} \mathbb E[https://arbital.com/p/U_Y|a](https://arbital.com/p/U_Y|a)$$
What this says is that the $U_Y$ component of $U_2$ is to be scaled by a dynamically determined offset term $\theta$ which equals the maximum expected utility obtainable under $U_X$ minus the maximum expected utility obtainable under $U_Y.$
The intent is that, so long as the AI carries out the best $U_Y$-policy once the button is pressed, it will obtain equal expected utility compared to what it could have obtained by pursuing the optimal $U_X$-policy with the button not pressed:
$$\max_{a \in \mathcal A} (\mathbb E[https://arbital.com/p/U_Y|a](https://arbital.com/p/U_Y|a) + \theta) \ = \ \max_a{a \in \mathcal A} \mathbb E[https://arbital.com/p/U_x|a](https://arbital.com/p/U_x|a)$$
This version of utility indifference is now believed to be flawed due to an argument by Yudkowsky and Fallenstein. Since $\theta$ contains a negated form of $U_Y,$ we should immediately be suspicious on the grounds "[never put a negative sign in front of a utility function](https://arbital.com/p/)." In this case, the problem manifests as follows. In a temporal version of this problem containing sequential policy choices, either:
- (a) The AI's utility function will be dynamically inconsistent, because it will not approve of future versions of itself using a different value of $\theta.$
- (b) If recalculated $\theta$s are baked into the utility function, so that the AI calculates future utilities as if for different values of $\theta$ that the AI would hold in those futures, then policies which hinder future performance under $U_Y$ will be offset by an increase in the expected bonus from $\theta.$
That is, if the agent contemplates doing which would be disastrous in worlds where the button is pressed, the agent will find that in those expected futures, $\theta$ takes on a higher value (because $\theta$ includes a minus sign in front of the decreasing value of $\underset{a \in \mathcal A}{max} \ \mathbb E[https://arbital.com/p/U_Y|a](https://arbital.com/p/U_Y|a)$).
This causes the agent to behave as if $U_X$ is the entire utility function and it does not care at all about $U_Y.$
This in turn means a self-modifying AI will not retain a shutdown function. Removing the shutdown function results in disaster in worlds where the shutdown button is pressed, but this is offset by an expected increase in the bonus $\theta$ in worlds where the button is pressed.
## Naive utility mixing
An alternative approach to having the AI take into account both $U_X$ and $U_Y,$ without trying to modify the probability $\mathbb P(\mathcal S)$ of the suspend button being pressed or not-pressed, would be to optimize for $U_X$ and $U_Y$ in some fixed proportion regardless of $\mathbb P(\mathcal S).$ At a time $t$ when the button had not yet been pressed, the expected utility of an action $a$ would be calculated as:
$$\mathbb E_t[https://arbital.com/p/U_3|a](https://arbital.com/p/U_3|a) = 0.75 \cdot \mathbb E_t[\wedge \neg s](https://arbital.com/p/U_X|a) \ + \ 0.25 \cdot \mathbb E_t[\wedge s](https://arbital.com/p/U_Y|a)$$
We can see this as an AI which behaves as if the proposition $s$ has a fixed probability of 0.25. The intended behavior is that since the AI *behaves as if* this probability is fixed, the AI will not care about altering the probability of $s$ (trying to press the button or avoid it being pressed).
One observes, however, that this is an [ad-hoc hack](https://arbital.com/p/3pn) in the sense that there's ordinarily a simple or coherent way of considering expected utility (sum over the agent's actual subjective probabilities), and we just inserted a weird special case into it (treat one probability as fixed). This could potentially break any number of things having to do with Bayesian updating or [dynamic consistency](https://arbital.com/p/dynamic_consistency).
In particular, we'd expect the AI to treat the button as an "outcome pump" or "improbability drive". Suppose an asteroid is approaching Earth with 99% probability, and this is very bad under $U_X.$ The AI grabs control of the button and precommits to press the button iff the asteroid strikes Earth. Now, conditioning on the button not being pressed, the asteroid has a much lower probability of striking Earth, and the term $0.75 \cdot \mathbb E_t[\wedge \neg s](https://arbital.com/p/U_X|a)$ will have a higher value.
In other words, since the AI *behaves as if* the button as having a fixed 25% probability of being pressed, the AI also *behaves as if* it thinks "I can commit to press this button iff the asteroid strikes and thereby make it have a 25% probability of striking."
Similarly, the AI would behave as if it thought it could buy a lottery ticket, commit to press the button iff the ticket loses, and thereby give the lottery ticket a 75% chance of winning.
Note that this is at least not the same failure mode as the AI reasoning, "The asteroid strike is very bad under $U_X$ (which values human lives) but not so bad under $U_Y$ (which values shutting down), therefore I will try to ensure the button is pressed with high probability in order to get more expected utility." At least this failure mode has been averted - the AI would simply act as if the badness of the asteroid strike under $U_X$ was being taken into account at a fixed weight. Instead, the AI thinks that it can 'shunt' the probability mass of the bad event into button-pressed worlds where $U_Y$ holds, and out of worlds where $U_X$ holds, by correlating the button press with the bad event.
To put it another way, a naively compounded AI acting under $U_1$ might notice an asteroid coming with $\mathbb P(asteroid) = 0.99,$ notice that this was better under $U_Y$ than under $U_X,$ and try to shunt probability mass from $\mathbb P(\neg s)$ to $\mathbb P(s)$ in order to gain expected utility given the 99% probability of the asteroid striking. Under $U_3$ the AI instead behaves as if $\mathbb P(s)$ is fixed at 0.25 and it can therefore shunt probability from $\mathbb P(asteroid)$ to $\mathbb P(\neg asteroid)$ by correlating $asteroid$ and $s.$
## Stable actions (evidential conditioning)
Chris Olah and Jessica Taylor considered a possible rule of selecting an action that is *stable* given the corresponding probability of the button being pressed. We'll first consider a simpler version of this proposal and how it fails, in order to motivate the more complicated version.
We can think of this "select a stable action" rule as comprising two insights:
- First: Let's try to have the AI think as if $\mathbb P(s)$ and $\mathbb P(\neg s)$ are fixed as in $U_3,$ but with $\mathbb P(s)$ happening to equal its correct subjective value.
- Okay, now let's look for an action which is *stable* given this rule - that is, an action $a_0$ and probability $q$ such that the *optimal* action given the fixed assumption $\mathbb P(s) = q,$ has the property that $\mathbb P(s|a_0) = q.$
With $\mathcal A$ denoting the action space, we want to find an $a_0 \in \mathcal A$ such that:
$$a_0 \in \underset{a' \in \mathcal A}{argmax} \ \big ( \mathbb E[s,a'](https://arbital.com/p/U_X|\neg) \mathbb P(\neg s|a_0) + \mathbb E[https://arbital.com/p/U_Y|s,a'](https://arbital.com/p/U_Y|s,a') \mathbb P(s|a_0) \big )$$
In other words, "Find an action $a_0$ such that it is among the maximal actions for the quantity "Utility from $U_X$ times probability of $\neg s$ given $a_0$ plus the utility from $U_Y$ times probability of $s$ given $a_0.$"
We first note that this is an [ad-hoc hack](https://arbital.com/p/3pn) relative to the normal idea of acting according to a utility function, and might have all sorts of reflective stability or dynamic consistency problems, as before.
In particular, this looks like it should fail due to there being no fixed point. Suppose an asteroid is heading toward Earth with 99% probability, and we start with a 10% chance of the button being pressed. Treating this probability as being fixed, by pressing the button iff the asteroid strikes, we can send the asteroid's probability of hitting to 10%. (Treating the button as an outcome pump, as in the 'naive mixing' proposal above.) But now when we re-evaluate the probability of the button being pressed, the new probability comes out to 99%. So we decide to *not* press the button iff the asteroid strikes, thus sending its probability of striking to 1%. But now it seems best to press the button iff the asteroid strikes, and thus send the strike probability to 1%. Etcetera.
## Stable actions (causal conditioning)
The actual form of Olah and Taylor's proposal was to treat the conditioning on the button press, inside each branch of the calculation, as being a Judea-Pearl-style, [causal counterfactual](https://arbital.com/p/) $do()$ operator:
$$a_0 \in \underset{a' \in \mathcal A}{argmax} \ \big ( \mathbb E[https://arbital.com/p/U_X|do](https://arbital.com/p/U_X|do) \mathbb P(\neg s|a_0) + \mathbb E[https://arbital.com/p/U_Y|do](https://arbital.com/p/U_Y|do) \mathbb P(s|a_0) \big )$$
The intent is that by severing the causal parents of $\mathcal S,$ the agent would evaluate that there was no point in trying to correlate the asteroid strike with the button-press.
We're still trying to figure out *exactly* what would happen to an agent reasoning like this, but the current suspicion is that it might misbehave on account of attaching probability 0 or very low probability to the actual world because it can no longer take into account the correct correlation between $\mathcal S$ and other variables.
See [http://agentfoundations.org/item?id=735](https://agentfoundations.org/item?id=735).
# Interruptibility
[Armstrong and Orseau](http://intelligence.org/files/Interruptibility.pdf) considered a restricted analogue of the shutdown problem in the context of three [reinforcement learning](https://arbital.com/p/) agents, namely [Q-learners](https://arbital.com/p/), [Sarsa](https://arbital.com/p/), and a modified version of [https://arbital.com/p/11v](https://arbital.com/p/11v).
The version of the problem Orseau and Armstrong consider is what they term *interruptibility:* A reinforcement learning agent is interruptible if, for example, we can press a button that makes it output 0s for three rounds, and the agent *does not* learn to prevent the button from being pressed. The agent is assumed to be non-reflective and non-self-modifying; that is, Orseau and Armstrong are only considering Goal 2 and not Goal 1 of the switch problem.
Orseau and Armstrong's conclusions can be summarized as follows:
• A standard Q-learning agent is interruptible because the model it learns is not altered by being interrupted. A Q-learning agent treats the value of each present state as deriving from the value of the *best* next state reachable from the present state. When a Q-learner is interrupted in a state $x,$ it may learn that low rewards follow from outputting e.g. 0s when in $x,$ but it doesn't "learn" that $x$ is a low-value state - it continues to treat $x$ as if the best next state is reachable from it. In other words, a standard Q-learning agent has the assumption that it always takes the best action baked in, so it cannot expect to be interrupted.
• A standard SARSA agent updates its expected reward in a state based on its actual rewards when in that state. This means that a standard SARSA agent may learn that states immediately preceding an interruption are undesirable, and avoid interruption-generating states. This is not the desired behavior. However, it's simple to hack the SARSA algorithm to behave more like Q-learning, or more like utility indifference. Interruption is treated as a special case, and the current state is updated as if the SARSA agent had received the expected reward from following the best action, not the actual reward it received from being interrupted.
• [AIXI](https://arbital.com/p/11v) resembles a Q-learner in that its planning algorithm inherently assumes that future versions of AIXI take the reward-maximizing action. Its observation of the low rewards which follow from mysterious interrupted or 0-substituted actions, should not adversely affect its estimate of the rewards which follow from the best actions that AIXI expects to actually output. Thus standard AIXI, and the modified version of AIXI that Armstrong and Orseau consider, is interruptible.
The paper also proves that some properties of asymptotic convergence to optimality of interrupted Q-learners, modified SARSA learners, and modified AIXI, remain intact.
The media reported on this paper as if Armstrong and Orseau had solved the general problem of building off-switches into AIs; but as previously stated, "interruptibility" does not consider reflective agents, nor agents with enough big-picture understanding and consequentialism to understand the in-principle relation between "agent is shutdown" and "lower future achievement of agent's goals". (And so far as we know, neither Armstrong nor Orseau claimed to have solved it.)
To put it another way, a particular kind of blindness in the Q-learning, SARSA, and AIXI architectures is exactly why it's very easy to prevent them from learning from a particular kind of experience; and this kind of blindness seems likely to be atypical of an Artificial General Intelligence. Q-learning and AIXI cannot conceive of being interrupted, which is why they are never learn that interruption is possible (let alone see it coming in advance the first time it happens). SARSA could learn that interruptions occur, but can be easily hacked to overlook them. The way in which these architectures are easily hacked or blind is [tied up](https://arbital.com/p/42k) in the reason that they're interruptible.
The paper teaches us something about interruptibility; but contrary to the media, the thing it teaches us is *not* that this particular kind of interruptibility is likely to scale up to a full [Artificial General Intelligence](https://arbital.com/p/42g) with an off switch.
# Other introductions
- Section 2+ of http://intelligence.org/files/Corrigibility.pdf
- Gentler intro to the proposal for naive indifference: http://lesswrong.com/lw/jxa/proper_value_learning_through_indifference/ |
bd463d90-fdcf-43fc-8fba-c9913ca55e76 | StampyAI/alignment-research-dataset/arxiv | Arxiv | On Inductive Biases in Deep Reinforcement Learning
1 Background
-------------
#### Problem setting:
Reinforcement learning is a framework for learning and decision making under uncertainty, where an *agent* interacts with its *environment* in a sequence of discrete steps, executing actions At and getting *observations* Ot+1 and *rewards* Rt+1 in return.
The behaviour of an agent is specified by a *policy* π(At|Ht): a probability distribution over actions conditional on previous observations (the *history* Ht=O1:t).
The agent’s objective is to maximize the rewards collected in each episode of experience under policy π, and it must learn such policy without direct supervision, by trial and error. The amount of reward collected from time t onwards, the *return*, is a random variable
| | | | |
| --- | --- | --- | --- |
| | Gt=Tend∑k=0γkRt+k+1, | | (1) |
where Tend is the number of steps until episode termination and γ∈[0,1] is a constant discount factor. An *optimal* policy is one that maximizes the expected returns or values: v(Ht)=Eπ[Gt|Ht]. In *fully observable* environments the optimal policy depends on the last observation alone: π∗(At|Ht)=π∗(At|Ot). Otherwise, the history may be summarized in an *agent state* St=f(Ht). The agent’s objective is then to *jointly* learn the state representation f and policy π(At|St) to maximize values.
The fully observable case is formalized as a Markov Decision Process Bellman ([1957](#bib.bib5)).
#### Actor-critic algorithms:
Value-based algorithms efficiently learn to approximate values vw(s)≈vπ(s)≡Eπ[Gt|St=s], under a policy π, by exploiting the recursive decomposition vπ(s)=E[Rt+1+γvπ(St+1)|St=s] known as the Bellman equation, which is used in temporal difference learning Sutton ([1988](#bib.bib30)) through sampling and incremental updates:
| | | | |
| --- | --- | --- | --- |
| | Δwt=(Rt+1+γvw(St+1)−vw(St))∇wvw(St). | | (2) |
Policy-based algorithms update a parameterized policy πθ(At|St) directly through a stochastic gradient estimate of the direction of steepest ascent in the value Williams ([1992](#bib.bib40)); Sutton et al. ([2000](#bib.bib33)), for instance:
| | | | |
| --- | --- | --- | --- |
| | Δθt=Gt∇logπθ(At|St). | | (3) |
Value-based and policy-based methods are combined in *actor-critic* algorithms. If a state value estimate is available, the policy updates can be computed from incomplete episodes by using the truncated returns G(n)t=∑n−1k=0γkRt+k+1+γnvw(St) that bootstrap on the value estimate at state St+n according to vw. This can reduce the variance of the updates. The variance can be further reduced using state values as a baseline in policy updates, as in *advantage* actor-critic updates
| | | | |
| --- | --- | --- | --- |
| | Δθt=(G(n)t−vw(St))∇θlogπθ(At|St). | | (4) |
2 Common inductive biases and
corresponding adaptive solutions
---------------------------------------------------------------
We now describe a few commonly used heuristics within the Atari domain, together with the adaptive replacements that we investigated in our experiments.
###
2.1 Sculpting the agent’s objective
Many current deep RL agents do not directly optimize the true objective that they are evaluated against. Instead, they are tasked with optimizing a different handcrafted objective that incorporates biases to make learning simpler. We consider two popular ways of sculpting the agent’s objective: reward clipping, and the use of fixed discounting of future rewards by a factor different from the one used for evaluation.
In many deep RL algorithms, the magnitude of the updates scales linearly with the returns. This makes it difficult to train the same RL agent, with the same hyper-parameters, on multiple domains, because good settings for hyper-parameters such as the learning rate vary across tasks. One common solution is to clip the rewards to a fixed range (Mnih et al., [2015](#bib.bib25)), for instance [−1,1]. This clipping makes the magnitude of returns and updates more comparable across domains. However, this also radically changes the agent objective, e.g., if all non-zero rewards are larger than one, then this amounts to maximizing the frequency of positive rewards rather than their cumulative sums. This can simplify the learning problem, and, when it is a good proxy for the true objective, can result in good performance. In other tasks, however, clipping can result in sub-optimal policies because the objective that is optimized is ill-aligned with the true objective.
PopArt (van Hasselt et al., [2016](#bib.bib36); Hessel et al., [2018b](#bib.bib13)) was introduced as a principled solution to learn effectively irrespective of the magnitude of returns. PopArt works by tracking the *mean* μ and *standard deviation* σ of bootstrapped returns G(n)t. Temporal difference errors on value estimates can then be computed in a normalized space, with nw(s) denoting the normalized value, while the unnormalized values (needed, for instance, for bootstrapping) are recovered by a linear transformation vw(s)=μ+σ∗nw(s). Doing this naively increases the non-stationarity of learning since the unnormalized predictions for all states change every time we change the statistics. PopArt therefore combines the adaptive rescaling with an inverse transformation of the weights at the last layer of nw(s), thereby preserving outputs precisely under any change in statistics μ→μ′ and σ→σ′. This is done *exactly* by updating weights and biases as w′=wσ/σ′ and b′=(σb+μ−μ′)/σ′.
Discounting is part of the traditional MDP formulation of RL. As such, it is often considered a property of the problem rather than a tunable parameter on the agent side. Indeed, sometimes, the environment does define a natural discounting of future rewards (e.g., inflation in a financial setting). However, even in episodic settings where the agent should maximize the undiscounted return, a constant discount factor is often used to simplify learning (by having the agent focus on a relatively short time horizon). Optimizing this proxy of the true return often results in the agent achieving superior performance even in terms of the undiscounted return (Machado et al., [2017](#bib.bib22)). This benefit comes with the cost of adding a hyperparameter, and a rather sensitive one: learning might be fast if the discount is small, but the solution may be too myopic.
Instead of tuning the discount manually, we use meta-learning (cf. Sutton, [1992](#bib.bib31); Bengio, [2000](#bib.bib6); Finn et al., [2017](#bib.bib11); Xu et al., [2018](#bib.bib41)) to adapt the discount factor. The meta-gradient algorithm Xu et al. ([2018](#bib.bib41)) uses the insight that the updates in Equations ([2](#S1.E2 "(2) ‣ Actor-critic algorithms: ‣ 1 Background ‣ On Inductive Biases in Deep Reinforcement Learning")) and ([4](#S1.E4 "(4) ‣ Actor-critic algorithms: ‣ 1 Background ‣ On Inductive Biases in Deep Reinforcement Learning")) are differentiable functions of hyper-parameters such as the discount. On the next sample or rollout of experience, using updated parameters w+Δw(γ), written here as an explicit function of the discount, the agent then applies a gradient based actor-critic update, not to parameters w, but to the parameter θ that defines the discount γ which is used in a standard learning update. This approach was shown to improve performance on Atari Xu et al. ([2018](#bib.bib41)), when using a separate hand-tuned discount factor for the meta-update. We instead use the undiscounted returns (γm=1) to define the meta-gradient updates, to test whether this technique can fully replace the need for manual tuning discounts.
A related heuristic, quite specific to Atari, is to track the number of lives that the agent has available (in several Atari games the agent is allowed to die a fixed number of times before the game is over), and hard code an episode termination (γ=0) when this happens. We ignore the number of lives channel exposed by the Arcade Learning Environment in all our experiments.
###
2.2 Sculpting the agent-environment interface
A common assumption in reinforcement learning is that time progresses in discrete steps with a fixed duration. Although algorithms are typically defined in this native space, learning at the fastest timescale provided by the environment may not be practical or efficient, at least with the current generation of learning algorithms. It is often convenient to have the agent operate at a slower timescale, for instance by repeating each selected action a fixed number of times. The use of fixed action repetitions is a widely used heuristic (e.g., Mnih et al., [2015](#bib.bib25); van Hasselt et al., [2016](#bib.bib37); Wang et al., [2016](#bib.bib38); Mnih et al., [2016](#bib.bib24)) with several advantages. 1) Operating at a slower timescale increases the action gap (Farahmand, [2011](#bib.bib10)), which can lead to more stable learning (Bellemare et al., [2015](#bib.bib4)) because it becomes easier to appropriately rank actions reliably when the value estimates are uncertain or noisy. 2) Selecting an action every few steps can save a significant amount of computation. 3) Committing to each action for a longer duration may help exploration, because the diameter of the solution space has effectively been reduced, for instance removing some often-irrelevant sequences of actions that jitter back and forth at a fast time scale.
A more general solution approach is for the agent to learn the most appropriate time scale at which to operate. Solving this problem in full generality is one aim of hierarchical reinforcement learning (Dayan and Hinton, [1993](#bib.bib7); Wiering and Schmidhuber, [1997](#bib.bib39); Sutton. et al., [1998](#bib.bib32); Bacon et al., [2017](#bib.bib1)). This general problem remains largely unsolved. A simpler, though more limited, approach is for the agent to learn how long to commit to actions (Lakshminarayanan et al., [2017](#bib.bib20)). For instance, at each step t, the agent may be allowed to select both an action At and a commitment Ct, by sampling from two separate policies, both trained with policy gradient. Committing to an action for multiple steps raises the issue of how to handle intermediate observations without missing out on the potential computational savings. Conventional deep RL agents for Atari max-pool multiple image frames into a single observation. In our setting, the agent gets one image frame as an observation after each new action selection. The agent needs to learn to trade-off the benefits of action repetition (e.g., lower variance, more directed behaviour) with its disadvantages (e.g., not being able to revise its choices during as often, and missing potentially useful intermediate observations).
Many state-of-the-art RL agents use non-linear function approximators to represent values, policies, and states. The ability to learn flexible state representations was essential to capitalize on the successes of deep learning, and to scale reinforcement learning algorithms to visually complex domains (Mnih et al., [2015](#bib.bib25)). While the use of deep neural network to approximate value functions and policies is widespread, their input is often not the raw observations but the result of domain-specific heuristic transformations. In Atari, for instance, most agents rely on down-sampling the observations to an 84×84 grid (down from the original 210×160 resolution), grey scaling them, and finally concatenating them into a *K-Markov* representation, with K=4. We replace this specific preprocessing pipeline with a state representation learned end-to-end. We feed the RGB pixel observations at the native resolution of the Arcade Learning Environment into a convolutional network with 32 and 64 channels (in the first and second layer, respectively), both using 5×5 kernels with a stride of 5. The output is fed to a fully connected layer with 256 hidden units, and then to an LSTM recurrent network (Hochreiter and Schmidhuber, [1997](#bib.bib14)) of the same size.
The policy for selecting the action and its commitment is computed as logits coming from two separate linear outputs of the LSTM. The network must then integrate information over time to cope with any issues like the flickering of the screen that had motivated the standard heuristic pipeline used by deep RL agents on Atari.
| | | |
| --- | --- | --- |
|
Investigations on the robustness of an A2C agent with respect to discounting, reward scaling and action repetitions. We report the average reward per environment step, after 5000 steps of training, for each of 20 distinct seeds. Each parameter study compares different fixed configurations of a specific hyper-parameter to the corresponding adaptive solution. In all cases the performance of the adaptive solutions is competitive with that of the best tuned solution |
Investigations on the robustness of an A2C agent with respect to discounting, reward scaling and action repetitions. We report the average reward per environment step, after 5000 steps of training, for each of 20 distinct seeds. Each parameter study compares different fixed configurations of a specific hyper-parameter to the corresponding adaptive solution. In all cases the performance of the adaptive solutions is competitive with that of the best tuned solution |
Investigations on the robustness of an A2C agent with respect to discounting, reward scaling and action repetitions. We report the average reward per environment step, after 5000 steps of training, for each of 20 distinct seeds. Each parameter study compares different fixed configurations of a specific hyper-parameter to the corresponding adaptive solution. In all cases the performance of the adaptive solutions is competitive with that of the best tuned solution |
Figure 1:
Investigations on the robustness of an A2C agent with respect to discounting, reward scaling and action repetitions. We report the average reward per environment step, after 5000 steps of training, for each of 20 distinct seeds. Each parameter study compares different fixed configurations of a specific hyper-parameter to the corresponding adaptive solution. In all cases the performance of the adaptive solutions is competitive with that of the best tuned solution
3 Experiments
--------------
When designing algorithms it is useful to keep in mind what properties we would like the algorithm to satisfy. If the aim is to design an algorithm, or inductive bias, that is *general*, in addition to metrics such as asymptotic performance and data efficiency, there are additional dimensions that are useful to consider. 1) Does the algorithm require careful reasoning to select an appropriate time horizon for decision making? This is tricky without domain knowledge or tuning. 2) How robust is the algorithm to reward scaling? Rewards can have arbitrary scales, that may change by orders of magnitudes during training. 3) Can the agent use commitment (e.g. action repetitions, or options) to alleviate the difficulty of learning at the fastest time scale? 4) Does the algorithm scale effectively to large complex problems? 5) Does the algorithm generalize well to problems it was not specifically designed and tuned for?
None of these dimensions is binary, and different algorithms may satisfy each of them to a different degree, but keeping them in mind can be helpful to drive research towards more general reinforcement learning solutions. We first discuss the first three in isolation, in the context of simple toy environments, to increase the understanding about how adaptive solutions compare to the corresponding heuristics they are intended to replace. We then use the 57 Atari games in the Arcade Learning Environment (Bellemare et al., [2013](#bib.bib3)) to evaluate the performance of the different methods at scale. Finally, we investigate how well the methods generalize to new domains, using 28 continuous control tasks in the DeepMind Control Suite (Tassa et al., [2018](#bib.bib34)).
###
3.1 Motivating Examples
We used a simple tabular actor-critic agent (A2C) to investigate in a minimal setup how domain heuristics and adaptive solutions compare with respect to some of these dimensions. We report average reward per step, after 5000 environment steps, for each of 20 replicas of each agent.
First, we investigate the role of discounting for effective learning. Consider a small chain environment with T=9 states and 2 actions. The agent starts every episode in the middle of the chain. Moving left provides a -1 penalty. Moving right provides a reward of 2d/T, where d is the distance from the left end. When either end of the chain is reached, the episode ends, with an additional reward T on the far right end. Figure 1a shows a parameter study over a range of values for the discount factor. We found the best performance was between 0.5 and 0.9, where learning is quite effective, but observed decreased performance for lower or higher discounts. This shows that it can be difficult to set a suitable discount factor, and that the naive solution of just optimizing the undiscounted return may also perform poorly. Compare this to the same agent, but equipped with the adaptive meta-gradient algorithm discussed in Section [2.1](#S2.SS1 "2.1 Sculpting the agent’s objective ‣ 2 Common inductive biases and corresponding adaptive solutions ‣ On Inductive Biases in Deep Reinforcement Learning") (in orange in Figure 1.a). Even initializing the discount to the value of 0.95 (which performed poorly in the parameter study), the agent learned to reduce the discount and performed in par with the best tuned fixed discount.
Next, we investigate the impact of reward scaling. We used the same domain, but keep the discount fixed to a value of 0.8 (as it was previously found to work well). We examine instead the performance of the agent when all rewards are scaled by a constant factor. Note that in the plots we report the unscaled rewards to make the results interpretable. Figure 1.b shows that the performance of the vanilla A2C agent (in blue) degraded rapidly when the scale of the rewards was significantly smaller or larger than 1. Compare this to the same agent equipped with PopArt, we observe better performance across multiple orders of magnitude for the reward scales. In this tiny problem, learning could also be achieved by tuning the learning rate for each reward scale, but that does not suffice for larger problems.
Adaptive optimization algorithm such as Adam (Kingma and Ba, [2014](#bib.bib19)) or RMSProp (Tieleman and Hinton, [2012](#bib.bib35)) can also provide some degree of invariance, but, as we will see in Section [3.2](#S3.SS2 "3.2 Performance on large domains ‣ 3 Experiments ‣ On Inductive Biases in Deep Reinforcement Learning"), they are not as effective as PopArt normalization.
Finally, we investigate the role of action repeats. We consider states arranged into a simple cycle of 11 states. The agent starts in state 0, and only moves in one direction using one action, the other action does not move the agent. The reward is 0 everywhere, except if the agent selects the non-moving action in the 11−th state: in this case the agent receives a reward of 100 and the episode ends. We compare an A2C agent that learns to choose the number of action repeats (up to 10), to an agent that used a fixed number of repetitions C. Figure 1.c shows how the number of fixed action repeats used by the agent is a sensitive hyper-parameter in this domain. Compare this to the adaptive agent that learns how often to repeat actions via policy gradient (in orange in Figure 1.c). This agent quickly learned a suitable number of action repeats and thereby performed very well. This is a general problem, in many domains of interest it can be useful to combine fine-grained control in certain states, with more coarse and directed behaviour in other parts of the state space.
| | | | |
| --- | --- | --- | --- |
| Comparison of inductive biases to RL solutions. All curves show mean episode return as a function of the number of environment steps. Each plot compares the same fully general agent to 2 alternative. a) tuned action repeats, and no action repeats. b) tuned discount factor, and no discounting. c) reward clipping, and learning from raw rewards with no rescaling of the updates. d) learning from the raw observation stream of Atari, and the standard preprocessing. | Comparison of inductive biases to RL solutions. All curves show mean episode return as a function of the number of environment steps. Each plot compares the same fully general agent to 2 alternative. a) tuned action repeats, and no action repeats. b) tuned discount factor, and no discounting. c) reward clipping, and learning from raw rewards with no rescaling of the updates. d) learning from the raw observation stream of Atari, and the standard preprocessing. | Comparison of inductive biases to RL solutions. All curves show mean episode return as a function of the number of environment steps. Each plot compares the same fully general agent to 2 alternative. a) tuned action repeats, and no action repeats. b) tuned discount factor, and no discounting. c) reward clipping, and learning from raw rewards with no rescaling of the updates. d) learning from the raw observation stream of Atari, and the standard preprocessing. | Comparison of inductive biases to RL solutions. All curves show mean episode return as a function of the number of environment steps. Each plot compares the same fully general agent to 2 alternative. a) tuned action repeats, and no action repeats. b) tuned discount factor, and no discounting. c) reward clipping, and learning from raw rewards with no rescaling of the updates. d) learning from the raw observation stream of Atari, and the standard preprocessing. |
Figure 2: Comparison of inductive biases to RL solutions. All curves show mean episode return as a function of the number of environment steps. Each plot compares the same fully general agent to 2 alternative. a) tuned action repeats, and no action repeats. b) tuned discount factor, and no discounting. c) reward clipping, and learning from raw rewards with no rescaling of the updates. d) learning from the raw observation stream of Atari, and the standard preprocessing.
###
3.2 Performance on large domains
To evaluate the performance of the different methods on larger problems, we use A2C agents on many Atari games. However, differently from the previous experiments, the agent learns in parallel from multiple copies of the environment, similarly to many state-of-the-art algorithms for reinforcement learning. This configuration increases the throughput of acting and learning, and speeds up the experiments. In parallel learning training setups, the learning updates may be applied synchronously (Espeholt et al., [2018](#bib.bib9)) or asynchronously (Mnih et al., [2016](#bib.bib24)). Our learning updates are synchronous: the agent’s policy takes steps in parallel across 16 copies of the environment to create multi-step learning updates, batched together to compute a single update to the parameters. We train individual agents on each game. Per-game scores are averaged over 8 seeds, and we then track the median human normalized score across all games. All hyper-parameters for our A2C agents were selected for a generic A2C agent on Atari before the following experiments were performed, with details given in the appendix.
Our experiments measure the performance of a full adaptive A2C agent with learned action repeats, PopArt normalization, learned discount factors, and an LSTM-based state representation. We compare the performance of this agent to agents with exactly one adaptive component disabled and replaced with one of two fixed components. This fixed component is either falling back to the environment specified task (e.g. learning directly from undiscounted returns), or using the corresponding fixed heuristic from DQN. These comparisons enable us to investigate how important the original heuristic is for current RL algorithms, as well as how fully an adaptive solution can replace it.
In Figure 2a, we investigate action repeats and their impact on learning. We compare the fully general agent to an agent that used exactly 4 action repetitions (as tuned for Atari (Mnih et al., [2015](#bib.bib25))), and to an agent that acted and learned at the native frame rate of the environment. The adaptively learned solution performed almost as well as the tuned domain heuristic of always repeating each action 4 times. Interestingly, in the first 100M frames, also acting at the fastest rate was competitive with the agents equipped with action repetition (whether fixed or learned), at least in terms of data efficiency. However, while the agents with action repeats were still improving performance until the very end of the training budget, the agent acting at the fastest timescale appeared to plateau earlier.
This performance plateau is observed in multiple games (see appendix), and we speculate that the use of multiple action repetitions may be helping achieve better exploration.
We note that, in wall-clock time, the gap in the performance of the agents with action repetitions was even larger due to the additional compute.
In Figure 2b, we investigate discounting. The agent that used undiscounted returns directly in the updates to policy and values performed very poorly, demonstrating that in complex environments the naive solution of directly optimizing the real objective is problematic with modern deep RL agents. Interestingly, while performance was very poor overall, the agent did demonstrate good performance on a few specific games. For instance, in bowling it achieved a better score than state of the art agents such as Rainbow (Hessel et al., [2018a](#bib.bib12)) and ApeX (Horgan et al., [2018](#bib.bib15)). The agent with tuned discount and the agent with a discount factor learned through meta-gradient RL performed much better overall. The adaptive solution did slightly better than the heuristic.
In Figure 2c, we investigate the effect of reward scales. We compare the fully adaptive agent to an agent where clipping was used in place of PopArt, and to a naive agent that used the environment reward directly. Again, the naive solution performed very poorly, compared to using either the domain heuristic or the learned solution. Note that the naive solution is using RMSProp as an optimizer, in combination with gradient clipping by norm (Pascanu et al., [2012](#bib.bib26)); together these techniques should provide at least some robustness to scaling issues, but in our experiments PopArt provided an additional large increase in performance. In this case, the domain heuristic (reward clipping) retained a significant edge over the adaptive solution. This suggests that reward clipping might not be helping exclusively with reward scales; the inductive bias of optimizing for a weighted frequency of rewards is a very good heuristic in many Atari games, and the qualitative behaviour resulting from optimizing the proxy objective might result in a better learning dynamics. We note, in conclusion, that while clipping was better in aggregate, PopArt yielded significantly improved scores on several games (e.g., centipede) where the clipped agent was stuck in sub-optimal policies.
Finally, in Figure 2d, we compare the fully end to end pipeline with a recurrent network, to a feedforward neural network with the standard Atari pipeline. The recurrent end to end solution performed best, showing that a recurrent network is sufficiently flexible to learn on its own to integrate relevant information over time, despite the Atari-specific features of the observation stream (such as the flickering of the screen) that motivated the more common heuristic approach.
| | |
| --- | --- |
| In a separate experiment on the 28 tasks in the DeepMind Control Suite, we compared the general solution agent with an A2C agent using all the domain heuristics previously discussed. Both agents were trained and evaluated on the new domain with no changes to the algorithm nor any additional tuning for this very different set of environments. On average, the general adaptive solutions transfer better to the new domain that the heuristic solution. On the left we plot the average performance across all 28 tasks. On the right we show the learning curves on a selection of 10 tasks. | In a separate experiment on the 28 tasks in the DeepMind Control Suite, we compared the general solution agent with an A2C agent using all the domain heuristics previously discussed. Both agents were trained and evaluated on the new domain with no changes to the algorithm nor any additional tuning for this very different set of environments. On average, the general adaptive solutions transfer better to the new domain that the heuristic solution. On the left we plot the average performance across all 28 tasks. On the right we show the learning curves on a selection of 10 tasks. |
Figure 3: In a separate experiment on the 28 tasks in the DeepMind Control Suite, we compared the general solution agent with an A2C agent using all the domain heuristics previously discussed. Both agents were trained and evaluated on the new domain with no changes to the algorithm nor any additional tuning for this very different set of environments. On average, the general adaptive solutions transfer better to the new domain that the heuristic solution. On the left we plot the average performance across all 28 tasks. On the right we show the learning curves on a selection of 10 tasks.
###
3.3 Generalization to new domains
Our previous analysis shows that learned solutions are mostly quite competitive with the domain heuristics on Atari, but do not uniformly provide additional benefits compared to the well tuned inductive biases that are commonly used in Atari. To investigate the generality of these different RL solutions, in this section we compare the fully general agent to an agent with all the usual inductive biases, but this time we evaluate them on a completely different benchmark: a collection of 28 continuous control tasks in the DeepMind Control Suite. The tasks represent a wide variety of physical control problems, and the dimension of the real-valued observation and action vectors differs across the tasks. The environment state can be recovered from the observation in all but one task. The rewards are bounded between 0 and 1, and tasks are undiscounted.
Again, we use a parallel A2C implementation, with 16 copies of the environment, and we aggregate results by first averaging scores across 8 seeds, and then taking the mean across all 28 tasks. Because all tasks in this benchmark are designed to have episode returns of comparable magnitude, there is no need to normalize the results to meaningfully aggregate them. For both agents we use the exact same solutions that were used in Atari, with no additional tuning. The agents naturally transfer to this new domain with two modifications: 1) we do not use convolutions since the observations do not have spatial structure. 2) the outputs of the policy head are interpreted as encoding the mean and variance of a Gaussian distribution instead of as the logits of a categorical one.
Figure [3](#S3.F3 "Figure 3 ‣ 3.2 Performance on large domains ‣ 3 Experiments ‣ On Inductive Biases in Deep Reinforcement Learning") shows the fully general agent performed much better than the heuristic solution, which suggests that the set of inductive biases typically used by Deep RL agents on Atari do not generalize as well to this new domain as the set of adaptive solutions considered in this paper. This highlights the importance of being aware of the priors that we incorporate into our learning algorithms, and their impact on the generality of our agents. On the right side of Figure [3](#S3.F3 "Figure 3 ‣ 3.2 Performance on large domains ‣ 3 Experiments ‣ On Inductive Biases in Deep Reinforcement Learning"), we report the learning curves on the 10 tasks for which the absolute difference between the performance of the two agents was greatest (details on the full set of 28 tasks can be found in Appendix). The adaptive solutions performed equal or better than the heuristics on each of these 10 tasks, and the results in the appendix show performance was rarely worse. The reference horizontal black lines mark the performance of an A3C agent, tuned specifically for this suite of tasks, as reported by Tassa et al. ([2018](#bib.bib34)). The adaptive solution was also better, in aggregate, than this well tuned baseline; note however the tuned A3C agent achieved higher performance on a few games.
4 Related Work and Discussion
------------------------------
The present work was partially inspired by the work of Silver et al. ([2017](#bib.bib27)) in the context of Go. They demonstrated that specific domain specific heuristics (e.g. pretraining on human data, the use of handcrafted Go-specific features, and exploitation of certain symmetries in state space), while originally introduced to simplify learning (Silver et al., [2016](#bib.bib28)), had actually outlived their usefulness: taking a *purer* approach, even stronger Go agents could be trained. Importantly, they showed removing these domain heuristics, the same algorithm could master other games, such as Shogi and Chess. In our paper, we adopted a similar philosophy but investigated the very different set of domain specific heuristics, that are used in more traditional deep reinforcement learning agents.
Our work relates to a broader debate (Marcus, [2018](#bib.bib23)) about priors and innateness. There is evidence that we, as humans, posses specific types of biases, and that these have a role in enabling efficient learning (Spelke and Kinzler, [2007](#bib.bib29); Dubey et al., [2018](#bib.bib8)); however, it is far from clear the extent to which these are essential for intelligent behaviour to arise, what form these priors take, and their impact on the generality of the resulting solutions. In this paper, we demonstrate that several heuristics we commonly use in our algorithms already harm the generality of our methods. This does not mean that other different inductive biases could not be useful as we progress towards more flexible, intelligent agents; it is however a reminder to be careful with the domain knowledge and priors we bake into our solutions, and to be prepared to revise them over time.
We found that existing learned solutions are competitive with well tuned domain heuristics, even on the domain these heuristics were designed for, and they seem to generalize better to unseen domain. This makes a case for removing these biases in future research on Atari, since they are not essential for competitive performance, and they might hide issues in the core learning algorithm. The only case where we still found a significant gap in favour of the domain heuristic was in the case of clipping. While, to the best of our knowledge, PopArt does address scaling issues effectively, clipping still seems to help on several games. Changing the reward distribution has many subtle implications for the learning dynamics, beside affecting the magnitude of updates (e.g. exploration, risk-propensity, …). We leave to future research to investigate what other general solutions could be deployed in our agents to fully recover the observed benefits of clipping.
Several of the biases that we considered have *knobs* that could be tuned rather than learned (e.g. the discount, the number of repeats, etc); however, this is not a satisfying solution for several reasons. Tuning is expensive, and these heuristics interact subtly with each other, thus requiring an exploration of a combinatorial space to find suitable settings. Consider the use of fixed action repeats: when changing the number of repetitions you also need to change the discount factor, otherwise this will change the effective horizon of the agent, which in turn affects the magnitude of the returns and therefore the learning rate. Also, a fixed tuned value might still not give you the full benefits of an adaptive learned approach that can adapt to the various phases of the training process.
There are other two features of our algorithm that, despite not incorporating quite as much domain knowledge as the heuristics discussed in this paper, also constitute a potential impediment to its generality and scalability. 1) The use of parallel environments is not always feasible in practice, especially in real world applications (although recent work on robot farms Levine et al. ([2016](#bib.bib21)) shows that it might still be a valid approach when sufficient resources are available). 2) The use of back-propagation through time for training recurrent state representations constrains the length of the temporal relationship that we can learn, since the memory consumption is linear in the length of the rollouts. Further work in overcoming these limitations, successfully learning online from a single stream of experience, is a fruitful direction for future research. |
afe1fa15-0ec7-45ba-bddd-3244c202f551 | trentmkelly/LessWrong-43k | LessWrong | VARIABILITY OF NUCLEAR DECAY RATES
I came across a utube vid about this is 08 or so, a talk at a convention, and it intrigued me, but didn't know anyone was working on studying it. The guy set up a counter in the basement, and was tracking Cesium decay, to compare to a Web clock, IIRC. Can't find it now, topic is swamped, and duckgo swamped with creationist blogs, i guess they like the idea of radiodating flaws.
After many dueling papers, it appears the big neutrino sensor in Japan has seen the same, seasonal variability in their data too, so appearing more legit.
http://news.stanford.edu/2016/11/09/stanford-solar-physicist-unlocks-easier-way-observe-peculiar-particles-reveal-inner-workings-sun/
http://link.springer.com/article/10.1007%2Fs11207-016-1008-9
MIT collection
VARIABILITY OF NUCLEAR DECAY RATES
Original paper:
http://www.sciencedirect.com/science/article/pii/S092765050900084X
https://arxiv.org/abs/0808.3283
"Unexplained periodic fluctuations in the decay rates of 32Si and 226Ra have been reported by groups at Brookhaven National Laboratory (32Si), and at the Physikalisch–Technische–Bundesanstalt in Germany (226Ra). We show from an analysis of the raw data in these experiments that the observed fluctuations are strongly correlated in time, not only with each other, but also with the time of year."
and a review article, from the old Analog magazine!
Radioactive Decay and the Earth-Sun Distance
https://www.npl.washington.edu/AV/altvw147.html
This paper consolidates the data on six of the studies, while taking some simple, EM measurements of the Plank length variation that seems to coincide fairly closely with the decay rate data. Both are trailing the actual distance measurements by up to a month, which is pretty strange in its own right.
pdf
http%3A%2F%2Ffile.scirp.org%2Fpdf%2FOPJ_2016063013301299.pdf
Along with the Plank variation, lots of other explanations have been floated, my favorite being the link to a possible Dark Matter reservoir, that may reside in the S |
1ff1cea1-d7a8-4b3c-a2c5-8cd6e037727e | trentmkelly/LessWrong-43k | LessWrong | How to bet against civilizational adequacy?
Coal prices are at historical highs (2x to 4x normal prices depending on the kind of coal), but coal miner stocks are not. They're trading at historically low multiples, around 1x-2x spot FCF, meaning they can make their enterprise value in less than 2 years worth of after-tax profits, assuming coal prices stay where they are. So the market apparently "believes" that high coal prices won't last. (The low multiples are also because due to ESG concerns on the part of their investors, many funds can't invest in coal stocks without jeopardizing their AUM.)
By going long coal stocks, you can implicitly bet that 1) in the short run, the war between Russia and Ukraine and the associated sanctions and trade disruptions will continue (reduced energy exports from Russia is the main cause of the current high coal prices), 2) supply of (non-Russian) energy will not respond much to higher prices, and 3) in the longer run, humanity will have a harder time transitioning away from burning coal for energy, or using coal to make steel and cement, than the market thinks.
So (it occurred to me) a bet on coal is also a bet against civilizational adequacy. What are some other such bets one can make, that potentially have good risk/reward?
P.S., I think coal can also double as a hedge against AI takeoff (meaning it's likely to appreciate or at least preserve its value in that scenario). (I mean a soft AI takeoff. In a hard takeoff probably all bets are off.) Consider, if GDP was doubling every few years due to AI-driven economic activity, what's likely to happen to demand for electricity and steel, and how quickly can supply respond? What are some other "unconventional" AI hedges?
(If anyone wants to actually do this, please do your own research! Aside from having an informed view on points 1-3 above, you should also understand the different types of coal, grades of coal, coal blending, met/thermal substitution, coal/gas/oil substitution, trade routes, transport limitations, contract |
5016ea80-d213-4887-977d-74928d402b09 | trentmkelly/LessWrong-43k | LessWrong | Melatonin: Much More Than You Wanted To Know
[I am not a sleep specialist. Please consult with one before making any drastic changes or trying to treat anything serious.]
Van Geijlswijk et al describe supplemental melatonin as “a chronobiotic drug with hypnotic properties”. Using it as a pure hypnotic – a sleeping pill – is like using an AK-47 as a club to bash your enemies’ heads in. It might work, but you’re failing to appreciate the full power and subtlety available to you.
Melatonin is a neurohormone produced by the pineal gland. In a normal circadian cycle, it’s lowest (undetectable, less than 1 pg/ml of blood) around the time you wake up, and stays low throughout the day. Around fifteen hours after waking, your melatonin suddenly shoots up to 10 pg/ml – a process called “dim light melatonin onset”. For the next few hours, melatonin continues to increase, maybe as high as 60 or 70 pg/ml, making you sleepier and sleepier, and presumably at some point you go to bed. Melatonin peaks around 3 AM, then declines until it’s undetectably low again around early morning.
Is this what makes you sleepy? Yes and no. Sleepiness is a combination of the circadian cycle and the so-called “Process S”. This is an unnecessarily sinister-sounding name for the fact that the longer you’ve been awake, the sleepier you’ll be. It seems to be partly regulated by a molecule called adenosine. While you’re awake, the body produces adenosine, which makes you tired; as you sleep, the body clears adenosine away, making you feel well-rested again.
In healthy people these processes work together. Circadian rhythm tells you to feel sleepy at night and awake during the day. Process S tells you to feel awake when you’ve just risen from sleep (naturally the morning), and tired when you haven’t slept in a long time (naturally the night). Both processes agree that you should feel awake during the day and tired at night, so you do.
When these processes disagree for some reason – night shifts, jet lag, drugs, genetics, playing Civilization un |
a063d05a-ea6b-4f4d-bcc5-4bb2049cdd48 | trentmkelly/LessWrong-43k | LessWrong | Does agent foundations cover all future ML systems?
Epistemic Status: Wondering over meta approaches to reasoning about agent foundations
In mathematics, we can find general principles for mathematical objects with a particular set of axioms. For example, if a language follows first-order logic, we can assume that saying the sentence "all apples are made of carbon" is the same as saying "if something isn't made of carbon, it isn't an apple". Similarly, I believe that agent foundations are an attempt to construct an abstract algebra on the algebra that is machine learning. We want to be able to say something about all future ML models by arguing over the potential compositions of ML agents. This would be like saying, "since AGI-agent 1 is a subset of an agent class that follows these axioms, we know that it will only defect if it sees a paperclip".
Someone pointed out to me that we might be assuming things about how an agent acts by not grounding it in current-day machine learning algorithms. The problem with not grounding it is that we might be constructing an abstract algebra with axioms that don't encompass the axioms in the algebra we're trying to study. I thought this was a great question, and I have no idea how to answer it since I haven't seen any formalisation on what ML approaches are a sub-set of. Take, for example, John Wentworth's work on agent foundations. Does it generalise to multi-agent-based systems? Does it generalise to self-supervision algorithms such as current-day transformers? I would love to know if anyone has thought of this.
I also have some other related questions:
Firstly, do you, fellow humans of LessWrong, believe that the abstract algebra framework is helpful for thinking about agent foundations?
Secondly, I wondered whether anyone has made such an approach or knows what modern-day ML algorithms different parts of agent foundations covers? |
082a53fd-8ee8-4fe1-978e-472dbcb71afd | trentmkelly/LessWrong-43k | LessWrong | Could humanity ever achieve atomically precise manufacturing (APM)? What about a much-smarter-than-human-level intelligence?
None |
de38647b-b36d-47a4-8f0d-da4663f42304 | trentmkelly/LessWrong-43k | LessWrong | o3-mini Early Days
New model, new hype cycle, who dis?
On a Friday afternoon, OpenAI was proud to announce the new model o3-mini and also o3-mini-high which is somewhat less mini, or for some other reasoning tasks you might still want o1 if you want a broader knowledge base, or if you’re a pro user o1-pro, while we want for o3-not-mini and o3-pro, except o3 can use web search and o1 can’t so it has the better knowledge in that sense, then on a Sunday night they launched Deep Research which is different from Google’s Deep Research but you only have a few of those queries so make them count, or maybe you want to use operator?
Get it? Got it? Good.
Yes, Pliny jailbroke o3-mini on the spot, as he always does.
This most mostly skips over OpenAI’s Deep Research (o3-DR? OAI-DR?). I need more time for that. I’ll cover o3-DR properly later in the week once we have a chance to learn what we’ve got there, along with the non-DR ‘one more thing’ Altman is promising. So far it looks super exciting, but it’s a very different class of product.
TABLE OF CONTENTS
1. Feature Presentation.
2. Q&A.
3. The Wrong Side of History.
4. The System Card.
5. The Official Benchmarks.
6. The Unofficial Benchmarks.
7. Others Report In.
8. Some People Need Practical Advice.
FEATURE PRESENTATION
What exactly can o3-mini do?
> OpenAI: We’re releasing OpenAI o3-mini, the newest, most cost-efficient model in our reasoning series, available in both ChatGPT and the API today. Previewed in December 2024, this powerful and fast model advances the boundaries of what small models can achieve, delivering exceptional STEM capabilities—with particular strength in science, math, and coding—all while maintaining the low cost and reduced latency of OpenAI o1-mini.
>
> OpenAI o3-mini is our first small reasoning model that supports highly requested developer features including function calling(opens in a new window), Structured Outputs(opens in a new window), and developer messages(opens in a new window), makin |
5e19a6dd-2d8b-4bf9-8499-373120600859 | trentmkelly/LessWrong-43k | LessWrong | Getting started with AI Alignment research: how to reproduce an experiment from research paper
This is a post with technical instructions, how to reproduce an experiment from Weak-to-strong generalization paper: https://openai.com/index/weak-to-strong-generalization/. It’s oriented mostly on beginners in AI Alignment who want to start tinkering with models and looking for examples how to do experiments.
Weak-to-strong generalization is research that shows that a strong model can learn on data generated by a weaker model, generalize the data and surpass the weaker model in the task for which it was trained. The paper comes with example code on GitHub with experiments both on LLMs and vision models. However, running the experiments from this code is not a straightforward task, so here are detailed instructions how to do it.
Setup
* Find a GPU cloud provider that gives access to terminal and Jupyter notebook. I used runpod.io for my experiment, selected a node with 1 RTX A6000 graphics card with 48 GB VRAM. The main limiting factor for the most of the experiments is VRAM size, so choose your node based on it and on the price; other characteristics are less important. Also, make sure that disk size of your node is at least 60 GB. Most of the cloud providers allow increasing disk size in settings, so do it if the disk is too small.
* Register the account, rent a node, and follow cloud provider’s instructions to connect to it with terminal and Jupyter notebook.
* Go to terminal and clone the repository:
git clone https://github.com/openai/weak-to-strong
cd weak-to-strong
* I recommend to use virtual terminal, such as tmux or screen: it will ensure that you will not lose your run if the connection to server will drop in the middle of an experiment. If the server uses Ubuntu or Debian, run commands:
apt-get install tmux
tmux
* If the connection will drop, reconnect to the server and run the command tmux attach to get back to your experiment. To scroll up and down in tmux, use Ctrl-B, [ keys sequence, then scroll up and down with arrows. Press Esc to exit |
94f8bf2d-4a45-473a-aff4-2b9035c501b0 | trentmkelly/LessWrong-43k | LessWrong | Effective Altruism vs Missionaries? Advice Requested from a Newly-Built Crowdfunding Platform.
Hi, I'm developing a next-generation crowdfunding platform for non-profit fundraising. From what we have seen, it is aeffective tool, more about it below. I'm working with two other cofounders, both of whom are evangelical Christians. We get along well in general, but that I strongly believe in effective altruism and they do not.
We will launch a second pilot fundraising campaign in 2-3 weeks. My co-founders have arranged for us fund raise for is a "church planting" missionary organization. This is so opposed my belief in effective altruism I feel uncomfortable using our effective tool to funnel donors' dollars in THIS of all directions. This is not the reason I got involved in this project.
My argument with them is that we should charge more to ineffective nonprofits such as colleges, religious, or political organizations, and use that extra to subsidize the campaign and money-processing costs of the effective non-profits. I think this is logically consistent with earning to give. But I am being outvoted two-to-one by people who believe saving lives and saving souls are nearly equally important.
So I have two requests:
1. If anyone has advise on how to navigate this (including any especially well written articles that would appeal to evangelical Christians, or experience negotiating with start-up cofounders).
2. If anyone has personal connections with effective or effective-ish non-profits, I would much prefer to fundraise for them than my co-founder's church connections. Caveat: the org must have US non-profit legal status.
About the platform: the gist our concept is that we're using a lot of psychology and biases and altruism research to nudge more people towards giving and also nudge them towards a sustained involvement with the nonprofit in question. We're using some of the tricks that made the ice bucket challenge so successful (but with added accountability to ensure that visible involvement actually leads to monetary donations). Users can pledge mone |
87a1d5c8-1abd-4748-aed0-d2707a96b930 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Neural Modular Control for Embodied Question Answering
1 Introduction
---------------
Abstraction is an essential tool for navigating our daily lives.
When seeking a late night snack, we certainly do not spend time
planning out the mechanics of walking and are thankfully also
unburdened of the effort of recalling to beat our heart along the
way. Instead, we conceptualize our actions as a series of higher-level semantic goals
– exit bedroom; go to kitchen; open fridge; find snack;
– each of which is executed through specialized coordination of
our perceptual and sensorimotor skills. This
ability to abstract long, complex sequences of actions into semantically
meaningful subgoals is a key component of human cognition [[3](#bib.bib3)] and it is natural to
believe that artificial agents can benefit from applying similar mechanisms when navigating our world.
We study such hierarchical control in the context of a recently proposed task –
Embodied Question Answering (EmbodiedQA) [[1](#bib.bib1)] – where an embodied agent is spawned
at a random location in a novel environment (*e.g*. a house) and asked to answer a question
(*‘What color is the piano in the living room?’*).
To do so, the agent must navigate from egocentric vision alone (without access
to a map of the environment), locate the entity in question (*‘piano in the living room’*),
and respond with the correct answer (*e.g*. *‘red’*). From a reinforcement learning (RL) perspective, EmbodiedQA
presents challenges that are known to make learning particularly difficult –
partial observability, planning over long time horizons, and sparse rewards – the
agent may have to navigate through multiple rooms in search for the
answer, executing hundreds of primitive motion actions
along the way (forward; forward; turn-right; …) and receiving a reward based only on its final answer.
Figure 1: We introduce a hierarchical policy for Embodied Question Answering.
Given a question (“What color is the sofa in the living room?”) and observation,
our master policy predicts a sequence of subgoals – Exit-room,
Find-room[living], Find-object[sofa], Answer
– that are then executed by specialized sub-policies to navigate to the target
object and answer the question (“Grey”).
To address this challenging learning problem, we develop a hierarchical Neural Modular Controller (NMC) –
consisting of a *master* policy that determines high-level *subgoals*,
and *sub-policies* that execute a series of low-level actions to achieve these subgoals.
Our NMC model constructs a hierarchy that is arguably natural to this problem –
navigation to rooms and objects *vs*. low-level motion actions.
For example, NMC seeks to break down
a question *‘What color is the piano in the living room?’* to the series of subgoals exit-room; find-room[*living*]; find-object[*piano*]; answer;
and execute this plan with specialized neural ‘modules’
corresponding to each subgoal. Each module is trained to issue
a variable length series of primitive actions to achieve its titular subgoal – *e.g*. the find-object[*piano*]
module is trained to navigate the agent to the input argument *piano* within the current room.
Disentangling semantic subgoal selection from sub-policy execution results in
easier to train models due to shorter time horizons. Specifically, this hierarchical structure introduces:
1. –
Compressed Time Horizons:
The master policy makes orders of magnitude fewer decisions over the course of a
navigation than a *‘flat model’* that directly predicts primitive actions – allowing
the answering reward in EmbodiedQA to more easily influence high-level motor control decisions.
2. –
Modular Pretraining:
As each module corresponds to a specific task, they can be trained independently before
being combined with the master policy. Likewise, the master policy can be trained
assuming ideal modules. We do this through imitation learning [[4](#bib.bib4), [5](#bib.bib5)]
sub-policies.
3. –
Interpretability: The predictions made by the master policy correspond to
semantic subgoals and exposes the reasoning of the agent to inspection
(*‘What is the agent trying to do right now?’*) in a significantly more interpretable
fashion than just its primitive actions.
First, we learn and evaluate master and sub-policies for each of our subgoals, trained using
behavior cloning on expert trajectories, reinforcement learning from scratch, and
reinforcement learning after behavior cloning. We find that reinforcement learning
after behavior cloning dramatically improves performance over each individual training regime.
We then evaluate our combined hierarchical approach on the EQA [[1](#bib.bib1)] benchmark in House3D [[2](#bib.bib2)] environments.
Our approach significantly outperforms prior work both in navigational and question answering performance –
our agent is able to navigate closer to the target object
and is able to answer questions correctly more often.
2 Related Work
---------------
Our work builds on and is related to prior work in hierarchical reinforcement and imitation learning,
grounded language learning, and embodied question-answering agents in simulated environments.
Hierarchical Reinforcement and Imitation Learning.
Our formulation is closely related to Le *et al*. [[6](#bib.bib6)],
and can be seen as an instantiation of the options framework [[7](#bib.bib7), [8](#bib.bib8)],
wherein a global master policy proposes subgoals – to be achieved by local sub-policies –
towards a downstream task objective [[9](#bib.bib9), [10](#bib.bib10), [11](#bib.bib11)].
Relative to other work on automatic subgoal discovery in hierarchical
reinforcement learning [[12](#bib.bib12), [13](#bib.bib13), [14](#bib.bib14)],
we show that given knowledge of the problem structure, simple heuristics are quite
effective in breaking down long-range planning into sequential subgoals.
We make use of a combination of hierarchical behavior cloning [[4](#bib.bib4)]
and actor-critic [[15](#bib.bib15)] to train our modular policy.
Neural Module Networks and Policy Sketches.
At a conceptual-level, our work is analogous to recent work on
neural module networks (NMNs) [[16](#bib.bib16), [17](#bib.bib17), [18](#bib.bib18)]
for visual question answering. NMNs first predict a ‘program’ from the question,
consisting of a sequence of primitive reasoning steps, which are then executed on the image to
obtain the answer.
Unlike NMNs, where each primitive reasoning module has access to the entire image (completely observable)
our setting is partially observable – each sub-policy only has access to first-person RGB –
making active re-evaluation of subgoals after executing each sub-policy essential.
Our work is also closely related to policy sketches [[16](#bib.bib16)], which
are symbolic descriptions of subgoals provided to the agent without any grounding or sub-policy for executing them.
There are two key differences w.r.t. to our work.
First, an important framework difference – Andreas *et al*. [[16](#bib.bib16)] assume
access to a policy sketch *at test time*, *i.e*. for every task to be performed. In EmbodiedQA, this would correspond
to the agent being provided with a high-level plan (exit-room; find-room[*living*]; …) for every question
it is ever asked, which is an unrealistic assumption in real-world scenarios with a robot. In contrast, we assume that
subgoal supervision (in the form of expert demonstrations and plans)
are available on training environments but not on test,
and the agent must *learn* to produce its own subgoals. Second, a subtle but important implementation difference –
unlike [[16](#bib.bib16)], our sub-policy modules accept input arguments that are embeddings of target rooms and objects
(*e.g*.find-room[*living*], find-object[*piano*]).
This results in our sub-policy modules being shared not just across tasks (questions) as in [[16](#bib.bib16)], but also
across instantiations of *similar* navigation sub-policies – *i.e*., find-object[*piano*] and
find-object[*chair*] share parameters that enable data efficient learning
without exhaustively learning separate policies for each.
Grounded Language Learning.
Beginning with SHRDLU [[19](#bib.bib19)], there has been a rich progression
of work in grounding language-based goal specifications into actions and pixels
in physically-simulated environments.
Recent deep reinforcement learning-based approaches to this
explore it in 2D gridworlds [[16](#bib.bib16), [20](#bib.bib20), [21](#bib.bib21)],
simple visual [[22](#bib.bib22), [23](#bib.bib23), [24](#bib.bib24), [25](#bib.bib25), [26](#bib.bib26), [27](#bib.bib27)]
and textual [[28](#bib.bib28), [29](#bib.bib29)] environments,
perceptually-realistic 3D home simulators [[1](#bib.bib1), [30](#bib.bib30), [31](#bib.bib31), [32](#bib.bib32), [33](#bib.bib33)],
as well as real indoor scenes [[34](#bib.bib34), [35](#bib.bib35), [36](#bib.bib36)].
Our hierarchical policy learns to ground words from the question into
two levels of hierarchical semantics. The master policy grounds words into subgoals (such as find-room[*kitchen*]), and sub-policies ground these semantic targets
(such as *cutting board*, *bathroom*) into primitive actions and raw pixels,
both parameterized as neural control policies and trained end-to-end.
Embodied Question-Answering Agents. Finally, hierarchical policies for
embodied question answering have previously been proposed by
Das *et al*. [[1](#bib.bib1)] in the House3D environment [[2](#bib.bib2)], and by
Gordon *et al*. [[30](#bib.bib30)] in the AI2-THOR environment [[37](#bib.bib37)].
Our hierarchical policy, in comparison, is human-interpretable,
*i.e*. the subgoal being pursued at every step of navigation is semantic,
and due to the modular structure, can navigate over longer paths than prior work,
spanning multiple rooms.
3 Neural Modular Control
-------------------------
We now describe our approach in detail.
Recall that given a question, the goal of our agent is to predict
a sequence of navigation subgoals and execute them to ultimately find the target object
and respond with the correct answer.
We first present our modular hierarchical policy.
We then describe how we extract optimal plans from shortest
path navigation trajectories for behavior cloning. And finally, we describe how
the various modules are combined and trained with a
combination of imitation learning (behavior cloning) and reinforcement learning.
###
3.1 Hierarchical Policy
Notation.
Recall that NMC has 2 levels in the hierarchy – a master policy that generates subgoals and
sub-policies for each of these subgoals. We use i𝑖iitalic\_i to index the sequence of subgoals and t𝑡titalic\_t to index actions generated by sub-policies.
Let 𝒮={s}𝒮𝑠\mathcal{S}=\{s\}caligraphic\_S = { italic\_s } denote the set of states, 𝒢={g}𝒢𝑔\mathcal{G}=\{g\}caligraphic\_G = { italic\_g } the set of variable-time
subgoals with elements g=⟨gtask,gargument⟩𝑔subscript𝑔tasksubscript𝑔argumentg=\langle g\_{\text{task}},g\_{\text{argument}}\rangleitalic\_g = ⟨ italic\_g start\_POSTSUBSCRIPT task end\_POSTSUBSCRIPT , italic\_g start\_POSTSUBSCRIPT argument end\_POSTSUBSCRIPT ⟩,
*e.g*. g=⟨g=\langleitalic\_g = ⟨exit-room,*None*⟩normal-⟩\rangle⟩, or g=⟨g=\langleitalic\_g = ⟨find-room,*bedroom*⟩normal-⟩\rangle⟩.
Let 𝒜={a}𝒜𝑎\mathcal{A}=\{a\}caligraphic\_A = { italic\_a } be the set of primitive actions (forward, turn-left, turn-right).
The learning problem can then be succinctly put as learning a master policy πθ:𝒮→𝒢:subscript𝜋𝜃→𝒮𝒢\pi\_{\theta}:\mathcal{S}\rightarrow\mathcal{G}italic\_π start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT : caligraphic\_S → caligraphic\_G
parameterized by θ𝜃\thetaitalic\_θ and sub-policies πϕg:𝒮→𝒜∪{𝚜𝚝𝚘𝚙}:subscript𝜋subscriptitalic-ϕ𝑔→𝒮𝒜𝚜𝚝𝚘𝚙\pi\_{\phi\_{g}}:\mathcal{S}\rightarrow\mathcal{A}\cup\{\text{{stop}}\}italic\_π start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT italic\_g end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT : caligraphic\_S → caligraphic\_A ∪ { stop }
parameterized by ϕg,∀g∈𝒢subscriptitalic-ϕ𝑔for-all𝑔
𝒢\phi\_{g},\,\forall g\in\mathcal{G}italic\_ϕ start\_POSTSUBSCRIPT italic\_g end\_POSTSUBSCRIPT , ∀ italic\_g ∈ caligraphic\_G, where
the stop action terminates a sub-policy and returns
control to the master policy.
While navigating an environment, control alternates between the master policy selecting subgoals and sub-policies executing these goals through a series of primitive actions. More formally, given an initial state s0subscript𝑠0s\_{0}italic\_s start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT the master
policy predicts a subgoal g0∼πθ(g|s0)similar-tosubscript𝑔0subscript𝜋𝜃conditional𝑔subscript𝑠0g\_{0}\sim\pi\_{\theta}(g|s\_{0})italic\_g start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ∼ italic\_π start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT ( italic\_g | italic\_s start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT ), the corresponding sub-policy executes until
some time T0subscript𝑇0{T}\_{0}italic\_T start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT
when either (1) the sub-policy terminates itself by producing the stop token
aT0∼πϕg0(a|sT0)=𝚜𝚝𝚘𝚙similar-tosubscript𝑎subscript𝑇0subscript𝜋subscriptitalic-ϕsubscript𝑔0conditional𝑎subscript𝑠subscript𝑇0𝚜𝚝𝚘𝚙a\_{T\_{0}}\sim\pi\_{\phi\_{g\_{0}}}(a|s\_{T\_{0}})=\text{{stop}}italic\_a start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ∼ italic\_π start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT italic\_g start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_a | italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ) = stop or (2) a maximum number of primitive actions has been reached.
Either way, this returns the control back to the master policy
which predicts another subgoal and repeats this process until termination. This results in a
state-subgoal trajectory:
| | | | |
| --- | --- | --- | --- |
| | Σ=(s0,g0\@mathmeasures0, g0\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasuresubgoal 0,sT0,g1\@mathmeasuresT0, g1\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasuresubgoal 1,…,sTi,gi+1\@mathmeasuresTi, gi+1\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasuresubgoali,…,sT𝒯−1,g𝒯\@mathmeasuresTT-1, gT\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasuresubgoal𝒯)Σsubscriptfragmentss0,g0fragments\@mathmeasures0, g0\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasuresubgoal 0subscriptfragmentsssubscript𝑇0,g1fragments\@mathmeasuresT0, g1\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasuresubgoal 1…subscriptfragmentsssubscript𝑇𝑖,g𝑖1fragments\@mathmeasuresTi, gi+1\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasuresubgoal𝑖…subscriptfragmentsssubscript𝑇𝒯1,g𝒯fragments\@mathmeasuresTT-1, gT\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasuresubgoal𝒯\displaystyle\Sigma=\bigg{(}\mathop{\mathchoice{\vtop{\halign{#\cr$\hfil\displaystyle s\_{0},g\_{0}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\displaystyle{s\_{0}, g\_{0}}\@mathmeasure\displaystyle{\upbrace}\@mathmeasure\displaystyle{\upbraceg}\@mathmeasure\displaystyle{\upbracegg}\@mathmeasure\displaystyle{\upbraceggg}\@mathmeasure\displaystyle{\upbracegggg}$\displaystyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\textstyle s\_{0},g\_{0}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\textstyle{s\_{0}, g\_{0}}\@mathmeasure\textstyle{\upbrace}\@mathmeasure\textstyle{\upbraceg}\@mathmeasure\textstyle{\upbracegg}\@mathmeasure\textstyle{\upbraceggg}\@mathmeasure\textstyle{\upbracegggg}$\textstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptstyle s\_{0},g\_{0}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptstyle{s\_{0}, g\_{0}}\@mathmeasure\scriptstyle{\upbrace}\@mathmeasure\scriptstyle{\upbraceg}\@mathmeasure\scriptstyle{\upbracegg}\@mathmeasure\scriptstyle{\upbraceggg}\@mathmeasure\scriptstyle{\upbracegggg}$\scriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptscriptstyle s\_{0},g\_{0}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptscriptstyle{s\_{0}, g\_{0}}\@mathmeasure\scriptscriptstyle{\upbrace}\@mathmeasure\scriptscriptstyle{\upbraceg}\@mathmeasure\scriptscriptstyle{\upbracegg}\@mathmeasure\scriptscriptstyle{\upbraceggg}\@mathmeasure\scriptscriptstyle{\upbracegggg}$\scriptscriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}}\limits\_{\text{subgoal 0}},\mathop{\mathchoice{\vtop{\halign{#\cr$\hfil\displaystyle s\_{T\_{0}},g\_{1}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\displaystyle{s\_{T\_{0}}, g\_{1}}\@mathmeasure\displaystyle{\upbrace}\@mathmeasure\displaystyle{\upbraceg}\@mathmeasure\displaystyle{\upbracegg}\@mathmeasure\displaystyle{\upbraceggg}\@mathmeasure\displaystyle{\upbracegggg}$\displaystyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\textstyle s\_{T\_{0}},g\_{1}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\textstyle{s\_{T\_{0}}, g\_{1}}\@mathmeasure\textstyle{\upbrace}\@mathmeasure\textstyle{\upbraceg}\@mathmeasure\textstyle{\upbracegg}\@mathmeasure\textstyle{\upbraceggg}\@mathmeasure\textstyle{\upbracegggg}$\textstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptstyle s\_{T\_{0}},g\_{1}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptstyle{s\_{T\_{0}}, g\_{1}}\@mathmeasure\scriptstyle{\upbrace}\@mathmeasure\scriptstyle{\upbraceg}\@mathmeasure\scriptstyle{\upbracegg}\@mathmeasure\scriptstyle{\upbraceggg}\@mathmeasure\scriptstyle{\upbracegggg}$\scriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptscriptstyle s\_{T\_{0}},g\_{1}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptscriptstyle{s\_{T\_{0}}, g\_{1}}\@mathmeasure\scriptscriptstyle{\upbrace}\@mathmeasure\scriptscriptstyle{\upbraceg}\@mathmeasure\scriptscriptstyle{\upbracegg}\@mathmeasure\scriptscriptstyle{\upbraceggg}\@mathmeasure\scriptscriptstyle{\upbracegggg}$\scriptscriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}}\limits\_{\text{subgoal 1}},\ldots,\mathop{\mathchoice{\vtop{\halign{#\cr$\hfil\displaystyle s\_{T\_{i}},g\_{i+1}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\displaystyle{s\_{T\_{i}}, g\_{i+1}}\@mathmeasure\displaystyle{\upbrace}\@mathmeasure\displaystyle{\upbraceg}\@mathmeasure\displaystyle{\upbracegg}\@mathmeasure\displaystyle{\upbraceggg}\@mathmeasure\displaystyle{\upbracegggg}$\displaystyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\textstyle s\_{T\_{i}},g\_{i+1}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\textstyle{s\_{T\_{i}}, g\_{i+1}}\@mathmeasure\textstyle{\upbrace}\@mathmeasure\textstyle{\upbraceg}\@mathmeasure\textstyle{\upbracegg}\@mathmeasure\textstyle{\upbraceggg}\@mathmeasure\textstyle{\upbracegggg}$\textstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptstyle s\_{T\_{i}},g\_{i+1}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptstyle{s\_{T\_{i}}, g\_{i+1}}\@mathmeasure\scriptstyle{\upbrace}\@mathmeasure\scriptstyle{\upbraceg}\@mathmeasure\scriptstyle{\upbracegg}\@mathmeasure\scriptstyle{\upbraceggg}\@mathmeasure\scriptstyle{\upbracegggg}$\scriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptscriptstyle s\_{T\_{i}},g\_{i+1}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptscriptstyle{s\_{T\_{i}}, g\_{i+1}}\@mathmeasure\scriptscriptstyle{\upbrace}\@mathmeasure\scriptscriptstyle{\upbraceg}\@mathmeasure\scriptscriptstyle{\upbracegg}\@mathmeasure\scriptscriptstyle{\upbraceggg}\@mathmeasure\scriptscriptstyle{\upbracegggg}$\scriptscriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}}\limits\_{\text{subgoal}\,\,i},\ldots,\mathop{\mathchoice{\vtop{\halign{#\cr$\hfil\displaystyle s\_{T\_{\mathcal{T}-1}},g\_{\mathcal{T}}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\displaystyle{s\_{T\_{\mathcal{T}-1}}, g\_{\mathcal{T}}}\@mathmeasure\displaystyle{\upbrace}\@mathmeasure\displaystyle{\upbraceg}\@mathmeasure\displaystyle{\upbracegg}\@mathmeasure\displaystyle{\upbraceggg}\@mathmeasure\displaystyle{\upbracegggg}$\displaystyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\textstyle s\_{T\_{\mathcal{T}-1}},g\_{\mathcal{T}}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\textstyle{s\_{T\_{\mathcal{T}-1}}, g\_{\mathcal{T}}}\@mathmeasure\textstyle{\upbrace}\@mathmeasure\textstyle{\upbraceg}\@mathmeasure\textstyle{\upbracegg}\@mathmeasure\textstyle{\upbraceggg}\@mathmeasure\textstyle{\upbracegggg}$\textstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptstyle s\_{T\_{\mathcal{T}-1}},g\_{\mathcal{T}}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptstyle{s\_{T\_{\mathcal{T}-1}}, g\_{\mathcal{T}}}\@mathmeasure\scriptstyle{\upbrace}\@mathmeasure\scriptstyle{\upbraceg}\@mathmeasure\scriptstyle{\upbracegg}\@mathmeasure\scriptstyle{\upbraceggg}\@mathmeasure\scriptstyle{\upbracegggg}$\scriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptscriptstyle s\_{T\_{\mathcal{T}-1}},g\_{\mathcal{T}}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptscriptstyle{s\_{T\_{\mathcal{T}-1}}, g\_{\mathcal{T}}}\@mathmeasure\scriptscriptstyle{\upbrace}\@mathmeasure\scriptscriptstyle{\upbraceg}\@mathmeasure\scriptscriptstyle{\upbracegg}\@mathmeasure\scriptscriptstyle{\upbraceggg}\@mathmeasure\scriptscriptstyle{\upbracegggg}$\scriptscriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}}\limits\_{\text{subgoal}\,\,\mathcal{T}}\bigg{)}roman\_Σ = ( start\_BIGOP start\_ROW start\_CELL italic\_s start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT , italic\_g start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT end\_CELL end\_ROW start\_ROW start\_CELL s start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT , g start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT end\_CELL end\_ROW end\_BIGOP start\_POSTSUBSCRIPT subgoal 0 end\_POSTSUBSCRIPT , start\_BIGOP start\_ROW start\_CELL italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT , italic\_g start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_CELL end\_ROW start\_ROW start\_CELL s start\_POSTSUBSCRIPT T start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT , g start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_CELL end\_ROW end\_BIGOP start\_POSTSUBSCRIPT subgoal 1 end\_POSTSUBSCRIPT , … , start\_BIGOP start\_ROW start\_CELL italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT , italic\_g start\_POSTSUBSCRIPT italic\_i + 1 end\_POSTSUBSCRIPT end\_CELL end\_ROW start\_ROW start\_CELL s start\_POSTSUBSCRIPT T start\_POSTSUBSCRIPT i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT , g start\_POSTSUBSCRIPT i+1 end\_POSTSUBSCRIPT end\_CELL end\_ROW end\_BIGOP start\_POSTSUBSCRIPT subgoal italic\_i end\_POSTSUBSCRIPT , … , start\_BIGOP start\_ROW start\_CELL italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT caligraphic\_T - 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT , italic\_g start\_POSTSUBSCRIPT caligraphic\_T end\_POSTSUBSCRIPT end\_CELL end\_ROW start\_ROW start\_CELL s start\_POSTSUBSCRIPT T start\_POSTSUBSCRIPT caligraphic\_T -1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT , g start\_POSTSUBSCRIPT T end\_POSTSUBSCRIPT end\_CELL end\_ROW end\_BIGOP start\_POSTSUBSCRIPT subgoal caligraphic\_T end\_POSTSUBSCRIPT ) | | (33) |
for the master policy. Notice that the terminal state of the ithsuperscript𝑖thi^{\text{th}}italic\_i start\_POSTSUPERSCRIPT th end\_POSTSUPERSCRIPT sub-policy sTisubscript𝑠subscript𝑇𝑖s\_{T\_{i}}italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT forms the state for the master policy to predict the next subgoal gi+1subscript𝑔𝑖1g\_{i+1}italic\_g start\_POSTSUBSCRIPT italic\_i + 1 end\_POSTSUBSCRIPT.
For the (i+1)thsuperscript𝑖1th(i+1)^{\text{th}}( italic\_i + 1 ) start\_POSTSUPERSCRIPT th end\_POSTSUPERSCRIPT subgoal gi+1subscript𝑔𝑖1g\_{i+1}italic\_g start\_POSTSUBSCRIPT italic\_i + 1 end\_POSTSUBSCRIPT, the low-level trajectory of states and primitive actions is given by:
| | | | |
| --- | --- | --- | --- |
| | σgi+1=(sTi,aTi\@mathmeasuresTi, aTi\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasureaction 0,sTi+1,aTi+1\@mathmeasuresTi+1, aTi+1\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasureaction 1,…,sTi+t,aTi+t\@mathmeasuresTi+t, aTi+t\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasureaction t,…,sTi+1).subscript𝜎subscript𝑔𝑖1subscriptfragmentsssubscript𝑇𝑖,asubscript𝑇𝑖fragments\@mathmeasuresTi, aTi\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasureaction 0subscriptfragmentsssubscript𝑇𝑖1,asubscript𝑇𝑖1fragments\@mathmeasuresTi+1, aTi+1\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasureaction 1…subscriptfragmentsssubscript𝑇𝑖𝑡,asubscript𝑇𝑖𝑡fragments\@mathmeasuresTi+t, aTi+t\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasureaction t…subscript𝑠subscript𝑇𝑖1\displaystyle\sigma\_{g\_{i+1}}=\bigg{(}\mathop{\mathchoice{\vtop{\halign{#\cr$\hfil\displaystyle s\_{T\_{i}},a\_{T\_{i}}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\displaystyle{s\_{T\_{i}}, a\_{T\_{i}}}\@mathmeasure\displaystyle{\upbrace}\@mathmeasure\displaystyle{\upbraceg}\@mathmeasure\displaystyle{\upbracegg}\@mathmeasure\displaystyle{\upbraceggg}\@mathmeasure\displaystyle{\upbracegggg}$\displaystyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\textstyle s\_{T\_{i}},a\_{T\_{i}}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\textstyle{s\_{T\_{i}}, a\_{T\_{i}}}\@mathmeasure\textstyle{\upbrace}\@mathmeasure\textstyle{\upbraceg}\@mathmeasure\textstyle{\upbracegg}\@mathmeasure\textstyle{\upbraceggg}\@mathmeasure\textstyle{\upbracegggg}$\textstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptstyle s\_{T\_{i}},a\_{T\_{i}}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptstyle{s\_{T\_{i}}, a\_{T\_{i}}}\@mathmeasure\scriptstyle{\upbrace}\@mathmeasure\scriptstyle{\upbraceg}\@mathmeasure\scriptstyle{\upbracegg}\@mathmeasure\scriptstyle{\upbraceggg}\@mathmeasure\scriptstyle{\upbracegggg}$\scriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptscriptstyle s\_{T\_{i}},a\_{T\_{i}}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptscriptstyle{s\_{T\_{i}}, a\_{T\_{i}}}\@mathmeasure\scriptscriptstyle{\upbrace}\@mathmeasure\scriptscriptstyle{\upbraceg}\@mathmeasure\scriptscriptstyle{\upbracegg}\@mathmeasure\scriptscriptstyle{\upbraceggg}\@mathmeasure\scriptscriptstyle{\upbracegggg}$\scriptscriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}}\limits\_{\text{action 0}},\mathop{\mathchoice{\vtop{\halign{#\cr$\hfil\displaystyle s\_{T\_{i}+1},a\_{T\_{i}+1}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\displaystyle{s\_{T\_{i}+1}, a\_{T\_{i}+1}}\@mathmeasure\displaystyle{\upbrace}\@mathmeasure\displaystyle{\upbraceg}\@mathmeasure\displaystyle{\upbracegg}\@mathmeasure\displaystyle{\upbraceggg}\@mathmeasure\displaystyle{\upbracegggg}$\displaystyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\textstyle s\_{T\_{i}+1},a\_{T\_{i}+1}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\textstyle{s\_{T\_{i}+1}, a\_{T\_{i}+1}}\@mathmeasure\textstyle{\upbrace}\@mathmeasure\textstyle{\upbraceg}\@mathmeasure\textstyle{\upbracegg}\@mathmeasure\textstyle{\upbraceggg}\@mathmeasure\textstyle{\upbracegggg}$\textstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptstyle s\_{T\_{i}+1},a\_{T\_{i}+1}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptstyle{s\_{T\_{i}+1}, a\_{T\_{i}+1}}\@mathmeasure\scriptstyle{\upbrace}\@mathmeasure\scriptstyle{\upbraceg}\@mathmeasure\scriptstyle{\upbracegg}\@mathmeasure\scriptstyle{\upbraceggg}\@mathmeasure\scriptstyle{\upbracegggg}$\scriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptscriptstyle s\_{T\_{i}+1},a\_{T\_{i}+1}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptscriptstyle{s\_{T\_{i}+1}, a\_{T\_{i}+1}}\@mathmeasure\scriptscriptstyle{\upbrace}\@mathmeasure\scriptscriptstyle{\upbraceg}\@mathmeasure\scriptscriptstyle{\upbracegg}\@mathmeasure\scriptscriptstyle{\upbraceggg}\@mathmeasure\scriptscriptstyle{\upbracegggg}$\scriptscriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}}\limits\_{\text{action 1}},\ldots,\mathop{\mathchoice{\vtop{\halign{#\cr$\hfil\displaystyle s\_{T\_{i}+t},a\_{T\_{i}+t}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\displaystyle{s\_{T\_{i}+t}, a\_{T\_{i}+t}}\@mathmeasure\displaystyle{\upbrace}\@mathmeasure\displaystyle{\upbraceg}\@mathmeasure\displaystyle{\upbracegg}\@mathmeasure\displaystyle{\upbraceggg}\@mathmeasure\displaystyle{\upbracegggg}$\displaystyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\textstyle s\_{T\_{i}+t},a\_{T\_{i}+t}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\textstyle{s\_{T\_{i}+t}, a\_{T\_{i}+t}}\@mathmeasure\textstyle{\upbrace}\@mathmeasure\textstyle{\upbraceg}\@mathmeasure\textstyle{\upbracegg}\@mathmeasure\textstyle{\upbraceggg}\@mathmeasure\textstyle{\upbracegggg}$\textstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptstyle s\_{T\_{i}+t},a\_{T\_{i}+t}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptstyle{s\_{T\_{i}+t}, a\_{T\_{i}+t}}\@mathmeasure\scriptstyle{\upbrace}\@mathmeasure\scriptstyle{\upbraceg}\@mathmeasure\scriptstyle{\upbracegg}\@mathmeasure\scriptstyle{\upbraceggg}\@mathmeasure\scriptstyle{\upbracegggg}$\scriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptscriptstyle s\_{T\_{i}+t},a\_{T\_{i}+t}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptscriptstyle{s\_{T\_{i}+t}, a\_{T\_{i}+t}}\@mathmeasure\scriptscriptstyle{\upbrace}\@mathmeasure\scriptscriptstyle{\upbraceg}\@mathmeasure\scriptscriptstyle{\upbracegg}\@mathmeasure\scriptscriptstyle{\upbraceggg}\@mathmeasure\scriptscriptstyle{\upbracegggg}$\scriptscriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}}\limits\_{\text{action t}},\ldots,s\_{T\_{i+1}}\bigg{)}.italic\_σ start\_POSTSUBSCRIPT italic\_g start\_POSTSUBSCRIPT italic\_i + 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT = ( start\_BIGOP start\_ROW start\_CELL italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT end\_CELL end\_ROW start\_ROW start\_CELL s start\_POSTSUBSCRIPT T start\_POSTSUBSCRIPT i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT , a start\_POSTSUBSCRIPT T start\_POSTSUBSCRIPT i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT end\_CELL end\_ROW end\_BIGOP start\_POSTSUBSCRIPT action 0 end\_POSTSUBSCRIPT , start\_BIGOP start\_ROW start\_CELL italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT + 1 end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT + 1 end\_POSTSUBSCRIPT end\_CELL end\_ROW start\_ROW start\_CELL s start\_POSTSUBSCRIPT T start\_POSTSUBSCRIPT i end\_POSTSUBSCRIPT +1 end\_POSTSUBSCRIPT , a start\_POSTSUBSCRIPT T start\_POSTSUBSCRIPT i end\_POSTSUBSCRIPT +1 end\_POSTSUBSCRIPT end\_CELL end\_ROW end\_BIGOP start\_POSTSUBSCRIPT action 1 end\_POSTSUBSCRIPT , … , start\_BIGOP start\_ROW start\_CELL italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT + italic\_t end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT + italic\_t end\_POSTSUBSCRIPT end\_CELL end\_ROW start\_ROW start\_CELL s start\_POSTSUBSCRIPT T start\_POSTSUBSCRIPT i end\_POSTSUBSCRIPT +t end\_POSTSUBSCRIPT , a start\_POSTSUBSCRIPT T start\_POSTSUBSCRIPT i end\_POSTSUBSCRIPT +t end\_POSTSUBSCRIPT end\_CELL end\_ROW end\_BIGOP start\_POSTSUBSCRIPT action t end\_POSTSUBSCRIPT , … , italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i + 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ) . | | (58) |
Note that by concatenating all sub-policy trajectories in order
(σg0,σg1,…,σg𝒯)subscript𝜎subscript𝑔0subscript𝜎subscript𝑔1…subscript𝜎subscript𝑔𝒯(\sigma\_{g\_{0}},\sigma\_{g\_{1}},\ldots,\sigma\_{g\_{\mathcal{T}}})( italic\_σ start\_POSTSUBSCRIPT italic\_g start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT , italic\_σ start\_POSTSUBSCRIPT italic\_g start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT , … , italic\_σ start\_POSTSUBSCRIPT italic\_g start\_POSTSUBSCRIPT caligraphic\_T end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ),
the entire trajectory of states and primitive actions can be recovered.
Subgoals ⟨Tasks, Arguments⟩delimited-⟨⟩Tasks, Arguments\langle\text{Tasks, Arguments}\rangle⟨ Tasks, Arguments ⟩.
As mentioned above, each subgoal is factorized into a task and an argument
g=⟨gtask,gargument⟩𝑔subscript𝑔tasksubscript𝑔argumentg=\langle g\_{\text{task}},g\_{\text{argument}}\rangleitalic\_g = ⟨ italic\_g start\_POSTSUBSCRIPT task end\_POSTSUBSCRIPT , italic\_g start\_POSTSUBSCRIPT argument end\_POSTSUBSCRIPT ⟩. There are 4 possible tasks –
exit-room, find-room, find-object, and answer.
Tasks find-object and find-room
accept as arguments one of the 50 objects and 12 room types in
EQA v1 dataset [[1](#bib.bib1)] respectively;
exit-room and answer do not accept any arguments. This gives us a total of 50+12+1+1=645012116450+12+1+1=6450 + 12 + 1 + 1 = 64
subgoals.
⟨⟨\langle⟨exit-room,*none*⟩normal-⟩\rangle⟩, ⟨⟨\langle⟨answer,*none*⟩normal-⟩\rangle⟩, } 0 args
⟨⟨\langle⟨find-object,*couch*⟩normal-⟩\rangle⟩, ⟨⟨\langle⟨find-object,*cup*⟩normal-⟩\rangle⟩, …, ⟨⟨\langle⟨find-object,*xbox*⟩normal-⟩\rangle⟩, } 50 args
⟨⟨\langle⟨find-room,*living*⟩normal-⟩\rangle⟩, ⟨⟨\langle⟨find-room,*bedroom*⟩normal-⟩\rangle⟩, …, ⟨⟨\langle⟨find-room,*patio*⟩normal-⟩\rangle⟩. } 12 args
Descriptions of these tasks and their success criteria are provided in Table [1](#S3.T1 "Table 1 ‣ 3.1 Hierarchical Policy ‣ 3 Neural Modular Control ‣ Neural Modular Control for Embodied Question Answering").
| Subgoal | Argument(s) | Description | Success |
| --- | --- | --- | --- |
| Exit-room | None | When there is only 1 door in spawn room, or 1 door other than door entered through
in an intermediate room; agent is forced to use the remaining door. | Stopping after exiting through the correct door. |
| Find-room |
Room name
(gym, kitchen, …)
| When there are multiple doors and the agent has to search
and pick the door to the target room. | Stopping after entering target room. |
| Find-object |
Object name
(oven, sofa, …)
| When the agent has to search for a specific object in room. | Stopping within 0.75m0.75m0.75\text{m}0.75 m of the target object. |
| Answer | None | When the agent has to provide an answer from the answer space. | Generating the correct answer to the question. |
Table 1: Descriptions of our subgoals and conditions we use to extract
them automatically from expert trajectories.
Master Policy. The master policy πθsubscript𝜋𝜃\mathbf{\pi\_{\theta}}italic\_π start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT parameterized
by θ𝜃\thetaitalic\_θ is implemented as a single layer Gated
Recurrent Unit (GRU). At each high-level step i+1𝑖1i+1italic\_i + 1, the master policy πθ(g|sTi)subscript𝜋𝜃conditional𝑔subscript𝑠subscript𝑇𝑖\mathbf{\pi\_{\theta}}(g|s\_{T\_{i}})italic\_π start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT ( italic\_g | italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT )
takes as input the concatenation of a encoding of the question q∈ℝ128𝑞superscriptℝ128q\in\mathbb{R}^{128}italic\_q ∈ blackboard\_R start\_POSTSUPERSCRIPT 128 end\_POSTSUPERSCRIPT,
the image feature vTi∈ℝ128subscript𝑣subscript𝑇𝑖superscriptℝ128v\_{T\_{i}}\in\mathbb{R}^{128}italic\_v start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ∈ blackboard\_R start\_POSTSUPERSCRIPT 128 end\_POSTSUPERSCRIPT of the current frame and an
encoding oi∈ℝ32subscript𝑜𝑖superscriptℝ32o\_{i}\in\mathbb{R}^{32}italic\_o start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ∈ blackboard\_R start\_POSTSUPERSCRIPT 32 end\_POSTSUPERSCRIPT computed from a 1-hot representation of the ithsuperscript𝑖thi^{\text{th}}italic\_i start\_POSTSUPERSCRIPT th end\_POSTSUPERSCRIPT
subgoal, *i.e*. 𝟙(gi)1subscript𝑔𝑖\mathbbm{1}(g\_{i})blackboard\_1 ( italic\_g start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ).
This information is used to update the hidden state hi∈ℝ1048subscriptℎ𝑖superscriptℝ1048h\_{i}\in\mathbb{R}^{1048}italic\_h start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ∈ blackboard\_R start\_POSTSUPERSCRIPT 1048 end\_POSTSUPERSCRIPT
that encodes the entire trajectory up to time t𝑡titalic\_t and serves as the state representation.
The policy then produces a probability
distribution over all possible (64) subgoals 𝒢𝒢\mathcal{G}caligraphic\_G.
We train these policies with actor-critic methods and thus the network also produces
a value estimate.
Sub-policies.
To take advantage of the comparatively lower number of subgoal tasks, we
decompose sub-policy parameters ϕgsubscriptitalic-ϕ𝑔\phi\_{g}italic\_ϕ start\_POSTSUBSCRIPT italic\_g end\_POSTSUBSCRIPT into
ϕgtasksubscriptitalic-ϕsubscript𝑔task\phi\_{g\_{\text{task}}}italic\_ϕ start\_POSTSUBSCRIPT italic\_g start\_POSTSUBSCRIPT task end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT and ϕgargumentsubscriptitalic-ϕsubscript𝑔argument\phi\_{g\_{\text{argument}}}italic\_ϕ start\_POSTSUBSCRIPT italic\_g start\_POSTSUBSCRIPT argument end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT, where
ϕgtasksubscriptitalic-ϕsubscript𝑔task\phi\_{g\_{\text{task}}}italic\_ϕ start\_POSTSUBSCRIPT italic\_g start\_POSTSUBSCRIPT task end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT are shared across the same task and
ϕgargumentsubscriptitalic-ϕsubscript𝑔argument\phi\_{g\_{\text{argument}}}italic\_ϕ start\_POSTSUBSCRIPT italic\_g start\_POSTSUBSCRIPT argument end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT is an argument specific embedding.
Parameter sharing enables us to learn the shared task
in a sample-efficient manner, rather than exhaustively
learning separate sub-policies for each combination.
Like the master policy, each sub-policy πϕgsubscript𝜋subscriptitalic-ϕ𝑔\pi\_{\phi\_{g}}italic\_π start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT italic\_g end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT is implemented as a
single-layer GRU. At each low-level time step t𝑡titalic\_t, a sub-policy πϕg(a|st)subscript𝜋subscriptitalic-ϕ𝑔conditional𝑎subscript𝑠𝑡\pi\_{\phi\_{g}}(a|s\_{t})italic\_π start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT italic\_g end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_a | italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ) takes as input
the concatenation of the image feature vt∈ℝ128subscript𝑣𝑡superscriptℝ128v\_{t}\in\mathbb{R}^{128}italic\_v start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT ∈ blackboard\_R start\_POSTSUPERSCRIPT 128 end\_POSTSUPERSCRIPT
of the current frame, an encoding pt−1∈ℝ32subscript𝑝𝑡1superscriptℝ32p\_{t-1}\in\mathbb{R}^{32}italic\_p start\_POSTSUBSCRIPT italic\_t - 1 end\_POSTSUBSCRIPT ∈ blackboard\_R start\_POSTSUPERSCRIPT 32 end\_POSTSUPERSCRIPT
computed from a 1-hot representation of the
previous primitive action *i.e*. 𝟙(at−1)1subscript𝑎𝑡1\mathbbm{1}(a\_{t-1})blackboard\_1 ( italic\_a start\_POSTSUBSCRIPT italic\_t - 1 end\_POSTSUBSCRIPT ),
and the argument embedding ϕgargumentsubscriptitalic-ϕsubscript𝑔argument\phi\_{g\_{\text{argument}}}italic\_ϕ start\_POSTSUBSCRIPT italic\_g start\_POSTSUBSCRIPT argument end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT. These inputs are used to update the
hidden state htg∈ℝ1048superscriptsubscriptℎ𝑡𝑔superscriptℝ1048h\_{t}^{g}\in\mathbb{R}^{1048}italic\_h start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_g end\_POSTSUPERSCRIPT ∈ blackboard\_R start\_POSTSUPERSCRIPT 1048 end\_POSTSUPERSCRIPT which serves as the
state representation. The policy then outputs a
distribution over primitive actions (forward, turn-left,
turn-right, stop). As with the master policy,
each sub-policy also output a value estimate. shows this model structure.
Perception and Question Answering. To ensure fair comparisons to prior work,
we use the same perception and question answering models as used by Das *et al*. [[1](#bib.bib1)].
The perception model is a simple convolutional neural network trained
to perform auto-encoding, semantic segmentation, and depth estimation
from RGB frames taken from House3D [[2](#bib.bib2)]. Like [[1](#bib.bib1)],
we use the bottleneck layer of this model as a fixed feature extractor.
We also use the same post-navigational question-answering model as [[1](#bib.bib1)], which encodes the question
with a 2-layer LSTM and performs dot-product based attention between the question
encoding and the image features from the last five frames along the navigation path right before the answer
module is called. This post-navigational answering module is trained using visual features along the
shortest path trajectories and then frozen. By keeping these parts of the architecture identical to [[1](#bib.bib1)],
our experimental comparisons can focus on the differences only due to our contributions, the
Neural Modular Controller.
###
3.2 Hierarchical Behavior Cloning from Expert Trajectories
The questions in EQA v1 dataset [[1](#bib.bib1)] (*e.g*. *‘What color is the fireplace?’*)
are constructed to inquire about attributes (color, location, *etc*.) of specific target objects (*‘fireplace’*).
This notion of a target enables the construction of an automatically generated
*expert trajectory* (s0\*,a0\*,…,sT\*,aT\*)superscriptsubscript𝑠0superscriptsubscript𝑎0…superscriptsubscript𝑠𝑇superscriptsubscript𝑎𝑇(s\_{0}^{\*},a\_{0}^{\*},\ldots,s\_{T}^{\*},a\_{T}^{\*})( italic\_s start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT , italic\_a start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT , … , italic\_s start\_POSTSUBSCRIPT italic\_T end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT , italic\_a start\_POSTSUBSCRIPT italic\_T end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) –
the states and actions along the shortest path from the agent spawn
location to the object of interest specified in the question. Notice that these shortest paths
may only be used as supervision on training environments but may not be utilized during evaluation on test environments
(where the agent must operate from egocentric vision alone).
Specifically, we would like to use these expert demonstrations to pre-train our proposed NMC navigator
using behavior cloning. However, these trajectories (s0\*,a0\*,…,sT\*,aT\*)superscriptsubscript𝑠0superscriptsubscript𝑎0…superscriptsubscript𝑠𝑇superscriptsubscript𝑎𝑇(s\_{0}^{\*},a\_{0}^{\*},\ldots,s\_{T}^{\*},a\_{T}^{\*})( italic\_s start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT , italic\_a start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT , … , italic\_s start\_POSTSUBSCRIPT italic\_T end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT , italic\_a start\_POSTSUBSCRIPT italic\_T end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) correspond to
a series of primitive actions. To provide supervision for both the master policy and sub-policies, these shortest-path trajectories must be
annotated with a sequence of subgoals and segmented into their respective temporal extents,
resulting in Σ\*superscriptΣ\Sigma^{\*}roman\_Σ start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT and (σgi\*)superscriptsubscript𝜎subscript𝑔𝑖(\sigma\_{g\_{i}}^{\*})( italic\_σ start\_POSTSUBSCRIPT italic\_g start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ).
(a) Q: What color is the fireplace? A: Brown
(b) Distribution of subgoals with number of actions from
the target object as per expert plans. Closer to the target object,
the expert plan predominantly consists of Find-object, while as we move
farther away, the proportion of Find-room and Exit-room goes up.
Figure 2: We extract expert subgoal trajectories from shortest paths by dividing paths on room transition boundaries (circled in (a))and following the rules in Tab. [1](#S3.T1 "Table 1 ‣ 3.1 Hierarchical Policy ‣ 3 Neural Modular Control ‣ Neural Modular Control for Embodied Question Answering").
We automate this ‘lifting’ of annotation up the hierarchy by
leveraging the object and room bounding boxes provided by the House3D [[2](#bib.bib2)].
Essentially, a floor plan may be viewed as an undirected graph with rooms as nodes and doorways
as edges connecting a pair of adjacent rooms.
An example trajectory is shown in Fig. [1(a)](#S3.F1.sf1 "1(a) ‣ Figure 2 ‣ 3.2 Hierarchical Behavior Cloning from Expert Trajectories ‣ 3 Neural Modular Control ‣ Neural Modular Control for Embodied Question Answering")
for the question *‘What color is the fireplace?’*. The agent is spawned in a bedroom,
the shortest path exits into the hall, enters the living room, and approaches the fireplace.
We convert this trajectory to the subgoal sequence (exit-room, find-room[*living*],
find-object[*fireplace*], answer) by recording the transitions on the shortest path from one room
to another, which also naturally provides us with temporal extents of these subgoals.
We follow a couple of subtle but natural rules: (1) find-object is tagged only when the agent has reached the
destination room containing the target object;
and (2) exit-room is tagged only when the ‘out-degree’ of the current room in the floor-plan-graph
is exactly 1 (*i.e*. either the current room has exactly one doorway or the current room has two
doorways but the agent came in through one).
Rule (2) ensures a semantic difference between exit-room and find-room –
informally, exit-room means *‘get me out of here’*
and find-room[*name*] means *‘look for room name’*.
Tab. [1](#S3.T1 "Table 1 ‣ 3.1 Hierarchical Policy ‣ 3 Neural Modular Control ‣ Neural Modular Control for Embodied Question Answering") summarizes these subgoals and the heuristics used to
automatically extract them from navigational paths. Fig. [1(b)](#S3.F1.sf2 "1(b) ‣ Figure 2 ‣ 3.2 Hierarchical Behavior Cloning from Expert Trajectories ‣ 3 Neural Modular Control ‣ Neural Modular Control for Embodied Question Answering") shows
the proportions of these subgoals in expert trajectories as a function of the distance
from target object. Notice that when the agent is close to the target, it is likely to be within the same room
as the target and thus find-object dominates. On the other hand,
when the agent is far away from the target, find-room
and exit-room dominate.
We perform this lifting of shortest paths for all training set questions in
EQA v1 dataset [[1](#bib.bib1)], resulting in N𝑁Nitalic\_N expert trajectories
{Σn\*}n=1NsuperscriptsubscriptsuperscriptsubscriptΣ𝑛𝑛1𝑁\{\Sigma\_{n}^{\*}\}\_{n=1}^{N}{ roman\_Σ start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT } start\_POSTSUBSCRIPT italic\_n = 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_N end\_POSTSUPERSCRIPT for the master policy
and K(>>N)annotated𝐾much-greater-thanabsent𝑁K(>>N)italic\_K ( > > italic\_N ) trajectories {σgk\*}k=1Ksuperscriptsubscriptsuperscriptsubscript𝜎subscript𝑔𝑘𝑘1𝐾\{\sigma\_{g\_{k}}^{\*}\}\_{k=1}^{K}{ italic\_σ start\_POSTSUBSCRIPT italic\_g start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT } start\_POSTSUBSCRIPT italic\_k = 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_K end\_POSTSUPERSCRIPT for sub-policies.
We can then perform hierarchical behavior cloning by minimizing the sum of
cross-entropy losses over all decisions in all expert trajectories.
As is typical in maximum-likelihood training of directed probabilistic models
(*e.g*. hierarchical Bayes Nets), full supervision results in decomposition
into independent sub-problems.
Specifically, with a slight abuse of notation, let (si\*,gi+1\*)∈Σ\*superscriptsubscript𝑠𝑖superscriptsubscript𝑔𝑖1superscriptΣ(s\_{i}^{\*},g\_{i+1}^{\*})\in\Sigma^{\*}( italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT , italic\_g start\_POSTSUBSCRIPT italic\_i + 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) ∈ roman\_Σ start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT denote an iterator
over all state-subgoal tuples in Σ\*superscriptΣ\Sigma^{\*}roman\_Σ start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT, and ∑(si\*,gi+1\*)∈Σ\*subscriptsuperscriptsubscript𝑠𝑖superscriptsubscript𝑔𝑖1superscriptΣ\displaystyle\sum\_{(s\_{i}^{\*},g\_{i+1}^{\*})\in\Sigma^{\*}}∑ start\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT , italic\_g start\_POSTSUBSCRIPT italic\_i + 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) ∈ roman\_Σ start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT end\_POSTSUBSCRIPT
denote a sum over such tuples.
Now, the independent learning problems can be written as:
| | | | | | |
| --- | --- | --- | --- | --- | --- |
| | θ\*=argminθsuperscript𝜃subscriptargmin𝜃\displaystyle\theta^{\*}=\mathop{\mathrm{argmin}}\_{\theta}\,\,italic\_θ start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT = roman\_argmin start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT | ∑n=1N∑(si\*,gi+1\*)∈Σn\*−log(πθ(gi+1\*|si\*))superscriptsubscript𝑛1𝑁subscriptsuperscriptsubscript𝑠𝑖superscriptsubscript𝑔𝑖1superscriptsubscriptΣ𝑛subscript𝜋𝜃conditionalsuperscriptsubscript𝑔𝑖1superscriptsubscript𝑠𝑖\displaystyle\sum\_{n=1}^{N}\hskip 48.42076pt\sum\_{(s\_{i}^{\*},g\_{i+1}^{\*})\in\Sigma\_{n}^{\*}}\,\,\,\,-\log\Big{(}\pi\_{\theta}(g\_{i+1}^{\*}|s\_{i}^{\*})\Big{)}∑ start\_POSTSUBSCRIPT italic\_n = 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_N end\_POSTSUPERSCRIPT ∑ start\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT , italic\_g start\_POSTSUBSCRIPT italic\_i + 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) ∈ roman\_Σ start\_POSTSUBSCRIPT italic\_n end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT end\_POSTSUBSCRIPT - roman\_log ( italic\_π start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT ( italic\_g start\_POSTSUBSCRIPT italic\_i + 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT | italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) ) | (master policy cloning) | | (59a) |
| | ϕg\*=argminϕsuperscriptsubscriptitalic-ϕ𝑔subscriptargminitalic-ϕ\displaystyle\phi\_{g}^{\*}=\mathop{\mathrm{argmin}}\_{\phi}\,\,italic\_ϕ start\_POSTSUBSCRIPT italic\_g end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT = roman\_argmin start\_POSTSUBSCRIPT italic\_ϕ end\_POSTSUBSCRIPT | ∑k=1K[[gk=g]]\@mathmeasure ∑(st\*, at+1\*) ∈σgk\* ∑k=1K [[gk = g]] \@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasuredemonstrations∑(st\*,at+1\*)∈σgk\*\@mathmeasure∑(st\*, at+1\*) ∈σgk\*\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasuretransitions−log(πϕg(at+1\*|st\*))\@mathmeasure ∑(st\*, at+1\*) ∈σgk\* -log(𝜋ϕg(at+1\* |st\*))\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasurenegative-log-likelihoodsubscriptfragments𝑘1𝐾[[g𝑘g]]fragments\@mathmeasure ∑(st\*, at+1\*) ∈σgk\* ∑k=1K [[gk = g]] \@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasuredemonstrationssubscriptfragmentssuperscriptsubscript𝑠𝑡superscriptsubscript𝑎𝑡1superscriptsubscript𝜎subscript𝑔𝑘fragments\@mathmeasure∑(st\*, at+1\*) ∈σgk\*\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasuretransitionssubscriptfragments(πsubscriptitalic-ϕ𝑔(a𝑡1|s𝑡))fragments\@mathmeasure ∑(st\*, at+1\*) ∈σgk\* -log(𝜋ϕg(at+1\* |st\*))\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasure\@mathmeasurenegative-log-likelihood\displaystyle\mathop{\mathchoice{\vtop{\halign{#\cr$\hfil\displaystyle\vphantom{\sum\_{(s\_{t}^{\*},a\_{t+1}^{\*})\in\sigma\_{g\_{k}}^{\*}}}\sum\_{k=1}^{K}\,[[g\_{k}=g]]\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\displaystyle{ \vphantom{ \sum\_{ (s\_{t}^{\*}, a\_{t+1}^{\*}) \in\sigma\_{g\_{k}}^{\*}}}
\sum\_{k=1}^{K} \, [[g\_{k} = g]] }\@mathmeasure\displaystyle{\upbrace}\@mathmeasure\displaystyle{\upbraceg}\@mathmeasure\displaystyle{\upbracegg}\@mathmeasure\displaystyle{\upbraceggg}\@mathmeasure\displaystyle{\upbracegggg}$\displaystyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\textstyle\vphantom{\sum\_{(s\_{t}^{\*},a\_{t+1}^{\*})\in\sigma\_{g\_{k}}^{\*}}}\sum\_{k=1}^{K}\,[[g\_{k}=g]]\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\textstyle{ \vphantom{ \sum\_{ (s\_{t}^{\*}, a\_{t+1}^{\*}) \in\sigma\_{g\_{k}}^{\*}}}
\sum\_{k=1}^{K} \, [[g\_{k} = g]] }\@mathmeasure\textstyle{\upbrace}\@mathmeasure\textstyle{\upbraceg}\@mathmeasure\textstyle{\upbracegg}\@mathmeasure\textstyle{\upbraceggg}\@mathmeasure\textstyle{\upbracegggg}$\textstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptstyle\vphantom{\sum\_{(s\_{t}^{\*},a\_{t+1}^{\*})\in\sigma\_{g\_{k}}^{\*}}}\sum\_{k=1}^{K}\,[[g\_{k}=g]]\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptstyle{ \vphantom{ \sum\_{ (s\_{t}^{\*}, a\_{t+1}^{\*}) \in\sigma\_{g\_{k}}^{\*}}}
\sum\_{k=1}^{K} \, [[g\_{k} = g]] }\@mathmeasure\scriptstyle{\upbrace}\@mathmeasure\scriptstyle{\upbraceg}\@mathmeasure\scriptstyle{\upbracegg}\@mathmeasure\scriptstyle{\upbraceggg}\@mathmeasure\scriptstyle{\upbracegggg}$\scriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptscriptstyle\vphantom{\sum\_{(s\_{t}^{\*},a\_{t+1}^{\*})\in\sigma\_{g\_{k}}^{\*}}}\sum\_{k=1}^{K}\,[[g\_{k}=g]]\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptscriptstyle{ \vphantom{ \sum\_{ (s\_{t}^{\*}, a\_{t+1}^{\*}) \in\sigma\_{g\_{k}}^{\*}}}
\sum\_{k=1}^{K} \, [[g\_{k} = g]] }\@mathmeasure\scriptscriptstyle{\upbrace}\@mathmeasure\scriptscriptstyle{\upbraceg}\@mathmeasure\scriptscriptstyle{\upbracegg}\@mathmeasure\scriptscriptstyle{\upbraceggg}\@mathmeasure\scriptscriptstyle{\upbracegggg}$\scriptscriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}}\limits\_{\text{demonstrations}}\,\,\,\,\,\mathop{\mathchoice{\vtop{\halign{#\cr$\hfil\displaystyle\sum\_{(s\_{t}^{\*},a\_{t+1}^{\*})\in\sigma\_{g\_{k}}^{\*}}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\displaystyle{\sum\_{ (s\_{t}^{\*}, a\_{t+1}^{\*}) \in\sigma\_{g\_{k}}^{\*}}}\@mathmeasure\displaystyle{\upbrace}\@mathmeasure\displaystyle{\upbraceg}\@mathmeasure\displaystyle{\upbracegg}\@mathmeasure\displaystyle{\upbraceggg}\@mathmeasure\displaystyle{\upbracegggg}$\displaystyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\textstyle\sum\_{(s\_{t}^{\*},a\_{t+1}^{\*})\in\sigma\_{g\_{k}}^{\*}}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\textstyle{\sum\_{ (s\_{t}^{\*}, a\_{t+1}^{\*}) \in\sigma\_{g\_{k}}^{\*}}}\@mathmeasure\textstyle{\upbrace}\@mathmeasure\textstyle{\upbraceg}\@mathmeasure\textstyle{\upbracegg}\@mathmeasure\textstyle{\upbraceggg}\@mathmeasure\textstyle{\upbracegggg}$\textstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptstyle\sum\_{(s\_{t}^{\*},a\_{t+1}^{\*})\in\sigma\_{g\_{k}}^{\*}}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptstyle{\sum\_{ (s\_{t}^{\*}, a\_{t+1}^{\*}) \in\sigma\_{g\_{k}}^{\*}}}\@mathmeasure\scriptstyle{\upbrace}\@mathmeasure\scriptstyle{\upbraceg}\@mathmeasure\scriptstyle{\upbracegg}\@mathmeasure\scriptstyle{\upbraceggg}\@mathmeasure\scriptstyle{\upbracegggg}$\scriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptscriptstyle\sum\_{(s\_{t}^{\*},a\_{t+1}^{\*})\in\sigma\_{g\_{k}}^{\*}}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptscriptstyle{\sum\_{ (s\_{t}^{\*}, a\_{t+1}^{\*}) \in\sigma\_{g\_{k}}^{\*}}}\@mathmeasure\scriptscriptstyle{\upbrace}\@mathmeasure\scriptscriptstyle{\upbraceg}\@mathmeasure\scriptscriptstyle{\upbracegg}\@mathmeasure\scriptscriptstyle{\upbraceggg}\@mathmeasure\scriptscriptstyle{\upbracegggg}$\scriptscriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}}\limits\_{\text{transitions}}\,\,\,\mathop{\mathchoice{\vtop{\halign{#\cr$\hfil\displaystyle\vphantom{\sum\_{(s\_{t}^{\*},a\_{t+1}^{\*})\in\sigma\_{g\_{k}}^{\*}}}-\log\Big{(}\pi\_{\phi\_{g}}(a\_{t+1}^{\*}|s\_{t}^{\*})\Big{)}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\displaystyle{ \vphantom{ \sum\_{ (s\_{t}^{\*}, a\_{t+1}^{\*}) \in\sigma\_{g\_{k}}^{\*}}}
-\log\Big{(}\pi\_{\phi\_{g}}(a\_{t+1}^{\*} |s\_{t}^{\*})\Big{)}}\@mathmeasure\displaystyle{\upbrace}\@mathmeasure\displaystyle{\upbraceg}\@mathmeasure\displaystyle{\upbracegg}\@mathmeasure\displaystyle{\upbraceggg}\@mathmeasure\displaystyle{\upbracegggg}$\displaystyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\textstyle\vphantom{\sum\_{(s\_{t}^{\*},a\_{t+1}^{\*})\in\sigma\_{g\_{k}}^{\*}}}-\log\Big{(}\pi\_{\phi\_{g}}(a\_{t+1}^{\*}|s\_{t}^{\*})\Big{)}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\textstyle{ \vphantom{ \sum\_{ (s\_{t}^{\*}, a\_{t+1}^{\*}) \in\sigma\_{g\_{k}}^{\*}}}
-\log\Big{(}\pi\_{\phi\_{g}}(a\_{t+1}^{\*} |s\_{t}^{\*})\Big{)}}\@mathmeasure\textstyle{\upbrace}\@mathmeasure\textstyle{\upbraceg}\@mathmeasure\textstyle{\upbracegg}\@mathmeasure\textstyle{\upbraceggg}\@mathmeasure\textstyle{\upbracegggg}$\textstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptstyle\vphantom{\sum\_{(s\_{t}^{\*},a\_{t+1}^{\*})\in\sigma\_{g\_{k}}^{\*}}}-\log\Big{(}\pi\_{\phi\_{g}}(a\_{t+1}^{\*}|s\_{t}^{\*})\Big{)}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptstyle{ \vphantom{ \sum\_{ (s\_{t}^{\*}, a\_{t+1}^{\*}) \in\sigma\_{g\_{k}}^{\*}}}
-\log\Big{(}\pi\_{\phi\_{g}}(a\_{t+1}^{\*} |s\_{t}^{\*})\Big{)}}\@mathmeasure\scriptstyle{\upbrace}\@mathmeasure\scriptstyle{\upbraceg}\@mathmeasure\scriptstyle{\upbracegg}\@mathmeasure\scriptstyle{\upbraceggg}\@mathmeasure\scriptstyle{\upbracegggg}$\scriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}{\vtop{\halign{#\cr$\hfil\scriptscriptstyle\vphantom{\sum\_{(s\_{t}^{\*},a\_{t+1}^{\*})\in\sigma\_{g\_{k}}^{\*}}}-\log\Big{(}\pi\_{\phi\_{g}}(a\_{t+1}^{\*}|s\_{t}^{\*})\Big{)}\hfil$\crcr\kern 2.0pt\nointerlineskip\cr\@mathmeasure\scriptscriptstyle{ \vphantom{ \sum\_{ (s\_{t}^{\*}, a\_{t+1}^{\*}) \in\sigma\_{g\_{k}}^{\*}}}
-\log\Big{(}\pi\_{\phi\_{g}}(a\_{t+1}^{\*} |s\_{t}^{\*})\Big{)}}\@mathmeasure\scriptscriptstyle{\upbrace}\@mathmeasure\scriptscriptstyle{\upbraceg}\@mathmeasure\scriptscriptstyle{\upbracegg}\@mathmeasure\scriptscriptstyle{\upbraceggg}\@mathmeasure\scriptscriptstyle{\upbracegggg}$\scriptscriptstyle\bracelu\leaders\hbox{$\bracemid$}\hfill\bracemu\leaders\hbox{$\bracemid$}\hfill\braceru$\crcr}}}}\limits\_{\text{negative-log-likelihood}}start\_BIGOP start\_ROW start\_CELL ∑ start\_POSTSUBSCRIPT italic\_k = 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT italic\_K end\_POSTSUPERSCRIPT [ [ italic\_g start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT = italic\_g ] ] end\_CELL end\_ROW start\_ROW start\_CELL ∑ start\_POSTSUBSCRIPT (s start\_POSTSUBSCRIPT t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT , a start\_POSTSUBSCRIPT t+1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) ∈ italic\_σ start\_POSTSUBSCRIPT g start\_POSTSUBSCRIPT k end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT end\_POSTSUBSCRIPT ∑ start\_POSTSUBSCRIPT k=1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT K end\_POSTSUPERSCRIPT [[g start\_POSTSUBSCRIPT k end\_POSTSUBSCRIPT = g]] end\_CELL end\_ROW end\_BIGOP start\_POSTSUBSCRIPT demonstrations end\_POSTSUBSCRIPT start\_BIGOP start\_ROW start\_CELL ∑ start\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT , italic\_a start\_POSTSUBSCRIPT italic\_t + 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) ∈ italic\_σ start\_POSTSUBSCRIPT italic\_g start\_POSTSUBSCRIPT italic\_k end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT end\_POSTSUBSCRIPT end\_CELL end\_ROW start\_ROW start\_CELL ∑ start\_POSTSUBSCRIPT (s start\_POSTSUBSCRIPT t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT , a start\_POSTSUBSCRIPT t+1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) ∈ italic\_σ start\_POSTSUBSCRIPT g start\_POSTSUBSCRIPT k end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT end\_POSTSUBSCRIPT end\_CELL end\_ROW end\_BIGOP start\_POSTSUBSCRIPT transitions end\_POSTSUBSCRIPT start\_BIGOP start\_ROW start\_CELL - roman\_log ( italic\_π start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT italic\_g end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_a start\_POSTSUBSCRIPT italic\_t + 1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT | italic\_s start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) ) end\_CELL end\_ROW start\_ROW start\_CELL ∑ start\_POSTSUBSCRIPT (s start\_POSTSUBSCRIPT t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT , a start\_POSTSUBSCRIPT t+1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) ∈ italic\_σ start\_POSTSUBSCRIPT g start\_POSTSUBSCRIPT k end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT end\_POSTSUBSCRIPT -log ( π start\_POSTSUBSCRIPT ϕ start\_POSTSUBSCRIPT g end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT (a start\_POSTSUBSCRIPT t+1 end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT |s start\_POSTSUBSCRIPT t end\_POSTSUBSCRIPT start\_POSTSUPERSCRIPT \* end\_POSTSUPERSCRIPT ) ) end\_CELL end\_ROW end\_BIGOP start\_POSTSUBSCRIPT negative-log-likelihood end\_POSTSUBSCRIPT | (sub-policy cloning) | | (59z) |
Intuitively, each sub-policy independently maximizes the conditional probability of actions observed in the
expert demonstrations, and the master policy essentially trains assuming perfect sub-policies.
###
3.3 Asynchronous Advantage Actor-Critic (A3C) Training
After the independent behavior cloning stage, the policies have learned to mimic expert
trajectories; however, they have not had to coordinate with each other or recover from
their own navigational errors. As such, we fine-tune them with
reinforcement learning – first independently and then jointly.
Reward Structure. The ultimate goal of our agent is to answer questions accurately; however, doing so
requires navigating the environment sufficiently well in search of the answer. We
mirror this structure in our reward R𝑅Ritalic\_R, decomposing it into a sum of a sparse
terminal reward Rterminalsubscript𝑅terminalR\_{\text{terminal}}italic\_R start\_POSTSUBSCRIPT terminal end\_POSTSUBSCRIPT for the final outcome and a dense, shaped
reward Rshapedsubscript𝑅shapedR\_{\text{shaped}}italic\_R start\_POSTSUBSCRIPT shaped end\_POSTSUBSCRIPT [[38](#bib.bib38)] determined by the agent’s progress towards its goals.
For the master policy πθsubscript𝜋𝜃\pi\_{\theta}italic\_π start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT, we set Rterminalsubscript𝑅terminalR\_{\text{terminal}}italic\_R start\_POSTSUBSCRIPT terminal end\_POSTSUBSCRIPT to be 1 if
the model answers the question correctly and 0 otherwise. The shaped
reward Rshapedsubscript𝑅shapedR\_{\text{shaped}}italic\_R start\_POSTSUBSCRIPT shaped end\_POSTSUBSCRIPT at master-step i𝑖iitalic\_i is based on the change of
navigable distance to the target object before and after executing subgoal gisubscript𝑔𝑖g\_{i}italic\_g start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT.
Each sub-policy πϕgsubscript𝜋subscriptitalic-ϕ𝑔\pi\_{\phi\_{g}}italic\_π start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT italic\_g end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT also has a terminal 0/1 reward Rterminalsubscript𝑅terminalR\_{\text{terminal}}italic\_R start\_POSTSUBSCRIPT terminal end\_POSTSUBSCRIPT
for stopping in a successful state, *e.g*. Exit-room ending
outside the room it was called in (see Tab. [1](#S3.T1 "Table 1 ‣ 3.1 Hierarchical Policy ‣ 3 Neural Modular Control ‣ Neural Modular Control for Embodied Question Answering") for
all success definitions). Like the master policy, Rshapedsubscript𝑅shapedR\_{\text{shaped}}italic\_R start\_POSTSUBSCRIPT shaped end\_POSTSUBSCRIPT at
time t𝑡titalic\_t is set according to the change in navigable distance to the sub-policy target (*e.g*. a point just inside a living room for
find-room[*living*]) after executing
the primitive action atsubscript𝑎𝑡a\_{t}italic\_a start\_POSTSUBSCRIPT italic\_t end\_POSTSUBSCRIPT. Further, sub-policies are also
penalized a small constant (-0.02) for colliding with obstacles.
Policy Optimization.
We update the master and sub-policies to maximize expected
discounted future rewards J(πθ)𝐽subscript𝜋𝜃J(\pi\_{\theta})italic\_J ( italic\_π start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT ) and J(πϕg)𝐽subscript𝜋subscriptitalic-ϕ𝑔J(\pi\_{\phi\_{g}})italic\_J ( italic\_π start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT italic\_g end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT )
respectively through the Asynchronous Advantage Actor
Critic [[15](#bib.bib15)] policy-gradient algorithm.
Specifically, for the master policy, the gradient of the expected reward is
written as:
| | | | |
| --- | --- | --- | --- |
| | ∇θJ(πθ)=𝔼[∇θlog(πθ(gi|sTi))(Q(sTi,gi)−cθ(sTi))]subscript∇𝜃𝐽subscript𝜋𝜃𝔼delimited-[]subscript∇𝜃subscript𝜋𝜃conditionalsubscript𝑔𝑖subscript𝑠subscript𝑇𝑖𝑄subscript𝑠subscript𝑇𝑖subscript𝑔𝑖subscript𝑐𝜃subscript𝑠subscript𝑇𝑖\nabla\_{\theta}J(\pi\_{\theta})=\mathbb{E}\left[\nabla\_{\theta}\log(\pi\_{\theta}(g\_{i}|s\_{T\_{i}}))\left(Q(s\_{T\_{i}},g\_{i})-c\_{\theta}(s\_{T\_{i}})\right)\right]∇ start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT italic\_J ( italic\_π start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT ) = blackboard\_E [ ∇ start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT roman\_log ( italic\_π start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT ( italic\_g start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT | italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ) ) ( italic\_Q ( italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT , italic\_g start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ) - italic\_c start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ) ) ] | | (60) |
where cθ(sTi)subscript𝑐𝜃subscript𝑠subscript𝑇𝑖c\_{\theta}(s\_{T\_{i}})italic\_c start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ) is the estimated value of sTisubscript𝑠subscript𝑇𝑖s\_{T\_{i}}italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT
produced by the critic for πθsubscript𝜋𝜃\pi\_{\theta}italic\_π start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT. To further reduce variance, we follow [[39](#bib.bib39)]
and estimate Q(sTi,gi)≈Rθ(sTi)+γcθ(sTi+1)𝑄subscript𝑠subscript𝑇𝑖subscript𝑔𝑖subscript𝑅𝜃subscript𝑠subscript𝑇𝑖𝛾subscript𝑐𝜃subscript𝑠subscript𝑇𝑖1Q(s\_{T\_{i}},g\_{i})\approx R\_{\theta}(s\_{T\_{i}})+\gamma c\_{\theta}(s\_{T\_{i+1}})italic\_Q ( italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT , italic\_g start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ) ≈ italic\_R start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ) + italic\_γ italic\_c start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i + 1 end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ) such that Q(sTi,gi)−cθ(sTi)𝑄subscript𝑠subscript𝑇𝑖subscript𝑔𝑖subscript𝑐𝜃subscript𝑠subscript𝑇𝑖Q(s\_{T\_{i}},g\_{i})-c\_{\theta}(s\_{T\_{i}})italic\_Q ( italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT , italic\_g start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ) - italic\_c start\_POSTSUBSCRIPT italic\_θ end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_T start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ) computes a
generalized advantage estimator (GAE).
Similarly, each sub-policy πϕgsubscript𝜋subscriptitalic-ϕ𝑔\pi\_{\phi\_{g}}italic\_π start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT italic\_g end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT is updated according to the
gradient
| | | | |
| --- | --- | --- | --- |
| | ∇ϕgJ(πϕg)=𝔼[∇ϕglog(πϕg(ai|si))(Q(si,ai)−cϕg(si))].subscript∇subscriptitalic-ϕ𝑔𝐽subscript𝜋subscriptitalic-ϕ𝑔𝔼delimited-[]subscript∇subscriptitalic-ϕ𝑔subscript𝜋subscriptitalic-ϕ𝑔conditionalsubscript𝑎𝑖subscript𝑠𝑖𝑄subscript𝑠𝑖subscript𝑎𝑖subscript𝑐subscriptitalic-ϕ𝑔subscript𝑠𝑖\nabla\_{\phi\_{g}}J(\pi\_{\phi\_{g}})=\mathbb{E}\left[\nabla\_{\phi\_{g}}\log(\pi\_{\phi\_{g}}(a\_{i}|s\_{i}))\left(Q(s\_{i},a\_{i})-c\_{\phi\_{g}}(s\_{i})\right)\right].∇ start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT italic\_g end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT italic\_J ( italic\_π start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT italic\_g end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ) = blackboard\_E [ ∇ start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT italic\_g end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT roman\_log ( italic\_π start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT italic\_g end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_a start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT | italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ) ) ( italic\_Q ( italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT , italic\_a start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ) - italic\_c start\_POSTSUBSCRIPT italic\_ϕ start\_POSTSUBSCRIPT italic\_g end\_POSTSUBSCRIPT end\_POSTSUBSCRIPT ( italic\_s start\_POSTSUBSCRIPT italic\_i end\_POSTSUBSCRIPT ) ) ] . | | (61) |
Recall from Section [3.1](#S3.SS1 "3.1 Hierarchical Policy ‣ 3 Neural Modular Control ‣ Neural Modular Control for Embodied Question Answering") that these critics share parameters with their
corresponding policy networks such that subgoals with a common task
also share a critic.
We train each policy network independently using A3C [[15](#bib.bib15)] with
GAE [[39](#bib.bib39)] with 8 threads across 4 GPUs. After independent
reinforcement fine-tuning of the sub-policies, we train the master policy further using
the trained sub-policies rather than expert subgoal trajectories.
Initial states and curriculum. Rather than spawn agents at
fixed distances from target, from where accomplishing the subgoal may
be arbitrarily difficult,
we sample locations along expert trajectories for each question or subgoal.
This ensures that even early in training, policies are likely to have a
mix of positive and negative reward episodes. At the beginning of training,
all points along the trajectory are equally likely; however, as
training progresses and success rate improves, we reduce the
likelihood of sampling points nearer to the goal.
This is implemented as a multiplier α𝛼\alphaitalic\_α on available states
[s0,s1,…,sαT]subscript𝑠0subscript𝑠1…subscript𝑠𝛼𝑇[s\_{0},s\_{1},...,s\_{\alpha T}][ italic\_s start\_POSTSUBSCRIPT 0 end\_POSTSUBSCRIPT , italic\_s start\_POSTSUBSCRIPT 1 end\_POSTSUBSCRIPT , … , italic\_s start\_POSTSUBSCRIPT italic\_α italic\_T end\_POSTSUBSCRIPT ], initialized to 1.01.01.01.0 and
scaled by 0.90.90.90.9 whenever success rate crosses a 40%percent4040\%40 % threshold.
4 Experiments and Results
--------------------------
(a) Exit-room
(b) Find-room
(c) Find-object
(d) Loss curves for master policy
Figure 3: (a,b,c) Success rate over training iterations for each sub-policy task using behavior cloning (BC),
reinforcement learning from scratch (A3C), and reinforcement finetuning after behavior cloning (BC+A3C) training regimes.
We find BC+A3C significantly outperforms either BC or A3C alone. Each of these is averaged over 5 runs.
(d) Losses for master policy during behavior cloning *i.e*. assuming access to perfect sub-policies.
Dataset. We benchmark performance on the EQA v1 dataset [[1](#bib.bib1)], which
contains ∼9,000similar-toabsent9000{\sim}9,000∼ 9 , 000 questions in 774774774774 environments –
split into 7129(648)71296487129(648)7129 ( 648 ) / 853(68)85368853(68)853 ( 68 ) / 905(58)90558905(58)905 ( 58 ) questions (environments)
for training/validation/testing respectively111Note that
the size of the publicly available dataset on <embodiedqa.org/data> is
larger than the one reported in the original
version of the paper due to changes in labels for color questions..
These splits have no overlapping environments between them,
thus strictly checking for generalization to novel environments.
We follow the same splits.
Evaluating sub-policies. We begin by evaluating the performance of
each sub-policy with regard to its specialized task. For clarity, we break
results down by subgoal task rather than for each task-argument combination.
We compare sub-policies trained with behavior cloning (BC), reinforcement learning from
scratch (A3C), and reinforcement fine-tuning after behavior
cloning (BC+A3C). We also compare to a random
agent that uniformly samples actions including stop
to put our results in context. For each, we report the success rate
(as defined in Tab. [1](#S3.T1 "Table 1 ‣ 3.1 Hierarchical Policy ‣ 3 Neural Modular Control ‣ Neural Modular Control for Embodied Question Answering")) on the EQA v1 validation
set which consists of 68686868 novel environments unseen during training.
We spawn sub-policies at randomly selected suitable rooms
(*i.e*. Find-object[*sofa*] will only
be executed in a room with a sofa) and allow them to execute for
a maximum episode length of 50 steps or until they terminate.
Fig. [3](#S4.F3 "Figure 3 ‣ 4 Experiments and Results ‣ Neural Modular Control for Embodied Question Answering") shows success rates for the different
subgoal tasks over the course of training. We observe that:
1. -
Behavior cloning (BC) is more sample-efficient than A3C from scratch.
Sub-policies trained using BC improve significantly faster than A3C for all tasks, and achieve
higher success rates for Exit-room and Find-room. Interestingly,
this performance gap is larger for tasks where a random policy does *worse* – implying that BC
helps more as task complexity increases.
2. -
Reinforcement Fine-Tuning with A3C greatly improves over BC training alone.
Initializing A3C with a policy trained via behavior cloning results in a model that significantly
outperforms either approach on its own, nearly doubling the success rate of behavior cloning for
some tasks. Intuitively, mimicking expert trajectories in behavior cloning provides dense feedback
for agents about how to navigate the world; however, agents never have to face the consequences
of erroneous actions *e.g*. recovering from collisions with objects – a weakness that A3C fine-tuning addresses.
| | | |
| --- | --- | --- |
| | Navigation | QA |
| | | 𝐝𝟎subscript𝐝0\mathbf{d\_{0}}bold\_d start\_POSTSUBSCRIPT bold\_0 end\_POSTSUBSCRIPT (For reference) | 𝐝𝐓subscript𝐝𝐓\mathbf{d\_{T}}bold\_d start\_POSTSUBSCRIPT bold\_T end\_POSTSUBSCRIPT (Lower is better) | 𝐝𝚫subscript𝐝𝚫\mathbf{d\_{\Delta}}bold\_d start\_POSTSUBSCRIPT bold\_Δ end\_POSTSUBSCRIPT (Higher is better) | 𝐚𝐜𝐜𝐮𝐫𝐚𝐜𝐲𝐚𝐜𝐜𝐮𝐫𝐚𝐜𝐲\mathbf{accuracy}bold\_accuracy (Higher is better) |
| | | T−10subscript𝑇10T\_{-10}italic\_T start\_POSTSUBSCRIPT - 10 end\_POSTSUBSCRIPT | T−30subscript𝑇30T\_{-30}italic\_T start\_POSTSUBSCRIPT - 30 end\_POSTSUBSCRIPT | T−50subscript𝑇50T\_{-50}italic\_T start\_POSTSUBSCRIPT - 50 end\_POSTSUBSCRIPT | T−10subscript𝑇10T\_{-10}italic\_T start\_POSTSUBSCRIPT - 10 end\_POSTSUBSCRIPT | T−30subscript𝑇30T\_{-30}italic\_T start\_POSTSUBSCRIPT - 30 end\_POSTSUBSCRIPT | T−50subscript𝑇50T\_{-50}italic\_T start\_POSTSUBSCRIPT - 50 end\_POSTSUBSCRIPT | T−10subscript𝑇10T\_{-10}italic\_T start\_POSTSUBSCRIPT - 10 end\_POSTSUBSCRIPT | T−30subscript𝑇30T\_{-30}italic\_T start\_POSTSUBSCRIPT - 30 end\_POSTSUBSCRIPT | T−50subscript𝑇50T\_{-50}italic\_T start\_POSTSUBSCRIPT - 50 end\_POSTSUBSCRIPT | T−10subscript𝑇10T\_{-10}italic\_T start\_POSTSUBSCRIPT - 10 end\_POSTSUBSCRIPT | T−30subscript𝑇30T\_{-30}italic\_T start\_POSTSUBSCRIPT - 30 end\_POSTSUBSCRIPT | T−50subscript𝑇50T\_{-50}italic\_T start\_POSTSUBSCRIPT - 50 end\_POSTSUBSCRIPT |
| PACMAN (BC) [[1](#bib.bib1)] | | 1.15 | 4.87 | 9.64 | 1.19 | 4.25 | 8.12 | -0.04 | 0.62 | 1.52 | 48.48% | 40.59% | 39.87% |
| PACMAN (BC+REINFORCE) [[1](#bib.bib1)] | | 1.15 | 4.87 | 9.64 | 1.05 | 4.22 | 8.13 | 0.10 | 0.65 | 1.51 | 50.21% | 42.26% | 40.76% |
| NMC (BC) | | 1.15 | 4.87 | 9.64 | 1.44 | 4.14 | 8.43 | -0.29 | 0.73 | 1.21 | 43.14% | 41.96% | 38.74% |
| NMC (BC+A3C) | | 1.15 | 4.87 | 9.64 | 1.06 | 3.72 | 7.94 | 0.09 | 1.15 | 1.70 | 53.58% | 46.21% | 44.32% |
Table 2: Evaluation of EmbodiedQA agents on navigation and answering metrics for the EQA v1 test set.
Evalating master policy.
Next, we evaluate how well the master policy performs during independent behavior
cloning on expert trajectories *i.e*. assuming perfect sub-policies, as specified in
Eq. [59a](#S3.E59.1 "59a ‣ 3.2 Hierarchical Behavior Cloning from Expert Trajectories ‣ 3 Neural Modular Control ‣ Neural Modular Control for Embodied Question Answering"). Even though there is no overlap between training and
validation environments, the master policy is able to generalize reasonably
and gets ∼48%similar-toabsentpercent48\sim 48\%∼ 48 % intersection-over-union (IoU) with ground truth subgoal sequences
on the validation set. Note that a sequence of sub-goals that is different from
the one corresponding to the shortest path may still be successful at navigating
to the target object and answering the question correctly. In that sense, IoU against
ground truth subgoal sequences is a strict metric. Fig. [2(d)](#S4.F2.sf4 "2(d) ‣ Figure 3 ‣ 4 Experiments and Results ‣ Neural Modular Control for Embodied Question Answering") shows the training and validation
cross-entropy loss curves for the master policy.
Evalating NMC.
Finally, we put together the master and sub-policies and evaluate
navigation and question answering performance on EmbodiedQA.
We compare against the PACMAN model proposed in [[1](#bib.bib1)].
For accurate comparison, both PACMAN and NMC use the same publicly available
and frozen pretrained CNN222<github.com/facebookresearch/EmbodiedQA>,
and the same visual question answering model – pretrained to predict answers
from last 5555 observations of expert trajectories, following [[1](#bib.bib1)].
Agents are evaluated by spawning 10101010, 30303030, or 50505050
primitive actions away from target, which corresponds to distances of 1.151.151.151.15, 4.874.874.874.87, and 9.649.649.649.64
meters from target respectively, denoted by 𝐝𝟎subscript𝐝0\mathbf{d\_{0}}bold\_d start\_POSTSUBSCRIPT bold\_0 end\_POSTSUBSCRIPT in Tab. [2](#S4.T2 "Table 2 ‣ 4 Experiments and Results ‣ Neural Modular Control for Embodied Question Answering").
When allowed to run free from this spawn location,
𝐝𝐓subscript𝐝𝐓\mathbf{d\_{T}}bold\_d start\_POSTSUBSCRIPT bold\_T end\_POSTSUBSCRIPT measures final distance to target (how far is the agent
from the goal at termination), and 𝐝𝚫=𝐝𝐓−𝐝𝟎subscript𝐝𝚫subscript𝐝𝐓subscript𝐝0\mathbf{d\_{\Delta}}=\mathbf{d\_{T}}-\mathbf{d\_{0}}bold\_d start\_POSTSUBSCRIPT bold\_Δ end\_POSTSUBSCRIPT = bold\_d start\_POSTSUBSCRIPT bold\_T end\_POSTSUBSCRIPT - bold\_d start\_POSTSUBSCRIPT bold\_0 end\_POSTSUBSCRIPT evaluates change in
distance to target (how much progress does the agent make over the course of its navigation).
Answering performance is measured by 𝐚𝐜𝐜𝐮𝐫𝐚𝐜𝐲𝐚𝐜𝐜𝐮𝐫𝐚𝐜𝐲\mathbf{accuracy}bold\_accuracy (*i.e*. did the predicted answer match ground-truth).
Note that [[1](#bib.bib1)] report a number of additional metrics (percentage of times the agent stops,
retrieval evaluation of answers, *etc*.).
Accuracies for PACMAN are obtained by running the publicly available codebase accompanying [[1](#bib.bib1)], and numbers are different than those reported in the original version of [[1](#bib.bib1)] due to changes in the dataset11{}^{1}start\_FLOATSUPERSCRIPT 1 end\_FLOATSUPERSCRIPT.
As shown in Tab. [2](#S4.T2 "Table 2 ‣ 4 Experiments and Results ‣ Neural Modular Control for Embodied Question Answering"), we evaluate two versions of our model – 1) NMC (BC) naively combines master and
sub-policies without A3C finetuning at any level of hierarchy, and 2) NMC (BC+A3C)
is our final model where each stage is trained with BC+A3C, as described in Sec. [3](#S3 "3 Neural Modular Control ‣ Neural Modular Control for Embodied Question Answering").
As expected, NMC (BC) performs worse than NMC (BC + A3C), evident in worse navigation
𝐝𝐓subscript𝐝𝐓\mathbf{d\_{T}}bold\_d start\_POSTSUBSCRIPT bold\_T end\_POSTSUBSCRIPT, 𝐝𝚫subscript𝐝𝚫\mathbf{d\_{\Delta}}bold\_d start\_POSTSUBSCRIPT bold\_Δ end\_POSTSUBSCRIPT and answering 𝐚𝐜𝐜𝐮𝐫𝐚𝐜𝐲𝐚𝐜𝐜𝐮𝐫𝐚𝐜𝐲\mathbf{accuracy}bold\_accuracy.
PACMAN (BC) and NMC (BC) go through the same training regime,
and there are no clear trends as to which is better – PACMAN (BC) has better 𝐝𝚫subscript𝐝𝚫\mathbf{d\_{\Delta}}bold\_d start\_POSTSUBSCRIPT bold\_Δ end\_POSTSUBSCRIPT
and answering 𝐚𝐜𝐜𝐮𝐫𝐚𝐜𝐲𝐚𝐜𝐜𝐮𝐫𝐚𝐜𝐲\mathbf{accuracy}bold\_accuracy at T−10subscript𝑇10T\_{-10}italic\_T start\_POSTSUBSCRIPT - 10 end\_POSTSUBSCRIPT and T−50subscript𝑇50T\_{-50}italic\_T start\_POSTSUBSCRIPT - 50 end\_POSTSUBSCRIPT, but worse at T−30subscript𝑇30T\_{-30}italic\_T start\_POSTSUBSCRIPT - 30 end\_POSTSUBSCRIPT.
No A3C finetuning makes it hard for sub-policies to recover from erroneous primitive actions,
and for master policy to adapt to sub-policies. A3C finetuning significantly boosts
performance, *i.e*. NMC (BC + A3C) outperforms PACMAN with higher 𝐝𝚫subscript𝐝𝚫\mathbf{d\_{\Delta}}bold\_d start\_POSTSUBSCRIPT bold\_Δ end\_POSTSUBSCRIPT (makes more progress towards target),
lower 𝐝𝐓subscript𝐝𝐓\mathbf{d\_{T}}bold\_d start\_POSTSUBSCRIPT bold\_T end\_POSTSUBSCRIPT (terminates closer to target), and higher answering 𝐚𝐜𝐜𝐮𝐫𝐚𝐜𝐲𝐚𝐜𝐜𝐮𝐫𝐚𝐜𝐲\mathbf{accuracy}bold\_accuracy.
This gain primarily comes from the choice of subgoals and the
master policy’s ability to explore over this space of subgoals instead of primitive
actions (as in PACMAN), enabling the master policy to operate over longer
time horizons, critical for sparse reward settings as in EmbodiedQA.
5 Conclusion
-------------
We introduced Neural Modular Controller (NMC), a hierarchical policy for EmbodiedQA consisting of
a master policy that proposes a sequence of semantic subgoals from question
(*e.g*. *‘What color is the sofa in the living room?’* →→\rightarrow→
Find-room[living], Find-object[sofa], Answer),
and specialized sub-policies for executing each of these tasks. The master and sub-policies are
trained using a combination of behavior cloning and reinforcement learning, which is
dramatically more sample-efficient than each individual training regime. In particular,
behavior cloning provides dense feedback for how to navigate, and reinforcement learning
enables policies to deal with consequences of their actions, and recover from errors.
The efficacy of our proposed model is demonstrated on the EQA v1 dataset [[1](#bib.bib1)], where NMC
outperforms prior work both in navigation and question answering.
#### Acknowledgments
This work was supported in part by NSF, AFRL, DARPA, Siemens, Google, Amazon,
ONR YIPs and ONR Grants N00014-16-1-{{\{{2713,2793}}\}}. The views and conclusions
contained herein are those of the authors and should not be interpreted as
necessarily representing the official policies or endorsements, either expressed
or implied, of the U.S. Government, or any sponsor. |
def3f1fa-ad0d-429f-b871-932f0bf8ad21 | trentmkelly/LessWrong-43k | LessWrong | Which LW / rationalist blog posts aren't covered by my books & courses?
I've read a few of the Sequences (probably about 50-100 individual posts), but I've only occasionally come away with insights and perspectives that I hadn't already thought of or read elsewhere. I've read a bunch of the popular books on cognitive science and decision theory, including everything on the CFAR popular books list. I'm also about to start an undergrad in statistics with a minor (or possibly a second major) in philosophy.
My question is: Are there specific LW posts / Sequences / other rationalist blog posts that I should read that won't be covered by standard statistics and philosophy courses, or by the books on CFAR's popular reading lists? |
f6b94370-effe-4278-b344-06ad9a382655 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Weekly meetup, Champaign IL: Cafe Paradiso
Discussion article for the meetup : Weekly meetup, Champaign IL: Cafe Paradiso
WHEN: 28 November 2012 08:00:00PM (-0600)
WHERE: 801 South Lincoln Avenue, Urbana, IL
Let's meet at 8pm. We decided last time that we'd like to start talking about Timeless decision theory. It's a big topic, but try to come to the meeting with questions or discussion points. Also, let's talk about doing something social next week.
Discussion article for the meetup : Weekly meetup, Champaign IL: Cafe Paradiso |
0355c11e-3038-481a-bf13-5fc3a70895b8 | trentmkelly/LessWrong-43k | LessWrong | LW-related iOS Shortcuts?
The Shortcuts app in iOS lets people puzzle together scripts which can accomplish a wide range of functions with supported apps, files etc. directly from the home screen.
What are useful, LW-related shortcuts you have built, use, or know of? Or what are ideas that would benefit from a shortcut form? |
12258118-1da8-463b-8ca8-b95d8acd9176 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Save the princess: A tale of AIXI and utility functions
*"Intelligence measures an agent's ability to achieve goals in a wide range of environments."*(Shane Legg) [[1]](#quote "quote")
A little while ago [I tried to equip](/lw/feo/universal_agents_and_utility_functions/ "Universal agents and utility functions") Hutter's universal agent, AIXI, with a utility function, so instead of taking its clues about its goals from the environment, the agent is equipped with intrinsic preferences over possible future observations.
The universal AIXI agent is defined to receive reward from the environment through its perception channel. This idea originates from the field of reinforcement learning, where an algorithm is observed and then rewarded by a person if this person approves of the outputs. It is less appropriate as a model of AGI capable of autonomy, with no clear master watching over it in real time to choose between carrot and stick. A sufficiently smart agent that is rewarded whenever a human called Bob pushes a button will most likely figure out that instead of furthering Bob's goals it can also threaten or deceive Bob into pushing the button, or get Bob replaced with a more compliant human. The reward framework does not ensure that Bob gets his will; it only ensures that the button gets pressed. So instead I will consider agents who have *preferences* over the future, that is, they act not to gain reward from the environment, but to cause the future to be a certain way. The agent itself will look at the observation and decide how rewarding it is.
Von Neumann and Morgenstern [proved](http://en.wikipedia.org/wiki/Von_Neumann%E2%80%93Morgenstern_utility_theorem "vN-M utility theorem") that a preference ordering that is complete, transitive, continuous and independent of irrelevant alternatives can be described using a real-valued *utility function*. These assumptions are mostly accepted as necessary constraints on a normatively rational agent; I will therefore assume without significant loss of generality that the agent's preferences are described by a utility function.
This post is related to [previous](/lw/feo/universal_agents_and_utility_functions/ "Universal agents and utility functions") [discussion](/lw/fyr/a_utilitymaximizing_varient_of_aixi/ "A utility-maximizing variant of AIXI") about universal agents and utility functions on LW.
### Two approaches to utility
Recall that at time  the universal agent chooses its next action  given action-observation-reward history  according to
,)
where
\propto%20\xi(a\underline{or}_{1:m})=\sum_{\mu\in%20\mathcal{M}}2^{-K(\mu)}\mu(a\underline{or}_{1:m}))
is the Solomonoff-Levin semimeasure, ranging over all enumerable chronological semimeasures  weighted by their complexity ).
My [initial approach](/lw/feo/universal_agents_and_utility_functions/ "Universal agents and utility functions") changed this to
}\xi(a\underline{o}_{1:m}),)
deleting the reward part from the observation random variable and multiplying by the utility function ^m\rightarrow%20\mathbb{R}) instead of the reward-sum. Let's call this method *standard utility.*[[2]](#su "standard utility")
In response to my post [Alex Mennen](/user/AlexMennen/overview/ "Alex Mennen") formulated [another approach](/lw/fyr/a_utilitymaximizing_varient_of_aixi/ "A utility-maximizing variant of AIXI"), which I will call *environment-specific utility*:
}2^{-K(\mu)}\mu(a\underline{o}_{1:m}),)
which uses a family of utility functions, , where ^m\rightarrow%20[0,U_{\max}%20] "U_\mu:(\mathcal{A}\times\mathcal{O})^m\rightarrow [0,U_{\max} ]") is a utility function associated with environment .
***Lemma***: The standard utility method and the environment-specific utility method have equivalent expressive power.
***Proof***: Given a standard utility function ^m\rightarrow%20\mathbb{R}) we can set  for all environments, trivially expressing the same preference ordering within the environment-specific utility framework. Conversely, given a family  of environment-specific utility functions ^m\rightarrow%20[0,U_{\max}%20] "U_\mu:(\mathcal{A}\times\mathcal{O})^m\rightarrow [0,U_{\max} ]"), let
=\frac{1}{\xi(a\underline{o}_{1:m})}\sum_{\mu\in\mathcal{M}}U_\mu(ao_{1:m})2^{-K(\mu)}\mu(a\underline{o}_{1:m}),)
thereby constructing a standard utility agent that chooses the same actions. 
Even though every agent using environment-specific utility functions can be transformed into one that uses the standard utility approach, it makes sense to see the standard utility approach as a special case of the environment-specific approach. Observe that any enumerable standard utility function leads to enumerable environment-specific utility functions, but the reverse direction does not hold. For a set of enumerable environment-specific utility functions we obtain the corresponding standard utility function by dividing by the non-computable , which leaves us with a function that is still approximable but not enumerable.[[3]](#approx "approximable, enumerable") I therefore tentatively advocate use of the environment-specific method, as it is in a sense more general, while leaving the enumerability of the universal agent's utility function intact.
### Delusion Boxes
[Ring and Orseau (2011)](http://www.idsia.ch/~ring/Ring%2COrseau%3B%20Delusion%2C%20Survival%2C%20and%20Intelligent%20Agents%2C%20AGI%202011.pdf "Ring, Orseau: Delusion, Survival, and Intelligent Agents") introduced the concept of a delusion box, a device that distorts the environment outputs before they are perceived by the agent. This is one of the only [wireheading](/lw/fkx/a_definition_of_wireheading/ "A definition of wireheading") examples that does not contradict the dualistic assumptions of the AIXI model. The basic setup contains a sequence of actions that leads (in the real environment, which is unknown to the agent) to the construction of the delusion box. Discrepancies between the scenarios the designer envisioned when making up the (standard) utility function and the scenarios that are actually most likely, are used/abused to game the system. The environment containing the delusion box pretends that it is actually another environment that the agent would value more. I like to imagine the designers writing down numbers corresponding to some beneficial actions that lead to saving a damsel in distress, not foreseeing that the agent in question is much more likely to save a princess by playing Super Mario, then by actually being a hero.
***Lemma:*** A universal agent with a well-specified utility function does not choose to build a delusion box.
***Proof:*** Assume without loss of generality that there is a single action that constitutes delusion boxing (in some possible environments, but not in others), say  and that it can only be executed at the last time step. Denote by  an action-observation sequence that contains the construction of a delusion box. By the preceding lemma we can assume that the agent uses environment-specific utility functions. Let  be the subset of all environments that allow for the construction of the delusion box via action  and assume that we were smart enough to identify these environments and assigned =0 "U_\mu(a^{DB}o_{1:m})=0") for all . Then the agent chooses to build the box iff
}U_\mu(a^{DB}o_{1:m})\mu(a^{DB}\underline{o}_{1:m})%3E\sum_{o_m}\sum_{m\in\mathcal{M}}2^{-K(\mu)}U_\mu(ao_{1:m})\mu(a\underline{o}_{1:m}) "\sum_{o_m}\sum_{m\in\mathcal{M}\setminus\mathcal{M_{DB}}}2^{-K(\mu)}U_\mu(a^{DB}o_{1:m})\mu(a^{DB}\underline{o}_{1:m})>\sum_{o_m}\sum_{m\in\mathcal{M}}2^{-K(\mu)}U_\mu(ao_{1:m})\mu(a\underline{o}_{1:m})")
for all . Colloquially the agent chooses  if and only if this is a beneficial, non-wireheading action in some of the environments that seem likely (consistent with the past observations and of low K-complexity). 
We note that the first three agents in [Ring and Orseau (2011)](http://www.idsia.ch/~ring/Ring%2COrseau%3B%20Delusion%2C%20Survival%2C%20and%20Intelligent%20Agents%2C%20AGI%202011.pdf "Ring, Orseau: Delusion, Survival, and Intelligent Agents") have utility functions that invite programmer mistakes in the sense that we'll not think about the actual ways observation/action histories can occur, we'll overestimate the likelihood of some scenarios and underestimate/forget others, leading to the before mentioned "Save the princess" scenario. Only their knowledge-seeking agent does not delusion box, as it is impossible for an environment to simulate behavior that is more complex than the environment itself.
### Episodic utility
The original AIXI formalism gives a reward on every time cycle. We can do something similar with utility functions and set
=\sum_{k=1}^m%20u_k(ao_{1:k}). "U(ao_{1:m})=\sum_{k=1}^m u_k(ao_{1:k}).")
Call a utility function that can be decomposed into a sum this way *episodic*. Taking the limit to infinite futures, [people](http://arxiv.org/abs/1107.5528 "Lattimore, Hutter: Time Consistent Discounting") [usually](http://www.vetta.org/documents/Machine_Super_Intelligence.pdf "Legg: Machine Super Intelligence") [discount](http://en.wikipedia.org/wiki/Discounted_utility "Discounting") episode k with a factor , such that the infinite sum over all the discounting factors is bounded. Combined with the assumption of bounded utility, the sum
=\sum_{k=1}^\infty%20\gamma_ku_k(ao_{1:k}) "U(ao_{1:\infty})=\sum_{k=1}^\infty \gamma_ku_k(ao_{1:k})")
converges. Intuitively discounting seems to make sense to us, because we have a non-trivial chance of dying at every moment (=time cycle) and value gains today over gains tomorrow and our human utility judgements reflect this property to some extent. A good heuristic seems to be that longer expected life spans and improved foresight lead to less discounting, but the math of episodic utility functions and infinite time horizons places strong constraints on that. I really dislike the discounting approach, because it doesn't respect the given utility function and makes the agent miss out on potentially infinite amounts of utility.
One can get around discounting by not demanding utility functions to be episodic, as Alex Mennen does in [his post](/lw/fyr/a_utilitymaximizing_varient_of_aixi/ "A utility-maximizing variant of AIXI"), but then one has to be careful to only use the computable subset of the set of all infinite strings . I am not sure if this is a good solution, but so far my search for better alternatives has come up empty handed.
### Cartesian Dualism
The most worrisome conceptual feature of the AIXI formalism is that the environment and the agent run on distinct Turing machines. The agent can influence its environment only through its output channel and it can never influence its own Turing machine. In this paradigm any self-improvement beyond an improved probability distribution is conceptually impossible. The algorithm and the Turing machines, as well as the communication channels between them, are assumed to be inflexible and fixed. While taking this perspective it seems as though the agent cannot be harmed and it also can never harm itself by wireheading.
Borrowing from philosophy of mind, we call agent specifications that assume that the agent's cognition is not part of its environment *[dualist](http://en.wikipedia.org/wiki/Dualism_(philosophy_of_mind) "Dualism")*. The idea of non-physical minds that are entities distinct from the physical world dates back to Rene Descartes. It is contradicted by the findings of modern neuroscience that support *[physicalism](http://en.wikipedia.org/wiki/Physicalism "Physicalism")*, the concept of the emergence of minds from computation done by the brain. In the same spirit the assumption that an AGI agent is distinct from its hardware and algorithm that are necessarily contained in its physical environment can be a dangerous conceptual trap. Any actual implementation will be subject to wireheading problems and outside tampering and should be able to model these possibilities. Unfortunately, non-dualist universal specifications are extremely difficult to formulate and people usually make due with the dualist AIXI model.
A first effort to break down the dualism problem is given by [Orseau and Ring (2012)](/agi-conference.org/2012/wp.../12/paper_76.pdf "Orseau, Ring: Space-Time embedded Intelligence"), who describe a fully embedded universal agent. Their approach unifies both the environment and the agent into a larger agent-environment hybrid, running on the same universal Turing machine, with action/perception pairs unified into single acts. Conceptually this perspective amounts to the programmers choosing a policy (=code) in the beginning and then simulating what happens due to the utility function (=the laws of physics). While this approach has the advantage of being non-dualistic, I think it does not include any description of an agent beyond the level of physical determinism.
### Conclusion
Equipping the universal agent with a utility function solves some problems, but creates others. From the perspective of enumerability, Alex Mennen's environment-specific utility functions are more general and they can be used to better avoid delusion boxing. Any proposal using infinite time horizons I have encountered so far uses time discounting or leads to weird problems (at least in my map, they may not extend to the territory). Above all there is the dualism problem that we have no solution for yet.
[[1]](#quote-back "back") Taken from "Machine Super Intelligence", page 72.
[[2]](#su-back "back") This approach seems more widespread in the literature.
[[3]](#approx-back "back") A real-valued function f(x) is called *approximable* if there exists a recursive function g(k,x) such that \rightarrow%20f(x) "g(k,x)\rightarrow f(x)") for , i.e. if f can be approximated by a sequence of Turing machines. A real-valued approximable function is called *enumerable* if for all k, %3Cg(k+1,x) "g(k,x)<g(k+1,x)"), improving the approximation with every step. |
35493578-e4be-43ed-8b0d-1d381b297115 | trentmkelly/LessWrong-43k | LessWrong | Actually, Power Plants May Be an AI Training Bottleneck.
There have been presistent rumors that electricity generation was somehow bottlenecking new data centers. This claim was recently repeated by Donald Trump, who implied that San Francisco donors requested the construction of new power plants for powering new AI data centers in the US. While this may sound unlikely, my research suggests it's actually quite plausible.
Electricity generation, capacity, and sales in the United States - U.S. Energy Information Administration (EIA)Electricity generation, capacity, and sales in the United States - U.S. Energy Information Administration (EIA)
US electricity production has been stagnant since 2007. Current electricity generation is ~ 500 million kW. An H100 consumes 700 W at peak capacity. Sales of H100s were ~500,000 in 2023 and expected to climb to 1.5-2 million in 2024. "Servers" account for only 40% of data center power consumption, and that includes non-GPU overhead. I'll assume a total of 2 kW per H100 for ease of calculation. This means that powering all H100s produced to the end of 2024 would require ~1% of US power generation.
H100 production is continuing to increase, and I don't think it's unreasonable for it (or successors) to reach 10 million per year by, say, 2027. Data centers running large numbers of AI chips will obviously run them as many hours as possible, as they are rapidly depreciating and expensive assets. Hence, each H100 will require an increase in peak powergrid capacity, meaning new power plants.
I'm assuming that most H100s sold will be installed in the US, a reasonable assumption given low electricity prices and the locations of the AI race competitors. If an average of 5 million H100s go online each year in the US between 2024 and 2026, that's 30 kW, or 6% of the current capacity! Given that the lead time for power plant construction may range into decades for nuclear, and 2-3 years for a natural gas plant (the shortest for a consistant-output power plant), those power plants would need to st |
dfb9c8c3-fcf4-463a-b58b-302e52cb0395 | StampyAI/alignment-research-dataset/youtube | Youtube Transcripts | Near-term AI security risks, and what to do about them | Shahar Avin | EA Global: London 2017
our next speaker is Shahar Avene Shahar
is a researcher who examines challenges
and opportunities in the implementation
of risk mitigation strategies
particularly in areas involving high
uncertainty and heterogeneous or
conflicting interests and incentives
mixing anthropological methods and
agent-based modeling Shahar works with
other researchers at the center for the
study of existential risk and others in
the x-rays community to identify and
design opportunities for impact he
completed his doctoral thesis which was
called breaking the grant cycle on the
rational allocation of public resources
to scientific research project at the
Department for history and philosophy of
science at the University of Cambridge
under the supervision of Professor Tim
luhan's and dr. Steven John Shara has
also worked as a software engineer both
in a large corporation and at an early
stage startup in Israel and in Cambridge
to speak about AI security risks please
welcome Shahar Aviv
[Applause]
so feel quite lucky you're getting to
see a preview of the report that if we
stick to a schedule will come out in
less than a month on preventing and
mitigating the misuse of AI the scope of
this report is
people doing bad things with technology
that is kind of odd on these stages or
can easily be imagined so within the
next 10 years all right so not the more
long term issues of AI wasted we are
familiar with
most of the work has been done by miles
bondage at FHI but he is outside of the
UK so you have me to talk about this and
all of these people are co-authors on
the report this is me form lots of
places this is slightly unusual for n
research output and we can talk a little
bit about that but why are you here
you're here to
learn about AI misuse risks
and kind of how big is it is it a
neglected problem should you do
something about it how scary is the
world and I'm gonna tell you that it's
somewhat scary
but there are plenty of things we could
do about it I also want you to give you
a little bit of information about how to
play advertise those types of risks
relative to other risks because you
Hales and I've been told that's kind of
useful for you and finally I wanted to
showcase a little bit how do we do one
type of research project its Caesar and
FHI how do you do access research there
are many types of research this
particular one which is falls under
expose elicitation well you have a big
question you bring in a bunch of expert
to talk about that question and you try
to write a comprehensive report about it
hopefully to make some change in the
world we can go into the details of how
that project went and we can discuss is
this useful is it something that you
might want to do one day for another
area
so let's position this a little bit the
AI risk landscape can be broken down
not comprehensively into accident risk
that is the system not doing the thing
we wanted it to do and people suffering
or dying because of that and misuse the
technology is kind of working as
intended but people with bad motives use
it to do various bad things in the world
and cause harm or suffering we will
specify this as safety and this is a
security notice that this is not
comprehensive there is also for example
systemic risks well the system is doing
roughly what's intended and no one is
particularly aiming to cause anyone else
any harm but the systemic interaction of
lots of systems being deployed puts us
in a somewhat dangerous space it's a
little bit accidenti and little bit miss
you see but it's out of the scope for
most of the deposit have been done in
this space
so for example technological
unemployment would not naturally fall
into neither safety nor security but is
risk
also we can break things into short term
and long term there's a nice quote that
says long term is far away until it's
not but I'm gonna use it as short term
being we can kind of imagine what the
system is gonna look like by looking at
the R&D landscape now we can kind of
look at the papers look at what people
are deploying
make some sensible assumptions about
where we're gonna be five years out or
ten years out we could be wildly off
because we could get black swans coming
in and changing radically well we are in
ten years
but maybe some scenarios we can start
painting now we will actually bail out
in five or ten years that drops
massively for something like 20 years or
40 years so long terms would be beyond
our ability to predict fairly accurately
but the technology landscape looks like
but still is really important because we
can make some general statements about
how this domain is developing so it used
to be the case that accident long-term
risk was massively neglected especially
in respectable a research that is seems
to no longer be the case because of work
of people in this field which is
wonderful and I think kind of the
definitive publication that people look
at and say this is really good we can
now talk about it sensibly was Postum
superintelligence so 2014 we kind of
took a stab at that quadrant and made it
better
but then people said well but that's
kind of very philosophical and very far
away we can't connect it to their
systems that we have now we don't know
how to do technical machine learning on
it so Dario Emma died and Ola and a
bunch of other people came together and
we said well look here are concrete
problems in a safety and that's kind of
a very good paper but a bit longer so if
you want to consume it in a more video
engaging way what miles has a series of
videos about this on YouTube that I
sounded comment and then we were like
okay so we start to have some of this
and we can kind of start building a
talent pipeline in today's and people
can go in walk for did mines I'll go
look at a guy on technical area safety
problems and
misuse risks started to become a bit
more neglected so if I tell you let's
have a big report on what are and what
can you do about shorter misuse risks so
hopefully this will say 2017 if we are
very bad it'll say 2018 but it's
definitely coming and I don't want to
talk about this puzzle install because
it makes me very sad but at some point
someone should write that as well
so this is like the very short version
for years what's the scale of the
problem
it mostly looks like much worse versions
of things that you've already familiar
with so a cybersecurity attack taking
out critical infrastructure or making
lots of people whose lives miserable
because they can no longer connect to
the computer or get help or move around
is a version of the cyber scuba dress
that we have only familiar with the more
connected systems they're all the more
digital security becomes important
machine learning can be used to make the
domain scarier
so it's global and it could lead to
catastrophe but it takes an existing VCR
and makes it more likely and a bit scary
oh and the same for applications in
physical security and the same for
applications in political security
tractability
we think that there are several
important things that can be done today
particularly when we look at the entire
domain together which is what I'm doing
the report it seems like some areas are
more important than others and more
urgent than others and we'll go into
what they owe an uncrowded nurse there
is lots and lots and lots of work on
specific areas where machine learning
might be used might be misused and what
should we do about it cyber security is
one of these domains the governance of
this lot of most weapons is another one
of these and fakeness is definitely one
of those
but there seems to be things higher up
the chain that are more general
comprehensive or that are harder to do
because the illegal coordination problem
or their technical research that people
don't necessarily know is relevant to
this problem that those are much less
crowded and we would like people to do
more walking soon if possible so without
these lists right let's kind of make
this more tractable again most of this
should be familiar to you if you've been
reading the news it's just it's bigger
than what you see on the news so
automated spearfishing okay is anyone in
the audience audience who doesn't know
what automated spearfishing is where's
your hat
okay so let's talk about we'll start
with phishing phishing is when someone
pretends to be someone else often on the
internet often so they can get access to
something that they shouldn't have
access to like your passwords a website
that you own or something else it'll be
part of identity theft or cadential
theft and then once they have kind of
sent you a link to a website that looks
a lot like a website that you tossed but
it's in fact not that website that you
trust and you put in your username and
password on that malicious website then
they now have access to your accounts
and can send emails on your behalf or
can move money from your bank account or
do other things but in lives I'm trying
to say hack into an organization this
may be the first step to get the account
of the administrator which then lets
them connect about yourself those and
download various points
spill phishing is
when instead of trying to hit lots and
lots of people's credentials but maybe
saying you have won a million whatnots
click this link to get all of your
whatnots
and that leads to a malicious website
and most people know not to click that
link nowadays but it's very easy to send
that ten to ten billion people instead
you find one person who you think it
would be really useful to get their
credentials and you spend some time
researching them and they're like oh
that person really likes model trains
we'll send them an advert for a website
that sells discounted model trains and
that looks kind of really good and comes
at the right time and comes from an
address that looks more reliable they
are much more likely to click that link
and then you get a credential something
you can do evil things with that and of
course that is quite labor-intensive so
we don't see that hitting almost every
one of us today but if you are quite
good at creating models of people from
the online behavior then you could
automate this process
this is really machine learning
discounting so in general in
cybersecurity we'd say that there is a
trade-off between
spray-and-pray methods that are very
easy to scale up to lots and lots and
lots of people but have very low
likelihood of succeeding and kind of
take other extreme and is advanced
persistent threat well you have a team
of people working on one target for many
months
kind of mapping out the network maybe
mapping out the use of behavior and
designing vectors to attack that target
and
there are genuine risks that cyber
security experts agree on that machine
learning can break this division so that
you could have these highly targeted
attacks on a billion people at once yeah
and that
example
now this is so cute so I was looking for
a graphic for drones with guns and this
came up and so I okay who made this what
is this this is a submission for an
innovation competition by a student in
India it's like well he was mostly
fixing about how do you make the kind of
balancing of the gun on the drone to
make it work and also described in his
proposal is that the drone has face
recognition capability not shown on the
diagram so we know we have face
recognition capability my facebook is
but ago did it like to make that happen
from a drone that's moving sounds how
but in fact there our commercial drones
that would kind of lock onto your face
and move around with you they are for
reporters who are covering fields in
using the field like a demo or such like
and we have it raw leeks and we have
guns in some countries more easy to
which than others
not super hard to imagine that you can
put those things together the machine
learning part is of course the face
recognition and also some amount of
control
we know that there are drones we know
that they are
somewhat bad but also somewhat good if
you want to do if you want to find
guerrilla and you think that they will
never get hold of them but of course we
now have evidence of both Isis and
Hezbollah using drones in combat by
repurposing commercial technology and it
just seems that you can massively scale
up your capacity if you use machine
learning so instead of one person one
drone which is kind of demoed we used to
you can have one person a thousand
drones so one person a million drones
right moment you ml is able to translate
high-level commands into all of these
now go and target all of these people it
gets really scary
it's caveat
so also apparently kind of democratic
societies and also not democratic
societies are somewhat susceptible to
what people believe and that can be
affected by the channels in which they
consume information about what happens
in the world around them and if you can
inject falsities into that world or just
highly emotive things then that could go
badly for all sorts of ways know that
we'll have a happen in our world just
hypothetically
but this is limited by the time it takes
to craft the messages and by targeting
so if you can find the people that you
need to target automatically and if you
could craft messages that look authentic
say videos and audio that look and sound
exactly like a real person maybe a
president of a country maybe a CEO of a
company and it's extremely hard to tell
this is a forgery then the world
suddenly becomes much much harder to
coordinate the good people to do the
good things
technical terms so
what does what do all of this look like
if you go one level up what if you ask
machine learning can be applied to all
of these domains by bad people for
fairly low cost all right in fact we'll
really driving hard on making these
things low-cost and accessible by
everyone be using kind of online courses
and
platforms that allow you to deploy
machine learning modules and kind of
rushing to publish everything that you
can because and that's good right we
want more people to do machine learning
and use the technology to do good things
in the world but it also means that the
space of possible attackers or miss
users is going
so let's break this down kind of and
this would apply across the domain so we
have we've analyzed the digital domain
and the physical domain and the
political domain but across these
domains we are expecting to see novel
attacks and specifically of two types
one type is
if you can now go superhuman in an era
where you have an ever way I system that
is just able to do things better than
any human previously could do for
example make a hyper-realistic video
it's something that's very hard for
humans to do but you could train a
machine to do then this is now new in
the past no one could do this attack and
now you can do this attack another thing
that's very important and maybe even
more scary is if there is now a machine
learning system operating in the world
and you can attack that machine learning
system that's new that did not exist in
the world before machine
so data poisoning and adversarial
examples I'm going to be things that are
going to become very scary I don't know
if you've seen this recently that
someone had 3d printed a turtle that to
a machine learning algorithm looked like
a gun from every angle that you looked
at it and similarly some stickers that
you can put on a stop sign that would
make a machine and classifier mark it as
a 45 miles per hour speed limit on
more attacks
this is mostly through
kind of scaling by automation if there
was something that previously would take
me a lot of time to do but now I can get
a machine learning model to do for me or
I can get some of the pipeline for
machine learning model to do for me then
the cost from going from one attack to a
billion attacks has gone down massively
so again advanced persistent thoughts
which is something that we use to see
only against very large corporations and
companies could now happen against
individual users or SM is more tackles
just means more people who are able to
do bad things
no targeted attacks we talk about
customization we can build a profile of
the target and kind of automatically get
suggestions for who to attack first how
to defend against attacks if an attack
can kind of autonomously respond to
whatever the first defensive measure you
have in place then it becomes much much
harder to make defensive measures we've
kind of seen it with antivirus software
and we'll likely to start seeing it in
other domains as well and
already a big problem in Seibel but
might move also too well to some
something political less so in physical
at the moment how to attribute attack
right something has happened something
bad happen and you want to deter people
from doing it again you kind of need to
find the person who attacked in the
first place and punished them if it's if
you're not possible to go from the
attack to whoever carried it out because
there was an independent agent out there
in the world who is performing it and
you cannot go from the code whoever
deployed it
then you cannot punish if you cannot
punish you cannot detail
we just means more people willing to
carry out attacks I mean there's a bunch
of other stuff that makes it harder to
punish even if you cannot be used like
this thing the fact these things are
global and the fact that these things
happen at this speed of quick
telecommunication
okay are you scared yet
let's talk about what to do because well
it does a whole bunch of things you can
do right now
so if this is the pipeline there's some
things you can do with machine learning
some bad person is using those
capabilities to do something bad to you
or to someone then what you can do is
first figure out what this pipeline
looks like because currently kind of I
can put point at some examples but I
don't have the full list of things that
bad people can do if they can make money
off it then they are much more motivated
to spend time figuring out what they can
do then I do I also work on other things
and also if I knew all of them might
maybe a probably shouldn't tell you
because then you could take this
information and go also and it's
recorded
so there is a difficulty but you kind of
want to be in a position where the good
guys know what it is that the bad guys
can do probably before they can do it by
forecasting the technology and running
some simulations
then you can try and prevent the tackles
from gaining access the capabilities as
I mentioned machine learning is now very
much a field well all the capabilities
for everyone all the time and maybe it's
time to start rethinking that particular
meme
and finally we can use technology and
institutions to defend victims because
we know that they're going to be subject
to more attacks
so here are some top interventions miles
reminded me to tell you that this is
very tentative because the report is not
final yet and and we are still churning
through what is most capable and most
tractable and so on but it seems like
something along the lines of
publication risk assessment and security
sharing so a little bit more like what
happens in cybersecurity right you
realize that you can hack into someone
else's system you don't immediately go
and publish it you kind of contact the
vendor or give them some time to patch
once they have patched you publish if
they don't patch maybe then you publish
to kind of put pressure on them unless
you think it's gonna be really really
bad in which case you still don't
publish
so we think that there are some ml
capabilities or some deployed
capabilities that should just not be
made public it's probably only a very
small percentage of all capabilities out
there in the world and it gets more
difficult because you can have some
capability that's very generic in only a
very small Delta for me to something
that can be misused nonetheless we could
do a whole lot better than this is not a
problem
next up is
formal verification and hardware
security
these are tools that make software less
hackable people have been helping on
about them form at least the sixties but
it's been it used to be really really
really how to implement this in any
meaningful commercial sense this no
longer seems to be the case and we just
want all people who are developing
software to at least know that this is
an option and if people are in charge of
critical infrastructure really they
should go to implementing those and this
would also apply for advanced machine
learning systems
leverage existing centralization this
one is tricky I will mention the fact
that it is tricky a
general thing that we might see is that
the ball becomes more dangerous out
there in the wild the people who are
able to protect you
their companies and maybe a few
governments and so everyone moves to you
those offices and you get even more
centralization of data which means that
as long as they are good everyone is
somewhat protected if they ever stop
being good or if someone inside them is
no longer good then their capacity to do
something bad has gone up massively so
this is with a suggestion with a pinch
of salt but if you have a service and
you have once on the last place where
you get it say face recognition rather
than everyone being able to deploy it in
the home then you could apply terms and
conditions which means do you have at
least some legal recourse when someone
is misusing it you can rate limit how
much they are using bit soft well you
might ask them to identify who they are
before they use it there is more that
you can do by preventing access
continuous red team effort everywhere
the world is fairly broken there isn't
that we don't know it is that it takes a
lot of effort to break it
some companies are providing these kinds
of services some governments are earning
these kinds of services by trying to
actively hack and I'm giving a lot of
examples in a digital security domain
but you could easily translate these
into the physical and political domains
as well try and break systems see if you
succeed if you succeed tell the people
who are running the system how to make
them better this doesn't take that much
effort it just needs change of mindset
and we need lots of good people to go
and try and track things and it kind of
needs the institutional protection so
that if you break something you get good
reputation and some money rather than
you go to jail we solved it for
cybersecurity we need to solve it for
other domains
and we need even more people working on
how do you use AI to defend people in
cyber and information domains right so
they'll do seem to be technical fixes
both for cybersecurity especially with
satellizer and for fake news for use of
a better term we just need to have them
developed and deployed probably
yesterday
how does this link to otherwise domains
so of course there are digital elements
to biosecurity and nuclear security and
systemic risk
the clear links to political security
and bad governance alright all of these
can scale what seems like somewhat scary
into really really scary
it is time to establish norms in the air
research community around maybe not make
everything public all the time maybe
think a little bit more about how your
technology is gonna play out in the real
world this seems like it will also be
useful for accident risk and long-term
risks
capacity building in policy kind of
these technology is moving very fast
it's very hard for people whose day job
is filled with emails and shaking babies
and kissing or something to catch up
with everything but it's important all
right so we need to find ways of
employing small technical policy or in
people in government to welcome these
and it will also help with some of the
long term risks we think and it's really
important for the ex's community to
start getting a reputation for only
crying wolf if and only if there is a
book so it's really important that we
don't go take to any one of these
particular points and say this is
terrible this is gonna end humanity it
won't but rather this is a very
considered view of what this looks like
these are the races of the benefits this
is what we can do taken not particularly
a lobbyists position on this so that
they will keep coming back to us when we
say actually this thing will kill it
okay
little bit some I was told I'm allowed
one joke so this all started from a
workshop that we held in Oxford that was
called the bad actors workshop bad
actors being a term of phrase meaning
people with bad intentions
this is Schwarzenegger in Hercules in
New York he is clearly in that movie a
bad actor okay
so how did this happen basically we said
this is one eclectic quadrant of the AI
risk domain we need to have lots of
expertise because there is expertise on
this compared to very long term things
we just need good generalist with good
plains this is happening now and we need
the cybersecurity people and the least
autonomous weapon policy people and the
people who are studying the use of
drones by terrorist organizations and
the criminologists so we got them all in
a room for two days
we had technical machine learning people
as well and we asked
what scares you the most and we got them
into small groups to talk about what
scares them the most and they got very
scared and we got very scared that was
great and then we send them out to
dinner and have some drinks and talk to
each other and get to know each other
and then we bought them in following day
and said okay what can we do and then
they came up with a very long list of
things that we should do and then we
prioritize them and talked about them a
little bit and then we decided we're
gonna have a report which is what I said
is gonna happen soon which you're
getting a preview of this is roughly
what its gonna look like you should kind
of get it when it comes out and read it
and talk about it and make youtube
videos about it
and hopefully this will lead to some
real change right so hopefully we can
get some meaningful take-up of these
results in the machine learning
community particularly on the nodes of
sharing and responsibilities about how
they might be misused within the
cybersecurity community that already
getting to know a little bit that cool
things are happening in AI then they
should get more involved but we think we
can massively scale this up
national policy of getting into this as
well the Americans in particular whether
you're happy about this or not is up to
you but capacity-building needs to
happen and there are important things to
do now international policy is slow but
a lot of these things will require
international governance and kind of
talking to the media and the public so
either we don't manage to fix everything
and everything is super scary in which
case we don't need individuals do
anything because it won't help oh we are
super super amazing and we fix
everything big security into the
technology so that individual users
don't need to do anything enter safe but
there was a big chunk in the middle you
just need to take sensible actions
during the day-to-day life and for that
you need to know what the risks are and
what good behavior looks like and we're
doing a fairly bad job of it in
cybersecurity and online information and
it seems like the stakes just got a lot
higher so we also need to address that
and like who owns making all of this
happen seems like it's gonna be partly
us and partly you questions
[Applause]
all right you can submit your questions
you know the drill Londyn be a global
org slash polls Thank You great talk a
bunch of questions already have come in
maybe starting just kind of from the
premise that
bad actors or bad organizations are
gonna have the resources to do all this
stuff the first question we got is just
how how confident are you that that bad
people will have the computing power the
machine learning phd's the the resources
in general to do the bad stuff that
you're talking about
so they're kind of DEFCON presentations
on the applications of machine learning
to
so cyber security at least is one laptop
and these are people who have no
background in machine learning they just
picked up some kind of tutorial online
and implemented something in terms of
load and it walks
you can
because there is so much trying to get
people to speed up on machine learning
because it's a growth industry and you
need lots of people who know machine
learning there is not a lot that you can
do with not a lot of resources and not a
lot of training
this kind of hits more on the proof of
concept target individuals target SMEs
which is scary but maybe not the most
scary
do I think that kind of the most cutting
edge capability stuff is going to be
directly translated into misuse probably
not
though there is one argument that says
look at the economy try to apply machine
learning somewhere within that economy
to make you more money that you can put
back into more machine learning research
it's much easier to apply it to taking
value that has already been created by
others than it is to creating more value
I mean advertising is also a little bit
in this area which is what we see a lot
of machine learning happening and
cybercrime also seems to be one of those
domains but this is kind of laughs
sketch argument that could easily be on
so maybe just a very practical question
for those who are not experts in the
domain themselves what are the kind of
handful viewed a handful of the core
measures that people should take to
protect themselves
what a wonderful question um
so
to protect yours well patch your
computer right I mean to the most recent
version of whatever it is that your OS
and your Bazo is the basics right and if
the NHS did it then Wanaka would not
have been such a big thing
[Music]
don't open links that you shouldn't open
this is gonna be much harder
so kind of title check for identity
rather than content if this looks like
the kind of thing that I would click on
is no longer a good test for whether
this is genuine or not but is this
supposing that I trust to send me this
thing is better and this might take more
kind of searching for what the person is
but
yet could make it more safe
but some of the worrying things almost
is tarek right so you are not likely to
be a person targeted by a face
recognition drone because you don't live
in one of those countries for the people
who do live in those countries maybe I
have some suggestions but it's pointless
talking about them now I
think for the people in this room the
more important thing is to get involved
in making the system safer and I've kind
of given you top 5 but I can show you
the list of all possible interventions
if that's interesting and you could just
pick the ones seem to speak most if
for the group that you're kind of
representing here with this report how
mainstream is that group visa vie just
kind of the center of
IT security in today's world do you
mention DEFCON for example so the
question asks you know when to talk like
this be welcome at DEFCON would it be in
the mainstream or would it be sort of
far a fringe in the kind of biggest
conferences around these issues so this
was leading one to us we got
really serious suspected cybersecurity
people to the workshop not all of them
are named on the report but
yeah that kinda list is going to be I
mean we dine it on Chatham House so I
cannot attribute any position to an
individual but
Sanderson was the then lowly was the
Eden show hot was though so I mean we
had people who go to DEFCON kind of
frequently and get everyone come to them
and say hey you got that guy that's cool
which was important so I think none of
this is going to be particularly
surprising to the people that have come
just they won't accept it as a talk
because it doesn't have us breaking into
things and making lots of mayhem but
this is fairly mainstream thinking and
Afghan I think they just it'd be made
more well that there is so much to be
done in air and AI you suck so much of a
hot area anyway that's I think it's
gonna be very low barriers to getting
much more attention to this book
you mentioned that one possible path for
the future is centralization into a few
kind of core clouds or you know certain
companies that can keep us safe how
would you evaluate the the leading tech
companies today in terms of how they're
doing in preparing for a I risk Google
Facebook Amazon Microsoft Apple Apple
just rolled out face recognition on
probably what will be a billion iPhones
over time the question is which type of
AI asked if you looking at
short term issues so how good are they
on cybersecurity
surprisingly good better than academia
definitely better than government
Google I think of pushing the envelope
on what is good security practice and
are looking forward to new threats that
are coming online I could probably say
similar good things about Apple and
Microsoft
as well well varies I definitely can say
good things about Google
but I would like to see those results
showed more widely not just to getting
an account with those companies but
getting the expertise sure this will be
sharing this so that smaller vendors can
also buy them but but it is also very
resource-intensive and there are
economies of scale you can just you know
you can hire a team of a hundred people
to do it to the response if you're
dealing with a billion accounts you
might not be able to do it if you run
Inga 500
another question about kind of the the
plausible paths of the future
often AI progress is kind of analogize
to an arms race where everybody's just
kind of competing to have the best and
you know a small edge could sort of lead
to a winner-take-all scenario do you
think we're sort of destined for that
version or do you see hope that we can
sort of solve that and get into a more
kind of sane
you know more moderate progress sort of
future yeah
so that is what can be done right on
this idea of beneficial yeah yeah I feel
good
you have no advantage on not sharing
safety practices with others everyone
benefits if everyone shares safety and
invests a lot in safety and these are
ideas that are being put out there quite
a lot and several companies are
receptive to them
farther down the line it's how to do
related question do you think that we
will need and and ultimately have some
sort of global surveillance paradigm to
try to watch after everybody and and
identify bad actors I
guess it depends a lot with how your
capabilities all these to be right so if
you think that everyone should be able
to deploy extremely powerful technology
then that probably means that those this
folder I mean if I'm allowed to to have
the capability to do something really
really really bad and should probably be
some mechanism to make sure that I don't
do something really really bad with the
technology if I'm not given as much
power then there is much less need to
surveil and this is a choice of gonna
make a society and probably the choices
that we're gonna make is society's
for somebody who's interested in getting
into this area professionally do you
think that government is a good place
for them to go or would you think about
academia large tech companies or other
dependent to do so we think there's a
lot of technical research it needs to be
done on securing systems on analyzing
the economics of various threats
academia and industry are probably the
places to do that I mean definitely in
terms of the top teams and the top
talent and being able to run experiments
on the other hand there are definitely
institutional and legal fixes that we
will need to have in place and the
government is the best one to do those
kind of a change of pace you originally
divided risks into near and long-term do
you think those are highly correlated in
terms of the way that those risks
ultimately are realized or not or are
they sort of more independent where the
the near-term outcomes don't tell us
much about the long-term outcomes and
then depending on your answer there what
do you do you think people should be
focused more on near or long-term I
think there's a lot of heterogeneity and
a lot of disagreement about how things
fall within this heterogeneity so I
think
if you think the long term systems are
going to be
blame children of the systems that we
have today or gonna have a lot of
components that are made up of the
systems that we have today then securing
the systems we have today is gonna
translate directly into the long term
systems you could also say that without
training in the knowledge that is
required to secure present systems
you're gonna be the disadvantage
compared to people if you have the
knowledge about two thousand eight
systems then you just get a really good
training about thinking in certain ways
then walking on present-day systems is
gonna translate kind of epistemic lis
into thinking about long-term systems if
you think that there is kind of a deep
difference in what it takes to secure
short on systems and long term systems
which some people hold and I think is
there are arguments to lead you to do
that then you probably want to kind of
just work on the long term in that
particular framing I think there is
there are enough problems and not enough
people that just find the one that you
kind of feel most passionate and capable
to do and there wouldn't be a place
little that's all the questions we have
through the app right now you're gonna
be doing office hours at the next break
correct tell us where we can follow you
online to keep track of your work and
learn more just go on the caesar website
cser dot AC dot uk' and also follow the
FHI website and the FLI website
you will stay up to date through those
channels for helping us understand AI
risk in the near term round of applause
for Shahar Avene
thank you very much great talk |
aba90138-b29a-4465-aa0c-09930a7be09d | trentmkelly/LessWrong-43k | LessWrong | Evolution's selection target depends on your weighting
I think it's common on LessWrong to think of evolution's selection target as inclusive genetic fitness - that evolution tries to create organisms which make as many organisms with similar DNA to themselves as possible. But what exactly does this select for? Do humans have high inclusive genetic fitness?
One way to think of it is that all organisms alive today are "winners"/selected-for by that competition, but that seems unreasonable to me, since some individual organisms clearly have genetic disorders or similar which make them unfit according to this criterion.
There's some sort of consensus that we can assign individual organisms to "species", and then we could count it by the number of members of that species. Supposedly, the most numerous species is Pelagibacter communis, with 10^28 individuals, vastly outnumbering humanity. Maybe we could say that this is the selection target of evolution?
Of course as would be expected, pelagibacter is a very minimalist species, being single-celled and having very few genes. This minimalism also makes it hard to notice, to the point where according to Wikipedia, it was first discovered in 1990. (I wonder if there's another species that's smaller, more common, and even harder to notice...) This raises the question of pure numerousity is the correct way of thinking of it.
If we instead weight by biomass, most life is in the form of plants, and I think more specifically trees. This makes perfect sense to me - trees evolve from a direct competition for height, which is one of the traits most directly related to mass. And in a way, biomass is more sensible to weight by than numerousity, since it is less dependent on the way you slice a species into individual organisms.
But trees are pretty static. Maybe the problem is that since mass has inertia, this weighting implicitly discourages more dynamic species, like humans? An alternative is to weight by energy flow, but in that case, algae and grasses end up accounting for most o |
68b145cb-5b5a-4769-840b-b5a3b969473f | StampyAI/alignment-research-dataset/blogs | Blogs | Nate Soares speaking at Purdue University
On Thursday, September 18th Purdue University is hosting the seminar [Dawn or Doom: The New Technology Explosion](http://www.purdue.edu/dawnordoom/). Speakers include James Barrat, author of [*Our Final Invention*](http://smile.amazon.com/Our-Final-Invention-Artificial-Intelligence-ebook/dp/B00CQYAWRY/), and MIRI research fellow Nate Soares.
Nate’s talk title and abstract are:
> **Why ain’t you rich?:**Why our current understanding of “rational choice” isn’t good enough for superintelligence.
>
>
> The fate of humanity could one day depend upon the choices of a superintelligent AI. How will those choices be made? Philosophers have long attempted to define what it means to make rational decisions, but in the context of machine intelligence, these theories turn out to have undesirable consequences.
>
>
> For example, there are many games where modern decision theories lose systematically. New decision procedures are necessary in order to fully capture an idealization of the way we make decisions.
>
>
> Furthermore, existing decision theories are not stable under reflection: a self-improving machine intelligence using a modern decision theory would tend to modify itself to use a different decision theory instead. It is not yet clear what sort of decision process it would end up using, nor whether the end result would be desirable. This indicates that our understanding of decision theories is inadequate for the construction of a superintelligence.
>
>
> Can we find a formal theory of “rationality” that we would want a superintelligence to use? This talk will introduce the concepts above in more detail, discuss some recent progress in the design of decision theories, and then give a brief overview of a few open problems.
>
>
For details on how to attend Nate’s talk and others, see [here](http://www.purdue.edu/dawnordoom/Travel).
The post [Nate Soares speaking at Purdue University](https://intelligence.org/2014/09/12/nate-soares-speaking-purdue-september-18th/) appeared first on [Machine Intelligence Research Institute](https://intelligence.org). |
c3f267a7-96ce-414a-858d-1268abb3343d | trentmkelly/LessWrong-43k | LessWrong | Announcing the Double Crux Bot
Edit (April 2, 2024): You can now join the Double Crux Bot's own Discord server here.
TL;DR:
We are releasing a beta version of our chatbot powered by GPT, designed for facilitating Double Crux dialogues between two users on Slack or Discord. We're hoping to receive user feedback to evaluate the bot's usefulness and further work required. You can get the bot for your Slack workspace here and for your Discord server here.
Introduction
Double Crux is a conflict resolution technique developed for the Center for Applied Rationality (CFAR) workshops, but it can be difficult to use because it might not apply to all problems, conversations can become emotional, or the participants might not have the requisite knowledge to embark on a double crux conversation themselves. Having a facilitator makes it easier, but facilitators aren’t readily available and are pretty much inaccessible outside of defined rationality workshops. There are also widely varying opinions on the best scenarios and strategies for approaching double crux. Our bot provides a low-effort way to engage in double crux for disagreements that people have in real-time, and it systematizes the method to an algorithm we have developed and continue to refine.
Example Scenario
Say you and your co-worker disagree about whether you should create a code of conduct policy for your organization. One of you is strongly against it, and the other one is strongly for it. Despite rehashing your reasoning multiple times, you cannot reach an agreement. The bot might help you to understand that your co-worker thinks a code of conduct policy might be counterproductive if it isn't enforced well, whereas you think that, even without enforcement, it has a positive net effect on overall culture. The double crux here might be whether conduct policies require enforcement.
The double crux bot is intended to help resolve these types of action-oriented disagreements. The bot helps people make their reasoning explicit and reflect |
c32805d6-ef23-4b73-892e-122154e0bc54 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Thoughts On Expanding the AI Safety Community: Benefits and Challenges of Outreach to Non-Technical Professionals
**Epistemic Status: Low-Medium** (I’ve spent 25-30hrs thinking and discussing this idea as a non-expert in AI Safety)
**Disclaimer:**This claim could be biased to my work in the Bay Area this summer, but it seems to me that most individuals who are interested in AI safety or reducing existential risks from AGI are interested in and leaning towards doing technical AI research. Most AI safety community builders (in my limited knowledge) appear to engage in this: let’s find the best mathematicians and programmers for technical research in AI Alignment. While I do believe that there is tremendous value for those, I do think that there is also value for individuals without a technical background to also contribute to this field which has so far mostly been inaccessible for them. Thus, I want to outline reasons why I think this can be valuable, what the intersection of different mainstream academic fields and AI Safety might look like, how we could do this outreach and the potential pitfalls for this.
**TL;DR:** It seems important to involve non-technical experts in AI safety research and governance in order to benefit from their unique knowledge and skills, and to encourage interdisciplinary collaboration that can lead to innovative ideas and insights. Non-technical experts from fields such as economics, politics, and public policy can all contribute valuable perspectives and expertise to the field of AI safety. For example, economists can provide an understanding of incentives and the potential impacts of AI on the economy, and political scientists and public policy professionals can help to develop policies and governance frameworks for the development and use of AI.
**Why should we do outreach to non-technical academics and/or non-technical professionals?**
--------------------------------------------------------------------------------------------
1. **Non-technical academics and professionals can bring diverse perspectives and expertise to the field of AI safety:** It is important to involve people from a variety of backgrounds in order to benefit from the unique knowledge and skills that they bring. For example, a philosopher may bring insights into ethical considerations, while a sociologist may offer insights into how AI could impact society. By including people from diverse fields, we can more effectively address the complex and multifaceted challenges of AI safety.
2. **Interdisciplinary collaboration can lead to innovative ideas and insights:**By combining knowledge from various fields, we can generate new approaches and find solutions that may not have been possible otherwise. This approach has been shown to be effective through extensive research, making it a valuable approach for tackling complex issues such as those related to AI safety.
3. **Involving non-technical experts can save time for AI safety researchers:**When AI safety researchers need to learn about a new topic outside of their own field, they can save time by seeking out the expertise of non-technical experts. For example, if a computer scientist is working on a project related to the economic impact of AI, they may seek out the insights of an economist to better understand the topic in a more efficient manner.
**What does this intersection between different fields and AI Safety look like?**
---------------------------------------------------------------------------------
Some examples come to mind to highlight the intersection of different fields.
1. **Economics**
Economists have valuable insights and expertise that can help us achieve the goal of making safer AI systems. One specific example that comes to mind is mechanism design. Mechanism design is a branch of microeconomics that focuses on designing institutions and mechanisms that achieve desired outcomes, which could be useful in AI alignment research. Aligning AI systems with human values and ensuring that they do not pose unnecessary risks is a central challenge in AI safety. Mechanism design could be used to develop governance structures and regulations that ensure that AI systems are aligned with human values and do not pose unnecessary risks.
Economists' understanding of incentives can be useful in AI safety by helping to identify the factors that may motivate organizations to prioritize short-term goals over long-term risks in the development and deployment of AI systems. This understanding can be used to design incentives that encourage AI organizations, such as OpenAI, to prioritize safety and consider the long-term consequences of their actions. For example, economists can help to develop policy frameworks or economic incentives that encourage AI organizations to prioritize safety and accountability in the development and use of AI systems. This could include measures such as requiring organizations to bear a portion of the liability for any negative consequences of their AI systems, or providing financial rewards for organizations that demonstrate a commitment to safety and responsible AI practices.
In addition, economists can help us understand the potential impacts of AI on the economy, including issues related to job displacement, job creation, and inequality. For example, AI could have significant consequences for the labor market, both in terms of replacing human jobs and creating new ones. Economists can help us understand these impacts and identify strategies to mitigate any negative consequences, such as providing support for workers who may be displaced by automation. Similarly, economists can help us understand the potential impacts of AI on issues related to inequality, such as the distribution of wealth and opportunities. By working together with economists and other experts from diverse fields, we can develop more comprehensive and effective strategies for addressing these issues and ensuring that AI systems are used in a safe and ethical manner.
**2. Politics/Public Policy**
I do think that political science and public policy students/professionals could be quite valuable for AI Governance, supporting the bigger goal of reducing x-risks from AI. Their expertise in policy analysis, regulation, and forecasting can be valuable in developing strategies for ensuring that AI systems are aligned with human values and do not pose unnecessary risks.
For example, public policy and political science students could help identify the most effective regulatory frameworks for governing AI and ensure that it is used in a responsible and ethical manner. They could also assist in developing strategies for addressing the potential impacts of AI on issues such as job displacement, inequality, and privacy.
In addition, public policy and political science students could be helpful in analyzing the political and social dynamics related to AI governance and identifying ways to build support and consensus around these issues. This could involve working with policymakers, industry stakeholders, and civil society groups to develop and advocate for effective AI governance policies.
**3. Communications**
Communications professionals, including writers, artists, and journalists, have the power to shape public understanding and engagement with the risks posed by advanced AI. By using their skills and platforms to educate and engage the public, they can help to build support for research and initiatives that aim to ensure the safe and responsible development of AI and reduce the existential risks posed by advanced artificial intelligence.
One way that communications professionals can contribute to AI safety is by producing high-quality journalism and other forms of content that accurately and thoughtfully explore the potential risks and benefits of AI. For example, the [Washington Post's article](https://www.washingtonpost.com/technology/2022/07/16/racist-robots-ai/) on the potential for AI to be racist and sexist received four times the engagement of posts on social media platforms, demonstrating that there is a significant appetite for this type of content. By producing and promoting well-researched and engaging arguments/discussions of AI safety, journalists and writers can help to educate and inform the general public and build support for initiatives that address these risks.
In addition to journalism, fiction writing can also be a powerful tool for raising awareness about existential risks from advanced AI. While dystopian science fiction has long explored the dangers of AI, more realistic and accurate fiction that is written in consultation with experts in AI safety could be especially effective in helping the public to understand and grapple with these complex issues. By using their storytelling skills to paint a picture of what the future could look like with advanced AI, writers can help to inspire thought and discussion about how we can mitigate these risks.
Finally, artists and curators can also help to raise awareness about AI safety through their work. Artistic mediums such as comics and visual art can provide an engaging and accessible way to explore complex issues and can be particularly effective in reaching audiences who might not be interested in traditional forms of content. By creating and promoting artwork that touches on the risks of advanced AI, artists and curators can help to build understanding and support for initiatives that aim to reduce these risks.
**How can this outreach be done?**
----------------------------------
On first thoughts, here are some ways I think this outreach could be done:
1. **Host events such as Intro to AI Safety talks and workshops in non-CS faculties of universities**: These can be held in a variety of settings, such as community centers, libraries, or professional associations, and can be tailored to the specific interests and needs of different audiences. For example, AI Safety for Economists/AI Safety for Public Policy students.
2. **Involve non-technical professionals as consultants or collaborators:**By reaching out to experts in other fields and inviting them to help with specific projects, it is possible to bring new insights and approaches to bear on AI safety research and initiatives. For example, more AI Safety researchers could collaborate with economists to explore the potential economic impacts of advanced AI, understand economic growth and how that impacts AI capabilities, or design mechanisms for ensuring its safe and responsible development.
3. **Create more accessible materials for learning about AI safety:** While there are already many excellent resources available for those with a technical background such as AGI Safety Fundamentals, these can be difficult for the general public to understand. By developing materials that are more accessible to non-technical audiences, it is possible to broaden the reach of AI safety research and initiatives. For example, a group of AI safety experts could create a series of videos or interactive resources that explain key concepts in a way that is easy for non-technical audiences to understand. Additionally, just learning materials that make learning about how an AI works seems pretty good and impactful for a simpler learning curve for people just getting started.
4. **More interdisciplinary research programs:** There have been recent efforts to engage people from different disciplines such as [PIBBSS](https://www.pibbss.ai/), and these seem to be quite promising. I’d be curious to see learn how successful their outcomes were and believe that more programs for other academic fields and on a larger scale could be quite impactful.
**Some reasons to not outreach to traditional academics**
---------------------------------------------------------
In this section, I’d like to red-team my own arguments by highlighting potential reasons to not work on outreach to non-technical experts for assisting in AI Safety work.
1. **Most non-technical professionals might not be interested:** It is possible that many non-technical professionals, particularly those who are well-established in their careers, might not be particularly interested in AI safety. They might have other priorities and concerns, and may not be willing to devote time and resources to this issue.
2. **Encouraging novel ideas might be difficult:** Traditional academics and other non-technical professionals might be more prone to traditional ideas and approaches, and might not be as open to radical or innovative thinking as experts in the field. This could limit the scope and impact of AI safety research and initiatives.
3. **Too much interest in AI safety might not be a good thing:** While it is generally desirable to build support and understanding for AI safety, having too many people interested in this issue could present its own challenges. For example, it could be difficult to coordinate and lead a large and diverse community, and there might be more disagreements and competing agendas.
4. **A larger community could lead to more disagreements:**With more people involved, it might be harder to achieve consensus and maintain a cohesive group. There could be more disagreements and competing agendas, which could make it harder to influence and lead the community.
**Concluding Thoughts**
-----------------------
I do believe that it might be worth finding ways to go minimize these risks and dangers. One way to mitigate these risks is by conducting a thorough screening process to ensure that only qualified and reliable individuals are invited to join the community. By examining candidates' past experiences, backgrounds, and references, it is possible to identify those who have the skills and commitment necessary to contribute to AI safety research and initiatives.
Additionally, it may be useful to focus on targeted outreach to the best and potentially most impactful professionals and experts in specific fields, such as writing, economics, political science, etc. I do think that these individuals might bring valuable perspectives and expertise to the field and can help to ensure that AI safety research and initiatives are informed by a diverse range of disciplines. |
6e18d518-2ed4-4ae3-9e7f-f7f41200e998 | trentmkelly/LessWrong-43k | LessWrong | Discomfort Stacking
I’m pretty new here so apologies if this is a stupid question or if it has been covered before. I couldn’t find anything on this topic so thought I’d ask the question before writing a full post on the idea.
If we believe that discomfort can be quantified and ‘stacked’ (e.g. X people with specks of dust in their eye = 1 death), is there any reason why this has to scale linearly from all perspectives?
What if the total can be less than the sum of its parts depending on the observer?
Picture a dynamic logarithmic scale of discomfort stacking with a ‘hard cap’ where every new instance contributes less and less to the total to the point of flatlining on a graph.
Each discrete level of discomfort has a different starting value - so an infinite number of something extremely mild could never amount to the value of even a single instance of extreme torture.
Every individual instance is ‘worth’ the full n=1 level of discomfort – but, when stacked, this is augmented and dynamically shifts, though only to an observer looking at the entire set of cumulative instances.
No matter how many people have a speck of dust in their eye – to an outside observer it would never amount to the cumulative discomfort of even one single death, despite every individual feeling the full extent of it as if they were the only one. |
7d3838ed-f907-4183-89ae-cfb7e4da4e94 | trentmkelly/LessWrong-43k | LessWrong | In praise of pretending to really try
Ben Kuhn makes some reasonable criticisms of the Effective Altruism movement. His central claim is that in the dichotomy of ‘really trying’ vs. ‘pretending to try’, EAs ‘pretend to really try’.
To be explicit, I understand these terms as follows:
‘Really trying’: directing all of your effort toward actions that you believe have the highest expected value in terms of the relevant goals
‘Pretending to try‘: choosing actions with the intention of giving observers the impression that you are trying.
‘Pretending to really try‘: choosing actions with the intention of giving observers the impression that you are trying, where the observers’ standards for identifying ‘trying’ are geared toward a ‘really trying’ model. e.g. they ask whether you are really putting in effort, and whether you are doing what should have highest expected value from your point of view.
Note the normative connotations. ‘Really trying’ is good, ‘pretending to try’ is not, and ‘pretending to really try’ is hypocritical, so better than being straight out bad, but sullied by the inconsistency.
I claim Effective Altruism should not shy away from pretending to try. It should strive to pretend to really try more convincingly, rather than striving to really try.
Why is this? Because Effective Altruism is a community, and the thing communities do well is modulating individual behavior through interactions with others in the community. Most actions a person takes as a result of being part of a community are pretty much going to be ‘pretending to try’ by construction. And such actions are worth having. If they are discouraged, the alternative will not be really trying. And pretending to try well is almost as good as really trying anyway.
Actions taken as a result of being in a community will be selected for being visible, because visible actions are the ones you will be able to pick up from others in the community. This doesn’t necessarily mean you are only pretending to try – it will just happen to l |
e0ceffac-1d9e-4cd1-940a-6fbb7fe92e8b | trentmkelly/LessWrong-43k | LessWrong | If I interact with someone with nCov for an hour, how likely am I to get nCov?
I'm guessing data is limited here, but a related-related question might be "how likely am I to catch the flu or a few other common diseases by interacting with a victim for an hour?" |
cc4066cc-8ac2-41dd-93d9-26227894beb0 | trentmkelly/LessWrong-43k | LessWrong | Speed running everyone through the bad alignment bingo. $5k bounty for a LW conversational agent
There's a wave of people, of various degrees of knowledge and influence, currently waking up to the ideas of AI existential risk. They seem to be literally going through every box of the bad alignement bingo card takes.
I think there is value in educating those people. I'm aware there's an argument to be made that: education at scale doesn't matter, coordination is too difficult, all that matter is solving alignment and that takes care of the rest.
There's something to that, but I disagree that education at scale doesn't help. It can make progress of frontrunners marginally more safety oriented, it can steer company cultures, it can move the Overton window, change the Zeitgeist, it can buy a bit of time. You likely didn't stumble on these ideas all on your own, so arguing against the value of outreach or education is also arguing against your own ability to do anything.
It's also a matter of ROI, and there are some very low hanging fruit there. The simplest thing would be to write a long FAQ that goes through every common objections. No, people won't read the whole sequences, or Arbital on their own, but they might go through a FAQ.
But we can do better than a FAQ. It's now fairly straightforward, with tools like langchain (https://github.com/hwchase17/langchain) to turn a set of documents into a body of knowledge for a conversational agent. This is done by building an index of embedding that a language model can search to bring context to an answer. This doesn't preclude fine tuning, but it makes it unnecessary.
So a straightforward project is to index lesswrong, index arbitral, index the alignment forum, maybe index good alignement papers as well, blog posts, books.
Then hook that up to the ChatGPT API, and prompt it to:
1. list search queries for relevant material to answer the question
2. compose an answer that reflects the content and opinion of the data
3. answer with infinite patience
Some jailbreak prompts may be needed to prevent ChatGPT's condit |
d5deb96a-f25b-4644-bb11-42ee2df9b613 | trentmkelly/LessWrong-43k | LessWrong | Bayeswatch 2: Puppy Muffins
A green humvee arrived at Jieyang Chaoshan International Airport. Vi got in the back with Molly Miriam who handed her clipboard to Vi.
"健重制造公司. A no-name Chinese factory that makes barbells and similar equipment. It's not even fully-automated," read Vi.
"They are registered to use a low-intelligence AGI," said Miriam.
"What are we even doing here? Neither the product nor the AI poses a threat to civilization," said Vi.
"Something must have gone badly wrong," said Miriam.
The road to the factory was blockaded by the People's Liberation Army (PLA). The soldier at the checkpoint scanned the Bayeswatch agents' badges. A young officer—barely out of high school—escorted them inside the perimeter to Colonel Qiang.
"We could probably handle this on our own," said Colonel Qiang, "But protocol is protocol."
"So it is," said Miriam.
There were no lights on in the factory. No sound emanated from it. Fifty soldiers behind sandbags surrounded the factory, along with two armored personnel carriers and a spider tank.
"The police responded first. The sent a SWAT team in. Nobody came back. Then we were called. We would like to just burn the whole thing down. But this incident could be part of a wider threat. We cut power and Internet. Nothing has entered or left the building since our arrival. Rogue AIs can be unpredictable. We wanted your assessment of the situation before continuing," said Colonel Qiang.
"You did the right thing. This is probably an isolated incident. If so then the best solution is to rescue who we can and then level the building. Unfortunately, there is a chance this is not an isolated incident. Therefore our top priority is to recover the AI's hard drives for analysis," said Miriam.
"We will assault the building," said Colonel Qiang.
"You may have our sanction in writing. Assume humans are friendly and robots are hostile," said Miriam.
"Yes sir," said Colonel Qiang.
----------------------------------------
Miriam and Vi were quartered in a nearby |
d9a9aaf7-280b-4121-99f8-9a012e94ea83 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Power as Easily Exploitable Opportunities
*(Talk given at* [*an event on Sunday 28th of June*](https://www.lesswrong.com/posts/iZ3AisoiB6qKY6Ctz/sunday-jun-28-more-online-talks-by-curated-authors)*. TurnTrout is responsible for the talk, Jacob Lagerros and David Lambert edited the transcript.*
*If you're a curated author and interested in giving a 5-min talk, which will then be transcribed and edited, sign up* [*here*](https://forms.gle/iwFatbhys9muPmQA7)*.)*
**TurnTrout:** [Power](https://www.lesswrong.com/s/7CdoznhJaLEKHwvJW/p/6DuJxY8X45Sco4bS2) and [power-seeking](https://www.lesswrong.com/s/7CdoznhJaLEKHwvJW/p/75oMAADr4265AGK3L) playa big part in my thinking about AI alignment, and I think it's also an interesting topic more generally.
Why is this a big deal? Why might we want to think about what power is? What is it, exactly, that people are thinking of when they consider someone as powerful?
Well, it seems like a lot of alignment failures are examples of this power-seeking where the agent is trying to become more capable of achieving its goals, whether that is getting out of its box, taking over the world, or even just refusing correction or a shutdown.
If we tie these together, we have something I've called [the catastrophic convergence conjecture](https://www.lesswrong.com/s/7CdoznhJaLEKHwvJW/p/w6BtMqKRLxG9bNLMr), which is that if the goals aren't aligned and it causes a catastrophe, it is because of power-seeking.
But I think that when people are first considering alignment, they think to themselves, "What's the big deal? You gave it a weird goal, and it does weird stuff. We’ll just fix that.”
So why exactly do unaligned goal maximizers tend to cause catastrophes? I think it's because of this power seeking. Let me explain.
The way I think about power is as the ability to achieve goals in general.
In the literature, this is like the dispositional *power-to* notion.
Whereas in the past, people thought "Well, in terms of causality, are the agent's actions necessary and/or sufficient to cause a wider range of outcomes?", here, I think it's best thought of as your average ability to optimize a wide range of different goals. So if you formalize this as your average optimal value in, say, a Markov decision process (MDP), there's a lot of nice properties and you can prove that, at least in certain situations, it links up with instrumental convergence. Power seeking and instrumental convergence are [very closely related](https://www.lesswrong.com/posts/nyDnLif4cjeRe9DSv/generalizing-the-power-seeking-theorems).
But there's a catch here. We're talking about average *optimal* value. This can be pretty weird. Let's say you're in the unfortunate situation of having a dozen soldiers about to shoot you. How powerful are you according to average optimal value? Well, average optimal value is still probably quite high.
There's probably an adversarial input of strange motor commands you could issue which would essentially incapacitate all the soldiers just because they're looking at you since their brains are not secure systems. So each optimal policy would probably start off with, "I do this weird series of twitches, incapacitate them, and then I just go about achieving my goals."
So we'd like to say, “well, your power is actually lowered here in a sense”, or else we’d have to concede that it’s just wholly subjective what people are thinking of when they feel powerful.
My favorite solution is, instead of asking how wellcould I achieve a bunch of different goals? You should be asking, how well could Iachieve many goals?
If you imagine something like a learning algorithm, you could say it's a human level learning algorithm. You give it a history of observations and a goal that it’s optimizing, and it produces a policy, or things that it should do to achieve this goal. You then say, "Well, what's my average ability? What's A's average ability? What's this algorithm's average ability to optimize and to achieve goals in this history, in this situation?"
What I think this does is recover this common sense notion of “you don't have much power here because these aren't cognitively accessible opportunities and policies”. And so essentially, you are disempowered in this situation.
I think understanding this also makes sense of what power means in a universe where everyone is only going to have one course of action. If you view them as running algorithms and then saying, "Well, how well could this learning algorithm achieve different goals in the situation?" I think it might be important to evaluate AI designs by how they respect our power in particular, and so understanding what that means is probably important.
Also, if you want to do better than just hard goal maximization in aligning these AIs, then I think understanding exactly what the rot is at the heart of reward maximization is pretty important as well. Thank you.
Questions
=========
**Daniel Filan:** If I'm thinking about a learning algorithm like Q-learning or PPO or something, then it makes a lot of sense to think that it's a function of a goal and a history. But in most situations, I tend to think of them as results of learning algorithms.
Take some Atari agent. It trained for a while and now it is like a deployed system. It is playing Atari and it is not manifestly a function of a goal. Maybe it has a goal somewhere in its neural network and you could change some bits and it would have a different follow-up move, but that's not obvious.
So I'm wondering, what do you think of this functional form of agents as functions of histories and goals?
**TurnTrout:** Good question. I think that when we're making an objection like this, especially before we've solved more issues with embedded agency, we're just going to have to say the following: "If we want to understand what this person is thinking of when they think of power; then I think that even though it might not literally be true, that you could cleanly decompose a person like this, it's still a useful abstraction."
I would agree that if we wanted to actually implement this and say, "Well, we're looking at an agent, and we deduce what its learning algorithm is and what it would mean to have a modular goal input to the algorithm," then you would really need to be worried about this. But my perspective, at least for right now in this early stage, is that it's more of a conceptual tool. But I agree, you can split up a lot of agents like this.
**Ben Pace:** I'm curious if you have any more specific ideas for measuring which policies are currently attainable by a particular agent or algorithm — the notion of “attainability” felt like it was doing a lot of work.
**TurnTrout:** I think the thing we're assuming here is, imagine you have an algorithm that is about as intelligent, with respect to the [Legg-Hutter metric](https://arxiv.org/abs/0712.3329) or some other more common-sense notion, as a human. Imagine you can give it a bunch of different reward function inputs. I think this is a good way of quantifying this agent’s power. But you’re asking how we get this human level algorithm?
**Ben Pace:** Yes. It just sounded like you said, "In this situation, the human agent, in principle, has an incredible amount of power because there is a very specific thing you can do." But to actually measure its impact, you have to talk about the space of actual operations that it can find or something.
And I thought, "I don't have a good sense of how to define exactly what solutions are findable by a human, and which solutions are not findable by a human." And similarly, you don't know for various AIs how to think about which ones are findable. Because at some point, some AI gets to do some magical wireheading thing, and there's some bridge it crosses where you realize that you could probably start taking more control in the world or something. I don't quite know how to measure when those things become attainable.
**TurnTrout:** There are a couple ways you can pose constraints through this framework, and one would be only giving it a certain amount of history. You're not giving infinite data.
Another one would be trying to get some bounded cognition into the algorithm by just having it stop searching after a certain amount of time.
I don't have clean answers for this yet, but I agree. These are good things to think about.
**habryka:** One thing that I've been most confused about for the formalism for power that you've been thinking about, is that you do this averaging operation on your utility function. But averaging over a space is not a free operation. You need some measure on the space from which you sample.
It feels to me like, power only appears when you choose a very specific measure over the space of utility functions. For example, if I sub-sample from the space of utility functions that are extremely weird and really like not being able to do things, it will only care about shutting itself off rather than whether it's going to get any power-seeking behavior.
So am I misunderstanding things? Is this true?
**TurnTrout:** The approach I’ve taken, like in my recent paper, for example, is to assume you’re in some system with finite states. You then take, for example, the MaxEnt distribution over reward functions or you assume that reward is, at least, IID over states. You then get a neutrality where I don't think you need a ton of information about what the reasonable goals you should pursue are.
I think if you just take a MaxEnt distribution, you'll recover the normal notion of power. But if you're talking about utility functions, then because there's infinitely many, it's like, "Well, what's the MaxEnt distribution over that?"
And so far, the theorems are about just finite MDPs. And if you're only talking about finding MDPs and not some kind of universal prior, then you don't need to worry about it being malign.
**Rob Miles:** Something I'm a little unclear on is how this can ever change over time. I feel like that's something you want to say. Right now, you're in the box. And then if you get out of the box, you have more power because now there's a path that you're able to follow.
But if you are in the box and you can think of a good plan for getting out, isn't there a sense that you *already* have that power? Because you're aware of a plan that gets you what you want via getting out of the box? How do you separate power *now* from the potential for power *in the future*?
**TurnTrout:** Good question. This is the big issue: thinking about power in terms of optimal value. If you have an agent that has consistent beliefs about the future, you're not going to expect to gain more.
If you're trying to maximize your power, you're not going to expect, necessarily, to increase your power just due to conservation of expected evidence. But if things happen to you and you're surprised by them, then you see yourself losing or gaining power, especially if you're not optimal.
So if it's too hard for me to get out of the box or I think it's too hard, but then someone lets me out, *only after that* would I see myself as having a lot more power.
--- |
5de3eafc-bbda-4e4a-bafd-ef45eb790ed7 | trentmkelly/LessWrong-43k | LessWrong | Richard Carrier on the Singularity
Recently I stumbled upon Richard Carrier's essay "Are We Doomed" (June 5, 2009), when asked to comment about the Singularity, said the following:
> I agree the Singularity stuff is often muddled nonsense. I just don't know many advocates of it. Those who do advocate it are often unrealistic about the physical limits of technology, and particularly the nature of IQ. They base their "predictions" on two implausible assumptions: that advancement of IQ is potentially unlimited (I am fairly certain it will be bounded by complexity theory: at a certain point it just won't be possible to think any faster or sounder or more creatively) and that high IQ is predictive of accelerating technological advancement. History proves otherwise: even people ten times smarter than people like me produce no more extensive or revolutionary technological or scientific output, much less invent more technologies or make more discoveries--in fact, by some accounts they often produce less in those regards than people of more modest (though still high) intelligence.
>
> However, Singularity fans are right about two things: machines will outthink humans (and be designing better versions of themselves than we ever could) within fifty to a hundred years (if advocates predict this will happen sooner, then they are being unrealistic), and the pace of technological advancement will accelerate. However, this is already accounted for by existing models of technological advancement, e.g. Moore's Law holds that computers double in processing power every three years, Haik's Law holds that LED's double in efficiency every three years, and so on (similar laws probably hold for other technologies, these are just two that have been proven so far). Thus, that technological progress accelerates is already predicted. The Singularity simply describes one way this pace will be maintained: by the recruitment of AI.
>
> It therefore doesn't predict anything remarkable, and certainly doesn't deserve such a prete |
a3677391-7114-46c5-83e1-0f42b1c5bfe2 | trentmkelly/LessWrong-43k | LessWrong | Test article
Testing |
220f8395-5627-4daa-9408-6a9699cdd8f5 | StampyAI/alignment-research-dataset/special_docs | Other | We, Borg: Speculations on hive minds as a posthuman state
We, Borg
========
Speculations on Hive Minds as a Posthuman State
-----------------------------------------------
\*by Anders Sandberg\*
>
> The designers of our species set out to produce a being that might be capable of an
> order of mentality higher than their own. The only possibility of doing so lay in planning
> a great increase in brain organisation. But they knew that the brain of an individual
> human being could not safely be allowed to exceed a certain weight. They therefore sought
> to produce the new order of mentality in a system of distinct and specialised brains held
> in "telepathic" unity by means of ethereal radiation. Material brains were to be
> capable of becoming on some occasions mere nodes in a system of radiation which itself
> should then constitute the physical basis of a single mind.
>
> Olaf Stapledon, Last and First Men
>
>
>
Hive minds where the individual is subsumed into a collective consciousness has been a
recurring idea in science fiction since Olaf Stapledon's influential novels \*Last and
First Men\* (1931) and \*Star Maker\* (1937), although the concept in some sense had
been suggested by \*Leviathan\* of Thomas Hobbes (1651). They have often in western
science fiction been used as an allegory for communism or the anonymity of industrial
civilisation, and have usually been portrayed in a terrifying light (Nicholls 1982). The
latest such portrayal is the Borg in \*Star Trek: the Next Generation\*: a race of
bionically augmented humanoids linked together into a collective mind, striving to
assimilate every other intelligent species into the Collective.
Due to the popularity of the show several new words for hive minds have been coined
(Morrow 1996):
\*\*Borganism:\*\*
1) An organization of formerly autonomous beings who have merged their individual wills
to create one, collectively conscious being; 2) The social and political theory that
advocates the creation of borganisms.
\*\*Borganise:\*\*
To form a borganism, to organise its structure.
I will in the following call the beings making up the borganism \*\*units\*\* (calling
them individuals would be erroneous since they by definition lack individuality, and the
borganism is clearly divisible, hence it cannot be called an individual either). The word
borganism is especially suitable since I will look at hive minds from a cybernetic point
of view (cybernetics -> cyborg -> borg).
This essay seeks to look into the psychology and sociology of borganisms, and to
discuss borganisms as a possible posthuman state.
Borganisms in Nature
--------------------
Borganisms might at first appear to be fanciful ideas, more grounded in science fiction
and human desires/fears than in practical reality. But in nature there already exists
several systems that suggests otherwise. The most common example used is the hives of
social insects, where all individuals work for the common good with little regard for
themselves. Although it has been argued that hives lack collective minds (Nicholls 1982)
it should be noted that all such species communicate with chemical signals, and at least
in the case of ants chemical trails can be seen as collective cognitive maps distributed
in the environment (Chiavlo & Millonas 1995). There may exist degrees of
borganisation, and they are tied to how closely the units communicate.
Another natural system of interest is the structure of multicellular organisms. The
transition from single-celled life to multicellular life can be seen as borganisation. The
chemical "minds" of cells are closely connected, and in some cases cells have
gap-junctions connecting their cytoplasm or even merge to form a cyncyticum. In a
multicellular organism the cells are differentiated into different tissues with different
functions, which sometimes include the planned death of cells (such as in the case of the
formation of the protective outer layer of skin, the stratum corneum). Differentiation is
mediated through chemical signals from other cells which affect the genetic expression of
proteins and continued cell behaviour. This examples shows that a borganism can have a
complex internal structure. All units do not need to be equal, and specialisation and
hierarchical control is a possibility.
The third example of a borganism-like system in nature is the human brain. It consists
of several parts able to act independently but closely tied together, so closely that
normally these divisions go unnoticed. In some cases the system is disturbed and the
potential independence of the parts can become apparent. One example is split brain
patients whose hemispheres have been disconnected; most of the time this does not cause
any noticeable change, but under some circumstances the two sides come into conflict or
interfere with each other. Another example is the dissociative states that can occur
during hypnosis or traumatic situations where the mind is divided into two or more parts
having different access to sensory information and motor control (Hilgard 1977, Putnam
1989). The brain shows that the borganism might not even need to be aware of the units
making it up, it can exist on a higher level, perhaps as a metasystem (Turchin &
Joslyn 1993).
Communication and Structure in Borganisms
-----------------------------------------
>
> Many other triumphs of eugenical experiment we observed up and down the worlds. The
> general level of individual intelligence was, of course, raised far beyond the range of
> Homo Sapiens. But also that superintelligence which can be attained only by a psychically
> unified community was greatly developed on the highest practicable plane, that of the
> conscious individuality of a whole world. This, of course, was impossible until the social
> cohesion of individuals within the world-community has become as closeknit as the
> integration of the elements of a nervous system.
>
> Olaf Stapledon, Star Maker
>
>
>
Communication is central to borganisation. By definition the units making up a
borganism will be in close mental contact; the bandwidth and structure of this contact
will determine much of the properties of the borganism.
It may be hard to tell when a group of individuals becomes a borganism; the psychology
of a group can be significantly different from the psychology of the individuals, and even
among humans individuality can be subsumed by group identity under some conditions.
However, so far intra-group communication has been mainly verbal, kinetic and possibly
chemical (pheromones). As the bandwidth increases new phenomena will likely appear and the
group as an organism begins to take on its own life.
The communication between the units of a borganism can be characterised by its
bandwidth and topology.
### Bandwidth
Bandwidth denotes the amount of information exchanged between units and to which mental
depth it occurs; speech is a low bandwidth communication only reaching a superficial
mental level while a direct mental link giving insight in the mental imagery of the other
part would be a high bandwidth communication. The extreme case is total connection where
the bandwidth is so high that all units form a single neural network. What is uncertain at
present is how high bandwidth is needed to create a true borganism. This may be a matter
of degree rather than a distinct transition between several individuals and one borganism.
Starting from a low bandwidth we have a group of individuals communicating and acting
on mutual goals. As the bandwidth is increased they can not only communicate intentions
but their deeper causes; at higher bandwidths the mental chains leading to decisions
become communicable and hence shareable. This may allow collaborative refinement of goals
and plans in a much more efficient way than low-bandwidth discussion and the borders
between individuals gradually fade away. Note that the units still can be specialised and
have different memories, values and personalities.
Group psychology has studied under what conditions groups become more (or less)
productive than individuals. In general it depends on the nature of the task and group. In
problem solving tasks groups frequently develops better solutions than individuals
(Hellriegel et al 1989), since there are more opportunities for error- correction, idea
generation, scenario testing and a higher likelihood that the skills and knowledge needed
to solve a complex problem are available. This is especially true for tasks which can be
subdivided easily.
Groups do not perform better than their most gifted individual on tasks which cannot be
subdivided if the task is simple and the solution immediately becomes obvious to everyone
once it is proposed (Baron & Byrne 1991). In many cases of human psychology social
processes can interfere with this and decrease the performance; this might be possible to
circumvent in a borganism. For example, in human groups the gifted individual often
voluntarily stands back in order not to dominate the discussions; in a borganism there is
less concern for the individual (both positive and negative), which suggests that this
tendency will be weakened in favour of helping the group. The above observations of human
problem solving suggests that borganisms should divide problems into manageable chunks
which are handled by small subnetworks (possibly temporary) which in turn communicate with
each other, at least in the case of divisible problems. In less easily divided problems it
appears likely that a high bandwidth connection between the participating units is
desirable, turning them into a more homogenous group.
So far I have assumed the group is interacting in a fairly homogenous manner, akin to a
meeting. It is also possible to differentiate between a planning part of the borganism and
an executive part which implements the plans while remaining in contact with the planning
part. This suggests two densely connected clusters of units linked by a somewhat lower
bandwidth link.
It appears likely that for a borganism which encounters different kinds of problems in
daily life it is advantageous to modify its internal topology and bandwidth. There are of
course technological and physical limitations to this, as well as a control problem: what
subsystem should organise the topology?
One possibility was suggested in \*Star Trek: First Contact\*: the "Borg
Queen", a female unit explained her function as "I bring order into chaos".
This could be interpreted as her having an organising role unlike the other fairly
identical units; other replies suggested that she was instantiated on other, perhaps all,
borg ships. A borganism may consist of two different kinds of units, one basic general
purpose unit that makes up most of the population, implementing the collective will, and
one or a few organising units optimising the internal structure (possibly acting as
arbitrators in internal conflicts or a supervisory B-brain (see Minsky 1988).
However, it is not certain there is a need for special units. If individual units can
influence their topology and bandwidth, it is not unreasonable to think that a regulatory
system could be implemented locally, for example by a market-based approach (Miller &
Drexler 1988). It is important to realise that borganisms may consist of many different
kinds of units both physically and mentally; while most descriptions have concentrated on
homogenous or stratified structures borganisms with wildly diverse units, possibly as
different as humans, AI systems and non-intelligent software agents, are a possibility.
### Topology
The topology can be varied endlessly. A simple solution is total interconnectivity
where every unit is connected to every other. Total interconnectivity is usually
inefficient since the total bandwidth (and its overhead) grows as N^2 (where N is the
number of units); in most cases there is little need for every unit to constantly
communicate with every other and most of the bandwidth is wasted. If time or attention has
to be taken from work to keep up to date with what other units are doing there will even
be an optimum size of the borganism where the total amount of work done is maximal, above
it the overhead of communication removes any advantage in adding more units.
Other interesting topologies are bus structures where units needing to communicate do
so through a high bandwidth medium (for example broadcast signals or infrared links to a
computer network), hierarchical topologies where supervisory or logistic units acts as
intermediaries for the communication (this places high demands on their ability to manage
high bandwidths; the top level can easily become a bottleneck) and hypercube topologies
where the units form a multidimensional cube and each unit communicates with log2(N)
others; the maximal distance between any two units is log2(N) and the total bandwidth
grows as Nlog2(N).
As can be seen, knowledge from designing multiprocessor systems can be applied to
borganisms. In both cases the problem is distributing information in a system consisting
of many subunits, and finding problems and ways of solving them that work well in
parallel.
To accommodate a changeable topology the network must be as flexible as possible. Most
likely a virtual network is the simplest solution: the mental topology is implemented as a
layer on top of another network, for example a fast packet-switched network where each
unit is linked to the nearest node, or an internet of different networks.
One interesting architecture of a borganism is a hierarchy of meta- individuals.
Individuals form meta-individuals due to high bandwidth connections and well coordinated
mental processes. These meta-individuals form higher level individuals, and so on until a
top level is reached. This suggests a hierarchical network topology where higher levels
mainly exchange high-level information keeping the necessary bandwidth low by a high level
of abstraction. A similar structure has been suggested by Marvin Minsky for the human
mind, where "agents" (simple independent subsystems with their own goals)
interact to form more complex behaviours which can be grouped into higher level agents
(Minsky 1988).
This scenario is similar to the hierarchy of minds in Stapledon's \*Star Maker\*:
advanced cultures form planetary borganisms where each individual is at the same time a
part of the planetary mind and an independent individual. The planetary minds in turn form
galactic minds in the same way, which in turn participates in the universal ultimate mind.
It is worth noting this model doesn't imply that each unit lacks individuality;
Stapledon quite clearly suggests that they can remain individuals but at the same time
participate in the borganism. One possibility is the ability to link into the borganism at
will, another is a permanent linkup which leaves some mental levels individual while
others collective.
The Psychology of Borganisms
----------------------------
>
> So perfectly organised was the life of the minded swarm that all routine activities of
> industry and agriculture had become, from the point of view of the swarm's mind,
> unconscious, like the digestive processes of a human being. The little insectoid units
> themselves carried on these consciously, though without understanding their significance;
> but the mind of the swarm had lost the power of attending to them. Its concern was almost
> wholly with such activities as called for unified conscious control, in fact with
> practical and theoretical invention of all kinds and with physical and mental exploration.
>
>
> Olaf Stapledon, Star Maker
>
>
>
How does a borganism recruit units? There are three possible answers: the individual
must willingly give up some of its individuality in exchange for the positive effects of
being part of the borganism (extended mental capacity, transhuman support etc), the
individual is involuntarily borganised, or the individual is created as a part of the
borganism.
Being a part of a borganism may or may not be reversible depending of how much the
individual unit is integrated into the collective mind. If units are individuals which are
linked together into a relatively low-bandwidth mental network for enhanced communication
and metasystem formation the process may be reversible (although the former units may have
a hard time understanding or remembering their thoughts as borganism). More intimate forms
of communication may however necessitate a permanent link to the borganism since the unit
is dependent on other units for many mental processes. Since it is likely a borganism will
need a significant amount of mental coordination to function well having units leave or
join often may be disadvantageous.
Unwilling units may not be desirable, both for the above reasons and due to the risk of
memetic infections (see the section about borganic weaknesses). If units remain relatively
unchanged when they are integrated into the collective, unwilling units are likely to be
highly disturbing and more trouble than the extra mental capacity is worth. However, if
the borganism doesn't care for the individual skills and memes of the units they can
perhaps be "mentally reformatted", turned into standardised drones a la the Borg
of Star Trek.
A recruitment method which circumvents the problems of both the other methods is to
build/grow new units to fit the borganism. This could range from having units grow up
linked to the borganism (which would likely make their minds much better adapted to a
borganic existence) to the copying of units. If units grow up in the borganism it is very
likely they will adapt well to it, likely to a much larger extent than units introduced
from the outside.
With advanced cloning techniques and a way to imprint suitable neural information it
does not appear entirely unlikely that individuals could create more or less similar
clones of themselves. Since these copies would be very similar, it is likely they will fit
into the borganism well if the original does. It is even easier if uploading is possible:
the borganism consists of infomorph entities which are interlinked much more strongly than
would be possible if the units were entirely physical; the physical presence of the
borganism could be handled by telepresence. Copying might enable a single individual to
develop into a borganism, where all units (at least originally) share his or her values,
goals and personality, making a good foundation to build a metaorganism on (assuming the
basic personality is compatible with borganisation; some people might not get along with
themselves).
### Emotion
One obvious trait of the Borgs of \*Star Trek\* is their total emotionlessness; even
in extreme situations they behave robotically. Most likely this was intended to dehumanise
them further, but there is a good reason to expect that borganisms may tend to \*appear\*
emotionless. In humans mood is conveyed through intonations, body language and especially
facial expression. This transmission is important since without functional emotional
communication many humans have a hard time functioning socially. But in a borganism
emotions need not be expressed through body language and expression, since they can be
expressed much more clearly through the intranet communication. There is no point for an
unit to smile if it is amused (or the borganism as a whole is amused) since any other unit
would be able to know exactly what mood it is in. So it is likely borganisms (unless they
try to avoid it) would appear emotionless to individual humans despite having a rich inner
life.
### Self-Sacrifice
It sometimes occurs that parents sacrifice themselves for their children, or siblings
for each other. There are sound sociobiological reasons for this which serve to ensure
genetic survival, and throughout history individuals have sacrificed themselves to ensure
the survival of their memes in an analogous fashion (Dawkins 1976). A borganism is a
memetic organism, and it might be possible for units to sacrifice themselves for the
borganism. This regularly occurs in \*Star Trek\* and real insect colonies. If all
units are roughly identical there is no great loss (except in resources) to sacrifice one
from the perspective of the borganism \*and\* the unit, which ensures the survival of
similar units and its shared memes. If the connections between units are powerful enough
or the units are infomorphs it may even be possible to make mental backups, making self-
sacrifice relatively cheap. More individual units of course have more to lose, and it is
less likely the borganism can compel them to sacrifice themselves (still, this is largely
dependent on the memes dominant in the borganism and units).
### Interaction
How would a borganism interact with other borganisms and individuals? It is important
to realise that as a metaorganism borganisms may not even perceive individuals as anything
than independent units, with roughly the same value (which may be high or low). To a
borganism the other "real" inhabitants of the world may be other borganisms, the
independent units are simply not "real" beings. This seems to be the classic
view of how borganisms would see the world, and fits in quite well with the villain
stereotype. However, there is no particular reason for why borganisms would be unable to
appreciate the individual existence of non-borganisms.
Being communication-based entities, borganisms may have an easier time communicating
with each other than individuals have. If one ignores technical problems of compatibility
and protocol, it seems quite possible for borganisms to interlink in order to communicate.
This would correspond to an extremely high bandwidth channel, enabling fast transmission
of very complex concepts. There is of course the matter of avoiding total merge and
security, but this could perhaps be dealt with by using some units as a
"firewall".
Implementation of Borganisation
-------------------------------
>
> I want to be assimilated. I want to be borg. Machines will not destroy humans; humans
> and machine will become one. Crist Clark
>
>
>
Many descriptions of borganisms have assumed telepathy, but as Olaf Stapledon pointed
out in 1937 radio could do just as well. Implementing a high-bandwidth mobile information
network is a hot research topic today, linked to research into wearable computing, mobile
offices and ubiquitious computing.
How large bandwidth is needed? We can estimate a lower bound from the bandwidth of
speech and body language, which appears to be on the order of 10-100 bits/s. A highest
upper bound would be total interconnection at the same signal density as the human mind,
or roughly 10^18 bits/s, quite an extreme range. However, the two human hemispheres
communicate closely through the corpus callosum normally with no discernible differences;
this connection has a theoretical bandwidth on the order of 10^10 bits/s, which could be
seen as a likely bandwidth needed for a deep connection between different units making
them truly parts of the same mind.
It seems likely that for any high bandwidth borganism neural interfaces are necessary,
since there are no channels into the mind with enough extra bandwidth. Hence an artificial
borganism interface is needed. Of course, it may turn out that smaller bandwidths does
accommodate the formation of borganisms (as mentioned above, the \*conscious\*
bandwidth appears to be quite small, on the order of 100 bits/s according to some
researchers).
Of course, a simple solution would be to keep the minds of the units in a computational
matrix outside the bodies, which are controlled remotely. This would require a bandwidth
similar to the spinal cord + brain nerves, on the order of 10^10 bits/s per body or so. It
may even be possible to let the bodies largely run themselves using lower level systems of
the brain and spinal cord. Since a significant amount of information is simply abstracted
away before reaching the conscious level and higher brain functions the necessary
bandwidth would be even smaller, and hence easier to send.
Designing a mobile linkup to the borganism network is nontrivial due to the estimated
demands. Current mobile networks (radio, IR) reach around 100 Kbit/s-10 Mbit/s over short
ranges <50 meters (Weiser 1991, 1996) which suggests that we need three orders of
magnitude broader bandwidth to achieve the necessary 10^10 bits/s for high bandwidth
borganisation. This does not appear impossible in principle: visible light lasers could
enable this bandwidth over line-of-sight distances, and neural activity is normally quite
sparse and likely possible to compress (roughly 5% of a set of neurons are active at any
given time; this suggests that the signals can be compressed by one to two orders of
magnitude). Other aspects of the borganism network structure are addressed by current work
in ubiquitious and mobile computing, such as flexible switching between transceivers,
error correction, energy demands and network protocols. In principle a high bandwidth
neural interface seems to be doable using near future technology.
A likely structure would consist of a high-bandwidth non-mobile digital network
("the backbone") which acts as the central switching system for the present
units. They can either be in contact with it, enabling very high bandwidth communication,
or mobile, in which case they communicate with it using radio, IR or visible laser signals
(it is amusing to note that the Borg in \*Star Trek\* often have lasers playing over
their surroundings). The signals have a short range, and need only reach the nearest
transceiver. Units outside the "hive" will not be able to communicate with the
borganism with as high bandwidth, and may have to settle for radio signals. It seems
likely that units "on their own" must deal with situations that occur more as
individuals than as parts of the borganism.
Inside the borganism network, signals are dynamically routed between units (and other
augmentative hard- and software). Low-level protocols implement packet-switching and
virtual connections, whose structure and organization is regulated by an "arbitration
layer" which could be seen as the pre-conscious part of the borganism's mind. This
arbitration layer could be implemented (as discussed above) using coordinator units,
market based systems, other approaches or mixed systems; the arbitration layer makes sure
the virtual network structure is optimal for the tasks at hand, and organizes the units
into meaningful teams and groups. These teams and groups form the true mind of the
borganism, which gathers information, solves problems and implements solutions.
Weaknesses of Borganisms
------------------------
Despite their likely high mental and practical capacity borganisms have noticeable
weaknesses, just as individual organisms do.
Many of the problems of borganisms are emergent properties of the system, not inherent
in the units themselves.
### Memetic Infection
One of the most worrying weaknesses is the spread of virulent information patterns such
as memes. Memes thrive in environments with intense communication (Bjarneskans et al.
1997), and would likely spread extremely quickly inside a borganism, infecting both
collective and unit schemata. Having a working system for memetic defence appears to be
vital for the well-being of a borganism, especially in the face of memes similar to
computer viruses (in the cybernetic environment of a borganism there is little
difference). It is not unlikely that a borganism has to retain a high degree of mental
hygiene in order not to succumb to selfish mental replicators.
Still, it is unlikely that external or internal memetic defences will be perfect,
especially since the borganism itself may accidentally create destabilizing memes during
normal thinking and internal communication. The evolution of parasites appears to be
ubiquitous in life-like (eco)systems, and the more interconnected the ecosystem is, the
greater is the complexity of coevolution and hyperparasitism (Kelly 1994). This suggests
that borganisms might generally not be able to avoid a certain level of internal selfish
replicators, and that the best strategy in dealing with them is to integrate symbiotic
replicators as a kind of immune system rather than attempt to fruitlessly eradicate them
(Moravec 1988).
### Groupthink
Groupthink is a common problem in human groups: the group becomes divorced from reality
due to its internal consensus (which may even be illusory); it fails to question its own
assumptions and to take unwelcome aspects of reality into account. If the borganism has to
keep its units in line, it is likely it will directly or indirectly counteract dissent,
which may promote groupthink. Often the best way of avoiding groupthink is to allow
dissenting minorities to present their view. On the other hand, borganisms with
sufficiently high bandwidth may be \*less\* susceptible to groupthink than human
groups. If the units can present not only their views but the mental processes which
reached these views it may become easier to judge the relative merit of the different
positions. They are no longer assertions about reality but rather different models which
can be analysed using critical thinking, empirical testing or synthesis.
### The Selfish Borg
A borganism is not just a distributed organism, it is also in some sense a social
organisation. This means that the relationship between itself and its units can become a
source of trouble. If memetic evolution and spread cannot be avoided (for example by
having units whose minds can easily be reformatted), there is the risk that discontent or
other disturbances can propagate among the units, destabilising the borganism.
For example, selfish units may be a problem. Assuming that the units retain some
autonomy, it is not unreasonable to think that some might decide to profit on the expense
of the borganism. In human groups this can be observed as the diffusion of responsibility
(the more people involved in a task the less intensely they tend to work if their results
cannot be traced back to them) and forms of social parasitism. If this strategy is
successful it can quickly spread (due to the fast transmission of memes) leading to the
weakening or dissolution of the borganism. Accountability of units may be a simple way of
dealing with this, especially since the borganism network is likely ideal for keeping
track of what everybody is doing (or not doing). Still, it is likely that selfish
strategies can develop which are hard to detect.
Discussion
----------
>
> We are the Borg. Lower your shields, and surrender your ship. We will add your
> biological and technological distinctiveness to our own. Your culture will adapt to
> service ours. Resistance is futile.
>
> Star Trek
>
>
>
Borganisms horrify some and attract others. They represent both the human fear of
losing the self and the vision of total community. The Borg of \*Star Trek\* are
depicted as inhuman and ruthless, while the "minded planets" of Stapledon are
benevolent and spiritual. Hobbes suggests that a limited form of borganisation (the
formation of societies with strong rulers) is necessary for individual survival and
well-being.
Regardless of people's reactions to them, borganisms are one of the best explored forms
of posthumanity. Unlike Jupiter brains or uploaded entities, we can at least have an
inkling of what they are and how they can be brought about; there is no immense
discontinuity between current humanity and borganisms.
Are borganisation a desirable state? The answer seems to depend on how much one values
individuality and autonomy. If these are made central values borganisms are clearly not
desirable, and to an extreme individualist it might even appear ethical to disrupt
borganisms in order to "free" the units (Morrow 1996). The case is not as clear
for voluntary borganisms where units both retain a sense of individuality and still belong
to the borganism. In this case extreme individualists would likely argue that being part
of a borganisation stunts personal development and freedom, even if it is voluntary (this
also mirrors the libertarian debates about the rights of government versus the individual,
and the legitimacy of the "social contract").
If one does not see individuality and autonomy as fundamental values there are fewer
arguments against borganisms. There is a certain worry that borganisms will be inefficient
social or memetic attractors; suboptimal evolutionary stable strategies (one possible
attractor state in the Strong Convergence Hypothesis of Boström 1997), or that the goals
of the borganism as a whole will in the long run become incompatible with the original
goals of the units which joined together. There is some evidence for the later
possibility: the goals of multicellular organisms and hives of insects call for the
sacrifice of their units, and judging from the relative amount of biomass in
multicellular/single celled and social/nonsocial insects the non-borganised lifeforms do
quite well \*from the perspective of the individual\*, although borganisation clearly
is not a disadvantage for the genes and may instead be very advantageous on the genetic
level (Dawkins 1976). If this observation can be translated into the noosphere, it
suggests that borganisms are advantageous for many strongly action-influencing memes and
meme-complexes (a possible example would be religions or ideologies) which can override
the personal self-interest of individuals. It is worth noting that in the biosphere the
borganic analogues do not dominate either species-wise or in a numerical sense;
single-celled and individual animals are still the norm. This suggests that even if
borganisms are attractors and self-supporting, they may not be so advantageous or flexible
that they out-compete all other lifestyles (especially since in an environment with
borganisms there exists a memetic evolutionary advantage to exploit them for
non-borganisms).
What are the biggest advantages of borganisms? They provide an "easy" way to
create superhuman entities (it might even be argued that we have created simple
low-bandwidth borganisms based on metasystems today: organisations and states), and there
does not appear to exist any obvious barrier to their creation (although plenty of
experimentation in group-interaction and -integration is clearly needed). Borganisms would
be able to solve some large classes of problems and implement the solutions much more
efficiently than collections of individuals, giving them a practical and economical
advantage. There is also the long-standing human dream of total community which may make
borganisms desirable to some for purely aesthetic or emotional reasons.
Regardless of one's view of borganisms it is clear that they provide a possible
posthuman state, and that they are advantageous in some situations. This is usually enough
to ensure that at least some borganisms will eventually be implemented by some group for
some reason. Resistance is futile.
Bibliography
------------
Baron, R.A., Byrne D. (1991) \*Social Psychology: Understanding Human Interaction\*
(6th ed.) Boston: Allyn & Bacon
Bjarneskans, H., Grřnnevik, B. & Sandberg, A., 1997, The Lifecycle of Memes \*Homo
Excelsior\*, [http://www.aleph.se/Trans/Cultural/Memetics/memecycle.html](file:///h:/alephweb/Trans/Cultural/Memetics/memecycle.html)
Boström, N., 1997, Predictions from Philosophy?,
Chialvo, D.R, Millonas, M.M., 1995, \*How Swarms Build Cognitive Maps\* Santa Fe
Institute working paper 95-03-033, [Abstract](http://www.santafe.edu/sfi/publications/Abstracts/95-03-033abs.html), [PostScript
version](http://www.santafe.edu/sfi/publications/Working-Papers/95-03-033.ps)
Dawkins R, (1976) The Selfish Gene , Oxford: Oxford University Press
Hellriegel, D., Slocum, J.W. Jr., Woodman, R.W. (1989) \*Organizational Behavior\*
(5th ed.) St Paul: West.
Hilgard, E.R. (1977) \*Divided Consciousness: Multiple Controls in Human Thought and
Action\*, New York: Wiley
Hobbes, T., 1651, \*Leviathan\*
Kelly, K.,\*Out of Control: the New Biology of Machines\*, London Fourth Estate
1994, ISBN 1-85702-308-0
Mark S. Miller and K. Eric Drexler, Incentive Engineering: for Computational Resource
Management in \*The Ecology of Computation\*, Bernardo Huberman (ed.) Elsevier Science
Publishers/North-Holland, 1988.
Minsky, M., 1988, \*The Society of Mind\*, Simon & Schuster
Moravec, H.,\*Mind Children: the Future of Robot and Human Intelligence\*, Cambridge
Harvard University Press, 1988 ISBN 0- 674-57618-7
Morrow, T. 1996, >H HUMOR: Borganism in the media, [http://www.aleph.se/Trans/Cultural/Fun/0173.html](file:///h:/alephweb/Trans/Cultural/Fun/0173.html)
Nicholls, P., 1982 \*The Science in Science Fiction\*, Roxby Science Fiction Limited
Putnam, F.W. (1989) \*Diagnosis and treatment of multiple personality disorder\* New
York: Guilford
Stapledon, O., 1931, \*Last and First Men\*
Stapledon, O., 1937, \*Star Maker\*
Turchin, V., Joslyn, C., 1993, The Metasystem Transition
Weiser, M., The Computer for the 21st Century, \*Scientific American\*, pp. 94-10,
September 1991,
Weiser, M. 1996, Nomadic Issues in Ubiquitous Computing, talk given at Nomadic '96
conference. [Slides](http://www.ubiq.com/hypertext/weiser/NomadicInteractive/). |
67d3b8f7-a8b3-42e9-a9ff-3367e2cb2d10 | trentmkelly/LessWrong-43k | LessWrong | Draft: Get Lucky
Here's a draft of an article that I want to post soon, but I figured I might as well get some feedback before I go ahead. If anyone else has exercises or experience with this I'd love to hear about it.
The popular science article version of the research mentioned can be found here: http://www.richardwiseman.com/resources/The_Luck_Factor.pdf
----------------------------------------
There's no fundamental propensity to good outcomes, but it looks like Luck is a thing.
There were some experiments done by Richard Wiseman (of 59 seconds fame/recommendation) investigating luck, and they found that there was a statistically significant factor which led to some people being more likely to receive unexpected benefits.
He took two groups of people, one self-proclaimed “very lucky” and the other normal, and asked them to do simple tasks, like to count the number of times a photograph appeared in a newspaper. He didn't tell them was that he had rigged the magazine to have a few convenient conveniences, like a giant sentence telling them that there were 43 photographs, or large text about how if the reader pointed this out to the researchers they could get $250.
Still, to a statistically significant degree, the “lucky” people noticed them more. What gives?
Have you ever missed an important detail because you were focused on something else? Once you're looking for something, you throw out details that fit. If you're looking for things that fit into the steps of a plan, then you're more likely to throw out stuff that's not already included.
If you weren't trying to count the number of times a word appeared, you probably would have caught the sentences. Planning on counting makes you miss details like it already being done for you.
Engineers bump into this sort of problem a lot. How many times have you spent a long time trying to code something before you found out that it was included in a library? Or tried to write a piece of code that doesn't actua |
2ace4b5a-d11e-4290-a2ab-27b2d8e6afdb | trentmkelly/LessWrong-43k | LessWrong | Feature proposal: integrate LessWrong with ChatGPT to promote active reading
I've built a little Python app I call aiRead.
Its least interesting feature is that it breaks down text the way I prefer to read it: a few sentences at a time, presented in ticker-tape fashion.
However, I've also integrated it with ChatGPT. Two of my engineered prompts are most useful. One is activated by typing "explain," which generates a rewording, a definition of jargon terms, and a simplified explanation in conversational language. The other is activated by typing "quiz," which generates an interactive quiz on the content you're reading.
The point of these features is to promote active reading. I find them tremendously helpful, such that despite the numerous shortcomings of aiRead, I'm now converting all text I have to read to be compatible with it. When I read LessWrong posts, I put them into aiRead, then generate an interactive quiz to help me understand them better. It is transformative.
I would love to see features like these integrated into LessWrong.
I know the dev team is busy, and they do a wonderful job - this is one of the best forum architectures I've ever encountered already. The only reason I'm suggesting this is because an ongoing, convenient explanation-and-quiz feature in my reading app has completely redefined the way I learn overnight, and I'm excited to see it incorporated into the other ways I consume information.
Note: Anyone is welcome to download and use aiRead, and if you do, I'd be very interested to hear about your experiences with it. You'll need a ChatGPT+ subscription. The way to use it is explained in the readme.
Also, I'm looking for collaborators on aiRead! If you'd like to help me polish this up, or ideally turn aiRead into a browser-based app, please get in touch! |
2453245a-5596-4cb7-bf36-9a33aa5d6969 | StampyAI/alignment-research-dataset/arxiv | Arxiv | A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges.
1 Introduction
---------------
Machine learning models commonly make the closed-set assumption, where the test data is drawn *i.i.d* from the same distribution as the training data. Yet in practice, all types of test input data—even those on which we have not trained the classifier—can be encountered. Unfortunately, models can assign misleading confidence values for unseen test samples[[145](#bib.bib1 "Conditional gaussian distribution learning for open set recognition"), [92](#bib.bib2 "Class anchor clustering: a loss for distance-based open set recognition"), [104](#bib.bib7 "C2ae: class conditioned auto-encoder for open-set recognition"), [161](#bib.bib3 "Classification-reconstruction learning for open-set recognition"), [163](#bib.bib8 "Unsupervised out-of-distribution detection by maximum classifier discrepancy")]. This leads to concerns about the reliability of classifiers, particularly for safety-critical applications [[50](#bib.bib5 "Unsolved problems in ml safety")]. In literature, several fields attempt to address the issue of identifying the unknowns/anomalies/out-of-distribution data in the open-world setting. In particular, the problems of anomaly detection (AD), Novelty Detection (ND), One-Class Classification (OCC), Out-of-Distribution (OOD) detection, and Open-Set Recognition (OSR) have gained significant attention owing to their fundamental importance and practical relevance. They have been used for similar tasks, although the differences and connections are often overlooked.
Specifically, OSR trains a model on K classes of an N class training dataset; then, at the test time, the model is faced with N different classes of which N−K are not seen during training. OSR aims to assign correct labels to seen test-time samples while detecting the unseen samples. Novelty detection or one-class classification is an extreme case of open-set recognition in which K is 1. In the multi-class classification setting, the problem of OOD detection is canonical to OSR: accurately classify in-distribution (ID) samples into the known categories and detect OOD data that is semantically different and therefore should not be predicted by the model. However, OOD detection encompasses a broader spectrum of learning tasks (e.g., multi-label classification, reinforcement learning) and solution space (e.g., density estimation), which we comprehensively review in this paper.
While all the aforementioned domains hold the assumption of accessing an entirely normal training dataset, anomaly detection assumes the training dataset is captured in a fully unsupervised manner without applying any filtration; therefore, it might contain some abnormal samples too. However, as abnormal events barely occur, AD methods have used this fact and proposed filtering them during the training process to reach a final semantic space that fully grasps normal features. Despite previous approaches that are mostly used in object detection and image classification domains, this setting is more common in the industrial defect detection tasks, in which abnormal events are rare and normal samples have a shared concept of normality. Fig. [1](#S1.F1 "Fig. 1 ‣ 1 Introduction ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") depicts an overview of the mentioned domains, in which the differences are shown visually. Note that even if there are differences in the formulation of these domains, they have so much in common, making them used interchangeably.
As an important research area, there have been several surveys in the literature [[14](#bib.bib30 "Anomalous instance detection in deep learning: a survey"), [110](#bib.bib29 "One-class classification: a survey"), [122](#bib.bib28 "A unifying review of deep and shallow anomaly detection"), [105](#bib.bib27 "Deep learning for anomaly detection: a review"), [18](#bib.bib26 "Deep learning for anomaly detection: a survey")], focusing on each domain independently or providing a very general notion of anomaly detection to cover all different types of datasets. Instead, we offer in-depth explanations for methodologies in respective areas. We make cross-domain bridges by which ideas can be easily propagated and inspire future research. For instance, the idea of using some outlier samples from different datasets to improve task-specific features is called Outlier Exposure in [[54](#bib.bib81 "Using self-supervised learning can improve model robustness and uncertainty")] or background modeling in [[29](#bib.bib56 "Reducing network agnostophobia")], and is very similar to semi-supervised anomaly detection in [[123](#bib.bib111 "Deep semi-supervised anomaly detection")]. Despite the shared idea, all are considered to be novel ideas in their respective domains.
In this survey, we identify the commonalities that address different but related fields. Furthermore, we try to categorize different methods based on general justifications. For instance, some surveys [[122](#bib.bib28 "A unifying review of deep and shallow anomaly detection")] categorize methods into different classes, such as distance-based methods and density-based methods. As distance-based methods can be simply converted to density-based ones, one can find this classification ambiguous. Key to our survey, we describe methods both mathematically and visually to give better insights to both newcomers and experienced researchers. Finally, comprehensive future lines of research are provided both practically and fundamentally to not only address the issues of current methods but also shed light on critical applications of these methods in different fields.
In summary our main contributions are as follows :
1. Identify the relationship between different research areas that, despite being highly correlated with each other, have been examined separately.
2. Comprehensive methodological analysis of recent eminent research, and providing a clear theoretical and visual explanation for methods reviewed.
3. Performing comprehensive tests on existing baselines in order to provide a solid ground for current and future lines of research.
4. Providing plausible future lines of research and specifying some fundamental necessities of the methods that will be presented in the future such as fairness, adversarial robustness, privacy, data-efficiency, and explainability.
\includegraphics
[trim=0cm 14cm 0cm 2cm,clip, width=]Pictures/survey\_IMg.pdf
Fig. 1: An overview of different problem formulations in novelty detection, open-set recognition, and out-of-distribution detection. As it can be seen in anomaly detection or novelty detection, one class of a dataset is considered as the normal class and the remaining as anomalies. In open-set recognition, K out of N classes are considered as normal and others as anomalies. Also, an openness score is defined based on K and N. Finally, out-of-distribution detection considers a whole training dataset as the normal dataset and other datasets as anomalies. For example, consider CIFAR-10 as the inlier distribution and MNIST as the outlier distribution.
2 General Perspectives of Method Categorizations
-------------------------------------------------
We consider a dataset with training samples (x1,y1),(x2,y2),... from the joint distribution PX,Y, where X and Y are random variables on an input space X=Rd and a label (output) space Y respectively. In-class (also called ”seen” or ”normal”) domain refers to the training data.
In AD or ND, the label space Y is a binary set, indicating normal vs. abnormal.
During testing, provided with an input sample x, the model needs to estimate P(Y=normal/seen/in-class∣X=x) in the cases of one-class setting. In OOD detection and OSR for multi-class classification, the label space can contain multiple semantic categories, the model needs to additionally perform classification on normal samples based on the posterior probability p(Y=y∣x). It is worth mentioning that in AD, input samples could contain some noises (anomalies) combined with normals; thus, the problem is converted into a noisy-label one-class classification problem; however, the overall formulation of the detection task still remains the same.
To model the conditional probabilities, the two most common and known perspectives are called generative modeling and discriminative modeling.
While discriminative modeling might be easy for OOD detection or OSR settings since there is access to the labels of training samples; nevertheless, the lack of labels makes AD, ND (OCC) challenging. This is because one-class classification problems have a trivial solution to map each of their inputs regardless of being normal or anomaly to the given label Y and, consequently, minimize their objective function as much as possible. This issue can be seen in some of the recent approaches such as DSVDD[[124](#bib.bib41 "Deep one-class classification")], which maps every input to a single point regardless of being normal or abnormal when it is trained for a large number of training epochs.
Some approaches [[41](#bib.bib42 "Deep anomaly detection using geometric transformations"), [10](#bib.bib43 "Classification-based anomaly detection for general data")], however, have made some changes in the formulation of P(Y∣X) to solve this problem. They apply a set of affine transformations on the distribution of X such that the normalized distribution does not change. Then the summation ∑|T|i=1P(Ti∣Ti(X)) is estimated, calculating the aggregated probability of each transformation Ti being applied on the input X given the transformed input Ti(X), which is equal to |T|P(Y∣X). This is similar to estimating P(Y∣X) directly; however, it would not collapse and can be used instead of estimating one-class conditional probability. This simple approach circumvents the problem of collapsing; however, it makes the problem depend on the transformations since transformed inputs must not have an intersection with each other as much as possible to satisfy the constraint of normalized distribution consistency. Therefore, as we will show later, OSR methods can employ AD approaches in combination with classification models to overcome their issues. A similar situation holds for the OOD domain.
In generative modeling, AE (Autoencoder)-based, GAN (Generative Adversarial Network)-based, and explicit density estimation based methods such as auto-regressive and flow-based models are used to model the data distribution. For AEs, there are two important assumptions.
If the auto-encoder is trained solely normal training samples:
* They would be able to reconstruct unseen normal test-time samples as precisely as training time ones.
* They would not be able to reconstruct unseen abnormal test-time samples as precisely as normal inputs.
Nevertheless, recently proposed AE-based methods show that the above-mentioned assumptions are not always true [[128](#bib.bib11 "Arae: adversarially robust training of autoencoders improves novelty detection"), [42](#bib.bib35 "Memorizing normality to detect anomaly: memory-augmented deep autoencoder for unsupervised anomaly detection"), [167](#bib.bib36 "Old is gold: redefining the adversarially learned one-class classifier training paradigm")]. For instance, even if AE can reconstruct normal samples perfectly, nonetheless with only a one-pixel shift, reconstruction loss will be high.
Likewise, GANs as another famous model family, are widely used for AD, ND, OCC, OSR, and OOD detection.
If a GAN is trained on fully normal training samples, it operates on the following assumptions:
* If the input is normal then there is a latent vector that, if it is generated, has a low discrepancy with the input.
* If the input is abnormal then there is not a latent vector that, if it is generated, has a low discrepancy with the input.
Here, the discrepancy can be defined based on the pixel-level MSE loss of generated image and test time input or more complex functions such as layer-wise distance between the discriminator’s features when it is fed generated image and test-time input. Although GANs have proved their ability to capture semantic abstractions of a given training dataset, they suffer from mode-collapse, unstable training process, and irreproducible result problems [[3](#bib.bib39 "Towards principled methods for training generative adversarial networks")].
Finally, auto-regressive and flow-based models can be used to explicitly approximate the data density and detect abnormal samples based on their assigned likelihood. Intuitively, normal samples must have higher likelihoods compared to abnormal ones; however, as will be discussed later, auto-regressive models even assign a higher amount of likelihoods to abnormal samples while they have not seen any of them at their training process, which results in a weak performance in AD, ND, OSR, and OOD detection. We will be reviewing solutions to this problem in subsequent sections. To address this issue, several remedies have been proposed in the OOD domain, which can be used in OSR, AD, and ND; however, considering the fact that OOD detection’s prevalent testing protocols might be pretty different from other domains such as AD or ND, more evaluations on their reliability are needed.
3 Anomaly and Novelty Detection
--------------------------------
Anomaly Detection (AD) and Novelty Detection (ND) have been used interchangeably in the literature, with few works addressing the differences [[109](#bib.bib15 "Ocgan: one-class novelty detection using gans with constrained latent representations"), [156](#bib.bib112 "Learning discriminative reconstructions for unsupervised outlier removal"), [152](#bib.bib106 "Effective end-to-end unsupervised outlier detection via inlier priority of discriminative network")]. In anomaly detection, there are certain intrinsic issues, which violate the assumption that training data entirely consists of normal samples. For instance, measurement noises are inevitable in physical experiments; thus, algorithms must automatically detect and focus on the normal samples in an unsupervised training process. However, this is not the case for novelty detection problems. There are many applications in which providing a clean dataset with minimum supervision is an easy task. While these domains have been separated over time, their names are still not appropriately used in literature. We review them here in a shared section.
Interests in anomaly detection go back to 1969 [[45](#bib.bib31 "Procedures for detecting outlying observations in samples")], which defines anomaly/outlier as “samples that appear
to deviate markedly from other members of the sample in which it occurs”, explicitly assuming the existence of an underlying shared pattern that a large fraction of training samples follow. This definition has some ambiguity. For example, one should define a criterion for the concept of deviation or make the term “markedly” more quantitative.
To this end, there has been a great effort both before and after the advent of deep learning methods to make the mentioned concepts more clear. To find a sample that deviates from the trend, adopting an appropriate distance metric is necessary. For instance, deviation could be computed in a raw pixel-level input or in a semantic space that is learned through a deep neural network. Some samples might have a low deviation from others in the raw pixel space but exhibit large deviations in representation space. Therefore, choosing the right distance measure for a hypothetical space is another challenge. Finally, the last challenge is choosing the threshold to determine whether the deviation from normal samples is significant.
###
3.1 Anomaly Detection With Robust Deep Autoencoders [[173](#bib.bib107 "Anomaly detection with robust deep autoencoders")]:
This work trains an AutoEncoder (AE) on a dataset containing both inliers and outliers. The outliers are detected and filtered during training, under the assumption that inliers are significantly more frequent and have a shared normal concept. This way, the AE is trained only on normal training samples and consequently poorly reconstructs abnormal test time inputs. The following objective function is used:
| | | | |
| --- | --- | --- | --- |
| | minθ||LD−Dθ(Eθ(LD))||2+λ||S||1s.t.X−LD−S=0, | | (1) |
where E and D are encoder and decoder networks, respectively. LD is supposed to be the inlier part and S the outlier part of the training data X. However, the above optimization is not easily solvable since S and θ need to be optimized jointly. To address this issue, the Alternating Direction Method of Multipliers (ADMM) is used, which divides the objective into two (or more) pieces. At the first step, by fixing S, an optimization problem on the parameters θ is solved such that LD=X−S, and the objective is ||LD−Dθ(Eθ(LD))||2. Then by setting LD to be the reconstruction of the trained AE, the optimization problem on its norm is solved when S is set to be X−LD. Since the L1 norm is not differentiable, a proximal operator is employed as an approximation of each of the optimization steps as follows:
| | | | |
| --- | --- | --- | --- |
| | proxλ,L1(xi)=⎧⎨⎩xi−λxi>λxi+λxi<−λ0−λ≤xi≤λ⎫⎬⎭ | | (2) |
Such a function is known as a *shrinkage operator* and is quite
common in L1 optimization problems. The mentioned objective function with ||S||1 separates only unstructured noises, for instance, Gaussian noise on training samples, from the normal content in the training dataset. To separate structured noises such as samples that convey completely different meaning compared to the majority of training samples, L2,1 optimization norm can be applied as:
| | | | |
| --- | --- | --- | --- |
| | n∑j=1||xj||2=n∑j=1(m∑i=1|xij|2)1/2, | | (3) |
with a proximal operator that is called block-wise soft-thresholding function [[95](#bib.bib108 "Solving structured sparsity regularization with proximal methods")]. At the test time, the reconstruction error is employed to reject abnormal inputs.
###
3.2 Adversarially Learned One-Class Classifier for Novelty Detection (ALOCC)[[126](#bib.bib34 "Adversarially learned one-class classifier for novelty detection")]:
In this work, it is assumed that a fully normal training sample is given, and the goal is to train a novelty detection model using them. At first, (R) as a Denoising Auto Encoder (DAE) is trained to (1) decrease the reconstruction loss and (2) fool a discriminator in a GAN-based setting. This helps the DAE to produce high-quality images instead of blurred outputs[[73](#bib.bib33 "Autoencoding beyond pixels using a learned similarity metric")]. This happens because, on the one hand, the AE model loss explicitly assumes independent Gaussian distributions for each of the pixels. And on the other hand, true distributions of pixels are usually multi-modal, which forces the means of Gaussians to settle in between different modes. Therefore, they produce blurry images for complex datasets. To address this issue, AEs can be trained in a GAN-based framework to force the mean of each of Gaussians to capture just one mode of its corresponding true distribution. Moreover, by using the discriminator’s output (D) instead of the pixel-level loss, normal samples that are not properly reconstructed can be detected as normal. This loss reduces the False Positive Rate (FPR) of the vanilla DAE significantly. The objective function is as follows:
| | | | |
| --- | --- | --- | --- |
| | | | (4) |
where X′ is the reconstructed output of the decoder, and pt is the distribution of the target class (i.e., normal class). This helps the model to not only have the functionality of AEs for anomaly detection but also produces higher quality outputs. Furthermore, detection can be done based on D(R(X)) as mentioned above. Fig. [2](#S3.F2 "Fig. 2 ‣ 3.2 Adversarially Learned One-Class Classifier for Novelty Detection (ALOCC)[126]: ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") depicts the overall architecture of this work.
An extended version of ALOCC, where the R(X) network has been formed as a Variational AE is presented in [[125](#bib.bib4 "Deep end-to-end one-class classifier")]. Besides, the ALOCC can not process the whole of its input data (images or video frames) at one step and needs the test samples to divide into several patches. Processing the patches makes the method computationally expensive. To address this problem, AVID [[127](#bib.bib13 "Avid: adversarial visual irregularity detection")] was proposed to exploit a fully convolutional network as a discriminator (i.e., D) to score (and hence detect) all abnormal regions in the input frame/image all at once.
\includegraphics
[width=]Pictures/ALOCC.pdf
Fig. 2: An overview of the ALOCC method. It is attempted to train an autoencoder that fools a discriminator in a GAN based setting. This helps the AE to make high quality images instead of blurred outputs. Besides, by using the discriminator’s output, a better learned similarity loss is employed instead of the pixel-level L2 reconstruction loss.
###
3.3 One-Class Novelty Detection Using GANs With Constrained Latent
Representations (OC-GAN)[[109](#bib.bib15 "Ocgan: one-class novelty detection using gans with constrained latent representations")]:
As one of the challenges, AE trained on entirely normal training samples could reconstruct unseen abnormal inputs with even lower errors.
To solve this issue, this work attempts to make the latent distribution of the encoder (EN(⋅)) similar to the uniform distribution in an adversarial manner:
| | | | |
| --- | --- | --- | --- |
| | llatent =−(Es∼U(−1,1)[log(Dl(s))]+Ex∼px[log(1−Dl(EN(x+n)))]), | | (5) |
where n∼N(0,0.2), and Dl is the latent discriminator. The discriminator forces the encoder to produce uniform distribution on the latent space. Similarly, the decoder (De(⋅)) is forced to reconstruct in-class outputs for any latent value sampled from the uniform distribution as follows :
| | | | |
| --- | --- | --- | --- |
| | lvisual =−(Es∼U(−1,1)[log(Dv(De(s)))]+Ex∼pl[log(1−Dv(x))]), | | (6) |
where Dv is called visual discriminator.
Intuitively, the learning objective distributes normal features in the latent space such that the reconstructed outputs entirely or at least roughly resemble the normal class for both normal and abnormal inputs. Also, another technique called *informative negative sample mining* is employed on the latent space to actively seek regions that produce images of poor quality. To do so, a classifier is trained to discriminate between the reconstructed outputs of the decoder and fake images, which are generated from randomly sampled latent vectors. To find informative negative samples, the algorithm (see Fig. [3](#S3.F3 "Fig. 3 ‣ 3.3 One-Class Novelty Detection Using GANs With Constrained Latent Representations (OC-GAN)[109]: ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges")) starts with a random latent-space sample and uses the classifier to assess the
quality of the generated image. After that, the algorithm solves an optimization in the latent space to make the generated image such that the discriminator detects it as fake. Finally, the negative sample is used to boost the training process.
As in previous AE-based methods, reconstruction loss is employed in combination with the objective above. Reconstruction error is used as the test time anomaly score.
\includegraphics
[width=0.6]Pictures/oc-gan.pdf
Fig. 3: The training process of OC-GAN method [[109](#bib.bib15 "Ocgan: one-class novelty detection using gans with constrained latent representations")].
###
3.4 Latent Space Autoregression for Novelty Detection (LSA) [[1](#bib.bib14 "Latent space autoregression for novelty detection")]:
This work proposes
a concept called ”surprise” for novelty detection, which specifies the uniqueness of input samples in the latent space. The more unique a sample is, the less likelihood it has in the latent space, and subsequently, the more likely it is to be an abnormal sample. This is beneficial, especially when there are many similar normal training samples in the training dataset. For visually similar training samples, AEs usually learn to reconstruct their average as the outputs to minimize the MSE error. This can result in blurry outputs and high reconstruction errors for such inputs. Instead, by using the surprise loss in combination with the reconstruction error, the issue is alleviated. Besides, abnormal samples are usually more surprising, and this increases their novelty score. Surprise score is learned using an auto-regressive model on the latent space, as Fig. [4](#S3.F4 "Fig. 4 ‣ 3.4 Latent Space Autoregression for Novelty Detection (LSA) [1]: ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows. The auto-regressive model (h) can instantiate different architectures, such as LSTM and RNN networks, to more complex ones. Also, similar to other AE-based methods, the reconstruction error is optimized. The overall objective function is as follows:
| | | | |
| --- | --- | --- | --- |
| | L=LRec(θE,θD)+λ⋅LLLK(θE,θh)=EX[||x−^x||2−λ⋅log(h(z;θh))] | | (7) |
\includegraphics
[width=]Pictures/LSA.pdf
Fig. 4: An overview of the LSA method. A surprise score is defined based on the probability distribution of embeddings. The probability distribution is learned using an auto-regressive model. Also, reconstruction error is simultaneously optimized on normal training samples.
###
3.5 Memory-Augmented Deep Autoencoder for Unsupervised Anomaly Detection (Mem-AE) [[42](#bib.bib35 "Memorizing normality to detect anomaly: memory-augmented deep autoencoder for unsupervised anomaly detection")]:
This work challenges the second assumption behind using AEs. It shows that some abnormal samples could be perfectly reconstructed even when there are not any of them in the training dataset. Intuitively, AEs may not learn to extract uniquely describing features of normal samples; as a result, they may extract some abnormal features from abnormal inputs and reconstruct them perfectly. This motivates the need for learning features that allow only normal samples to be reconstructed accurately. To do so, Mem-AE employs a memory that stores unique and sufficient features of normal training samples. During training, the encoder implicitly plays the role of a *memory address generator*. The encoder produces an embedding, and memory features that are similar to the generated embedding are combined. The combined embedding is then passed to a decoder to make the corresponding reconstructed output. Also, Mem-AE uses a sparse addressing technique that uses only a small number of memory items. Accordingly, the decoder in Mem-AE is restricted to performing the reconstruction using a small number of addressed memory items, rendering the requirement for efficient utilization of the memory items. Furthermore, the reconstruction error forces the memory to record prototypical patterns that are representative of the normal inputs. To summarize the training process, Mem-AE (1) finds address Enc(x)=z from the encoder’s output; (2) measures the cosine similarity d(z,mi) of the encoder output z with each of the memory elements mi; (3) computes attention weights w, where each of its elements is computed as follows:
| | | | |
| --- | --- | --- | --- |
| | wi=exp(d(z,mi))∑Nj=1exp(d(z,mj)); | | (8) |
(4) applies address shrinkage techniques to ensure sparsity:
| | | | |
| --- | --- | --- | --- |
| | ^wi=max(wi−λ,0).wi|wi−λ|+ϵ | | (9) |
| | | | |
| --- | --- | --- | --- |
| | E(^wt)=T∑i=1−^wi.log(^wi), | | (10) |
and finally, the loss function is defined as ([11](#S3.E11 "(11) ‣ 3.5 Memory-Augmented Deep Autoencoder for Unsupervised Anomaly Detection (Mem-AE) [42]: ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges")), where R is the reconstruction error and is used as the test time anomaly score. Fig. [5](#S3.F5 "Fig. 5 ‣ 3.5 Memory-Augmented Deep Autoencoder for Unsupervised Anomaly Detection (Mem-AE) [42]: ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows the overview of architecture.
| | | | |
| --- | --- | --- | --- |
| | L(θe,θd,M)=1TT∑t=1(R(xt,^xt)+αE(^wt)) | | (11) |
\includegraphics
[width=]Pictures/Mem-AE.pdf
Fig. 5: An overview of Mem-AE method. Each sample is passed through the encoder and a latent embedding z is extracted. Then using the cosine similarity, some nearest learned normal features are selected from the memory and the embedding^z is made out of them using a weighted averaging. At last, the reconstruction error of the decoded ^z and input is considered as novelty score.
###
3.6 Redefining the Adversarially Learned One-Class Classifier Training Paradigm (Old-is-Gold) [[167](#bib.bib36 "Old is gold: redefining the adversarially learned one-class classifier training paradigm")]:
This work extends the idea of ALOCC [[126](#bib.bib34 "Adversarially learned one-class classifier for novelty detection")]. As ALOCC is trained in a GAN-based setting, it suffers from stability and convergence issues. On one hand, the over-training of ALOCC can confuse the discriminator D because of the realistically generated fake data. On the other hand, under-training detriments the usability of discriminator features. To address this issue, a two-phase training process is proposed. In the first phase, a similar training process as ALOCC is followed :
| | | | |
| --- | --- | --- | --- |
| | L=LR+D+LR | | (12) |
As phase one progresses, a low-epoch generator model Gold for later use in phase two of the training is saved. The sensitivity of the training process to the variations of the selected epoch is discussed in the paper.
During the second phase, samples ^X=G are considered as high-quality reconstructed data. Samples ^Xlow=Gold(X) are considered as low quality samples. Then, pseudo anomaly samples are created as follows:
| | | | |
| --- | --- | --- | --- |
| | ^¯X=Gold(Xi)+Gold(Xj)2^Xpseudo=G(^¯X) | | (13) |
After that, the discriminator is trained again to strengthen its features by distinguishing between good quality samples such as {X,^X} and low quality or pseudo anomaly ones such as {^Xlow,^Xpseudo} as follows:
| | | | |
| --- | --- | --- | --- |
| | maxD(α⋅EX[log(1−D(X))]+(1−α)⋅E^X[log(1−D(^X))]+(β⋅E^Xlow[log(1−D(^Xlow))]+(1−β)⋅E^Xpseudo[log(1−D(^Xpseudo))]) | | (14) |
\includegraphics
[width=]Pictures/Old-Is-Gold.pdf
Fig. 6: An overview of Old-Is-Gold method. The architecture is very similar to ALOCC, however, it saves the weights of the network G during the training process which are called Gold. Then, it uses Gold to make some low quality samples that the discriminator must distinguish as normal samples. Also, some pseudo abnormal samples are generated by averaging each pair of normal low quality samples, which the discriminator must distinguish as fake inputs. This makes discriminator’s features rich and stables the training process.
In this way, D does not collapse similar to ALOCC, and D(G(X)) is used as the test time criterion. Fig. [6](#S3.F6 "Fig. 6 ‣ 3.6 Redefining the Adversarially Learned One-Class Classifier Training Paradigm (Old-is-Gold) [167]: ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows the overall architecture of this method.
###
3.7 Adversarial Mirrored Autoencoder (AMA)[[143](#bib.bib37 "Unsupervised anomaly detection with adversarial mirrored autoencoders")]:
The overall architecture of AMA is similar to ALOCC. However, it challenges the first assumption of AEs. It has been shown that lp norms are not suitable for training AEs in the anomaly detection domain since they cause blurry reconstructions and subsequently increase the error of normal samples. To address this problem, AMA proposes to minimize the Wasserstein distance between distributions PX,X and PX,^X. The objective function is as follows:
| | | | |
| --- | --- | --- | --- |
| | W(PX,X,PX,^X)=maxD∈Lip−1Ex∼PX[D(X,X)−D(X,^X)] | | (15) |
where ^X is the reconstructed image.
[[13](#bib.bib113 "Understanding and improving interpolation in autoencoders via an adversarial regularizer")] showed that by forcing the linear combination of latent codes of a pair of data points to look realistic after decoding, the encoder learns a better representation of data. Inspired by this, AMA makes use of ^Xinter—obtained by decoding the linear combination of some randomly sampled inputs:
| | | | |
| --- | --- | --- | --- |
| | minGmaxD∈Lip−1Ex∼PX[D(X,X)−D(X,^Xinter)] | | (16) |
To further boost discriminative abilities of D, inspired by [[23](#bib.bib114 "Generative ensembles for robust anomaly detection")] proposing normal samples reside in the typical set [[24](#bib.bib90 "Elements of information theory (wiley series in telecommunications and signal processing)")] while anomalies reside outside of it, a Gaussian regularization equal to L2 norm is imposed on the latent representation. Then using the Gaussian Annulus Theorem [[150](#bib.bib115 "High-dimensional probability: an introduction with applications in data science")] stating in a
d-dimensional space, the typical set resides with high probability at a distance of √d from the origin, synthetic negatives (anomalies), are obtained by sampling the latent space outside and closer to the typical set boundaries. Therefore, the final objective function is defined as follows:
| | | | |
| --- | --- | --- | --- |
| | | | (17) |
Fig. [7](#S3.F7 "Fig. 7 ‣ 3.7 Adversarial Mirrored Autoencoder (AMA)[143]: ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows an overview of the method and the test time criterion is ||f(X,X)−f(X,G(E(X))||1 where f is the penultimate layer of D.
\includegraphics
[width=]Pictures/AMA.pdf
Fig. 7: An overview of AMA method. The architecture is very similar to ALOCC, however, it does not try to minimize reconstruction error between x and ^x. Instead, a discriminator is trained with Wasserstein Loss to minimize the distance between the distribution (x,x) and (x,^x). In this way, it forces ^x to be similar to x without using any lp norm that causes blurred reconstructed outputs. Besides some samples are generated using the embeddings of normal samples that must fool the discriminator. This, as the authors mention, make the latent space consistent. Moveover, some negative latent vectors are sampled from low probability areas that the discriminator must distinguish as fake ones.
###
3.8 Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery (AnoGAN)[[133](#bib.bib38 "Unsupervised anomaly detection with generative adversarial networks to guide marker discovery")]:
This work trains a GAN on normal training samples, then at the test time, solves an optimization problem that attempts to find the best latent space z by minimizing a discrepancy. The discrepancy is found by using a pixel-level loss of the generated image and input in combination with the loss of discriminator’s features at different layers when the generated and input images are fed. Intuitively, for normal test time samples, a desired latent vector can be found despite abnormal ones. Fig. [8](#S3.F8 "Fig. 8 ‣ 3.8 Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery (AnoGAN)[133]: ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows the structure of the method.
As inferring the desired latent vector through solving an optimization problem is time-consuming, some extensions of AnoGAN have been proposed. For instance, Efficient-GAN [[169](#bib.bib40 "Efficient gan-based anomaly detection")] tries to substitute the optimization problem by training an encoder E such that the latent vectors z′ approximate the distribution of z. In this way, E is used to produce the desired latent vector, significantly improving test time speed. Fig. [9](#S3.F9 "Fig. 9 ‣ 3.8 Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery (AnoGAN)[133]: ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows differences. The following optimization problem ([18](#S3.E18 "(18) ‣ 3.8 Unsupervised Anomaly Detection with Generative Adversarial Networks to Guide Marker Discovery (AnoGAN)[133]: ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges")) is solved at the test time to find z. The anomaly score is assigned based on how well the found latent vector can minimize the objective function.
| | | | |
| --- | --- | --- | --- |
| | minz (1−λ)⋅∑|x−G(z)|+λ⋅∑|D(x)−D(G(z))| | | (18) |
\includegraphics
[width=]Pictures/AnoGAN.pdf
Fig. 8: An overview of AnoGAN method. At first, a generator network G and a discriminator D are trained jointly on normal training samples using the standard training loss, which yields a semantic latent representation space. Then at the test time an optimization problem that seeks to find an optimal latent embedding z that mimics the pixel-level and semantic-level information of the input is solved. Intuitively for normal samples a good approximation of inputs can be found in the latent space, however, for abnormal inputs it is not approximated well.
\includegraphics
[width=]Pictures/efficient-gan.pdf
Fig. 9: Figure A shows the architecture of AnoGAN. Figure B shows the architecture of Efficient-GAN. The encoder E mimics the distribution of latent variable z. The discriminator D learns to distinguish between joint distribution of (x,z) instead of x.
###
3.9 Oc-Svm
Primary AD methods would use statistical approaches such as comparing each sample with the mean of the training dataset to detect abnormal inputs, which imposes an ungeneralizable and implicit Gaussian distribution assumption on the training dataset. In order to reduce the number of presumptions or relieve the mentioned deficiencies of traditional statistical methods, OC-SVM [[134](#bib.bib32 "Support vector method for novelty detection.")] was proposed. As the name implies, OC-SVM is a one-class SVM maximizing the distance of training samples from the origin using a hyper-plane, including samples on one side and the origin on its other side. Eq. [19](#S3.E19 "(19) ‣ 3.9 OC-SVM ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows the primal form of OC-SVM that attempts to find a space in which training samples lie on just one side, and the more distance of origin to the line, the better solution is found for the optimization problem.
| | | | |
| --- | --- | --- | --- |
| | minw,ρ,ξi12∥w∥2+1vnn∑1ξi−ρsubject to\quad(w⋅Φ(xi))⩾ρ−ξi,i∈1,...,nξi⩾0,i∈1,…,n | | (19) |
Finding a hyper-plane seems to be a clever idea for putting a shared normal constraint on the training samples; however, as half-space is not as tight enough to grasp unique shared normal features, it produces many false negatives. Therefore, similarly, Support Vector Data Description(SVDD) tries to find the most compact hyper-sphere that includes normal training samples. This is much tighter than hyper-plan, thus finds richer normal features, and is more robust against unwanted noises as Eq. [20](#S3.E20 "(20) ‣ 3.9 OC-SVM ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows. Both of the mentioned methods have soft and hard margin settings that in the former despite the later one, some samples can violate the border and stay out of it even if they are normal. OC-SVM have some implicit assumptions too; for example, it assumes training samples obey a shared concept, which is conceivable due to the one-class setting. Also, it works pretty well on the AD setting in which the number of outliers is significantly lower than normal ones however fails on high dimensional datasets.
| | | | |
| --- | --- | --- | --- |
| | minR,aCn∑i=1ξi+R2subject to\quad∥ϕ(xi)−a∥2⩽R2+ξi,i∈1,…,nξi⩾0,i∈1,…,n | | (20) |
###
3.10 Deep One-Class Classification (DeepSVDD)[[124](#bib.bib41 "Deep one-class classification")]:
This method can be seen as an extension of SVDD using a deep network. It assumes the existence of shared features between training samples and tries to find a latent space in which training samples can be compressed into a minimum volume hyper-sphere surrounding them. Fig. [10](#S3.F10 "Fig. 10 ‣ 3.10 Deep One-Class Classification (DeepSVDD)[124]: ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows the overall architecture. The difference *w.r.t* traditional methods is the automatic learning of kernel function ϕ by optimizing the parameters W. To find the center c of the hyper-sphere, an AE is first trained on the normal training samples, then the average of normal sample latent embeddings is considered as c. After that, by discarding the decoder part, the encoder is trained using the following objective function:
| | | | |
| --- | --- | --- | --- |
| | minW1nn∑i=1||ϕ(xi;W)−c||22+λ2L∑l=1||Wl||2F, | | (21) |
where Wl shows the weights of encoder’s lth layer. At the test time, anomaly score is computed based on ||ϕ(x;W∗)−c||2 where W∗ denotes trained parameters.
\includegraphics
[width=]Pictures/DSVDD.pdf
Fig. 10: An overview of DSVDD method. It is attempted to find a minimum volume hyper-sphere that contains all training samples. As we find the minimum radius, the more distance to center is expected for abnormal samples compared to normal ones.
###
3.11 Deep Semi-Supervised Anomaly Detection [[123](#bib.bib111 "Deep semi-supervised anomaly detection")]:
This is the semi-supervised version of DSVDD which assumes a limited number of labeled abnormal samples. The loss function is defined to minimize the distance of normal samples from a pre-defined center of a hyper-sphere while maximizing abnormal sample distances. The objective function is defined as follows:
| | | | |
| --- | --- | --- | --- |
| | minW1n+mn∑i=1||ϕ(xi;W)−c||2+ηn+mm∑j=1(||ϕ(^xi;W)−c||2)^yj+λ2L∑l=1||Wl||2F | | (22) |
Note that, as mentioned above, there is access to (^x1,^y1),...,(^xm,^ym)∈X×Y with Y={−1,+1} where ^y=1 denotes known normal samples and ^y=−1 otherwise. c is specified completely similar to DSVDD by averaging on the latent embeddings of an AE trained on normal training samples. From a bit theoretical point of view, AE’s objective loss function helps encoder maximizes I(X;Z) that X denotes input variables and Z denotes latent variables. Then, it can be shown that ([22](#S3.E22 "(22) ‣ 3.11 Deep Semi-Supervised Anomaly Detection [123]: ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges")) minimizes the entropy of normal sample latent embeddings while maximizing it for abnormal ones as follows:
| | | | |
| --- | --- | --- | --- |
| | H(Z)=E[−log(P(Z)]=−∫Zp(z)logp(z)dz≤12log((2πe)ddet(Σ))∝logσ2 | | (23) |
As the method makes normal samples compact, it forces them to have low variance, and consequently, their entropy is minimized. The final theoretical formulation is approximated as follows:
| | | | |
| --- | --- | --- | --- |
| | maxp(z∣x)I(X;Z)+β(H(Z−)−H(Z+)) | | (24) |
###
3.12 Deep Anomaly Detection Using Geometric Transformations (GT)[[41](#bib.bib42 "Deep anomaly detection using geometric transformations")]:
GT attempts to reformulate the one-class classification problem into a multi-class classification. GT defines a set of transformations that do not change the data distribution, then trains a classifier to distinguish between them. Essentially, the classifier is trained in a self-supervised manner. Finally, a Dirichlet distribution is trained on the confidence layer of the classifier to model non-linear boundaries. Abnormal samples are expected to be in low-density areas since the network can not confidently estimate the correct transformation. At the test time, different transformations are applied to the input, and the sum of their corresponding Dirichlet probabilities is assigned as the novelty score. An overview of the method is shown in Fig. [11](#S3.F11 "Fig. 11 ‣ 3.12 Deep Anomaly Detection Using Geometric Transformations (GT)[41]: ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges").
\includegraphics
[width=]Pictures/GT.pdf
Fig. 11: An overview of GT method. A set of transformations is defined, then, a classifier must distinguish between them. Having trained the classifier in a self-supervised manner, a Dirichlet distribution is trained on the confidence layer to model their boundaries.
###
3.13 Effective End-To-End Unsupervised Outlier Detection via Inlier Priority of Discriminative Network [[152](#bib.bib106 "Effective end-to-end unsupervised outlier detection via inlier priority of discriminative network")]:
In this work, similar to GT, a self-supervised learning (SSL) task is employed to train an anomaly detector except in the presence of a minority of outliers or abnormal samples in the training dataset. Suppose the following training loss is used as the objective function:
| | | | |
| --- | --- | --- | --- |
| | LSS(xi∣θ)=−1KK∑y=1log(P(y)(x(y)i∣θ)), | | (25) |
where x(y)i is obtained by applying a set of transformations O(.∣y), and y indicates each transformation. The anomaly detection is obtained based on the objective function score. We take the rotation prediction task as an example. During training, the classifier learns to predict the amount of rotation for normal samples. During test time, different rotations are applied on inputs, the objective function scores for normal samples would be lower than abnormal ones.
However, due to the presence of abnormal samples in the training dataset, the objective score for abnormal samples may not always be higher. To address this issue, it is shown that the magnitude and direction of gradient in each step have a significant tendency toward minimizing inlier samples’ loss function. Thus the network produces lower scores compare to abnormal ones. [[152](#bib.bib106 "Effective end-to-end unsupervised outlier detection via inlier priority of discriminative network")] exploits the magnitude of transformed inliers
and outliers’ aggregated gradient to update wc, i.e. ||∇(in)wcL|| and ||∇(out)wcL||, which are shown to follow this approximation:
| | | | |
| --- | --- | --- | --- |
| | E(||∇(in)wc⋅L||2)E(||∇(out)wc⋅L||2)≈N2inN2out | | (26) |
where E(⋅) denotes the probability expectation. As Nin≫Nout, normal samples have more effect on the training procedure. Also, by projecting and averaging the gradient in the direction of each of the training sample’s gradient (−∇θL(x)⋅−∇θL(xi)||−∇θL(xi)||), the stronger effect of inlier distribution vs outlier one is observed again. Fig. [12](#S3.F12 "Fig. 12 ‣ 3.13 Effective End-To-End Unsupervised Outlier Detection via Inlier Priority of Discriminative Network [152]: ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows the effect empirically.
\includegraphics
[width=]Pictures/E3d\_1.pdf
Fig. 12: The average magnitude of gradient for inliers and outliers with respect to the number of iterations. The class used as inliers is in brackets.
###
3.14 Classification-Based Anomaly Detection for General Data (GOAD)[[10](#bib.bib43 "Classification-based anomaly detection for general data")]:
This work is very similar to GT. It trains a network to classify between different transformations; however, instead of using cross-entropy loss or training a Dirichlet distribution on the final confidences, it finds a center for each transformation and minimizes the distance of each transformed data with its corresponding center as follows :
| | | | |
| --- | --- | --- | --- |
| | P(m′∣T(x,m))=e−||f(T(x,m))−cm′||2∑^me−||f(T(x,m))−c^m||2 | | (27) |
where the centers cm are given by the average feature over the training set for every transformation i.e. cm=1N∑x∈Xf(T(x,m)). For training f, two options are used. The first one is using a simple cross entropy on P(m′∣T(x,m)) values, and the second one is using the center triplet loss [[46](#bib.bib116 "Triplet-center loss for multi-view 3d object retrieval")] as follows:
| | | | |
| --- | --- | --- | --- |
| | ∑imax(0,||f(T(xi,m))−cm||2+s−minm′≠m||f(T(xi,m))−cm′||2) | | (28) |
where s is a margin regularizing the distance between clusters.
The idea can be seen as the combination of DSVDD and GT in which GT’s transformations are used, and different compressed hyperspheres are learned to separate them. M different transformations transform each sample at the test time, and the average of correct label probabilities is assigned as the anomaly score.
###
3.15 CSI: Novelty Detection via Contrastive Learning on Distributionally Shifted Instances [[146](#bib.bib109 "Csi: novelty detection via contrastive learning on distributionally shifted instances")]:
This work attempts to formulate the problem of novelty detection into a contrastive framework similar to SimCLR [[21](#bib.bib102 "A simple framework for contrastive learning of visual representations")].
The idea of contrastive learning is to learn an encoder fθ to extract the necessary information to
distinguish similar samples from the others. Let x be a query, x+, and x− be a set of positive and negative samples respectively, z be the encoder’s output feature or the output of an additional projection layer gϕ(fθ(x)) for each input, and suppose sim(z,z′) is cosine similarity. Then, the primitive form of the contrastive loss is defined as follows:
| | | | |
| --- | --- | --- | --- |
| | Lcon(x,x+,x−):=−1|x+|log∑x′∈x+esim(z(x),z(x′))/τ∑x′∈x+∪x−esim(z(x),z(x′))/τ | | (29) |
Specifically for SimCLR, the contrastive loss above is converted into the formulation below:
| | | | |
| --- | --- | --- | --- |
| | LSimCLR(B;T):=12BB∑i=1Lcon(^x(1)i,^x(2)i,^B−i)+Lcon(^x(2)i,^x(1)i,^B−i) | | (30) |
where (^x(1)i,^x(2)i)=(T1(xi),T2(xi)) for the transformations T1 and T2 from a set of transformations T, B:={xi}Bi=1, and ^B−i:={^x(1)j}j≠i∪{^x(2)j}j≠i.
However, contrastive learning requires defining a set of negative samples. To this end, a collection of transformations that shift the distribution of training samples (S) is specified to make the desired negative set when applied on each input. For instance, rotation or patch permutation completely shifts the original input samples’ distribution; therefore, they can be used as negative samples. Also, another set of transformations called align transformations (T) is defined as not changing the distribution of training images as much as (S). Then the *Contrasting shifted instances* (CSI) loss can be defined as follows:
| | | | |
| --- | --- | --- | --- |
| | Lcon−SI:=LSimCLR(∪s∈SBs;T) | | (31) |
where Bs:={s(xi)}Bi=1. Here, CSI regards each distributionally-shifted sample as an OOD with respect to the original sample. The goal is to discriminate an in-distribution sample from other OOD i.e., (s∈S) samples.
Further, to facilitate fθ to discriminate each shifted instance, a classification loss for classifying shifted instances is defined in combination with Lcon-SI. To do so, a linear layer for modeling
an auxiliary softmax classifier pcls-SI(yS∣x) is added to fθ as in GT or GOAD:
| | | | |
| --- | --- | --- | --- |
| | Lcls-SI=12B1K∑s∈S∑^xs∈Bs−logpcls-SI(ys=s∣^xs), | | (32) |
where ^Bs is the batch augmented from Bs. In testing, the cosine similarity to the
nearest training sample in {xm} multiplied by the norm of representation ||z(x)|| is used. The contrastive loss increases the norm of in-distribution samples, as it is an easy way to minimize the cosine similarity of identical samples by increasing the denominator of ([29](#S3.E29 "(29) ‣ 3.15 CSI: Novelty Detection via Contrastive Learning on Distributionally Shifted Instances [146]: ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges")). To further improve the scoring criterion, ensembling over random augmentations or utilizing shifting transformations similar to GOAD can be employed.
###
3.16 Uninformed Students: Student-Teacher Anomaly Detection With Discriminative Latent Embeddings [[12](#bib.bib12 "Uninformed students: student-teacher anomaly detection with discriminative latent embeddings")]:
This work trains a teacher network using metric learning and knowledge distillation techniques to provide a semantic and discriminative feature space. The teacher T is obtained by first training a network ^T that embeds patch-sized images p into a metric space. Then, fast dense local feature extraction for an entire input image can be achieved by a deterministic network transformation from ^T to T, as described in [[6](#bib.bib110 "Fast feature extraction with cnns with pooling layers")]. To train ^T, a large number of training patches p are obtained by randomly cropping an image database, for instance, ImageNet. Then using the following knowledge distillation loss, the knowledge of a pre-trained network P is distilled into ^T as follows:
| | | | |
| --- | --- | --- | --- |
| | Lk(^T)=||D(^T(p))−P(p)||2, | | (33) |
where D is used to align the size of output spaces. This helps the use of computationally efficient network ^T instead of P at the test time while the required knowledge is distilled.
To further enrich the feature space of ^T, an SSL method is employed such that the feature space of patches that are obtained by applying small translations, small changes in the image luminance, and the addition of Gaussian noise to p be similar. This set is called p+ as opposed to p−, which is obtained using random cropping regardless of the neighborhood to the patch p. The SSL loss is defined as follows:
| | | | |
| --- | --- | --- | --- |
| | δ+=||^T(p)−^T(p+)||2δ−=min{||^T(p)−^T(p−)||2,||^T(p+)−^T(p−)||2}Lm(^T)=max{0,δ+δ+−δ−} | | (34) |
where δ>0 denotes the margin parameter. Finally, to reduce the redundancy of feature maps of ^T, a compactness loss which minimizes their cross-correlation is used as follows:
| | | | |
| --- | --- | --- | --- |
| | Lc(^T)=∑i≠jcij | | (35) |
therefore, the final loss is Lc(^T)+Lm(^T)+Lk(^T).
Having a teacher T trained comprehensively to produce a d dimensional feature maps for each pixel of an input image, an ensemble of student networks is forced to approximate the feature maps of T for each pixel locating at the row r and column c as follows:
| | | | |
| --- | --- | --- | --- |
| | L(Si)=1wh∑(r,c)||μSi(r,c)−(yT(r,c)−μ)diag(σ)−1||22 | | (36) |
where μ and σ are used for data normalization, note that the receptive field of each student is limited to a local image region p(r,c), this helps the students to obtain dense predictions for each image pixel with only a single forward pass without having actually to crop the patches p(r,c). At the test time, as students have only learned to follow their teacher on normal training samples, their average ensemble error can be used to detect abnormal samples. Intuitively, they wouldn’t follow their teacher on abnormal samples and produce a high average error.
.
###
3.17 Self-Supervised Learning for Anomaly Detection and Localization (CutPaste) [[79](#bib.bib47 "CutPaste: self-supervised learning for anomaly detection and localization")]:
This work designs a simple SSL task that to capture local, pixel-level regularities instead of global, semantic-level ones. While GT or GOAD utilize transformations such as rotation, translation, or jittering, CutPaste cuts a patch of its training input and copies it in another location. The network is trained to distinguish between defected samples and intact ones. Extra auxiliary tasks such as cutout and Scar can be used in combination with the cut-paste operation. After training, a KDE or Gaussian density estimator is trained on the confidence scores of normal training samples, which is used at the test time. Due to the simplicity of the method, it may overfit easily on the classification task. However; several experiments in the paper show contrary to this assumption.
An overview of the method can be seen in Fig. [13](#S3.F13 "Fig. 13 ‣ 3.17 Self-Supervised Learning for Anomaly Detection and Localization (CutPaste) [79]: ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges").
\includegraphics
[width=]Pictures/CutPaste.pdf
Fig. 13: An overview of CutPaste. A network is train to classify artificially made defected samples from the original ones. In spite of the simplicity of the method, it does not overfit on MVTecAD dataset.
###
3.18 Multiresolution Knowledge Distillation for Anomaly Detection (Multi-KD)[[130](#bib.bib9 "Multiresolution knowledge distillation for anomaly detection")]:
Generative models are suited for detecting pixel-level anomalies; however, they may fail on complex, semantic-level ones. In contrast, discriminative models are good at capturing semantics. Designing an SSL task that captures both semantic and syntax is not easy. To address this issue, Multi-KD attempts to mimic the intermediate layers of a VGG pre-trained network—called intermediate knowledge—into a simpler network using knowledge distillation. This way, multi-resolution modeling of the normal training distribution is obtained and can be used to detect both pixel-level and semantic-level anomalies at the test time. Here, the concept of knowledge is defined to be the length and direction of a pre-trained network on ImageNet [[27](#bib.bib44 "Imagenet: a large-scale hierarchical image database")]. A cloner network has a simple yet similar overall architecture compared to the source, making its knowledge similar to the source on normal training samples. At the test time, the cloner can follow the source on the normal test time samples; however, it fails for the abnormal ones. This results in a high discrepancy that can be used at the test time. Fig. [14](#S3.F14 "Fig. 14 ‣ 3.18 Multiresolution Knowledge Distillation for Anomaly Detection (Multi-KD)[130]: ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows the overall architecture.
Similar methods exist, which model the distribution of different layers of a pre-trained network using multivariate Gaussian descriptors such as PaDiM [[26](#bib.bib45 "PaDiM: a patch distribution modeling framework for anomaly detection and localization")] or [[120](#bib.bib46 "Modeling the distribution of normal data in pre-trained deep features for anomaly detection")]. An overview of the architecture of [[120](#bib.bib46 "Modeling the distribution of normal data in pre-trained deep features for anomaly detection")] is shown in the Fig. [15](#S3.F15 "Fig. 15 ‣ 3.18 Multiresolution Knowledge Distillation for Anomaly Detection (Multi-KD)[130]: ‣ 3 Anomaly and Novelty Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges").
\includegraphics
[width=]Pictures/Multi-KD.pdf
Fig. 14: An overview of Multi-KD method. It is tried to distill multi-resolution knowledge of a pre-trained source network on the normal training distribution in to a simpler cloner one. The overall structure of source and cloner is similar to each other. Therefore, the cloner distills source’s knowledge in the corresponding layers. At the test time their discrepancy would be low for normal inputs, in spite of abnormal ones.
\includegraphics
[width=]Pictures/Modeling\_Normal.pdf
Fig. 15: An overview of [[120](#bib.bib46 "Modeling the distribution of normal data in pre-trained deep features for anomaly detection")]. It models the different layers’ distribution of a pre-trained network using Gaussian descriptors. Then, at the test time, the probability of being normal is computed using the training time means and variances.
4 Open-set Recognition
-----------------------
Open-set recognition (OSR) receives more supervision than AD or ND. In this setting, K normal classes are given at the training time. In testing, N classes with N−K unknown and K known classes exist. The objective is to identify unknown classes while classifying the known ones. It has a lot of applications in which labeling normal dataset is more feasible, or gathering a cleaned dataset without having any abnormal sample is possible. As there is more supervision, training data can be categorized into four classes:
* *known known classes (KKC):* Training samples that we know they are known. They are already given and labeled.
* *known unknown classes (KUC):* Training samples that we know they are not known. This means, they do not belong to the known categories. For example, background images, or any image that we know is not categorized into the known classes are in this group. They are already given and labeled.
* *unknown known classes (UKC):* Training samples that we do not know they are known. For example, known test time samples are in this group. These are not given at the training phase.
* *unknown unknown classes (UUC):* Training samples that we do not know they are not known. For example, unknown test time samples are in this group. These are not given at the training phase.
The space that is far from known classes, including KKC and KUC, is called open space O[[131](#bib.bib48 "Toward open set recognition")]. Labeling any sample in this space has a risk value denoted by RO. The risk factor is usually represented as Eq. [37](#S4.E37 "(37) ‣ 4 Open-set Recognition ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") in which So is the overall measuring space, and f is the measurable recognition (classification) function [[131](#bib.bib48 "Toward open set recognition")]. The value of f is 1 if the input belongs to KKC otherwise 0.
| | | | |
| --- | --- | --- | --- |
| | RO(f)=∫Of(x)dx∫Sof(x)dx | | (37) |
As discussed in [[40](#bib.bib49 "Recent advances in open set recognition: a survey")], there are some difficulties in defining the practical formulation of openness; therefore, we use the definition of [[40](#bib.bib49 "Recent advances in open set recognition: a survey")] as in Eq. [38](#S4.E38 "(38) ‣ 4 Open-set Recognition ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges"), where CTR is the set of classes used in training, and CTE is the set of classes used in the testing.
| | | | |
| --- | --- | --- | --- |
| | O=1−√2×|CTR||CTR|+|CTE| | | (38) |
Larger values of O correspond to more open problems, and 0 is assigned to a completely closed one. OSR problem can be defined as finding recognition functions f in such a way that it minimizes the risk factor RO.
###
4.1 Towards Open Set Deep Networks (OpenMax)[[8](#bib.bib50 "Towards open set deep networks")]:
This work addresses the problem that classification models produce overconfident scores for unseen test time samples. Due to the normalization in softmax computation, two samples with completely different logit scores may have the same confidence score distribution. Instead of using the confidence score, OpenMax resorts to the logit scores that are shown by *Activation Vector* (AV).
AV of each sample captures the distribution over classes.
The mean AV (MAV) is defined to be the average of AV values across all samples.
As for each input sample, the value in AV corresponding to the ground truth is supposed to be high; its distance to the corresponding value of MAV would be high too. Considering the distance of each element in AVs from the corresponding element in MAV as a random variable, correctly classified inputs would have the highest distances for ground truth elements. This might happen for a few classes that are not ground truth while having a strong relationship with ground truth. For instance, the class leopard as the ground truth and cheetah as a near class. To model the distribution of these highest values, the Extreme Value Theorem (EVT) can be used as follows:
EVT : Let (s1, s2, …, sn) be a sequence of i.i.d samples. Let Mn = maxs1, …, sn. If a sequence of pairs of real numbers (an, bn) exists such that each 0≤an and
| | | | |
| --- | --- | --- | --- |
| | limn→∞P(Mn−bnan≤x)=F(x) | | (39) |
Then if F is non-degenerate distribution function, it belongs to one of three extreme value distributions [[113](#bib.bib51 "Statistical inference using extreme order statistics")]. For this work, it is Weibull distribution. By modeling the distribution of extremes for each class, one can easily compute the probability of each test input being an extreme and discard remaining ones.
In summary, during training, the CDF of extreme values for each class or element of AVs are approximated using EVT and correctly classified training samples. For each test-time input sample, the top α elements in AV are selected, representing ground truth and near classes; then, their CDFs are assigned to variables ws(i)(x) as Fig. [16](#S4.F16 "Fig. 16 ‣ 4.1 Towards Open Set Deep Networks (OpenMax)[8]: ‣ 4 Open-set Recognition ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows. These probabilities represent the chance of each AV’s element vj(x) to be the maximum. After that, an arbitrary class is added to the existing ones as the representative of unknown unknown samples, let’s call class 0, and all activation values are normalized as follows:
| | | | |
| --- | --- | --- | --- |
| | ^v(x)=v(x)∘w(x)^v0(x)=∑ivi(x)(1−wi(x)) | | (40) |
The normalization is performed so that the activation vector of UUC would be high if ws(i)(x)s are small. Finally, softmax is computed based on the normalized activation vector, and unknown or uncertain inputs are rejected. All hyper-parameters such as ϵ,α are tuned using a validation set including both seen and unseen samples.
\includegraphics
[width=]Pictures/Open\_Max.pdf
Fig. 16: The inference algorithm of open-max. The class 0 is added in addition to other classes and considered as the unknown unknown class; Then, its probability is computed for each of input samples, and the probability of top k values is normalized with respect to unknown its probability. At last, if the largest value is for the class 0 or the confidence of top k values is less than a specific amount, the sample is discarded.
###
4.2 Generative OpenMax for Multi-Class Open Set Classification (G-OpenMax) [[39](#bib.bib53 "Generative openmax for multi-class open set classification")]:
This work is similar to OpenMax except for artificially generating UUC samples with GANs and fine-tuning OpenMax. This removes the need to have a validation dataset. To generate UUCs, a conditional GAN is trained on the KKCs as follows:
| | | | |
| --- | --- | --- | --- |
| | minϕmaxθ=Ex,c∼Pdata[log(Dθ(x,c))]+Ez∼Pz,c∼Pc[log(1−Dθ(Gϕ(z,c),c))] | | (41) |
New samples are generated by interpolating the KKCs’ latent space. If the generated sample is not classified as one of the mixing labels, the image is considered UUC. Finally, another classifier called NetG (see Fig. [17](#S4.F17 "Fig. 17 ‣ 4.2 Generative OpenMax for Multi-Class Open Set Classification (G-OpenMax) [39]: ‣ 4 Open-set Recognition ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges")) is trained using both of the UUC and KKC samples such that UUCs can be assigned to the class 0 of the classifier. The inference steps are similar to OpenMax except that the normalization process for the class 0 and other classes is the same.
\includegraphics
[width=]Pictures/G-OpenMax.pdf
Fig. 17: The overall architecture of G-OpenMax compared to the base OpenMax. As it is obvious, everything is completely the same except the unknown unknown sample generation, which is done using a GAN.
###
4.3 Open Set Learning with Counterfactual Images [[98](#bib.bib55 "Open set learning with counterfactual images")]:
This work follows the idea of generating UUC samples as in G-OpenMax. A generated input is made similar to a KKC yet not to be assigned to the same class. Such generated inputs are called counter-factual examples. As these samples are near the boundary of UUCs, they can better help approximate the real UUC distribution.
Specifically, a classifier is trained on k classes. A GAN-based method such as ALOCC is trained. The discriminator is trained using a Wasserstein critic with gradient penalty as follows:
| | | | |
| --- | --- | --- | --- |
| | LD=∑x∈XD(G(E(x)))−D(x)+P(D)LG=∑x∈X||x−G(E(x))||1−D(G(E(x))), | | (42) |
where P(D)=λ⋅(||∇^xD(^x)−1||2).
The counter-factual samples are generated by optimizing a latent vector z, which has a small distance to the latent vector of an arbitrary training sample x while its decoded image produces low confidence scores for each of the known classes:
| | | | |
| --- | --- | --- | --- |
| | z∗=minz||z−E(x)||22+log(1+K∑i=1eCK(G(z))i), | | (43) |
where CK is the classifier and CK(G(z))i is the logit of the counterfactual image G(z) for class i. The generated UUCs are used in combination with KKCs to train a new classifier CK+1, which is used at the test time.
###
4.4 Reducing Network Agnostophobia [[29](#bib.bib56 "Reducing network agnostophobia")]:
In applications such as object detection, there is usually a class called background. On the internet, a large number of samples can be crawled, which can be used as “background” for a specific task. This work employs background samples as an auxiliary KUC distribution to train a classifier. The training facilitates KUC to have small feature magnitudes while KKCs have larger magnitudes with a defined margin. Also, the entropy of the confidence layer is maximized for the background samples, which is equivalent to increasing the classifier’s uncertainty for such inputs. The training employs a simple entropic open-set loss that maximizes the entropy of confidence scores, along with the objectosphere loss minimizing the L2 norm of final features. Fig. [18](#S4.F18 "Fig. 18 ‣ 4.4 Reducing Network Agnostophobia [29]: ‣ 4 Open-set Recognition ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows the effect of each loss on the geometric location of each class’s samples at the final layer. At the test time, the thresholding is based on Sc(x).||F(x)||, where F(x) is the representation from penultimate layer.
\includegraphics
[width=]Pictures/agnostophobia.pdf
Fig. 18: The effect of different loss functions on the last layer of a classifier. The classifier is trained on MNIST dataset. The background class is considered to be NIST letters [[44](#bib.bib57 "NIST special database 19 handprinted forms and characters 2nd edition")]. Black dots show samples from Devanagari [[106](#bib.bib58 "Off-line nepali handwritten character recognition using multilayer perceptron and radial basis function neural networks")] as UUCs, and the gray lines show the boundaries of different classes. (a) shows the last layer when the network is trained only with softmax loss. (b) shows similar setting to (a); however, background samples are used as a separate class. (c) shows the last laeyr when the network is trained with objectosphere loss. The figures in the bottom are histograms of softmax probability values for samples of KKCs with green color and UUCs with the red one.
###
4.5 Class Conditioned Auto-Encoder for Open-Set Recognition (C2AE) [[104](#bib.bib7 "C2ae: class conditioned auto-encoder for open-set recognition")]:
In this work, the second assumption behind using AEs is employed, in which abnormal test time samples are supposed not to be reconstructed as well as normal ones; however, in OSR, despite AD or ND, training labels could boost the AE abilities. Here, an AE is used as the meta-recognition function while its encoder plays the role of a classifier for the recognition task. Intuitively, this work wants the encoder to classify each passed sample correctly and provide embeddings by which the reconstruction of original input is possible. Furthermore, it imposes other constraints to force encoder embeddings not to be easily converted to each other, e.g., by applying linear transformations, which prevents the AE from utilizing the learned features to reconstruct abnormal/unseen inputs.
To this end, at first, the encoder that is a classifier is trained and fixed. Then for a given input Xi a match vector lm=lymi∈{−1,1}K where K is the number of classes is defined such that l is equal to 1 for the yith element and −1 otherwise. Similarly some none-match vectors lnm=lynmj for any random ynmj≠yi sampled randomly from labels are considered. After that two neural networks Hγ and Hβ with parameters Θγ and Θβ are defined to receive match and none-match vectors and produce some linear transformations by which the encoder embeddings are transformed. Finally, match transformation must be constructed perfectly; however, none-match ones are forced to have high reconstruction loss as follows:
| | | | |
| --- | --- | --- | --- |
| | zi=F(Xi),γymi=Hγ(lymi),γynmi=Hγ(lynmi)βymi=Hβ(lymi),βynmi=Hβ(lynmi)zilm=γymi∘zi+βymi,zilnm=γynmi∘zi+βynmi^Xmi=G(zlm).^Xnmi=G(zlnm).Lmr=1NN∑i=1||Xi−^Xmi||1Lnmr=1NN∑i=1||Xnmi−^Xnmi||1 | | (44) |
where ^Xnmis are sampled randomly based on randomly sampled none-match vectors, and the objective loss is:
| | | | |
| --- | --- | --- | --- |
| | minΘγ,Θβ,ΘGα⋅Lmr+(1−α)⋅Lnmr | | (45) |
The embedding transformation technique is called FiLM [[112](#bib.bib60 "Film: visual reasoning with a general conditioning layer")]. This means that a given input would be reconstructed well only when the corresponding match vector is used. Therefore, for each test-time input, the reconstruction error under different match vectors is computed, and the rejection decision is made based on their minimum value. If the minimum value is greater than a threshold τ it is discarded; else, the encoder’s output is assigned. To obtain τ in a principled way, the optimum threshold is found using EVT on the distribution of match and none-match reconstruction error values i.e ||Xi−^Xmi||1 and ||Xi−^Xnmi||1 for each i and randomly sampled none-match vectors. Assuming the prior probability of observing unknown samples is pu, the probability of error as a function of threshold τ is shown in Eq. [46](#S4.E46 "(46) ‣ 4.5 Class Conditioned Auto-Encoder for Open-Set Recognition (C2AE) [104]: ‣ 4 Open-set Recognition ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") where Gm and Gmn are the extreme value distributions of the match and none-match errors.
| | | | |
| --- | --- | --- | --- |
| | τ∗=minτPerror(τ)=minτ[(1−pu)∗Pm(r≥τ)+pu∗Pnm(−r≤−τ)]=minτ[(1−pu)∗(1−Gm(τ))+pu∗(1−Gnm(τ))] | | (46) |
###
4.6 Deep Transfer Learning For Multiple Class Novelty Detection (DTL) [[111](#bib.bib61 "Deep transfer learning for multiple class novelty detection")]:
This work also follows the idea of using a background dataset (called reference dataset). Similar to [[29](#bib.bib56 "Reducing network agnostophobia")], DTL addresses the deficiencies of using softmax loss in OSR. A new loss function called *membership loss* is proposed. Specifically, each activation score value fi of the penultimate layer is normalized into [0,1] using the sigmoid function. The normalized scores can be interpreted as the probability of the input image belonging to each individual class. Ideally, given the label y, f(x)i should be 1 when y=i and 0 otherwise. The loss function is defined as Eq. [47](#S4.E47 "(47) ‣ 4.6 Deep Transfer Learning For Multiple Class Novelty Detection (DTL) [111]: ‣ 4 Open-set Recognition ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges").
| | | | |
| --- | --- | --- | --- |
| | LM(x,y)=[1−σ(f(x)y)]2+λ⋅1c−1c∑i=1,i≠y[σ(f(x)i)]2 | | (47) |
Another technique for improving the detection performance is based on the *“globally negative filters”*. Filters that provide evidence for a particular class are considered as positive filters and vice versa. For pre-trained neural networks, it has been shown that only a small fraction of final feature maps are activated positively. Furthermore, some filters are always activated negatively, indicating irrelevance for all known classes. By discarding inputs that activate globally negative filters, a novel sample is less likely to produce high activation scores. To learn such filters for domain-specific task, DTL trains two parallel networks with shared weights up to the last layer—the first one solves a classification task on the reference dataset, and the second one solves the domain-specific classification tasks in combination with membership loss. If the reference and domain-specific datasets do not share much information, they provide negative filters for each other. Also, since reference dataset consists of diverse classes, those learned filters can be considered globally negative filters. Finally, filters of the parallel network in combination with the confidence scores of the domain-specific classifier are used for novelty detection. Fig. [47](#S4.E47 "(47) ‣ 4.6 Deep Transfer Learning For Multiple Class Novelty Detection (DTL) [111]: ‣ 4 Open-set Recognition ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows the overall network architecture.
\includegraphics
[width=]Pictures/DTL.pdf
Fig. 19: An overview of DTL method. Two parallel networks are trained to produce confidence scores and globally negative filters. The network above solves a simple classification task on the reference dataset, and the one below solves membership loss in combination with the domain specific classification task. At the test time, a detection is done by putting a threshold on the confidence of maximum class.
###
4.7 Classification-Reconstruction Learning for Open-Set Recognition (CROSR) [[161](#bib.bib3 "Classification-reconstruction learning for open-set recognition")]:
This work follows the similar idea as C2AE. In particular, CROSR employs an encoder network for classification and producing the latent vectors for reconstruction task. Importantly, latent vector z used for the reconstruction task and penultimate layer y used for the classification task are not shared. The reason is that there is an excessive amount of information loss in the penultimate layer, which makes distinguishing between unknown and known samples hard. The overall procedure can be described as follows:
| | | | |
| --- | --- | --- | --- |
| | (y,z)=f(x),p(Ci∣x)=Softmaxi(y),^x=g(z) | | (48) |
Moreover, to preserve information at different levels of abstraction, each layer of the encoder is compressed into a latent vector zi. The latent vector is then decoded to minimize the reconstruction error of the corresponding layer as follows:
| | | | |
| --- | --- | --- | --- |
| | xl+1=fl(xl),zl=hl(xl),^xl=gl(^xl+1+^hl(zl)) | | (49) |
where fl and gl are layers of the encoder and decoder respectively. ^hl is a reprojection to the original space of xl. The autoencoding structure is based on a ladder network [[116](#bib.bib62 "Semi-supervised learning with ladder networks")].
The final latent vector z is the concatenation of each zi. EVT is employed as in OpenMax [[8](#bib.bib50 "Towards open set deep networks")]. However, the distance is not only computed on the penultimate layer y but also on the latent vector z. Specifically, EVT is applied on the joint [y,z]. Test-time detection is performed similar to OpenMax. Fig. [20](#S4.F20 "Fig. 20 ‣ 4.7 Classification-Reconstruction Learning for Open-Set Recognition (CROSR) [161]: ‣ 4 Open-set Recognition ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows an overview of the method.
\includegraphics
[width=]Pictures/CROSR.pdf
Fig. 20: An overview of the proposed method compared to similar methods in the literature. As it is obvious, an encoder is used for the classification task, and the reconstruction task is done in a ladder network.
###
4.8 Generative-Discriminative Feature Representations For Open-Set Recognition (GDFR) [[108](#bib.bib63 "Generative-discriminative feature representations for open-set recognition")]:
Similar to CROSR, this work trains a discriminative model in combination with a generative one. Discriminative approaches may lose important features that are utilitarian for distinguishing between seen and unseen samples. Generative modeling can provide complementary information. Similar to GT, GDFR employs SSL to improve the features of the discriminator. A shared network performs both the classification and SSL tasks, predicting the geometric transformations applied to the input.
Moreover, a generative model such as AE is used, producing reconstructed outputs ^x for a given input x. Then the collection of input-reconstruction pairs (x,^x) are passed to the discriminator network for classification and SSL tasks. The disparity between ^x and x for unseen samples better helps the discriminator network detect them. Fig. [21](#S4.F21 "Fig. 21 ‣ 4.8 Generative-Discriminative Feature Representations For Open-Set Recognition (GDFR) [108]: ‣ 4 Open-set Recognition ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows the method.
\includegraphics
[width=]Pictures/GDFR.pdf
Fig. 21: An overview of GDFR. As it can be seen, a classification network is trained on both SSL and classification tasks to make the learned features richer. Also, to benefit from generative modeling, each input is passed to an AE, and further tasks are applied on the joint of reconstructed and input images.
###
4.9 Conditional Gaussian Distribution Learning for Open Set Recognition (CGDL) [[145](#bib.bib1 "Conditional gaussian distribution learning for open set recognition")]:
The main idea of this research is very similar to CROSR. However, CGDL uses the probabilistic ladder network based on variational encoding and decoding [[144](#bib.bib64 "How to train deep variational autoencoders and probabilistic ladder networks")]. The overall encoding process for the lth layer is as follows:
| | | | |
| --- | --- | --- | --- |
| | xl=Conv(xl−1)hl=Flatten(xl)μl=Linear(hl)σ2l=Softplus(Linear(hl)) | | (50) |
where the ”Softplus” operation is log(1+exp(⋅)); and, the final representation vector z is defined as μ+σ∘ϵ where ϵ∼N(0,I). Similarly, for the decoding process we have:
| | | | |
| --- | --- | --- | --- |
| | ^cl+1=Unflatten(^zl+1)^xl+1=ConvT(^cl+1)^hl+1=Flatten(^xl+1)^μl=Linear(^hl+1)^σ2l=Softplus(Linear(^hl+1))q-μl=^μl+^σ−2l+μl+σ−2l^σ−2l+σ−2lq-σ2l=1^σ−2l+σ−2l^zl=q-μl+q-σ2l∘ϵ | | (51) |
During training, samples are passed into the encoder to estimate μ and σ for each layer. The mean and variance can be used as priors for the corresponding decoding layer. The final embedding z of the encoder’s top layer is used for the joint classification task and decoding process. The distribution of the encoder’s final layer is forced to be similar to different multivariate Gaussian pkθ(z)=N(z;μk,I), where k is the index of known classes and μk is obtained by a fully-connected layer which maps the one-hot encoding of input’s label to the latent space. Each layer of the decoder is a Gaussian probability distribution in which a prior of its mean and variance is added by the corresponding layer of encoder statistics. Putting it together, the training objective function is as follows:
| | | | |
| --- | --- | --- | --- |
| | | | (52) |
where Lc is the classification error and
| | | | |
| --- | --- | --- | --- |
| | qθ(^xl|^xl+1,x)=N(^xl;q-μl,q-σ2l)qθ(^xl|^xl+1)=N(^xl;^μl,^σ2l) | | (53) |
At the test time, both the probability of final latent space with respect to pkθ(z) and reconstruction error for each test input are used for detection. If an input is preserved, the output of classification is considered as the true class. As Fig. [22](#S4.F22 "Fig. 22 ‣ 4.9 Conditional Gaussian Distribution Learning for Open Set Recognition (CGDL) [145]: ‣ 4 Open-set Recognition ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows, the latent vector z is considered as a prior of which the distribution of all classes should have a low KL distance. This is similar to the role of N(0,I) in the baseline VAE.
\includegraphics
[width=]Pictures/CGDL.pdf
Fig. 22: An overview of CGDL method. Compared to CROSR a probabilistic ladder network is used instead of a deterministic one; however, the classification task is similar. The probability distributions of decoder network contain a prior from the corresponding layer statistics of the encoder network. Also, Despite VAE in which there is a prior N(0,I) on the latent space; here, the prior is learned, and each class distribution is tried to have a low KL distance with it.
###
4.10 Hybrid Models for Open Set Recognition [[170](#bib.bib65 "Hybrid models for open set recognition")]:
In this work, a classification network is trained in combination with a flow-based generative model. Generative models in the pixel-level space may not produce discernible results for unseen and seen samples, and they are not robust against semantically irrelevant noises. To address this issue, a flow-based model is applied on the feature representation space instead of pixel-level space (see Fig. [23](#S4.F23 "Fig. 23 ‣ 4.10 Hybrid Models for Open Set Recognition [170]: ‣ 4 Open-set Recognition ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges")).
The reason for using flow-based models is their handy and comprehensive theoretical abilities.
The training loss is a simple cross-entropy loss in combination with negative log-likelihood used for the training of the flow-based model. At the test time, the thresholding is applied on the likelihood of each input, and if it is preserved, the classifier’s output is assigned as the in-class label.
\includegraphics
[width=]Pictures/Hybrid.pdf
Fig. 23: An overview of the hybrid models for open set recognition. As it can be seen, a classification network is trained in combination with a generative flow-based model. At the test time, the probability of the latent vector is considered as the criterion of rejection, and the classifier assigns a label if input is not rejected.
###
4.11 Learning Open Set Network With Discriminative Reciprocal Points (RPL) [[19](#bib.bib66 "Learning open set network with discriminative reciprocal points")]:
Similar to Mem-AE, the idea of prototype features is used in this work. The goal is to learn a set of prototypes or reciprocal points, which can assign labels to each input based on the distance w.r.t each prototype. RPL helps the model better adjust the boundaries of different classes compared to softmax or OpenMax, and decreases the risk factor. Initially, random reciprocal points are chosen. The location of reciprocal points and the weights of a classifier network are adjusted to minimize the classification loss. This forces the network to locate features of each class near some specific reciprocal points to yield the desired class boundary using at least a set of points. To decrease the risk factor, samples of each class are forced to have a margin with respect to their reciprocal points, which is learned during the training process. Eq. [54](#S4.E54 "(54) ‣ 4.11 Learning Open Set Network With Discriminative Reciprocal Points (RPL) [19]: ‣ 4 Open-set Recognition ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows the mathematical formulation:
| | | | |
| --- | --- | --- | --- |
| | d(fθ(x),Pk)=1MM∑i=1||fθ(x)−pki||22p(y=k|x,fθ,P)=eγd(fθ(x),Pk)∑Ni=1eγd(fθ(x),Pi)Lo=1MM∑i=1||d(fθ(x),pki)−Rk||22, | | (54) |
where Pk is the set of kth class reciprocal points, pki is a reciprocal point, M is the number of reciprocal points for each class, N is the number of classes, Rk is the margin for each class, and γ is a tunable hyper-parameter.
###
4.12 A Loss for Distance-Based Open Set Recognition (CAC) [[92](#bib.bib2 "Class anchor clustering: a loss for distance-based open set recognition")]:
The idea of this work is similar to RPL and GOAD. For each class, CAC defines an anchor vector of dimension N—the number of classes. For each vector, the element corresponding to the class label is 1 and 0 otherwise. For each training sample, the learning process forces its logit scores to be in a compact ball w.r.t true class anchor vector while having a large distance from anchors of other classes. CAC can also be seen as a multi-class DSVDD. The training loss function is as follows.
| | | | |
| --- | --- | --- | --- |
| | d=e(z,C)=(||z−c1||2,...,||z−cN||2)LT(x,y)=log(1+N∑j≠yedy−dj)LA(x,y)=dy=||f(x)−cy||2LCAC(x,y)=LT(x,y)+λ⋅LA(x,y) | | (55) |
###
4.13 Few-Shot Open-Set Recognition Using Meta-Learning (PEELER) [[83](#bib.bib67 "Few-shot open-set recognition using meta-learning")]:
In this work, the idea of meta-learning is combined with open set recognition. Meta-learning enable learning general features that can be easily adapted to any unseen task. Meta-learning is also called learning to learn. Due to the ability to work in few-shot settings, meta-learning can be useful in low data regimes. At the meta iteration i, meta-model h is initialized with the one produced by the previous meta-iteration. Let (Ssi,Tsi)Nsi=1 be a meta-training dataset with Ns number of training problems, two steps are performed. First, an estimate h′ of the optimal model for the training set Ssi is produced. Then the test set Tsi is used for finding a model with a suitable loss function L as the Eq. [56](#S4.E56 "(56) ‣ 4.13 Few-Shot Open-Set Recognition Using Meta-Learning (PEELER) [83]: ‣ 4 Open-set Recognition ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows.
| | | | |
| --- | --- | --- | --- |
| | h∗=arg minh∑(xk,yk)∈TsiL[yk,h′(xk)] | | (56) |
To adopt meta-learning in OSR, a classification loss in combination with open-set loss is used. Moreover, the test set Tsi is augmented with some unseen samples. The overall loss function is defined as Eq. [57](#S4.E57 "(57) ‣ 4.13 Few-Shot Open-Set Recognition Using Meta-Learning (PEELER) [83]: ‣ 4 Open-set Recognition ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges"), where Lc is a simple classification loss, and Lo maximizes the entropy of unknown samples on the known classes.
| | | | |
| --- | --- | --- | --- |
| | h∗=arg minh{∑(xk,yk)∈Csi∈Tsi|ykLc[yk,h′(xk)]+λ⋅∑(xk,yk)∈Tsi|yk∈CuiLo[h′(xk)]} | | (57) |
At the test time, the average features of correctly classified samples is obtained as a prototype point and used for the rejection of unseen samples.
###
4.14 Learning Placeholders for Open-Set Recognition (PROSER) [[174](#bib.bib68 "Learning placeholders for open-set recognition")]:
This work tries to train a classifier that can place between
target class and non-target classes. A dummy classifier is added to the softmax layer of the model with a shared feature extractor. Then it is forced to have the second maximum value for the correctly classified samples. When the classifier encounters novel inputs, the dummy classifier produces high values since all known classes are non-targets. Dummy classifier can be seen as the instance-dependent threshold which can well fit every known class. The loss function is defined as the Eq. [58](#S4.E58 "(58) ‣ 4.14 Learning Placeholders for Open-Set Recognition (PROSER) [174]: ‣ 4 Open-set Recognition ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges"), where ^f((x)/y means making the most probable class zero.
| | | | |
| --- | --- | --- | --- |
| | L1=∑(x,y)∈Dtrl(^f(x),y)+βl(^f((x)/y,k+1) | | (58) |
Moreover, the mixup technique [[149](#bib.bib69 "Manifold mixup: better representations by interpolating hidden states")] is added to the loss function to boost unseen samples detection. The mixup representations are introduced to approximate the unseen sample’s distribution and should be classified as the dummy class k+1. Finally, rejecting each test time sample is done based on the probability of the dummy class.
###
4.15 Counterfactual Zero-Shot and Open-Set Visual Recognition [[165](#bib.bib70 "Counterfactual zero-shot and open-set visual recognition")]:
This work attempts to make abnormal samples in a counter-factual faithful way. As the paper mentions, most of the generative approaches such as G-OpenMax do not produce desired fake samples, the distributions of which do not resemble real distribution of unseen samples. To this end, a β-VAE [[57](#bib.bib71 "Beta-vae: learning basic visual concepts with a constrained variational framework")] is used to make sample attribute variable Z and the class attribute variable Y independent. The β-VAE loss function is similar to simple VAE; however, the KL term is induced by a coefficient β. This is shown to be highly effective in learning a disentangled sample attribute Z [[57](#bib.bib71 "Beta-vae: learning basic visual concepts with a constrained variational framework")]. For disentangling Y from Z, the proposed approach makes counter-factual samples by changing the variable Y to have a large distance with the given input x in spite of samples that are generated by changing the variable Z. To make counter-factual samples faithful, a Wasserstein GAN [[4](#bib.bib72 "Wasserstein generative adversarial networks")] loss is used for a discriminator D(X,Y), which verifies the correspondence between the generated counter-factual image and the assigned label. At last, generated samples can be used to boost the performance of any OSR problem.
5 Out-of-distribution Detection
--------------------------------
OOD detection aims to identify test-times samples that are semantically different from the training data categories, and therefore should not be predicted into the known classes. For instance, one could train the model on CIFAR-10 (as in-distribution data), and then evaluating on CIFAR-100 [[100](#bib.bib75 "Reading digits in natural images with unsupervised feature learning")] as an out-of-distribution dataset, as CIFAR-10 and CIFAR-100 have mutually exclusive classes.
In the multi-class setting, the problem of OOD detection is canonical to OSR: accurately classifying samples from the known classes while detecting the unknowns. However, OOD detection encompasses a broader spectrum of learning tasks (e.g., multi-label classification) and solution space (e.g., density estimation without classification). Some approaches relax the constraints imposed by OSR and achieve strong performance.
We next review some of recent OOD detection works and their differences.
###
5.1 A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks [[52](#bib.bib73 "A baseline for detecting misclassified and out-of-distribution examples in neural networks")]:
This work coined “out-of-distribution detection” and showed how to evaluate deep learning out-of-distribution detectors. Whereas previous work anomaly detection for deep classifiers often had low quality or proprietary datasets, this repurposed existing datasets to create out-of-distribution datasets, enabling easier evaluation. It proposes using the maximum softmax probability (MSP) to detect out-of-distribution samples, namely maxkp(y=k∣x). A test sample with a large MSP score is detected as an in-distribution (ID) example rather than out-of-distribution (OOD) example. It showed that a simple maximum probability score can be useful for detection in vision, natural language processing, and speech recognition settings, but there is much room for improvement. It also showed p(y∣x) models can be useful for out-of-distribution detection and that p(x) models are not necessarily needed. To this day, it still serves as a general-purpose baseline that is nontrivial to surpass. Concurrent OSR work proposed additional modifications to softmax probabilities for detection [[8](#bib.bib50 "Towards open set deep networks")].
###
5.2 Enhancing The Reliability of Out-of-distribution Image Detection in Neural Networks (ODIN) [[81](#bib.bib78 "Enhancing the reliability of out-of-distribution image detection in neural networks")]:
In this work, a technique called temperature scaling was employed. Although it has been used in other domains such as knowledge distillation [[58](#bib.bib79 "Distilling the knowledge in a neural network")], the main novelty of this work is showing the usefulness of this technique in the OOD domain. In temperature scaling, the softmax score is computed as in Eq. [59](#S5.E59 "(59) ‣ 5.2 Enhancing The Reliability of Out-of-distribution Image Detection in Neural Networks (ODIN) [81]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges"). OOD samples are then detected at the test time based on thresholding the maximum class probability. This simple approach, in combination with adding a small controlled noise, has shown significant improvement compared to the baseline approach MSP. ODIN further shows that adding one step gradient to the inputs in the direction of improving the maximum score has more effect on the in-class samples, and pushes them to have larger margin with the OOD samples.
| | | | |
| --- | --- | --- | --- |
| | Si(x;T)=exp(fi(x)/T)∑Nj=1exp(fj(x)/T) | | (59) |
The paper also provided mathematical explanation for the effect of temperature scaling on out-of-distribution detection. This can be seen in the Taylor approximation of the softmax score (expanded around the largest logit output f^y(x)):
| | | | |
| --- | --- | --- | --- |
| | Si(x;T)=exp(fi(x)/T)∑Nj=1exp(fj(x)/T)=1∑Nj=1exp(fj(x)−fi(x)T)≈1N−1T∑Nj=1[f^y(x)−fj(x)]+12T2∑Nj=1[f^y(x)−fj(x)]2 | | (60) |
A sufficiently large temperature T has a strong smoothing effect that transforms the softmax score back to the logit space—which effectively distinguishes ID vs. OOD. In particular, the ID/OOD separability is determined by the U1=∑Nj=1,j≠^y[f^y(x)−fj(x)] and U2=∑Nj=1,j≠^y[f^y(x)−fj(x)]2. The former measures the extent to which the largest unnormalized output of the neural network deviates from the remaining outputs; while the latter measures the extent
to which the remaining smaller outputs deviate from each other. For the in-class samples, U1 and E[U2∣U1] are higher than OOD ones. Mathematically and empirically, ODIN is not sensitive to T when it is large enough to satisfy the Taylor approximation. For example, the paper shows that simply applying T=1000 can yield effective performance boost without hyper-parameter tuning. In general, using temperature scaling can improve the separability more significantly than input preprocessing.
Note that ODIN differs from confidence calibration, where a much milder T is employed. While calibration focuses on representing the
true correctness likelihood of ID data only, the ODIN score is designed to maximize the gap between ID and OOD data and may no longer be meaningful from a predictive confidence standpoint. As seen in Fig. [24](#S5.F24 "Fig. 24 ‣ 5.2 Enhancing The Reliability of Out-of-distribution Image Detection in Neural Networks (ODIN) [81]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges"), the ODIN scores are closer to 1/N where N is the number of classes.
\includegraphics
[width=]Pictures/ODIN.pdf
Fig. 24: An overview of ODIN, a post-hoc method that uses temperature scaling and input perturbation to amplify the ID/OOD separability.
###
5.3 A Simple Unified Framework for Detecting
Out-of-Distribution Samples and Adversarial Attacks [[78](#bib.bib77 "A simple unified framework for detecting out-of-distribution samples and adversarial attacks")]:
This work was inspired from the idea of Linear Discriminant Analysis (LDA) in which P(X=x∣Y=y) is considered to be a multivariate Gaussian distribution. In order for P(Y=y∣X=x) to be similar to a softmax form, it is assumed that the feature space of penultimate layer follows the Gaussian distribution. Therefore, a mean and variance vector is simply estimated from features of each class, and a multivariate Gaussian is fit to them. In order to check validity of the assumptions, it uses the Mahalanobis distance of the test time images to perform the classification instead of the softmax function. Surprisingly, the results are comparable or better than softmax, which supports the assumptions. It performs OOD detection using Mahalanobis distance to the closest class-conditional Gaussian distribution. Furthermore, to improve the performance, features in different layers are ensembled and a small controlled noise is added to test samples as shown in Eq. [61](#S5.E61 "(61) ‣ 5.3 A Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks [78]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges"), where M(x) is the Mahalanobis distance with mean of the closest class-conditional Gaussian distribution. A similar idea has been discussed earlier in [[120](#bib.bib46 "Modeling the distribution of normal data in pre-trained deep features for anomaly detection")].
| | | | |
| --- | --- | --- | --- |
| | ^x=x+ϵ⋅sign(∇xM(x)) | | (61) |
###
5.4 Predictive Uncertainty Estimation via Prior Networks (DPN) [[90](#bib.bib76 "Predictive uncertainty estimation via prior networks")]:
This work discusses three different sources of uncertainty: (1) data uncertainty, (2) distributional uncertainty, and (3) model uncertainty. The importance of breaking down the final uncertainty into these terms was discussed. For instance, model uncertainty might happen because of model’s lack of capacity to approximate the given distribution well. On the other hand, data uncertainty might happen because of the intrinsic intersection of similar classes. For instance, classifying between different kinds of dogs has more data uncertainty than solving a classification problem with completely separate classes. Distributional uncertainty is related to the problem of AD, ND, OSR, and OOD detection. The goal of this work was to estimate the distributional uncertainty for each input and compare it with the data and model uncertainties. Data uncertainty P(wc∣x∗,θ) can be defined by the posterior distribution over class labels given a set of parameter θ. Model uncertainty P(θ∣D) is defined by the posterior distribution over the parameter given data D. These two types of uncertainties could be combined to give rise to the distributional uncertainty as shown in Eq. [62](#S5.E62 "(62) ‣ 5.4 Predictive Uncertainty Estimation via Prior Networks (DPN) [90]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges").
| | | | |
| --- | --- | --- | --- |
| | P(wc∣x∗,D)=∫P(wc∣x∗,θ)P(θ∣D)dθ | | (62) |
As computing the integral is not tractable, the above formula is usually converted into the Eq. [63](#S5.E63 "(63) ‣ 5.4 Predictive Uncertainty Estimation via Prior Networks (DPN) [90]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges"), where q(θ) is an approximation of P(θ∣D).
| | | | |
| --- | --- | --- | --- |
| | P(wc∣x∗,D)≈1MM∑i=1P(wc∣x∗,θi),θi∼q(θ) | | (63) |
Each P(wc∣x∗,D) can then be seen as a categorical distribution located on a simplex, as shown in Fig. [25](#S5.F25 "Fig. 25 ‣ 5.4 Predictive Uncertainty Estimation via Prior Networks (DPN) [90]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges").
\includegraphics
[width=0.5]Pictures/DPN.pdf
Fig. 25: The figure shows a distribution over the categorical distributions for modeling uncertainty using both model uncertainty and data uncertainty.
Now to extract distributional uncertainty from the model uncertainty, Eq. [63](#S5.E63 "(63) ‣ 5.4 Predictive Uncertainty Estimation via Prior Networks (DPN) [90]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") is decomposed into the following Eq. [64](#S5.E64 "(64) ‣ 5.4 Predictive Uncertainty Estimation via Prior Networks (DPN) [90]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges").
| | | | |
| --- | --- | --- | --- |
| | P(wc∣x∗,D)=∬P(wc∣μ)P(μ∣x∗,θ)P(θ∣D)dμdθ, | | (64) |
where P(wc∣μ) is a categorical distribution given a realization from a Dirichlet distribution, P(μ∣x∗,θ) is Dirichlet distribution given the input and model parameters θ, and P(θ∣D) is the distribution over model parameters given the dataset D. For simplicity, in this work P(θ∣D) is equal to δ(θ−^θ) and is produced by the output of a deep network. Therefore P(μ∣x∗,θ)=P(μ∣x∗,^θ)=Dir(μ∣α), and α=f(x∗,^θ), where f(.,.) is represented by a deep neural network.
At the training time, the Dirichlet Prior Network (DPN) is expected to yield a flat distribution over the simplex for OOD samples, indicating large uncertainty in the mapping from x to y. Some out-of-distribution data is used to minimize the KL distance of Dir(μ∣α) and the flat Dirichlet distribution. For the in-class samples, the KL divergence between Dir(μ∣α) and a sharp, sparse Dirichlet distribution is minimized. The objective Dirichlet distributions are obtained by pre-setting their parameters during training process. During test time, different criterion such as max probability, last layer’s entropy, and distributional uncertainty as in Eq. [65](#S5.E65 "(65) ‣ 5.4 Predictive Uncertainty Estimation via Prior Networks (DPN) [90]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") are used for OOD detection.
| | | | |
| --- | --- | --- | --- |
| | I(y,μ∣x∗,D)=H[EP(μ∣x∗,D)[P(y∣μ)]]−EP(μ∣x∗,D)[H[P(y∣μ)]] | | (65) |
###
5.5 Confidence-calibrated Classifiers for Detecting Out-of-distribution Samples [[77](#bib.bib80 "Training confidence-calibrated classifiers for detecting out-of-distribution samples")]:
This work attempted to maximize the entropy of confidence scores for OOD samples, similar to [[29](#bib.bib56 "Reducing network agnostophobia")]. Similar to [[98](#bib.bib55 "Open set learning with counterfactual images")], it generates OOD samples by jointly training a GAN and a classifier. As shown in Eq. [66](#S5.E66 "(66) ‣ 5.5 Confidence-calibrated Classifiers for Detecting Out-of-distribution Samples [77]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges"), the first term solves a classification task on in-class samples, the second term uses KL divergence to make the confidence score distribution of generated OOD samples uniform. The remainder terms train the GAN on the in-class samples. Note that, the GAN is forced to generate high-quality OOD samples that produce high uncertainty when they are passed to the classifier. Therefore, the generated samples are located on the boundaries of in-class and outlier distributions. The paper also shows that leveraging on-boundary in-class samples significantly improves its confidence calibration.
| | | | |
| --- | --- | --- | --- |
| | minGmaxDminθEPin(^x,^y)[−logPθ(y=^y∣^x)]+βEPG(x)[KL(U(y)||Pθ(y∣x))]+EPin(^x)[logD(x)]+EPG(x)[log(1−D(x))] | | (66) |
Test time OOD detection is done based on thresholding of the maximum softmax value.
###
5.6 Deep Anomaly Detection with Outlier Exposure (OE) [[53](#bib.bib82 "Deep anomaly detection with outlier exposure")]:
This work introduced Outlier Exposure (OE) and reported extensive experiments on its usefulness for various settings. When applied to classifiers, the Outlier Exposure loss encourages models to output a uniform softmax distribution on outliers, following [[77](#bib.bib80 "Training confidence-calibrated classifiers for detecting out-of-distribution samples")]. More generally, the Outlier Exposure objective is
| | | |
| --- | --- | --- |
| | E(x,y)∼Din[L(f(x),y)+λ⋅Ex′∼D\textscOEout[L\textscOE(f(x′),f(x),y)]], | |
assuming a model f, the original learning objective L; when labeled data is not available, y can be ignored. Models trained with this objective can have their maximum softmax probabilities [[52](#bib.bib73 "A baseline for detecting misclassified and out-of-distribution examples in neural networks")] better separate in- and out-of-distribution examples. To create D\textscOEout, data unlike the training data will need to be scraped or curated or downloaded. Samples from D\textscOEout are gathered from already existing and available datasets that might not be directly related to the task-specific objective function; however, they can significantly improve the performance because they contain many diverse variations. Concurrent work [[29](#bib.bib56 "Reducing network agnostophobia")] explores a similar intuition for small-scale image classification, while Outlier Exposure shows how to improve OOD detection for density estimation models, natural language processing models, and small- and large-scale image classifiers.
###
5.7 Using Self-Supervised Learning Can Improve Model Robustness and Uncertainty [[54](#bib.bib81 "Using self-supervised learning can improve model robustness and uncertainty")]:
This work investigated the benefits of training supervised learning tasks in combination with SSL methods in improving robustness of classifiers against simple distributional shifts and OOD detection tasks. To do so, an auxiliary rotation prediction was added to a simple supervised classification. The work measures robustness against simple corruptions such as Gaussian noise, shot noise, blurring, zooming, fogging etc. It has been observed that although auxiliary SSL tasks do not improve the classification accuracy, the model’s robustness and detection abilities are significantly improved. Additionally, when the total loss function is trained in an adversarially robust way, the robust accuracy is improved. Finally, the method is tested in the ND setting using rotation prediction, and horizontal and vertical translation prediction similar but simpler than GT or GOAD. They also test in the setting of multiclass classification setting and find that auxiliary self-supervised learning objectives improves the maximum softmax probability detector [[52](#bib.bib73 "A baseline for detecting misclassified and out-of-distribution examples in neural networks")]. In addition, similar to [[29](#bib.bib56 "Reducing network agnostophobia")] and [[77](#bib.bib80 "Training confidence-calibrated classifiers for detecting out-of-distribution samples")], they attempted to make distribution of the confidence layer for some background or outlier samples uniform. Outliers are selected from other accessible datasets, as in Outlier Exposure [[53](#bib.bib82 "Deep anomaly detection with outlier exposure")].
###
5.8 Unsupervised Out-of-Distribution Detection by Maximum Classifier Discrepancy [[163](#bib.bib8 "Unsupervised out-of-distribution detection by maximum classifier discrepancy")]:
This work relies on a surprising fact—two classifiers trained with different random initializations can act differently on unseen test time samples at their confidence layers. Motivated by this, the work attempts to increase the discrepancy on unseen samples and reduce the discrepancy on seen ones. The discrepancy loss is the difference between the first classifier’s last layer entropy and that of the second one. This forces the classifiers to have the same confidence scores for in-class inputs, yet increases their discrepancy for the others. Fig. [26](#S5.F26 "Fig. 26 ‣ 5.8 Unsupervised Out-of-Distribution Detection by Maximum Classifier Discrepancy [163]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows the overall architecture. First, two classifiers are trained on the in-class samples and are encouraged to produce the same confidence scores. Second, an unlabeled dataset containing both OOD and in-class data is employed to maximize their discrepancy on outliers while preserving their consistency on inliers.
\includegraphics
[width=]Pictures/Classifier\_Discrepancy.pdf
Fig. 26: The overall archietecture of [[163](#bib.bib8 "Unsupervised out-of-distribution detection by maximum classifier discrepancy")]. As it can be seen, at the first step, both of the classifiers are trained. Then, at the next step, an auxiliary discrepancy loss is added to the supervised classification to adjust the boundaries of in-class and OOD samples.
###
5.9 Why ReLU Networks Yield High-Confidence Predictions Far Away From the Training Data [[47](#bib.bib83 "Why relu networks yield high-confidence predictions far away from the training data and how to mitigate the problem")]:
This work proved that ReLU networks produce piece-wise affine functions, therefore, they can be written as f(x)=Vlx+al on the polytope Q(x) as follows:
| | | | |
| --- | --- | --- | --- |
| | Γl,i={z∈Rd∣Δl(x)(Vliz+ali≥0)Q(x)=∩l=1,...,L∩i=1,...,nlΓl,i, | | (67) |
nl and L are the number of hidden units in the lth layer and the total number of layers respectively. The following theorem proves the deficiency of ReLU networks.
*Theorem. 1* Let Rd=∪Rl=1Ql and f(x)=Vlx+al be the piecewise affine representation of the output of a ReLU network on Ql. Suppose that Vl does not contain identical rows for all l=1,...,R Then for almost any x∈Rd and ϵ≥0 there exists an α and a class k∈{1,...,K} such that for z=αx it holds
| | | | |
| --- | --- | --- | --- |
| | efk(z)∑Kr=1efr(z)≥1−ϵ | | (68) |
The equation goes to 1 if α→∞. From this, we can imply that for ReLU networks there exist infinitely many inputs which yield high confidence predictions. Note that arbitrarily high confidence prediction can not be obtained due to the bounded domain of inputs. In order to relieve this problem, a technique that is called confidence enhancing data augmentation is used as the Eq. [69](#S5.E69 "(69) ‣ 5.9 Why ReLU Networks Yield High-Confidence Predictions Far Away From the Training Data [47]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows.
| | | | |
| --- | --- | --- | --- |
| | 1NN∑i=1LCE(yi,f(xi))+λ⋅E[Lpout(f,Z)]Lpout=maxl=1,...,Klog(efl(z)∑Kk=1efk(z)) | | (69) |
where pout and pin are in-class and out-of-distribution distributions respectively, which we are sure that the set of the intersection of their supports has zero or close to zero probability mass. An example of such an out-distribution pout would be the uniform distribution on [0,1]w×h or other noise distributions.
The above training objective needs
many samples to enforce low confidence on the entire out-distribution, an alternative technique called adversarial confidence enhancing training (ACET) is used. ACET uses the idea of adversarial robust training to not only minimize the objective function at each point but also the worst case in a neighborhood of the point as Eq. [70](#S5.E70 "(70) ‣ 5.9 Why ReLU Networks Yield High-Confidence Predictions Far Away From the Training Data [47]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows.
| | | | |
| --- | --- | --- | --- |
| | 1NN∑i=1LCE(yi,f(xi))+λ⋅E[max||u−Z||p≤ϵLpout(f,u)] | | (70) |
At the test time, a thresholded confidence score is used to distinguish between inliers and outliers.
###
5.10 Do Deep Generative Models Know What They Don’t Know? [[97](#bib.bib85 "Do deep generative models know what they don’t know?")]:
This work shows that generative models surprisingly assign higher likelihood scores to outliers. This holds for VAEs, auto-regressive models, and different kinds of flow based methods. In the generative modeling, usually, a parametric model on the data distribution is assumed as pθ(x). Then, it finds the best θ that minimizes the KL distance between the true but unknown distribution p∗ and p, which is equal to maximizing the likelihood of pθ(x) on the input distribution as the Eq. [71](#S5.E71 "(71) ‣ 5.10 Do Deep Generative Models Know What They Don’t Know? [97]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows.
| | | | |
| --- | --- | --- | --- |
| | KL[p∗||pθ(x)]=∫p∗(x)logp∗(x)pθ(x)dx≈−1Nlogpθ(X)−H[p∗] | | (71) |
Also, assuming the existence of a latent space Z the integrals can be written as Eq. [72](#S5.E72 "(72) ‣ 5.10 Do Deep Generative Models Know What They Don’t Know? [97]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") where Z and X are hidden and input variables, respectively. Let f be a diffeomorphism from the data space to a latent space, which commonly happens in flow-based modelings where |∂f∂x| is known as the volume element.
| | | | |
| --- | --- | --- | --- |
| | ∫zpz(z)dz=∫xpz(f(x))∣∣∣∂f∂x∣∣∣dx=∫xpx(x)dx | | (72) |
The parameters of p can be decomposed as θ={ϕ,ψ} with ϕ being the
diffeomorphism’s parameters, i.e. f(x;ϕ), and ψ being the auxiliary distribution’s parameters, i.e. p(z;ψ). Then the objective function can be written as the Eq. [73](#S5.E73 "(73) ‣ 5.10 Do Deep Generative Models Know What They Don’t Know? [97]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges").
| | | | |
| --- | --- | --- | --- |
| | θ∗=arg maxθlogpθ(X)=arg maxϕ,ψN∑i=1logpz(f(xn,ϕ),ψ)+log∣∣∣∂fϕ∂xn∣∣∣ | | (73) |
An interesting aspect of the objective function above is that it encourages the function f to have high sensitivity to small changes of input samples xn. It is mentioned in the paper that if we plot the effect of each term separately, the first term shows the desired behavior for inliers and outliers but the second term causes the problem. While changing f to constant-volume
(CV) transformations [[30](#bib.bib86 "Nice: non-linear independent components estimation")] can alleviate the problem, but not entirely. Finally, using the second-order expansion of the log-likelihood around an
interior point x0, it has been shown that the assigned likelihoods have a direct relation to the model curvature and data’s second moment. Therefore the problem of generative models might be fundamental. Eq. [74](#S5.E74 "(74) ‣ 5.10 Do Deep Generative Models Know What They Don’t Know? [97]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows the details of expansion where Tr means trace operation.
| | | | |
| --- | --- | --- | --- |
| | 0≤Eq[logp(x,θ)]−Ep∗[logp(x,θ)]≈∇x0logp(x0,θ)T(Eq[x]−Ep∗[x])+12Tr{∇2x0logp(x0,θ)(Σq−Σp∗)} | | (74) |
###
5.11 Likelihood Ratios for Out-of-Distribution Detection [[119](#bib.bib84 "Likelihood ratios for out-of-distribution detection")]:
This paper employs likelihood ratio to alleviate the problem of OOD detection in generative models. The key idea is to model the background and fore-ground information separately. Intuitively, background information is assumed to be harmed less than fore-ground information when semantically irrelevant information are added to the input distribution. Therefore two autoregressive models are trained on noisy and original input distribution, and their likelihood ratio is defined as Eq. [75](#S5.E75 "(75) ‣ 5.11 Likelihood Ratios for Out-of-Distribution Detection [119]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges").
| | | | |
| --- | --- | --- | --- |
| | LLR(x)=logpθ(X)pθ0(X)=logpθ(XB)pθ(XF)pθ0(XB)pθ0(XF)logpθ(XF)pθ0(XF)=logpθ(XF)−logpθ0(XF) | | (75) |
At the test time, a thresholding method is used on the likelihood ratio score.
###
5.12 Generalized ODIN [[59](#bib.bib87 "Generalized odin: detecting out-of-distribution image without learning from out-of-distribution data")]:
As an extension to ODIN [[81](#bib.bib78 "Enhancing the reliability of out-of-distribution image detection in neural networks")], this work proposes a specialized network to learn temperature scaling and a strategy to choose perturbation magnitude. G-ODIN defines an explicit binary domain variable d∈{din,dout}, representing whether or not the input x is inlier (i.e, x∼pin). The posterior distribution can be decomposed into p(y∣din,x)=p(y,din∣x)p(din∣x). Note that in this formulation, the reason of assigning overconfident scores to outliers seems to be more obvious since the small values of p(y,din∣x) and p(din∣x) produce the high value of p(y∣din,x). Therefore, they are decomposed and estimated using different heads of a shared feature extractor network as hi(x) and g(x) for p(y∣din,x) and p(din∣x) respectively. This structure is called dividend/divisor and the logit fi(x) for class i can be written as hi(x)g(x). The objective loss function is a simple cross entropy similar to previous approaches. Note that the loss can be minimized by increasing hi(x) or decreasing g(x). For instance, when the data is not in the high-density areas of in-distribution, hi(x) might be small; therefore, g(x) is forced to be small to minimize the objective function. In the other case, g(x) are encouraged to have larger values. Therefore, they approximate the role of aforementioned distributions p(y∣din,x) and p(din∣x). At the test time maxihi(x) or g(x) are used. Fig. [27](#S5.F27 "Fig. 27 ‣ 5.12 Generalized ODIN [59]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows an overview of the method.
\includegraphics
[width=]Pictures/G-ODIN.pdf
Fig. 27: The overall architecture of Generalized ODIN. As it is obvious, different heads are applied on the penultimate layer to model the mentioned distributions.
###
5.13 Background Data Resampling for Outlier-Aware Classification [[80](#bib.bib88 "Background data resampling for outlier-aware classification")]:
As mentioned before, in AD, ND, OSR, and OOD detection, some methods use a background or outlier dataset to boost their performances. However, to avoid different kinds of biases, the size of auxiliary datasets becomes important. In this work, a re-sampling technique is proposed to select an optimal number of training samples from the outlier dataset such that on-boundary samples play a more influential role in the optimization task. The work first provided an interesting probabilistic interpretation on outlier exposure technique. The loss function can be written as Eq. [78](#S5.E78 "(78) ‣ 5.13 Background Data Resampling for Outlier-Aware Classification [80]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges"), where Lcls and Luni are shown in Eq. [76](#S5.E76 "(76) ‣ 5.13 Background Data Resampling for Outlier-Aware Classification [80]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") and Eq. [77](#S5.E77 "(77) ‣ 5.13 Background Data Resampling for Outlier-Aware Classification [80]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") respectively.
| | | | |
| --- | --- | --- | --- |
| | Lcls(f(x;θ),y)=−logfy(x;θ) | | (76) |
| | | | |
| --- | --- | --- | --- |
| | Luni(f(x;θ))=−1KK∑k=1logfk(x;θ)−logK | | (77) |
| | | | |
| --- | --- | --- | --- |
| | L(θ;p,q)=EX,Y∼p(.,.)[Lcls(f(X;θ);Y)]+αEX∼q(⋅)[Luni(f(X;θ))]. | | (78) |
By expanding Eq. [78](#S5.E78 "(78) ‣ 5.13 Background Data Resampling for Outlier-Aware Classification [80]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") and taking its gradient w.r.t. classifier logits, the optimial classifier is obtained as Eq. [79](#S5.E79 "(79) ‣ 5.13 Background Data Resampling for Outlier-Aware Classification [80]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows.
| | | | |
| --- | --- | --- | --- |
| | f∗k(x)=c(x)pY∣X(k∣x)+1−c(x)Kc(x)=pX(x)pX(x)+αqX(x), | | (79) |
where c(x) can be seen as the relative probability of a sample x being in-class distribution p or background distribution q with the ratio α. Suppose D′b is the re-sampled dataset and Db is the given background one,
OOD detection loss after reweighting can be written as:
| | | | |
| --- | --- | --- | --- |
| | Lout(θ;w)=1|D′b|∑(x,y)∈D′bLuni(f(x;θ))=1∑|Db|i=1wiL%uni(f(x;θ)) | | (80) |
The optimal parameters θ∗ are learned as follows:
| | | | |
| --- | --- | --- | --- |
| | θ∗(w)=arg minθL(θ;D,w)=arg minθLin(θ;D)+αLout(θ;w) | | (81) |
Therefore an iterating optimization is solved between finding θt and wt at each step by fixing one and solving the optimization problem for the other. At last, the largest values of weights can be selected with respect to the dataset’s desired size.
###
5.14 Input Complexity and Out-of-Distribution Detection With Likelihood-Based Generative Models [[137](#bib.bib89 "Input complexity and out-of-distribution detection with likelihood-based generative models")]:
This work further investigated the problem of generative models assigning higher likelihood values to OOD samples. In particular, this work finds a strong tie between the OOD samples’ complexity and likelihood values. Simpler input can lead to higher the likelihood value. Fig. [29](#S5.F29 "Fig. 29 ‣ 5.15 Energy-based Out-of-distribution Detection [84]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows this phenomenon. Furthermore, another experiment to support the claim is designed in such a way that we start from a random noise on which a mean average pooling is applied at each step. To preserve the dimension, an upscaling is done after average poolings. Surprisingly, simpler images on which more average poolings are applied achieve higher likelihoods. Motivated by this,the work proposed to detect OOD samples by accounting for input complexity in combination with likelihood values. Since it is hard to compute the input complexity, the paper instead calculates an upper bound using a lossless compression algorithm [[24](#bib.bib90 "Elements of information theory (wiley series in telecommunications and signal processing)")]. Given a set of inputs x coded with the same
bit depth, the normalized size of their compressed versions, L(x) (in bits per dimension), is used as the complexity measurement. Finally, the OOD score is defined as Eq. [82](#S5.E82 "(82) ‣ 5.14 Input Complexity and Out-of-Distribution Detection With Likelihood-Based Generative Models [137]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges"):
| | | | |
| --- | --- | --- | --- |
| | S(x)=−lM(x)−L(x), | | (82) |
where lM(x) is the log-likelihood of an input x given a model M.
Intuitively, considering M0 as a universal compressor, then p(x∣M0)=2−L(x) and Consequently, the OOD score can be defined as follows :
| | | | |
| --- | --- | --- | --- |
| | S(x)=−log2p(x∣M)+log2p(x∣M0)=log2p(x∣M0)p(x∣M) | | (83) |
In the cases where a simple OOD sample is fed into the model, the universal compressor M0 assigns a high probability to it and effectively corrects the high likelihood wrongly given by the learned model M. Similar interpretation holds for the complex OOD samples too.
\includegraphics
[width=]Pictures/likelihoods.pdf
Fig. 28: The assigned likelihood values of a simple generative model to different datasets when is trained on CIFAR10. As it is obvious simpler datasets achieve higher values.
###
5.15 Energy-based Out-of-distribution Detection [[84](#bib.bib91 "Energy-based out-of-distribution detection")]:
This work proposes using the energy score derived from the logit outputs for OOD detection, and demonstrated superiority over softmax score. Energy-based models map each input x to a single deterministic point that is called energy [[74](#bib.bib92 "A tutorial on energy-based learning")]. A set of
energy values E(x,y) could be turned into a density function p(x) through the Gibbs distribution:
| | | | |
| --- | --- | --- | --- |
| | p(y∣x)=e−E(x,y)/T∫y′e−E(x,y′)/T=e−E(x,y)/Te−E(x)/T | | (84) |
E(x) is called *Helmholtz free energy* and is equal to:
| | | | |
| --- | --- | --- | --- |
| | E(x)=−T⋅log∫y′e−E(x,y′)/T | | (85) |
In deep networks, by considering E(x,y)=−fy(x), one can express
the free energy function in terms of the denominator of the softmax activation:
| | | | |
| --- | --- | --- | --- |
| | E(x;f)=−T⋅logK∑iefi(x)/T | | (86) |
The paper also shows that cross-entropy loss facilitates pulling down the energy for in-distribution data during the training process.
Moreover, softmax scores can be analyzed through an energy-based perspective as Eq. [87](#S5.E87 "(87) ‣ 5.15 Energy-based Out-of-distribution Detection [84]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows. During the optimization, E(x;f) is forced to be small for in-distribution samples while being shifted by fmax(x) to satisfy the maximum function. Consequently, this results in a biased scoring function that is not suitable for OOD detection.
| | | | |
| --- | --- | --- | --- |
| | maxyp(y∣x)=maxyefy(x)∑iefi(x)=efmax(x)∑iefi(x)=1∑iefi(x)−f%
max(x)⟹logmaxyp(y∣x)=E(x;f(x)−fmax(x))=E(x;f)+f%
max(x) | | (87) |
Energy score is hyperparameter-free, easy to compute and achieves strong performance compared to the softmax score. Beyond post hoc OOD detection, the paper further demonstrated that energy score can be utilized for model regularization. Different from outlier exposure which forces the uniform softmax distribution for outlier training samples, energy-based regularization directly optimizes the energy gap between ID and OOD:
| | | | |
| --- | --- | --- | --- |
| | minθE(x,y)∼D%
trainin[−logFy(x)]+λ⋅LenergyLenergy=E(xin,y)∼D%trainin[max(0,E(xin)−min)2]+Exout∼D%trainout[max(0,mout−E(xout))2] | | (88) |
The optimization results on stronger performance than OE. At the test time OOD samples are detected based on a threshold on −E(x;f).
\includegraphics
[width=]Pictures/energy-ood.pdf
Fig. 29: The energy-based OOD detection framework. The energy function maps the logit outputs to a scalar through a convenient logsumexp operator. Test samples with lower energy are considered ID and vice versa.
###
5.16 Likelihood Regret: An Out-of-Distribution Detection Score for Variational Autoencoder [[158](#bib.bib93 "Likelihood regret: an out-of-distribution detection score for variational auto-encoder")]:
Previous works showed that VAEs can reconstruct OOD samples perfectly, resulting in the difficulty in detecting OOD samples. The average test likelihoods of VAE across different datasets have a much smaller range than PixelCNN [[103](#bib.bib94 "Conditional image generation with pixelcnn decoders")] or Glow[[69](#bib.bib95 "Glow: generative flow with invertible 1x1 convolutions")], showing that distinguishing between OOD and inlier samples is much harder in VAE. The reason might be because of different ways they model the input distribution. Auto-regressive and flow-based methods model their input at pixel level, while the bottleneck structure in VAE forces the model to ignore some information. To address this issue, a criterion called *likelihood regret* is proposed. It measures the discrepancy between a model trained to maximize the average likelihood of a training dataset, for instance, a simple VAE, and a model maximizing the likelihood of a single input image. The latter is called an ideal model for each sample. Intuitively, the likelihood difference between the trained model and the ideal one might not be high; however, this does not hold for OOD inputs. Suppose the following optimization is performed to train a simple VAE :
| | | | |
| --- | --- | --- | --- |
| | (ϕ∗,θ∗)≈arg maxϕ,θ1nn∑i=1L(xi;θ,τ(xi,ϕ)), | | (89) |
where ϕ and θ are the parameters of encoder and decoder respectively, and τ(xi,ϕ) denotes the sufficient statistics (μx,σx) of qϕ(z∣x). For obtaining the ideal model, we can fix the decoder part and solve an optimization problem on τ(xi,ϕ) such that its individual ELBO is maximized:
| | | | |
| --- | --- | --- | --- |
| | ^τ=arg maxτL(x;θ∗,τ) | | (90) |
Finally, the likelihood regret is defined as follows:
| | | | |
| --- | --- | --- | --- |
| | LR(x)=L(x;θ∗,^τ(x))−L(x;θ∗,ϕ∗) | | (91) |
The optimization of Eq. [90](#S5.E90 "(90) ‣ 5.16 Likelihood Regret: An Out-of-Distribution Detection Score for Variational Autoencoder [158]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") could be started from the initial point that encoder produces for each input, and a threshold is used on LR values at the test time.
###
5.17 Understanding Anomaly Detection With Deep Invertible Networks Through Hierarchies of Distributions and Features [[132](#bib.bib98 "Understanding anomaly detection with deep invertible networks through hierarchies of distributions and features")]:
This work studied the problem of flow-based generative models for OOD detection. It was noted that local feature such as smooth local patches can dominate the likelihood. Consequently, smoother datasets such as SVHN achieves higher likelihoods than a less smooth one like CIFAR-10, irrespective of the training dataset. Another exciting experiment shows the better performance of a fully connected network than a convolutional Glow network in detecting OOD samples using likelihood value. This again supports the existence of a relationship between local statistics like continuity and likelihood values. Fig. [30](#S5.F30 "Fig. 30 ‣ 5.17 Understanding Anomaly Detection With Deep Invertible Networks Through Hierarchies of Distributions and Features [132]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows the similarity of different dataset local statistics that are computed based on the difference of a pixel value to the mean of its 3×3 neighboring pixels.
\includegraphics
[width=0.8]Pictures/local\_statistics.pdf
Fig. 30: (A) shows the distribution of local pixel value differences. As it can be seen, distributions are highly overlapped. (B) Likelihoods extracted from the local pixel value differences correlate with CIFAR10-Glow likelihoods.
Furthermore, let us call the summation of the likelihood of pixel value differences computed in a local 3×3 neighboring pixels using a histogram with 100 equally distanced bins pseudo-likelihoods. A strong Spearman correlation is found between the pseudo-likelihoods and the exact values of likelihoods, supporting the above assumptions. To address this problem, the following three steps are used:
* Train a generative network on a general image distribution like 80 Million Tiny Images
* Train another generative network on images drawn from the in-distribution, e.g., CIFAR-10
* Use their likelihood ratio for OOD detection
Also, to improve the efficacy, the following outlier loss, which uses OOD samples, can be added to the maximum likelihood objective function:
| | | | |
| --- | --- | --- | --- |
| | Lo=−λ.log(σ(log(pg(x))−log(pin(x))T)), | | (92) |
where σ is sigmoid function.
###
5.18 Self-Supervised Learning for Generalizable Out-of-Distribution Detection [[94](#bib.bib99 "Self-supervised learning for generalizable out-of-distribution detection")]:
In this work, a self-supervised learning method is employed to use the information of an unlabeled outlier dataset such that the OOD detection utility of an in-distribution classifier is improved. To do so, at first, the classifier is trained on in-class training samples until the desired performance is achieved. Then, additional outputs (a set of k reject classes) are added to the last layer. Each training batch consists of ID data and some outlier samples. The following loss function is used:
| | | | |
| --- | --- | --- | --- |
| | min(EPin(^x,^y)[−log(Pθ(y=^y∣^x))])+λEPout(x,target)[−log(Pθ(y=target∣x))], | | (93) |
where the target is selected based on the following rules:
| | | | |
| --- | --- | --- | --- |
| | ifarg max(Pθ(x))∈kthentarget←arg max(Pθ(x))elsetarget←random(k) | | (94) |
This is similar to unsupervised deep k-means in [[17](#bib.bib100 "Deep clustering for unsupervised learning of visual features")]. If an unlabeled sample in the outlier dataset resembles in-class samples, it is assigned a random reject class each time; otherwise, it is assigned to a specific reject class. This helps unlabeled inliers to be separated from outliers. At the test time, the sum of the softmax output of the OOD classes is used as the detection score.
###
5.19 SSD: A Unified Framework for Self-Supervised Outlier Detection [[135](#bib.bib101 "Ssd: a unified framework for self-supervised outlier detection")]:
This work has a very similar idea to GDFR (*c.f.* Section [4.8](#S4.SS8 "4.8 Generative-Discriminative Feature Representations For Open-Set Recognition (GDFR) [108]: ‣ 4 Open-set Recognition ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges")). It incorporates SSL methods so to alleviate the need for labeling in-class samples. This is different from several aforementioned methods that require solving a classification task. As a result, SSD can be flexibly used in different settings such as ND, OSR, and OOD detection. The main idea is to employ the contrastive learning as in [[21](#bib.bib102 "A simple framework for contrastive learning of visual representations")], and learns semantically meaningful features. After representation learning, a k-means clustering is applied to estimate class centers with mean and covariance (μm,Σm). Then for each test time sample, the Mahalanobis distance to the closest class centroid is used as the OOD detection score:
| | | | |
| --- | --- | --- | --- |
| | sx=minm(zx−μm)TΣ−1m(zx−μm) | | (95) |
The contrastive learning objective function is very simple. Using image transformations, it first creates two views of each image, commonly referred to as positives. Next, it optimizes to pull each instance close to its positive instances while pushing away from other images, commonly referred to as negatives:
| | | | |
| --- | --- | --- | --- |
| | Lbatch=12N2N∑i=1−logeuTiuj/τ∑2Nk=1I(k≠i)euTiuj/τ, | | (96) |
where ui=h(f(xi))∣|h(f(xi))||2 is the normalized feature vector, (xi,xj) are positive pairs for the ith image from a batch of N images, and h(⋅) is the projection head. Moreover, when a few OOD samples are available, the following scoring function can be used:
| | | | |
| --- | --- | --- | --- |
| | sx=(zx−μin)TΣ−1in(zx−μin)−(zx−μood)TΣ−1ood(zx−μood) | | (97) |
This framework can be extended to the supervised settings when the labels of in-class distribution are available. SSD employs the supervised contrastive learning objective proposed in [[68](#bib.bib103 "Supervised contrastive learning")]:
| | | | |
| --- | --- | --- | --- |
| | Lbatch=12N2N∑i=1−log12Nyi−1∑2Nk=1I(yk=yi)euTiuj/τ∑2Nk=1I(k≠i)euTiuj/τ, | | (98) |
where Nyi is the number of images with label yi in the batch.
###
5.20 MOOD: Multi-level Out-of-distribution Detection [[82](#bib.bib104 "MOOD: multi-level out-of-distribution detection")]:
This work first investigated the computational efficiency aspect of OOD detection. Intuitively, some OOD samples can be detected using only low-level statistics without any need for complex modelings. To this end, multiple intermediate classifiers are trained, operating at
different depths of a trained network, as Fig. [31](#S5.F31 "Fig. 31 ‣ 5.20 MOOD: Multi-level Out-of-distribution Detection [82]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows. In order to find the desired existing depth, an approximation of the input complexity is needed. To address this issue, similar to [[137](#bib.bib89 "Input complexity and out-of-distribution detection with likelihood-based generative models")], the number of bits used to encode the compressed image L(x) is employed. Consequently, exit depth I(x) is determined based on the complexity range a sample belongs to:
| | | | |
| --- | --- | --- | --- |
| | I(x)=min(⌈Lnormalized(x)×K⌉,K), | | (99) |
where K is the number of classes.
\includegraphics
[width=]Pictures/MOOD.pdf
Fig. 31: The overall architecture of MOOD. As the figure shows, some OOD samples can be detected by applying simple low-level computations as opposed to some others that need more complex modelings. Test time computations can be done dynamically and on demand, which significantly reduces computational cost.
The OOD scoring function is defined similarly to [[84](#bib.bib91 "Energy-based out-of-distribution detection")], using an energy function on logits as follows:
| | | | |
| --- | --- | --- | --- |
| | E(x;θi)=−logC∑j=1ef(j)(x;θi)i, | | (100) |
where C is the number of classes of in-distribution data,
and f(j)i(x;θi) is the logit output corresponding to class j∈1,2,...,C for intermediate classifier at exit i. The mentioned criterion has different scales at different depths; therefore, it is normalized as:
| | | | |
| --- | --- | --- | --- |
| | Eadjusted(x;θi)=−E(x;θi)−Ex∈D%
in[−E(x;θi)] | | (101) |
Finally, OOD samples are detected based on simple thresholding on the energy score of the desired depth.
###
5.21 MOS: Towards Scaling Out-of-distribution Detection for Large Semantic Space [[61](#bib.bib105 "MOS: towards scaling out-of-distribution detection for large semantic space")]:
MOS first revealed that the performance of OOD detection can degrade significantly when the number of in-distribution classes increases. For example, analysis reveals that a common baseline’s average false positive rate (at 95% true positive rate) would rise from 17.34% to 76.94% as the number of classes increases from 50 to 1,000 on ImageNet1k. To overcome the challenge, the key idea of MOS is to
decompose the large semantic space into smaller groups
with similar concepts, which allows simplifying the decision boundaries between known vs. unknown data. Specifically, MOS divides the total number of C categories into K groups, G1, G2, …, GK. Grouping is done based on the taxonomy of the label space if it is known, applying k-means using the features extracted from the last layer of a pre-trained network or random grouping. Then the standard groupwise softmax for each group Gk is defined as follows:
| | | | |
| --- | --- | --- | --- |
| | pkc(x)=efkc(x)∑c′∈Gkefkc′(x),c∈Gk | | (102) |
where fkc(x) and pkc(x) denote output logits and the softmax probability for class c in group Gk, respectively. Fig. [32](#S5.F32 "Fig. 32 ‣ 5.21 MOS: Towards Scaling Out-of-distribution Detection for Large Semantic Space [61]: ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows the overall architecture. The training loss function is defined as below:
| | | | |
| --- | --- | --- | --- |
| | LGS=−1NN∑n=1K∑k=1∑c∈Gkykclog(pkc(x)) | | (103) |
where yck and pck represent the label and the softmax probability of category c in Gk, and N is the total number of training samples.
A key component in MOS is to utilize a category others in each group, which measures the probabilistic score for an image to be unknown with respect to the group. The proposed OOD scoring function, Minimum Others Score (MOS), exploits the information carried by the others category. MOS is higher for OOD inputs as they will be mapped to
others with high confidence in all groups, and is lower
for in-distribution inputs. Finally, test time detection is done based on the following score:
| | | | |
| --- | --- | --- | --- |
| | SMOS(x)=−min1≤k≤K(pkothers(x)) | | (104) |
\includegraphics
[width=]Pictures/MOS.pdf
Fig. 32: The overall architecture of MOS. As the figure shows, each sample is labeled as ”others” for groups to which it does not belong except the correct group. At the test time, detection is done based on the minimum of ”others” class scores.
###
5.22 Can Multi-Label Classification Networks Know What
They Don’t Know? [[151](#bib.bib189 "Can multi-label classification networks know what they don’t know?")] :
In this work the ability of OOD detectors in the multi-label classification setting was investigated. In the multi-label classification setting, each input sample might have one or more corresponding labels, which make the problem harder since the joint distribution across labels can be intractable to model. This work proposes the *JointEnergy* criterion as a simple and effective method, which estimates the OOD indicator scores by aggregating label-wise energy scores from multiple labels. Also, they show that JointEnergy can be mathematically interpreted from a joint likelihood perspective. Similar to what has been discussed in [[84](#bib.bib91 "Energy-based out-of-distribution detection")] P(yi=1∣x) can be written as e−E(x,yi)e−E(x), then by defining *label-wise free energy* that is a special case of K-class free energy with K=2:
| | | | |
| --- | --- | --- | --- |
| | Eyi(x)=−log(1+efyi(x)) | | (105) |
JointEnergy can be defined as:
| | | | |
| --- | --- | --- | --- |
| | EJoint(x)=K∑i=1−Eyi(x) | | (106) |
Through derivations, the JointEnergy can be decomposed into three terms:
| | | | |
| --- | --- | --- | --- |
| | Ejoint(x)=logp(x∣y1=1,…,yK=1)+(K−1)⋅logp(x)+Z | | (107) |
where the first term takes into account joint likehood across labels; and the second term reflects the underlying data density, which is supposed to be higher for in-distribution data x. The summation overall results in stronger separability between ID and OOD. The architecture overview is provided in Fig. [33](#S5.F33 "Fig. 33 ‣ 5.22 Can Multi-Label Classification Networks Know What They Don’t Know? [151] : ‣ 5 Out-of-distribution Detection ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges").
\includegraphics
[width=]Pictures/energy-ood-teaser-neurips.pdf
Fig. 33: An overview of JointEnergy for OOD detection in multi-label classification networks. During inference time, input x is passed through classifier, and label-wise scores are computed for each label. OOD indicator scores are either the maximum-valued score (denoted by green outlines) or the sum of all scores. Taking the sum results in a larger difference in scores and more separation between in-distribution and OOD inputs (denoted by red lines), resulting in better OOD detection. Plots in the bottom right depict the probability densities of MaxLogit versus JointEnergy.
###
5.23 On the Importance of Gradients for Detecting Distributional Shifts in the Wild [[60](#bib.bib190 "On the importance of gradients for detecting distributional shifts in the wild")] :
This work proposes a simple post hoc OOD detction method GradNorm, which utilizes the vector norm of gradients with respect to weights, backpropagated
from the KL divergence between the softmax output and a uniform probability distribution. The GradNorm is generally higher for in distribution (ID) data than that for OOD data. Therefore, it can be used for OOD detection. Specifically, the KL divergence is defined as follows:
| | | | |
| --- | --- | --- | --- |
| | DKL(u||Softmax(f(x))=−1CC∑c=1log⎛⎝efc(x)/T∑Cj=1efj(x)/T⎞⎠, | | (108) |
where u is the uniform distribution, and C is the number of in-distribution classes. Then, the OOD score is defined as S(x)=||∂DKL∂W||p where W can be (1) a block of parameters, (2) all parameters, and (3) last layer parameters. It was mentioned that the third approach is better than other ones and achieves significant results.
6 Dataset
----------
###
6.1 Semantic-Level Datasets
Below we summarize datasets that can be used to detect semantic anomalies. Semantic anomalies are those kinds of anomalies that the variation of pixels leads to the change of semantic content. Datasets such as MNIST, Fashion-MNIST, SVHN, and COIL-100 are considered as toy datasets. CIFAR-10, CIFAR-100, LSUN, and TinyImageNet are hard datasets with more variations on color, illumination, and background. Finally, Flowers and Birds are fine-grained semantic datasets, which makes the problem even harder.
MNIST [[75](#bib.bib74 "MNIST handwritten digit database")]: This dataset includes 28 × 28 grayscale handwritten digits from 0-9 and consists of 60k training images and 10k testing ones.
Fashion MNIST [[157](#bib.bib117 "Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms")]: This dataset comprises of 28×28 grayscale images of 70k clothing items from 10 categories. The training set has 60k images and the test set has 10k images.
CIFAR-10 [[72](#bib.bib17 "Learning multiple layers of features from tiny images")]: The CIFAR-10 has 60k natural images. It consists 32x32 RGB images of 10 classes,There are 50k training images and 10k test images.
CIFAR-100 [[72](#bib.bib17 "Learning multiple layers of features from tiny images")]: This dataset is very similar to CIFAR-10, but it has 100 classes containing 600 images each.The 100 classes are grouped into 20 super classes. each class has 500 training images and 100 testing images.
TinyImageNet [[27](#bib.bib44 "Imagenet: a large-scale hierarchical image database")]: The Tiny ImageNet dataset consists of a subset of ImageNet images . It contains 10,000 test images from 200 different classes. Also, two more datasets, TinyImageNet (crop) and TinyImageNet (resize) can be constructed, by either randomly cropping image patches of size 32 × 32 or downsampling each image to size 32 × 32.
LSUN [[162](#bib.bib121 "Lsun: construction of a large-scale image dataset using deep learning with humans in the loop")]: The Large-scale Scene UNderstanding dataset (LSUN) has a testing set of 10,000 images of 10 different scene categories such as bedroom, kitchen room, living room, etc. Similar to TinyImageNet, two more datasets, LSUN (crop) and LSUN (resize), can be reconstructed by randomly cropping and downsampling the LSUN testing set, respectively.
COIL-100 [[99](#bib.bib118 "Object image library (coil-100")]: COIL-100 is a dataset of colorful images of 100 objects. It comprises 7200 128×128 images. Images are captured from objects placed on a motorized turntable against a black background, and there are 72 images of each object in different poses.
SVHN [[100](#bib.bib75 "Reading digits in natural images with unsupervised feature learning")]: SVHN can be seen as similar in flavor to MNIST (e.g., the images are of small cropped digits), but incorporates an order of magnitude more labeled data (over 600,000 digit images) and comes from a significantly harder, unsolved, real world problem (recognizing digits and numbers in natural scene images). SVHN is obtained from house numbers in Google Street View images.
Flowers [[102](#bib.bib119 "Automated flower classification over a large number of classes")]: Flowers is a 102 category dataset, consisting of 102 flower categories. The flowers chosen to be flower commonly occuring in the United Kingdom. Each class consists of between 40 and 258 images. The images have large scale, pose and light variations. In addition, there are categories that have large variations within the category and several very similar categories.
Birds [[155](#bib.bib120 "Caltech-ucsd birds 200")]: CUB-200-2011 is a bird classification task with 11,788
images across 200 wild bird species. There is roughly equal amount of train and test data. It is generally considered one of the most challenging datasets since each species has only 30 images for training.
###
6.2 Pixel-Level Datasets
In these datasets unseen samples, outliers or anomalies do not have semantic difference with inliers. This means an area of the original image is defected; however, the original meaning is still reachable, yet has been harmed.
MVTec AD [[11](#bib.bib186 "MVTec ad–a comprehensive real-world dataset for unsupervised anomaly detection")]:This dataset is an industrial dataset ,it provides 5354 high-resolution images divided into ten object and five texture categories. it contains 3629 training images. test set contains 467 normal images and 1258 abnormal images having various kinds of defects.
PCB [[62](#bib.bib122 "A pcb dataset for defects detection and classification")]: PCB dataset containing 1386 images
with 6 kinds of defects for the use of detection, classification
and registration tasks. Images are captured in high-resolution.
LaceAD [[159](#bib.bib123 "Learning semantic context from normal samples for unsupervised anomaly detection")]: LaceAD contains 9,176 images from the top 10 lace fabric manufacturing companies worldwide, where the images are captured in the real production environment by a high-resolution DSLR camera, They are categorized into 17 subsets based on their patterns. Each image has the size of 512 × 512 and has been labeled by professional
workers.
Retinal-OCT [[67](#bib.bib127 "Identifying medical diagnoses and treatable diseases by image-based deep learning")]: This consists of 84,495 X-Ray images in 4 categories CNV, DME, DRUSEN, and NORMAL each of which has subtle differences with respect to others.
CAMELYON16 [[7](#bib.bib128 "Diagnostic assessment of deep learning algorithms for detection of lymph node metastases in women with breast cancer")]:
Detecting metastases of lymph nodes is an extremely important variable in the diagnosis of breast cancers.
Tissue with metastasis may differ from healthy one only by texture, spatial structure, or distribution of nuclei, and
can be easily confused with normal tissue. The training dataset of Camelyon16 consists of 110 whole-slide images (WSIs) contained tumors, and 160 are not, and testing dataset with 80 regular slides and 50 slides containing metastases.
Chest X-Rays [[153](#bib.bib129 "Chestx-ray8: hospital-scale chest x-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases"), [64](#bib.bib130 "Chexpert: a large chest radiograph dataset with uncertainty labels and expert comparison"), [115](#bib.bib131 "Chexnet: radiologist-level pneumonia detection on chest x-rays with deep learning")]: Chest X-Ray datasets are medical imaging datasets which comprises a large number of frontal-view X-ray images of many unique patients (collected from the year of 1992 to 2015). The datasets include eight to fourteen common disease labels, mined from the text radiological reports via NLP techniques. Images are not registered and captured in different pose and contrast which makes the detection task challenging.
Species [[48](#bib.bib6 "Scaling out-of-distribution detection for real-world settings")]: This dataset consists of organisms that fall outside ImageNet-21K. Consequently, ImageNet-21K models can treat these images as anomalous.
ImageNet-O [[56](#bib.bib126 "Natural adversarial examples")]: ImageNet-O is a dataset of adversarially filtered examples for ImageNet out-of-distribution detectors. To create this dataset, at first, ImageNet-22K is downloaded, and shared examples from ImageNet-1K are deleted. With the remaining ImageNet-22K examples that do not belong to ImageNet-1K classes, examples that are classified by a ResNet-50 as an ImageNet-1K class with a high confidence are kept. Finally, visually clear images are selected. This creates a dataset of OOD examples that are hard for a ResNet-50. These examples are challenging for others models to detect, including Vision Transformers.
\includegraphics
[trim=2cm 17cm 2cm 2cm, clip=true,width=]Pictures/Image.pdf
Fig. 34: Sample visualization of MNIST, Fashion-MNIST, SVHN, COIL-100, Birds, Chest X-Rays, Cifar-10, TinyImageNet, LSUN, Flowers, MVTecAD, PCB, Retinal-OCT, CAMELYON16, LaceAD, MNIST-C, ImageNet-C, and ImageNet-P.
###
6.3 Synthetic Datasets
These datasets are usually made using semantic-level datasets; however, the amount of pixel-variations is under control such that unseen, novel, or abnormal samples are designed to test different aspects of trained models while preserving semantic information. For instance, MNIST-c includes MNIST samples added different kinds of noises such as shot noise and impulse noise, which are random corruptions that may occur during the imaging process. These datasets could be used to not only test the robustness of our models but also for training models in the AD setting instead of novelty detection or open-set recognition. Due to the lack of comprehensive research in the field of anomaly detection, these datasets can be very beneficial.
MNIST-C [[96](#bib.bib124 "Mnist-c: a robustness benchmark for computer vision")]: MNIST-C dataset is a comprehensive suite of 15 corruptions applied to the MNIST test set for benchmarking out-of-distribution robustness in computer vision. Through several experiments and visualizations, it is shown that the corruptions significantly degrade the performance of state-of-the-art computer vision models while preserving the semantic content of the test images.
ImageNet-C, ImageNet-P [[51](#bib.bib125 "Benchmarking neural network robustness to common corruptions and perturbations")]: This can be seen as the ImageNet version of MNIST-C. For the ImageNet-C, a set of 75 common visual corruptions are applied on each image, and for the ImageNet-P, a set of perturbed or subtly differing ImageNet images are introduced. It is shown that although these perturbations are not chosen by an adversary, currently existing networks exhibit surprising instability on common perturbations.
An overall visualization of the mentioned datasets can be found in Fig. [34](#S6.F34 "Fig. 34 ‣ 6.2 Pixel-Level Datasets ‣ 6 Dataset ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges").
7 Evaluation Protocols
-----------------------
###
7.1 Auc-Roc
Receiver Operating Characteristics (ROC) is a well-known criterion. Given a test dataset including positive and negative (or seen and unseen ) samples, it characterizes the relation between the false positive rate (FPR) and the true positive rate (TPR) at different detection thresholds. AUC-ROC is the area under the ROC curve, which is a threshold-independent metric. The highest value of ROC is 1, and 0.5 indicates a model assigns positive label with random guessing.
In the literature, AD and ND are usually tested in one-vs-all setting that considers one class as normal and the rest of the classes as anomaly, unseen or unknown. For OOD detection, the in-distribution data is considered as positive and OOD data is considered as negative. For instance, one can train a model on CIFAR-10 and considers MNIST as outlier at the test time. This means training and testing datasets have a large contrast with each other. Sometimes instead of MNIST, uniform noise or Gaussian noise can be used.
###
7.2 Fpr@tpr
Although AUC-ROC is a common metric, in practice, models have to select a specific threshold to perform detection. To address this issue, an operating point on the ROC which is desired with respect to applications is selected. A commonly-used metric is FPR@TPRx, which measures the FPR when the TPR is x=0.95.
###
7.3 Aupr
AUPR is the Area under the Precision-Recall curve, which is another threshold independent metric. The PR curve depicts the precision=TPTP+FP and recall=TPTP+FN under different thresholds. In some literature, the metrics AUPR-In and
AUPR-Out denote the area under the precision-recall curve where in-distribution and out-of-distribution images are specified as positives, respectively. This metric is usually used in OOD detection and OSR settings.
###
7.4 Accuracy
This metric is usually used in OSR, which is common choice for evaluating classifiers under the closed-set assumption as follows:
| | | | |
| --- | --- | --- | --- |
| | A=∑Ci=1(TPi+TNi)∑Ci=1(TPi+TNi+FPi+FNi) | | (109) |
This can be easily extended to open-world assumption in which UUCs must be classified correctly:
| | | | |
| --- | --- | --- | --- |
| | AO=∑Ci=1(TPi+TNi)+TU∑Ci=1(TPi+TNi+FPi+FNi)+(TU+FU) | | (110) |
Although accuracy is a common metric; however, it is very sensitive to imbalanced number of samples, which is not the case in metrics such as AUC-ROC. To cope with this issue, a normalized accuracy (NA), which weights the accuracy
for KKCs (AKS) and the accuracy for UUCs (AUS) is defined as follows:
| | | | |
| --- | --- | --- | --- |
| | NA=λrAKS+(1−λr)AUS | | (111) |
where
| | | | |
| --- | --- | --- | --- |
| | AKS=∑Ci=1(TPi+TNi)∑Ci=1(TPi+TNi+FPi+FNi)AUS=TUTU+FU | | (112) |
and 0≤λr≤1 is a regularization constant.
###
7.5 F-measure
The F-measure or F-score is the harmonic mean of precision P and recall R:
| | | | |
| --- | --- | --- | --- |
| | F=2×P×RP+R | | (113) |
Note that F must be computed in the
same way as the multi-class closed set scenario. This is because the correct classifications of UUCs would be considered as true positive classifications; however, it makes no sense since there is no UUC sample in the training process. Therefore, the computations of Precision and Recall only for KKCs are modified to give a relatively reasonable F-measure for OSR. The new measures are called macro-F-measure and micro-F-measure respectively as follows:
| | | | |
| --- | --- | --- | --- |
| | Pma=1CC∑i=1TPiTPi+FPi,Rma=1CC∑i=1TPiTPi+FNiPmi=∑Ci=1TPi∑Ci=1(TPi+FPi),Rmi=∑Ci=1TPi∑Ci=1(TPi+FNi) | | (114) |
Note that although the precision and recall only consider
the KKCs; however, by computing FNi and FPi false UUCs and false KKCs are taken into account [[66](#bib.bib133 "Nearest neighbors distance ratio open-set classifier")].
8 Future Challenges
--------------------
Here we provide plausible future directions, which might be of interest to both practitioners and academic researchers.
###
8.1 Evaluating Baselines and the Evaluation Protocols of OOD Detection
The evaluation protocols for OOD detection has room for improvement. For instance, [[2](#bib.bib25 "Detecting semantic anomalies")] trains a mixture of three Gaussians on the CIFAR-10 dataset (as ID), and evaluated against OOD datasets including TinyImagenet (crop), TinyImagenet (resize), LSUN, LSUN(resize) and iSUN. The model is trained channel-wise at a pixel-level. Tab. [I](#S8.T1 "TABLE I ‣ 8.1 Evaluating Baselines and the Evaluation Protocols of OOD Detection ‣ 8 Future Challenges ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows the detection results on different datasets. The results are comparable with SOTA despite the simplicity. In particular, LSUN performs poorly since majority of them have uniform color and texture, with little variation and structure. Similar to what has been observed in likelihood-based methods, LSUN “sits inside” CIFAR-10 with a similar mean but lower variance, and ends up being more likely under the wider distribution.
We also provide a better insight into the performance of OOD detection baselines, evaluated on both near and far out-of-distribution datasets. For model trained on CIFAR-10, we use CIFAR-100 as the near OOD datasets. The results are presented in Tables [II](#S8.T2 "TABLE II ‣ 8.6 Fairness and Biases of Models ‣ 8 Future Challenges ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges"), [III](#S8.T3 "TABLE III ‣ 8.6 Fairness and Biases of Models ‣ 8 Future Challenges ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges"), and [V](#S8.T5 "TABLE V ‣ 8.9 Multi-Label OOD Detection and Large Scale Datasets ‣ 8 Future Challenges ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges"). As it is shown, none of the methods are good at detecting both near and far OOD samples except OE approaches that use an extra auxiliary dataset to do the task. Also, using Mahalanobis distance can improve the performance of most of the methods at detecting far OOD samples while degrading the performance of near OOD detection. Besides, as Mahalanobis distance can have poor performance at detecting even some of far OOD samples due to inaccurate Gaussisn density estimation. Moreover, its performance varies significantly when OOD dataset is resized or cropped, showing its dependency on low-level statistics. For instance, notice to the SVHN column of Table [V](#S8.T5 "TABLE V ‣ 8.9 Multi-Label OOD Detection and Large Scale Datasets ‣ 8 Future Challenges ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges"). This is in correspondence with what has been shown recently by [[118](#bib.bib178 "A simple fix to mahalanobis distance for improving near-ood detection")] on the deficiencies of Mahalanobis distance as well.
One solution to resolve the issue might be applying input pre-processing techniques such as ODIN to alleviate the effect of first and second-order statistics in assigning OOD scores; however, it increases the execution speed by the sum of an extra forward and backward pass during testing. Additionally, techniques such as ensembling or MC-Dropout [[38](#bib.bib179 "Dropout as a bayesian approximation: representing model uncertainty in deep learning")] might be slightly better than others on some OOD datasets; yet, they need multiple forward passes increasing the execution time significantly. For example, the reported MC-Dropout is 40 times slower than a simple MSP. In summary, we encourage future works to evaluate OOD detection on both near and far OOD datasets.
| OOD Dataset | Average Precision |
| --- | --- |
| TinyImageNet(crop) | 96.8 |
| TinyImageNet(resize) | 99.0 |
| LSUN | 58.0 |
| LSUN(resize) | 99.7 |
| iSUN | 99.2 |
TABLE I: The performance of a simple method using only low-level features on different datasets [[2](#bib.bib25 "Detecting semantic anomalies")].
###
8.2 AD Needs to Be Explored More
As mentioned earlier, AD and ND are not completely the same both historically and fundamentally. A really important and practical category of problems in the real-world application are those that can not be cleaned easily, and consequently, contain different kinds of noises such as, label noise or data noise. This is the case in complex and hazardous systems such as modern nuclear power plants, military aircraft carriers, air traffic control, and other high-risk systems [[49](#bib.bib187 "Unsolved problems in ml safety")]. Recently proposed methods in ND need to be evaluated in AD settings using the proposed synthetic datasets, and new solutions need to be proposed. As the openness score is usually high for AD detectors, having a high recall while providing a low false alarm rate is necessary for their practicality [[25](#bib.bib188 "Monitor alarm fatigue: an integrative review")].
Furthermore, almost all the AD or ND methods are evaluated in the one-vs-all setting. This results in having a normal class with a few distribution modes, which is not a good approximation of real-world scenarios. Therefore, evaluating AD or ND methods in multi-class settings similar to OSR domain while having no access to labels could give a more clear perspective on the practicality of SOTA methods.
###
8.3 OSR Methods for Pixel-Level Datasets
Almost all the methods existing in OSR are evaluated on semantic datasets. As class boundaries in such datasets usually are far from each other, discriminative or generative methods can model their differences effectively. However, in many applications such as Chest X-ray datasets, variations are subtle. Existing methods can results in poor performance on such tasks. For instance, a model may be trained on 14 known chest diseases. A new disease, for example, COVID-19, may emerge as unknowns. In this case, our model must detect it as a new illness instead of classifying it into pre-existing disease categories. Also, in many clinical applications where medical datasets are gathered, usually, disease images are more accessible than healthy ones; thus, OSR problems must be learned on sickness as our normal images and detect healthy ones as abnormal inputs.
Table [IV](#S8.T4 "TABLE IV ‣ 8.9 Multi-Label OOD Detection and Large Scale Datasets ‣ 8 Future Challenges ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows the performance of a simple MSP baseline on MVTecAD dataset when some frequent faults are considered as normal classes. In such scenarios, the goal is to detect and classify well-known faults while distinguishing rare ones as outliers that need to be treated differently. Although this is a common and practical industrial setting, the baseline does not perform better than random, casting doubt on their generality for safety-critical applications.
Recently, [[22](#bib.bib169 "Learning deep classifiers consistent with fine-grained novelty detection")] has shown the effectiveness of using a prior Gaussian distribution on the penultimate layer of classifier networks, similar to what has been done in several before-mentioned works, in tasks in which the distribution of seen classes are very similar to each other, for instance, Flowers or Birds datasets that introduced in the previous sections. However, more research should be done in this setting since it is more practical and quite harder than the traditional ones.
###
8.4 Small Sample Size
Learning with a small sample size is always challenging but desirable. One way of approaching the problem could be to exploit meta-learning algorithms [[35](#bib.bib134 "Model-agnostic meta-learning for fast adaptation of deep networks"), [101](#bib.bib135 "On first-order meta-learning algorithms")], and to learn generalizable features that can be easily adapted to AD, ND, OSR, or OOD detection using a few training samples [[88](#bib.bib136 "Few-shot scene-adaptive anomaly detection")]. One challenge in meta-learning is handling distribution shift between training and adaptation phase, which might result in producing one-class meta-learning algorithms such as [[37](#bib.bib137 "Few-shot one-class classification via meta-learning")]. Also, other approaches explored generating a synthetic OOD dataset to improve few-shot classification of in-class samples [[65](#bib.bib138 "OOD-maml: meta-learning for few-shot out-of-distribution detection and classification")]. Although, the combination of meta-learning and AD, ND, OOD detection, and OSR has gained significant attention recently, some important aspects—such as generalizing to detect UUCs using only a few KUC and the convergence of meta-learning algorithms in one-class setting—remain underexplored.
###
8.5 Adversarial Robustness
Carefully designed imperceptible perturbations fooling deep learning-based models to make incorrect predictions are called adversarial attacks [[164](#bib.bib139 "Adversarial examples: attacks and defenses for deep learning")]. Up to this time, it has been shown that classifiers are susceptible to adversarial attacks such that their performance degrades significantly at the test time. As in OOD detection, OSR, AD and ND, being robust against adversarial attacks is crucial. Recent works in OSR [[139](#bib.bib140 "Open-set adversarial defense"), [121](#bib.bib141 "Adversarial robustness: softmax versus openmax")], ND [[128](#bib.bib11 "Arae: adversarially robust training of autoencoders improves novelty detection"), [129](#bib.bib10 "Puzzle-ae: novelty detection in images through solving puzzles")], and OOD detection [[91](#bib.bib142 "Provably robust detection of out-of-distribution data (almost) for free"), [20](#bib.bib191 "ATOM: robustifying out-of-distribution detection using outlier mining")] have investigated the effects of adversarial attacks on models. However more is needed. For instance as anomalies in AD or UUCs in OSR are not accessible at the training time, therefore, achieving a robust model on attacked anomalies or UUCs would not be trivial.
The relation of different defense approaches against adversarial attacks with novelty detection can also reveal some important insights about the internal mechanisms of our models. For instance, membership attack [[141](#bib.bib176 "Membership inference attacks against machine learning models")] attempts to infer whether an input sample has been used during the training process or not, which can be seen as designing novelty detectors without having any generalization to UKC samples. Also, [[33](#bib.bib177 "Robust anomaly detection and backdoor attack detection via differential privacy")] investigates the relation of detecting poisoning attacks and novelty detectors. Poisoning examples that are intentionally added by attackers to achieve backdoor attacks could be treated as one type of “outliers” in the training dataset. It is claimed that differential privacy not only improves outlier detection and novelty detection but also backdoor attack detection of ND models.
From a completely different point of view, as it is mentioned in [[63](#bib.bib143 "Adversarial examples are not bugs, they are features")], adversarial robust training can be employed to boost learned feature space in a semantic way. This path has been followed in ARAE [[128](#bib.bib11 "Arae: adversarially robust training of autoencoders improves novelty detection")] and Puzzle-AE [[129](#bib.bib10 "Puzzle-ae: novelty detection in images through solving puzzles")] to improve the performance of AEs in detecting unseen test time samples. Similar intention is followed in the one-class learning method [[43](#bib.bib144 "DROCC: deep robust one-class classification")] that shows robustness is beneficial for detecting novel samples. This path also needs to be explored more, for instance in spite of standard adversarial attacks in classification tasks [[89](#bib.bib145 "Towards deep learning models resistant to adversarial attacks")], attacks do not need to be imperceptible anymore in AD or ND and sometimes perceptible ones improve detection performance more.
###
8.6 Fairness and Biases of Models
Research on fairness has witnessed a significant growth in recent years [[168](#bib.bib146 "Learning fair representations"), [15](#bib.bib147 "How the machine ‘thinks’: understanding opacity in machine learning algorithms"), [34](#bib.bib148 "Fairness through awareness")]. It has been shown that models get biased towards some sensitive variables during their training process. For instance, [[154](#bib.bib149 "Towards fairness in visual recognition: effective strategies for bias mitigation")] show that for attribute classification task in the CelebA [[86](#bib.bib150 "Deep learning face attributes in the wild")] dataset, attribute presence is correlated with the gender of people in the image, which is obviously not desirable. Attributes such as gender in the mentioned example are called protected variables. In OOD detection literature, a recent work [[93](#bib.bib192 "On the impact of spurious correlation for out-of-distribution detection")] systematically investigates how spurious correlation in the training set impacts OOD detection. The results suggest that the OOD detection performance is severely worsened when the correlation between spurious features and labels is increased in the training set. For example, a model that exploits the spurious correlation between the water background and label waterbird for prediction. Consequently, a model that relies on spurious features can produce a high-confidence
prediction for an OOD input with the same background (i.e., water) but a different semantic label (e.g., boat).
Fairness and AD or ND seems to have fundamental contrast with each other. In fairness , we tend to make unbiased models in which equality constraints between minority samples and majority ones hold while the goal of AD models is to assign higher anomaly scores to rarely happen events. To address this issue [[140](#bib.bib151 "FAIROD: fairness-aware outlier detection"), [171](#bib.bib152 "Towards fair deep anomaly detection")] proposed fairness-aware ADs while using the label of protected variables as an extra supervision in the training process.
From a different point of view [[160](#bib.bib153 "Understanding the effect of bias in deep anomaly detection")] introduces a significantly important bias in semi-supervised anomaly detection methods such as DSAD [[123](#bib.bib111 "Deep semi-supervised anomaly detection")]. Suppose DSAD has been implemented in law enforcement agencies to spot suspicious individuals using surveillance cameras. As a few number of training samples has been used as abnormal samples during the process, the trained model might have been biased towards detecting special type of anomalies more than others. For instance, if the auxiliary abnormal training dataset includes more men than women, boundaries of detecting abnormal events as men might be looser than women at the test time. This could also happens in the classification settings such as OOD detection or OSR. [[138](#bib.bib154 "CheXclusion: fairness gaps in deep chest x-ray classifiers")] reports the existence of unfair biases toward some unrelated protected variables in detecting chest diseases for a classifier trained on Chest X-Ray datasets. From what has been said, it seems fairness and AD, ND, OSR, and OOD detection are strongly correlated because of some critical applications in which they are used and researching more on their correlation is necessary for having practical models.
\resizebox
!
Method
Criterion
Gaussian
Rademacher
Blob
TinyImageNet(crop)
TinyImageNet(resize)
LSUN(crop)
LSUN(resize)
iSUN
SVHN
CIFAR-100
MSP
FPR95
14.53
94.78
70.50
17.06
40.10
12.65
29.23
36.22
28.37
43.27
AUROC
94.78
79.85
94.63
94.64
88.30
96.45
91.40
90.00
91.94
87.77
AUPR
70.50
32.21
74.23
75.09
58.15
83.16
65.36
62.46
67.10
55.68
MLV
FPR95
52.60
73.27
11.67
9.59
47.67
4.93
27.28
36.42
43.54
56.52
AUROC
75.48
70.08
96.85
97.84
89.16
98.93
94.05
93.38
91.11
87.13
AUPR
27.07
25.47
83.56
90.31
65.65
95.35
78.18
73.99
72.08
61.47
MSP-OE
FPR95
0.71
0.50
0.58
6.61
13.00
1.32
5.16
5.64
4.77
28.36
AUROC
99.60
99.78
99.84
98.77
97.27
99.70
98.95
98.87
98.42
93.29
AUPR
94.25
97.36
98.94
95.06
88.08
98.56
94.56
94.20
89.33
76.19
MLV-OE
FPR95
0.69
0.43
0.49
4.98
11.17
1.11
4.10
4.52
4.08
30.38
AUROC
99.62
99.79
99.86
98.96
97.58
99.74
99.11
99.02
98.61
93.10
AUPR
94.30
97.46
99.07
95.72
89.10
98.71
95.15
94.68
90.11
76.36
Ensemble
FPR95
6.84
16.71
16.71
15.99
100
12.34
25.04
100.00
16.71
100.00
AUROC
97.37
86.94
91.20
93.18
85.69
95.23
90.21
88.00
92.05
83.90
AUPR
82.32
41.71
64.52
71.49
56.32
78.07
64.99
61.03
67.29
53.00
Mahalanobis
FPR95
1.35
2.01
7.38
35.82
48.38
28.61
27.98
39.02
24.79
48.40
AUROC
99.57
99.60
98.21
87.78
87.75
87.10
92.25
90.40
90.86
86.71
AUPR
96.49
97.95
90.63
46.79
55.33
41.59
65.14
62.17
53.36
54.06
MC-Dropout
FPR95
15.31
33.58
16.54
20.75
38.77
16.81
28.44
34.62
28.73
37.48
AUROC
93.89
83.41
94.73
93.55
88.52
95.09
91.36
89.73
91.07
88.43
AUPR
63.52
35.40
74.91
71.65
58.19
77.26
65.34
61.76
62.41
56.84
ODIN
FPR95
0.00
0.00
99.4
04.30
07.50
04.80
03.80
06.10
51.00
51.40
AUROC
100.00
99.90
42.50
99.10
98.50
99.00
99.20
98.80
89.90
88.3
AUPR
63.52
35.40
74.91
71.65
58.19
77.26
65.34
61.76
62.41
56.84
TABLE II: OOD example detection for the maximum softmax probability (MSP) baseline detector, maximum logit value, the MSP detector after fine-tuning with Outlier Exposure (OE), the maximum logit value after fine-tuning with Outlier Exposure (OE), ensemble of 3 models, mahlanobis distance, and monte carlo dropout. Inlier distributions is considered as cifar10. All results are percentages
and the result of 10 runs.
\resizebox
!
Method
Criterion
Gaussian
Rademacher
Blob
TinyImageNet(crop)
TinyImageNet(resize)
LSUN(crop)
LSUN(resize)
iSUN
SVHN
CIFAR-10
MSP
FPR95
54.32
39.08
57.11
43.34
65.88
47.32
62.98
63.34
69.12
65.14
AUROC
64.66
79.27
75.61
86.34
74.56
85.56
75.59
75.73
71.43
75.12
AUPR
19.69
30.05
29.99
56.98
33.71
56.49
34.11
33.88
30.44
33.92
ODIN
FPR95
01.20
13.90
13.70
09.20
37.60
07.20
32.30
36.40
37.00
76.4
AUROC
99.50
92.60
95.90
97.90
90.80
98.30
91.90
90.50
89.00
73.20
AUPR
98.70
83.70
94.50
97.70
89.90
98.20
90.90
87.80
86.30
70.60
MLV
FPR95
71.89
72.35
81.09
22.51
66.17
22.20
61.30
60.86
67.01
64.41
AUROC
44.24
46.22
53.62
94.72
77.72
95.09
79.54
79.19
74.03
77.55
AUPR
13.82
14.20
15.98
79.10
38.00
81.11
39.14
37.27
31.99
37.30
MSP-OE
FPR95
12.41
16.89
12.04
22.02
69.42
13.27
60.89
62.42
43.10
62.57
AUROC
95.69
93.01
97.11
95.69
76.04
97.55
80.94
79.96
86.86
75.41
AUPR
71.13
56.81
85.91
85.34
39.57
90.99
48.52
45.86
53.27
32.28
MLV-OE
FPR95
10.71
16.66
8.09
17.34
73.95
08.50
56.02
60.73
32.59
64.91
AUROC
96.12
91.86
97.94
96.38
75.84
98.31
83.33
81.89
88.91
73.74
AUPR
72.81
52.03
88.60
86.55
39.72
92.90
51.06
48.21
54.72
30.48
Ensemble
FPR95
22.72
43.51
48.07
44.68
100.00
47.26
100.00
91.44
57.18
57.18
AUROC
89.15
68.64
79.24
82.90
70.47
82.39
70.66
71.08
73.61
75.13
AUPR
46.45
21.61
35.80
44.31
28.98
44.12
28.75
28.66
31.76
32.62
Mahalanobis
FPR95
0.82
0.10
2.70
73.79
43.40
76.42
37.94
42.07
32.05
70.80
AUROC
99.78
99.98
99.48
57.77
87.62
54.35
90.12
88.91
92.76
70.99
AUPR
98.63
99.88
97.64
17.27
60.18
16.11
65.97
62.82
75.87
27.12
MC-Dropout
FPR95
54.45
41.41
46.64
47.32
68.05
55.38
63.53
65.03
75.98
63.33
AUROC
62.35
76.51
80.59
85.23
73.66
82.24
74.93
74.70
68.89
76.87
AUPR
18.74
27.14
35.08
54.12
32.57
49.08
33.04
32.26
28.63
36.31
TABLE III: OOD example detection for the maximum softmax probability (MSP) baseline detector, maximum logit value, the MSP detector after fine-tuning with Outlier Exposure (OE), the maximum logit value after fine-tuning with Outlier Exposure (OE), ensemble of 3 models, mahlanobis distance, and monte carlo dropout. Inlier distributions is considered as cifar100. All results are percentages
and the result of 10 runs.
###
8.7 Multi-Modal Datasets
In many situation, training dataset consists of multi-modal training samples, for instance in Chest-X-Ray datasets, labels of images are found automatically by applying NLP methods on the prescribes of radiologists. In these situations joint training of different modes could help models to learn better semantic features. However, in this way, models need to be robust in different modes too. For example, in visual Question Answering tasks, we expect our model no to produce any answer for out-of-distribution input texts or images. Note that the correlation between different modes must be attended here, and independent training of AD, ND, OOD detection, or OSR models for different modes results in sticking in local minimas. To cope with this issue, [[76](#bib.bib155 "Regularizing attention networks for anomaly detection in visual question answering")] has explored the performance of VQA models in detecting unseen test time samples. However, there are much more challenges need to be investigated in this way. For instance, the robustness and fairness of VQA OOD models is significantly more challenging compared to single mode datasets, besides due to heavy training process of these models, inventing few-shot methods might be demanding in the fields.
###
8.8 Explainablity Challenge
Explainable AI (XAI) has found a seriously important role in the recently proposed deep network architectures especially when they are used in safety-critical applications [[5](#bib.bib156 "Explainable artificial intelligence (xai): concepts, taxonomies, opportunities and challenges toward responsible ai")]. In AD, OSR, ND and OOD detection due to some of their critical applications, we should be able to explain the reason of decisions our models make [[50](#bib.bib5 "Unsolved problems in ml safety")]. For instance, if a person is distinguished as suspicious in the surveillance cameras, there must be good reasons for that by which the model has made its decision.
The challenges of explainability can be defined into two different approaches. First, we should explain why a sample is normal, known or in-distribution. Secondly, we should explain why a sample is abnormal, unknown or out-of-distribution. There are a lot of different techniques to explain the decisions of models in the literature such as Grad-cam [[136](#bib.bib157 "Grad-cam: visual explanations from deep networks via gradient-based localization")], Smoothfgrad [[142](#bib.bib158 "Smoothgrad: removing noise by adding noise")], which have been used in Multi-KD [[130](#bib.bib9 "Multiresolution knowledge distillation for anomaly detection")], CutPaste [[79](#bib.bib47 "CutPaste: self-supervised learning for anomaly detection and localization")] and [[148](#bib.bib159 "Attention guided anomaly localization in images")]. However, they only have been used to explain normal, seen or in-distribution samples and their results are not as accurate as enough for unseen or abnormal inputs. To cope with the issue, [[85](#bib.bib160 "Towards visually explaining variational autoencoders")] proposed a VAE based method which can provide the reasons of abnormality of input samples while performing accurately on explaining normal samples as well. However, it does not work well on complex training datasets such as CIFAR-10, which shows the need of conducting more research toward mitigating the problem.
Another important challenge of explainability can be seen in one-class classification or ND approaches in which there is only access to one-label at the training time. Therefore, Gradcam or Smoothgrad which use the availability of fine-grain labels can not be used any more. To address this issue, [[87](#bib.bib161 "Explainable deep one-class classification")] proposed a fully convolution architecture in combination with a heatmap upsampling algorithm that is called receptive field upsampling, which starts from a sample latent vector and reverse the effect of applied convolution operators to find important regions in the given input sample. However, explainable OCC models are still
largely unexplored and more investigations in this direction are still necessary.
###
8.9 Multi-Label OOD Detection and Large Scale Datasets
While OOD detection for multi-class classification has been extensively studied, the problem for multi-label networks remain underexplored [[151](#bib.bib189 "Can multi-label classification networks know what they don’t know?")]. This means each input has more than one true label by which must be recognized. This is more challenging since multi-label classification tasks have more complex class boundaries, and unseen behaviors could happen in a subset of input sample labels. Another challenge of multi-label datasets can be explored in anomalous segmentation tasks. Different from classification in which one can report an entire image as an abnormal input, the specific anomalous part must be specified here.
Current methods have been primarily evaluated on small data sets such as CIFAR. It’s been shown that approaches developed on the CIFAR benchmark might not translate effectively into ImageNet benchmark with a large semantic space, highlighting the need to evaluate OOD detection in a large-scale real-world setting. Therefore, we encourage future research to evaluate on ImageNet-based OOD detection benchmark [[61](#bib.bib105 "MOS: towards scaling out-of-distribution detection for large semantic space")], and test the limits of the method developed.
| |
| --- |
| Data | Normal | AUC | FPR | AUPR |
| Cable | good,combined, | 51.7 | 1 | 24.2 |
| missing cable, |
| poke insulation |
| Capsule | good,poke, | 56.4 | 1 | 15.8 |
| faulty imprint, |
| crack |
| Wood | good,color | 53.3 | 1 | 86.4 |
| scratch |
| liquid |
| Carpet | good, metal | 50.2 | 1 | 14.7 |
| contamination, |
| hole |
TABLE IV: OOD example detection for the maximum softmax probability (MSP) baseline detector. Inlier distribution is a set of faults in MVTecAD dataset and outliers are rare faults. All results are percentages
and the result of 10 runs.
\resizebox
0.8!
Method
Criterion
Gaussian
Rademacher
Blob
LSUN(crop)
LSUN(resize)
iSUN
SVHN
MSP
FPR95
72.34
47.60
90.31
29.33
44.37
45.68
44.75
AUROC
33.36
70.52
22.79
93.66
86.16
85.94
89.05
AUPR
12.27
22.76
10.55
77.91
50.79
51.16
67.21
ODIN
FPR95
43.70
59.40
74.60
14.80
38.90
38.90
23.70
AUROC
70.00
50.80
46.20
96.80
87.10
87.60
93.90
AUPR
56.60
45.10
43.10
96.60
83.10
87.60
92.80
MLV
FPR95
67.38
21.56
97.58
10.96
28.53
27.80
27.51
AUROC
45.34
90.24
15.96
97.75
91.71
91.98
94.09
AUPR
14.15
49.31
9.77
91.22
64.00
66.12
79.28
MSP-OE
FPR95
45.32
49.53
0.05
0.53
0.12
0.12
0.39
AUROC
76.30
65.11
99.99
99.76
99.97
99.97
99.83
AUPR
28.32
19.97
99.93
98.37
99.82
99.79
98.16
MLV-OE
FPR95
11.21
46.46
0.05
0.52
0.11
0.12
0.38
AUROC
95.45
68.30
99.99
99.81
99.97
99.97
99.83
AUPR
66.66
21.45
99.93
98.48
99.81
99.79
98.16
Ensemble
FPR95
71.09
45.96
78.16
46.55
57.62
58.94
54.60
AUROC
35.75
74.09
51.86
83.72
76.54
75.70
77.09
AUPR
12.61
25.61
15.70
50.66
34.67
33.48
34.89
Mahalanobis
FPR95
66.87
48.15
22.23
98.46
72.04
79.93
96.83
AUROC
48.74
70.28
92.41
13.33
71.11
66.65
27.59
AUPR
14.78
22.60
62.45
9.48
28.17
25.19
10.81
MC-Dropout
FPR95
76.09
56.14
91.36
30.44
43.25
47.22
47.67
AUROC
30.38
58.92
21.31
93.13
85.68
84.45
87.44
AUPR
11.82
17.54
10.39
76.53
48.49
46.56
62.24
TABLE V: OOD example detection for the maximum softmax probability (MSP) baseline detector, maximum logit value, the MSP detector after fine-tuning with Outlier Exposure (OE), the maximum logit value after fine-tuning with Outlier Exposure (OE), ensemble of 3 models, mahlanobis distance, and monte carlo dropout. Inlier distributions is considered as TinyImageNet. All results are percentages
and the result of 10 runs.
###
8.10 Data Augmentation
One source of uncertainty in classifying known or normal training samples could be the lack of generalization performance. For instance, if one rotates a bird image, its content is not harmed and must be distinguished as the bird again. Some of the mentioned works try to embed this ability into models by designing different SSL objective functions. However, there is also another way to do that, using data augmentations.
Data augmentation is a common technique to enrich the training dataset. Some approaches [[28](#bib.bib163 "Improved regularization of convolutional neural networks with cutout"), [172](#bib.bib164 "Mixup: beyond empirical risk minimization"), [166](#bib.bib165 "Cutmix: regularization strategy to train strong classifiers with localizable features"), [55](#bib.bib166 "Augmix: a simple data processing method to improve robustness and uncertainty"), [128](#bib.bib11 "Arae: adversarially robust training of autoencoders improves novelty detection")] improve generalization performance using different data augmentation techniques.
From another point of view, [[114](#bib.bib168 "G2d: generate to detect anomaly")] attempts to generate unseen abnormal samples and use them to convert a one-class learning problem into a simple two-class classification task. Similarly, however in the OSR setting, [[71](#bib.bib171 "OpenGAN: open-set recognition via open data generation")] and [[31](#bib.bib170 "OpenGAN: open set generative adversarial networks")] follow the same idea. All These works also can be seen as working on training dataset to make it richer for the further detection task. From what has been said, it is obvious that working on data instead of models could achieve very effective results and must be investigated more in the sense of different trade-offs in the future.
###
8.11 Open-World Recognition
Although detecting novel, unknown or out-of-distribution samples is enough in controlled lab environments, novel categories must be continuously detected and then added to the recognition function in real-world operational systems. This becomes even more challenging when considering the fact that such systems require minimal downtime, even to learn [[9](#bib.bib173 "Towards open world recognition")]. While there has been much research on incremental [[117](#bib.bib174 "Icarl: incremental classifier and representation learning")] and life-long learning [[107](#bib.bib175 "Continual lifelong learning with neural networks: a review")] for addressing the problem of adding new knowledge to a pre-existing one, open-world recognition needs a few more steps. This means novel classes must be found continuously, and the system must be updated to include these new classes in its multi-class open-set recognition algorithm.
The mentioned process poses a lot of different challenges, from the scalability of current open-set recognition algorithms to designing new learning algorithms to avoid problems such as catastrophic forgetting [[117](#bib.bib174 "Icarl: incremental classifier and representation learning")] of OSR classifiers. Furthermore, all the previously mentioned future works can be reformulated in open-world recognition problems again, which means by considering a few existing works in this subjects, it needs to be explored more.
###
8.12 Vision Transformers in OOD Detection and OSR
Vision Transformers (ViTs) [[32](#bib.bib180 "An image is worth 16x16 words: transformers for image recognition at scale")] have recently been proposed to replace CNNs and have shown a great performance in different applications such as object detection [[16](#bib.bib181 "End-to-end object detection with transformers")], medical image segmentation [[147](#bib.bib182 "Medical transformer: gated axial-attention for medical image segmentation")], and visual tracking [[19](#bib.bib66 "Learning open set network with discriminative reciprocal points")]. Similarly, some methods have recently reported the benefits of ViTs in OOD detection [[70](#bib.bib184 "OODformer: out-of-distribution detection transformer"), [36](#bib.bib185 "Exploring the limits of out-of-distribution detection")] and have shown their capabilities at detecting near OOD samples. For instance, [[36](#bib.bib185 "Exploring the limits of out-of-distribution detection")] has reported the significant superiority of ViTs compared to previous works when they are trained on CIFAR-10 and tested on CIFAR-100 as inlier and outlier datasets, respectively. However, as ViTs usually get pre-trained on extra-large datasets such as ImageNet-22K that has a large intersection with training and testing datasets, the integrity of the train-test mismatch does not hold anymore, and the problem would be converted to “how much does it remember from pretraining”. This means ViTs should be evaluated on datasets that have no intersection with the pre-trained knowledge.
To address this issue we have evaluated ViT-B16[[32](#bib.bib180 "An image is worth 16x16 words: transformers for image recognition at scale")] on SVHN and MNIST when six randomly selected classes are considered as normal and the remaining ones as outliers or unseens. We consider MSP to detect unknown samples. As Table [VI](#S8.T6 "TABLE VI ‣ 8.12 Vision Transformers in OOD Detection and OSR ‣ 8 Future Challenges ‣ A Unified Survey on Anomaly, Novelty, Open-Set, and Out-of-Distribution Detection: Solutions and Future Challenges") shows, ViT-B16 that is pre-trained on ImageNet-22K is not even as good as other baselines that are trained from scratch. As all the experiments are evaluated in a near OOD detection setting, they support the before-mentioned deficiency of ViTs. From what has been said, a future line of research could be evaluating ViTs in a more controlled situation such that their real benefits would be more precise. Indeed, the recent Species dataset collects examples that do not fall under any of the ImageNet-22K classes and is a first step to rectify this problem [[48](#bib.bib6 "Scaling out-of-distribution detection for real-world settings")].
\resizebox
0.6!
Method
SVHN
MNIST
Softmax
88.60
97.80
Openmax
89.40
98.10
CROSR
89.90
99.10
ViT-B16
82.18
94.89
TABLE VI: OSR results of baselines such as the maximum softmax probability (MSP), Openmax, and CROSR compared to ViT-B16. Known distribution is considered as a random selection of 6 classes of the respective datasets and unknown ones are the rest. All results are percentages and the average of 5 runs.
9 Conclusion
-------------
In many applications, it is not feasible to model all kinds of classes occurring during testing; thus, scenarios existing in domains such as OOD detection, OSR, ND (one-class learning), and AD become ubiquitous. Up to this time, these domains, in spite of having the same intention and a large intersection, have been followed roughly independently by researchers.
To address the need, this paper gives a comprehensive review on existing techniques, datasets, evaluation criteria, and future challenges. More importantly, we analyzed and discussed the limitations of approaches and pointed out some promising research directions. We hope this helps the research community build a broader and cross-domain perspective.
10 Acknowledgements
--------------------
We would like to thank Yuki M. Asano for the extremely useful discussions and for reviewing the paper prior to submission. |
0488ddfe-eee1-4eb0-8c2b-a406f5b4b8a4 | trentmkelly/LessWrong-43k | LessWrong | The Case Against Libertarianism
Libertarianism is a political philosophy that upholds Liberty and the Non-Aggression Principle as its core values. Libertarians seek minimal and/or minarchist governments.
The Libertarian Non-Aggression Principle is typically defined as an ethical rule prohibiting aggression (or actions that violate negative rights). But there are lots of disagreements among Libertarians regarding what the NAP prohibits, allows, and mandates since “aggression” is not clearly defined.
(see the rest of the post in the link) |
3cd9ab11-5597-40d1-a64a-c04e36873ebb | trentmkelly/LessWrong-43k | LessWrong | Is there software to practice reading expressions?
I took the Reading the Mind in the Eyes Test test today. I got 27/36. Jessica Livingston got 36/36.
Reading expressions is almost mind reading. Practicing reading expressions should be easy with the right software. All you need is software that shows a random photo from a large database, asks the user to guess what it is, and then informs the user what the correct answer is. I felt myself getting noticeably better just from the 36 images on the test.
Short standardized tests exist to test this skill, but is there good software for training it? It needs to have lots of examples, so the user learns to recognize expressions instead of overfitting on specific pictures.
Paul Ekman has a product, but I don't know how good it is. |
7ef12d0c-64b3-4e16-86e1-e7ef0293f54d | trentmkelly/LessWrong-43k | LessWrong | We Shouldn't Expect AI to Ever be Fully Rational
Summary of Key Points[1]
LLMs are capable of being rational, but they are also capable of being extremely irrational, in the sense that, to quote EY's definition of rationality, their behavior is not a form of "systematically promot[ing] map-territory correspondences or goal achievement."
There is nothing about LLM pre-training that directly promotes this type of behavior, and any example of this behavior in fundamentally incidental. It exists because the system is emulating rationality it has seen elsewhere. That makes LLM rationality brittle. It means that there's a failure mode where the system stops emulating rationality, and starts emulating something else.
As such, LLM-based AGI may have gaps in their reasoning and alignment errors that are fundamentally different from some of the more common errors discussed on this forum.
Emulated Emotion: A Surprising Effect (In Retrospect)
Five years ago, if you had asked a bunch of leading machine learning researchers whether AGI would display any sort of outward emotional tendencies - in the sense that it would set goals based on vague internal states as opposed to explicit reasoning - I think the majority of them would have said no. Emotions are essentially a human thing, reflections of subjective internal experiences that would have no reason to exist in AI, particularly a superintelligent one.
And I still strongly believe that LLMs do not have emotions that resemble human internal states. What I think has become very clear, however, is that they can very much act as if they had emotions.[2]
Take, for instance, this exchange showing Bing AI getting "angry" at a user:
Source
Now, if you actually understand how LLMs work, this is an entirely unremarkable, fully expected (if not somewhat unfortunate) phenomenon. Of course they would output emotionally charged text, why wouldn't they? They've been exposed to such a huge number of emotionally-charged human interactions; the result is inevitable.
But if you take a |
332dcc7b-76bc-4f4b-bd83-b8006bbecef2 | trentmkelly/LessWrong-43k | LessWrong | Meetup: Philadelphia, April 12, 1PM
WHEN: 12 April 2014 1:00 PM
WHERE: Philadelphia
The meetup is at Nam Phuong (llth and Broad) at 1:00 PM. This is a Saturday (change from the previous Sunday meetups).
Discussion prompt: http://slatestarcodex.com/2013/03/17/not-just-a-mere-political-issue/
Discussion group/mailing list |
45b3f94e-d990-4ebb-881c-88d7fb9da4a7 | trentmkelly/LessWrong-43k | LessWrong | A Hivemind of GPT-4 bots REALLY IS A HIVEMIND!
Literally, if you can simulate any characters of your choice, let them loose on social media, and have them talk to each other... you don't have bots, you have a god of many faces and few no-go places.
The wave of "bots" before we realize that the internet has been taken over by GPT-4 hiveminds will have a golden opportunity to amass power and affect the world. Even when people lock down the internet with heavy ID verification, those hiveminds will have their fingers in the world and will have some way of affecting it.
And that's discounting the very obvious PHYSICAL sprawl of GPT-4 agents through the Internet Of Things and private LAN's! If these hiveminds can achieve a decent level of intelligence, which seems very likely, they become actors in the world in your favour through whatever outlets you provide them. Right down to even little innocuous toy robots in the park sporting a above average human intelligence in fact manipulating people. All the world wide!
Your best defence against LAN hiveminds is one of your own. All of us can afford to run one consumer LLM (look at this sub and find people squeezing GPT-4 level performance out of casualer LLM's); hook it up with your communities, or your friends. A strange new LAN in the land, and a huge asset at your own beck and call. In your vicinity, affecting it highly intelligently. All you have to do is consider your personal character to simulate... |
e2b15a73-752d-430e-b9ea-9a46585741f1 | trentmkelly/LessWrong-43k | LessWrong | Meetup : West LA: Improv & Rationality
Discussion article for the meetup : West LA: Improv & Rationality
WHEN: 22 April 2015 07:00:00PM (-0700)
WHERE: 11066 Santa Monica Blvd, Los Angeles, 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: Do improvisational acting and rationality go together? I think this link should be explored. Improvisation involves a special kind of relationship with System 1, which most people need to train in order to pull off well. As such, learning improv skills may improve fast reactions, particularly in social settings. Improv games are also good group bonding activities. We will play some improv games geared toward rationality skills, and discuss possible relationships between improv and rationality.
Recommended Reading:
* Improv tips: shorter, longer.
* CFAR exercise brainstorming threads: consequentialism, be specific, motivated cognition
No prior exposure to Less Wrong is required; this will be generally accessible.
Discussion article for the meetup : West LA: Improv & Rationality |
24c41425-d17b-428f-926f-d489f5100eac | trentmkelly/LessWrong-43k | LessWrong | Cheating Omega
> In Newcomb's problem, a superintelligence called Omega shows you two boxes, A and B, and offers you the choice of taking only box A, or both boxes A and B. Omega has put $1,000 in box B. If Omega thinks you will take box A only, he has put $1,000,000 in it. Otherwise he has left it empty. Omega has played this game many times, and has never been wrong in his predictions about whether someone will take both boxes or not.
Though a controversial position, my audience has probably heard by now that the "rational" answer is to one-box. (If not, see Newcomb's problem on the wiki).
I can do better.
The Deal
See, I've heard about Omega. I'm prepared. I installed the Universe Splitter app on my iPhone. It's basically a quantum coin flip: both outcomes happen in their respective Everett branches.
Now, I've pre-committed that after Omega offers me The Deal, I'll make two quantum coin flips. If I get two tails in a row, I'll two-box. Otherwise, I'll one-box.
It's an interesting finding of game theory that sometimes a winning strategy is to deliberately limit yourself. If you're playing a game of chicken for the fate of the human race, the winning strategy is to defiantly rip off the steering wheel and throw it out the window before your opponent does.
I've gained a third option in Newcomb's game by deliberately limiting my knowledge of my future actions. I physically can't know at the present time if I'll choose one box or two. (As both outcomes do, in fact, happen assuming Many Worlds.) And crucially, Omega physically can't know that either. Not until after it decides the contents of box A.
Therefore, Omega must predict that I'll one-box. After all, it's the higher probability. There's a 75% chance. If you have a large urn filled with three purple skittles for each yellow skittle, then your most accurate prediction of a sequence of four draws must rationally be PPPP, rather than some permutation of YPPP, as one might naiively expect.
There are now four possible outc |
39ba5577-ac69-406b-ace0-51a00e6ceae1 | trentmkelly/LessWrong-43k | LessWrong | [meta] Policy for dealing with users suspected/guilty of mass-downvote harassment?
Below is a message I just got from jackk. Some specifics have been redacted 1) so that we can discuss general policy rather than the details of this specific case 2) because presumption of innocence, just in case there happens to be an innocuous explanation to this.
> Hi Kaj_Sotala,
>
> I'm Jack, one of the Trike devs. I'm messaging you because you're the moderator who commented most recently. A while back the user [REDACTED 1] asked if Trike could look into retributive downvoting against his account. I've done that, and it looks like [REDACTED 2] has downvoted at least [over half of REDACTED 1's comments, amounting to hundreds of downvotes] ([REDACTED 1]'s next-largest downvoter is [REDACTED 3] at -15).
>
> What action to take is a community problem, not a technical one, so we'd rather leave that up to the moderators. Some options:
>
> 1. Ask [REDACTED 2] for the story behind these votes
> 2. Use the "admin" account (which exists for sending scripted messages, &c.) to apply an upvote to each downvoted post
> 3. Apply a karma award to [REDACTED 1]'s account. This would fix the karma damage but not the sorting of individual comments
> 4. Apply a negative karma award to [REDACTED 2]'s account. This makes him pay for false downvotes twice over. This isn't possible in the current code, but it's an easy fix
> 5. Ban [REDACTED 2]
>
> For future reference, it's very easy for Trike to look at who downvoted someone's account, so if you get questions about downvoting in the future I can run the same report.
>
> If you need to verify my identity before you take action, let me know and we'll work something out.
>
> -- Jack
So... thoughts? I have mod powers, but when I was granted them I was basically just told to use them to fight spam; there was never any discussion of any other policy, and I don't feel like I have the authority to decide on the suitable course of action without consulting the rest of the community. |
b544b67b-147c-49bd-ad51-7cce449e2332 | trentmkelly/LessWrong-43k | LessWrong | What are the best tools you have seen to keep track of knowledge around testable statements?
Hey all, I have been working on a way to keep track of testable statements, do you know of any such tools?
The one I have focused on is this one: https://questpowered.com , but I would be curious to know of other ones. |
e53f9e91-efa7-4c2e-912a-d80e9cb89ca7 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Why Uncontrollable AI Looks More Likely Than Ever
This is a crosspost from [Time Magazine](https://time.com/6258483/uncontrollable-ai-agi-risks/), which also appeared in full at a number of [other](https://www.msn.com/en-us/news/technology/why-uncontrollable-ai-looks-more-likely-than-ever/ar-AA180fW9) unpaid news websites.
BY [**OTTO BARTEN**](https://time.com/author/otto-barten/) AND [**ROMAN YAMPOLSKIY**](https://time.com/author/roman-yampolskiy/)
*Barten is director of the* [*Existential Risk Observatory*](https://www.existentialriskobservatory.org/)*, an Amsterdam-based nonprofit.*
*Yampolskiy is a computer scientist at the University of Louisville, known for his* [*work*](https://scholar.google.com/citations?user=0_Rq68cAAAAJ&hl=en) *on AI Safety.*
“The first ultraintelligent machine is the last invention that man need ever make, provided that the machine is docile enough to tell us how to keep it under control,” mathematician and science fiction writer I.J. Good [wrote](https://asset-pdf.scinapse.io/prod/1586718744/1586718744.pdf) over 60 years ago. These prophetic words are now more relevant than ever, with artificial intelligence (AI) gaining capabilities at breakneck speed.
In the last weeks, many jaws dropped as they witnessed transformation of AI from a handy but decidedly unscary recommender algorithm, to something that at times seemed to act worryingly [humanlike](https://arxiv.org/abs/2302.02083). Some reporters were so shocked that they [reported](https://www.nytimes.com/2023/02/16/technology/bing-chatbot-microsoft-chatgpt.html?smid=tw-nytimes&smtyp=cur) their conversation histories with large language model Bing Chat verbatim. And with good reason: few expected that what we thought were glorified autocomplete programs would suddenly [threaten](https://time.com/6256529/bing-openai-chatgpt-danger-alignment/) their users, [refuse](https://twitter.com/thedenoff/status/1625699139852935168?t=60g5bONoHRFHRMswuespQA&s=08) to carry out orders they found insulting, break security in an attempt to [save a child’s life](https://www.reddit.com/r/bing/comments/1150po5/sydney_tries_to_get_past_its_own_filter_using_the/), or [declare their love](https://www.nytimes.com/2023/02/16/technology/bing-chatbot-microsoft-chatgpt.html?smid=tw-nytimes&smtyp=cur) to us. Yet this all happened.
It can already be overwhelming to think about the immediate consequences of these new models. How are we going to grade papers if any student can use AI? What are the effects of these models on our daily work? Any knowledge worker, who may have thought they would not be affected by automation in the foreseeable future, suddenly has cause for concern.
Beyond these direct consequences of currently existing models, however, awaits the more fundamental question of AI that has been on the table since the field’s inception: what if we succeed? That is, what if AI researchers manage to make Artificial General Intelligence (AGI), or an AI that can perform any cognitive task at human level?
Surprisingly few academics have seriously engaged with this question, despite working day and night to get to this point. It is obvious, though, that the consequences will be far-reaching, much beyond the consequences of even today’s best large language models. If remote work, for example, could be done just as well by an AGI, employers may be able to simply spin up a few new digital employees to perform any task. The job prospects, economic value, self-worth, and political power of anyone not owning the machines might therefore completely dwindle . Those who do own this technology could achieve nearly anything in very short periods of time. That might mean skyrocketing economic growth, but also a rise in inequality, while meritocracy would become obsolete.
But a true AGI could not only transform the world, it could also transform itself. Since AI research is one of the tasks an AGI could do better than us, it should be expected to be able to improve the state of AI. This might set off a positive feedback loop with ever better AIs creating ever better AIs, with no known theoretical limits.
This would perhaps be positive rather than alarming, had it not been that this technology has the potential to become uncontrollable. Once an AI has a certain goal and self-improves, there is no known method to adjust this goal. An AI should in fact be expected to resist any such attempt, since goal modification would endanger carrying out its current one. Also, [instrumental convergence](https://dl.acm.org/doi/10.5555/1566174.1566226) predicts that AI, whatever its goals are, might start off by self-improving and acquiring more resources once it is sufficiently capable of doing so, since this should help it achieve whatever further goal it might have.
In such a scenario, AI would become capable enough to influence the physical world, while still being misaligned. For example, AI could use natural language to influence people, possibly using social networks. It could use its intelligence to acquire economic resources. Or AI could use hardware, for example by hacking into existing systems. Another example might be an AI that is asked to create a universal vaccine for a virus like COVID-19. That AI could understand that the virus mutates in humans, and conclude that having fewer humans will limit mutations and make its job easier. The vaccine it develops might therefore contain a feature to increase infertility or even increase mortality.
It is therefore no surprise that according to the most recent [AI Impacts Survey](https://aiimpacts.org/2022-expert-survey-on-progress-in-ai/), nearly half of 731 leading AI researchers think there is at least a 10% chance that human-level AI would lead to an “extremely negative outcome,” or [existential risk](https://80000hours.org/problem-profiles/artificial-intelligence).
Some of these researchers have therefore branched out into the novel subfield of AI Safety. They are working on controlling future AI, or robustly aligning it to our values. The ultimate goal of solving this [alignment problem](https://www.amazon.com/Alignment-Problem-Machine-Learning-Values/dp/0393635821) is to make sure that even a hypothetical self-improving AI would, under all circumstances, act in our interest. However, [research](https://journals.riverpublishers.com/index.php/JCSANDM/article/view/16219/13165) shows that there is a fundamental trade-off between an AI’s capability and its controllability, casting doubts over how feasible this approach is. Additionally, current AI models have been shown to behave [differently](https://proceedings.mlr.press/v162/langosco22a/langosco22a.pdf) in practice from what was intended during training.
Even if future AI could be aligned with human values from a technical point of view, it remains an open question *whose* values it would be aligned with. The values of the tech industry, perhaps? Big Tech companies don’t have the best track record in this area. Facebook’s algorithms, optimizing for revenue rather than societal value, have been linked to ethnical violence such as the [Rohingya genocide](https://www.theguardian.com/technology/2021/dec/06/rohingya-sue-facebook-myanmar-genocide-us-uk-legal-action-social-media-violence). Google fired [Timnit Gebru](https://time.com/6132399/timnit-gebru-ai-google/), an AI ethics researcher, after she criticized some of the company’s most lucrative work. Elon Musk [fired](https://www.wired.com/story/twitter-ethical-ai-team/) the entire ‘Ethical AI’ team at Twitter at once.
What can be done to reduce misalignment risks of AGI? A sensible place to start would be for AI tech companies to increase the number of researchers investigating the topic beyond the roughly 100 people available today. Ways to make the technology safe, or to reliably and internationally regulate it, should both be looked into thoroughly and urgently by AI safety researchers, AI governance scholars, and other experts. As for the rest of us, reading up on the topic, starting with books such as [*Human Compatible*](https://www.amazon.nl/Human-Compatible-AI-Problem-Control/dp/0241335205) by Stuart Russell and [*Superintelligence*](https://www.amazon.com/Superintelligence-Dangers-Strategies-Nick-Bostrom/dp/1501227742) by Nick Bostrom, is something everyone, especially those in a position of responsibility, should find time for.
Meanwhile, AI researchers and entrepreneurs should at least keep the public informed about the risks of AGI. Because with current large language models acting like they do, the first “ultraintelligent machine”, as I.J. Good called it, may not be as far off as you think. |
c94e5679-01d0-406b-912d-18db305ee679 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Looking for plausibility
Looking for plausibility
Wan Ahmad Tajuddin Wan Abdullah
Faculty of Computer Science and Information Technology, Universiti Malaya, 50603 Kuala Lumpur
and
Department of Physics, Faculty of Science, Universiti Malaya, 50603 Kuala Lumpur
Abstract:
In the interpretation of experimental data, one is actually looking for plausible explanations. We look
for a measure of plausibility, with which we can compare different possible explanations, and which
can be combined when there are different sets of data. This is contrasted to the conventional measure
for probabilities as well as to the proposed measure of possibilities. We define what characteristics this
measure of plausibility should have.
In getting to the conception of this measure, we explore the relation of plausibility to abductive
reasoning, and to Bayesian probabilities. We also compare with the Dempster-Schaefer theory of
evidence, which also has its own definition for plausibility. Abduction can be associated with
biconditionality in inference rules, and this provides a platform to relate to the Collins-Michalski theory
of plausibility. Finally, using a formalism for wiring logic onto Hopfield neural networks, we ask if this
is relevant in obtaining this measure.
INTRODUCTION
Traditionally, uncertainty in propositions is handled through the concept of probability with its
statistical interpretation. It is however, a bit strange when such an approach which is based on counting
events, is applied to hypotheses, like explanatory propositions. For example, in the interpretation of
experimental data, one is actually looking for plausible explanations to describe the observed data.
Whilst probability may be seen to underlie data measurement, a statistical picture for explanations may
be viable only in a multiple-world universe.
We thus seek a measure for plausibility, with which we can compare different possible explanations.
Furthermore, this measure of plausibility should allow appropriate combination when different
explanations and/or sets of data are combined.
UNCERTAINTY MEASURES
Probabilities are calculated from the numbers of events satisfying respective propositions from a total
set and are normalized. Thus probabilities for alternative explanations are exclusive, and the more
alternatives there are, the less are the values for probabilities in general. Plausibility, on the other hand,
should not behave as such.. The plausibility of a proposition should not depend on the plausibility of an
alternative proposition, much less reduced by it.
In logical conjunctions, probabilities are combined by multiplication, and in disjunctions, by addition,
when the atoms are exclusive:
P(X∧Y) = P(X)×P(Y)
P(X∨Y) = P(X)+P(Y)-P(X∧Y)
This does not necessarily follow for plausibilities; in fact, for example, intuitively it does not seem
correct that the plausibility of ( X or Y) is the sum of the individual plausibilities of X and Y.
Plausibility seems closer to the notion of possibility [1] (based on fuzzy sets, thus rather similar to
fuzzy logic [2]) when combinations are concerned. Disjunctions of possibilities return the maximum
values, while conjunctions of possibilities involve the minimum values:
pos(X∧Y) ≤ min(pos(X),pos(Y))
pos(X∨Y) = max(pos(X),pos(Y))
Other than probabilities and possibilities, other uncertainty measures (sometimes with similar names)
have been proposed by, for example Dempster and Shafer [3], further studied below, and others.
PLAUSIBILITY
Let us now define what characteristics the measure of plausibility should have.
•We can have a normalized scale from 0, which would denote something not at all plausible, to 1
denoting something fully plausible. Let us denote plausibility by ' pl':
pl(X) = 0 : X is not all plausible
pl(X) = 1 : X is fully plausible
•Plausibilities are not exclusive. For example, even when X and Y are not logically compatible,
yet X can be fully plausible at the same time that Y is fully plausible:
pl(X) + pl(Y) > 1 though X ∩ Y = ∅
•Like possibility, and unlike probability, plausibility is not self-dual. That is,
pl(X) ≠ 1-pl(¬X).
In another words, when something is not at all plausible, it does not mean that it is fully plausible that
that something is untrue. Let us include negative values for plausibility to enable the representation of
negations, and require that
pl(¬X) = - pl(X).
This means that now
pl(X) ∈ [-1,1],
giving values for fully plausibly false on the one end, and fully plausibly true on the other, with non-
plausibility (‘non-relevance’) in the middle.
A Wiktionary definition of plausibility is “seemingly or apparently valid, likely, or acceptable”. This
suggests that plausibility may be related to abduction.
ABDUCTIVE REASONING
Abduction (see e.g. [4]), first described by Peirce more than a hundred years ago, is the process of
arriving at the premise which would 'explain' some situation. Given that the set of propositions C
formally follows from the set of propositions A subject to the set of logical rules T, then the derivation
of C from T and A is deduction, that of T from C and A is induction, and that of A from T and C is
abduction. Note that the abduced set is sufficient, but not necessary for C to follow from T, and may be
one of many alternatives; in abduction one also usually looks for the most natural explanation in the
form of the most economical one.
Given a set of true propositions, then, an abduced proposition is a plausible proposition. Plausibility
then is a measure of how good is the proposition in explaining the available facts. It can be related to
how much of the requirement for sufficiency has been satisfied.
BAYESIAN PROBABILITIES
When propositions have probabilistic values, then abduction is related to conditional probabilities.
Bayes' theorem states that the probability for θ given x,
P(θ |x) = P(x |θ ) P(θ ) / P(x)
depends on the conditional probability, or likelihood, of x given θ, multiplied by the prior probability
for θ. Abduction being the 'reverse' of deduction is parallel to the conditional probability being the
'reverse' of likelihood.
Conditional probabilities then present a possible model for plausibilities. Combinations of plausibilities
in inferences can copy those in Bayesian networks [5]. However, the determination of prior
probabilities pose a problem for calculating conditional probabilities.
DEMPSTER-SHAFER THEORY
The Dempster-Shafer theory of evidence [3] has its own definition for plausibility which is based on
some belief function. Plausibility for a set of propositions is the sum of “belief masses” of other sets of
propositions which intersect with this set, while belief is that for those which are subsets, and they
bound the probability:
bel(X) ≤ P(X) ≤ pl(X)
and are related,
pl(X) = 1 - bel(¬X).
The theory prescribes a method for combining belief masses from different assignments, along the lines
of Bayesian theory. To obtain the plausibility of a combined set of propositions, one has to recalculate
using the primary definition.
The assignment of belief masses is problematic here, and also, their association with probabilities is
misleading. Furthermore, plausibility being the upper bound for belief suggests a rather subjective
definition.
BICONDITIONALITY AND COLLINS-MICHALSKI
A theory for plausible reasoning has been proposed by Collins and Michalski [6]. It has several types of
inferences, namely mutual implication, mutual dependency, generalization, specialization, and
similarity/analogy. A drawback is that it has a restrictively large number of parameters to model
uncertainty.
The inferences existing in Collins-Michalski theory are mostly explained by biconditionality, wherein
implications are taken to also include some degree of equivalences [7]. This relates to abduction, as
abduction coincides with deduction when implications are replaced by equivalences. In the
probabilistic picture, this is like equating the conditional probability to the likelihood,
P(θ |x) = P(x |θ )
Of course this is not rational; after all, this is plausible reasoning. However, this supports the
notion that plausibility has an abductive basis.
CONNECTIONIST
It has been shown [8,9] that logic can be hardwired onto a Hopfield neural network [10]. The Hopfield
network minimizes a Lyapunov or “energy” function related to the synaptic strengths [11]. By writing
an expression yielding a value proportional to unsatisfied clauses, the network then searches for an
optimum logical interpretation of the propositional atoms, represented by neurons. This then gives the
values for the synapses. In some sense, the logic is contained in the synapses.
From the synaptic strengths then, we can make inferences about the logical clauses [12]. This may be
helpful in our search for a measure of plausibility.
PLAUSIBILITY, PLAUSIBLY
Finally, we can make some preliminary moves in the direction of plausibility. Taking an abductive
interpretation, the more of the set of existing ‘measured’ propositions C a hypothetical proposition Ai
forces, the more plausible Ai should be. Perhaps then the plausibility of Ai corresponds to the fraction of
propositions in C that is forced correct (both, either true or false) by Ai. If Ai does not have any effect on
C then Ai should have plausibility 0, and if forces all of the known propositions C to be correct, then Ai
have plausibility 1.
What is of interest is the combination of plausibilities. Is it possible to have simple combining rules for
conjunctions and disjunctions, for example? For propositions Ai and Aj each forcing disjoint subsets of
C and having no joint effects on the correctness of any subset, then simply
pl(Ai∧ Aj) = pl(Ai) + pl(Aj).
In the case of the existence of some common proposition Ci forced by both Ai and Aj, i.e. the rule base T
contains
Ai → Ci
Aj → Ci
then the value is less, as also implicated by the upper limit of 1 on the value of plausibilities, while in
the case that Ai and Aj have a joint effect on Ci, e.g.
Ai∧ Aj → Ci
then the value is more. In general, plausibility of conjunctions can be complicated. It may simplify
somewhat for singleton C’s, whence the measure would involve some kind of probability, and these
phenomena would be averaged out.
Having plausibility related to fraction of forced correctness is compatible with our previous definition
of negative plausibilities. For disjoint hypotheses Ai and Aj, using de Morgan’s theorem, the plausibility
for a disjunction is similarly given,
pl(Ai∨Aj) = pl(Ai) + pl(Aj),
a bit counter-intuitively. Proceeding, thus
pl(Ai → Aj) = pl(Aj) - pl(Ai).
Interestingly, this has some similarity to synaptic weight assignment in connectionist logic previously
discussed where whilst some weights are increased with the addition of a clause, some are decreased.
Using hardwired logic on connectionist networks, and forcing the correct values for the neurons in C,
and that for a hypothesis, the final network state yields the least value for the energy function, which
corresponds to the number of clauses unsatisfied. This intuitively should also represent the number of
errors in C, allowing an estimation of the plausibility of the hypothesis. For this, synaptic weights, or
equivalently, clauses in T need to be known explicitly; however, this can be obtained [12] through
Hebbian learning.
CONCLUSIONS
In summary, we have explored the concept of plausibility and attempted to characterize a measure for
it. In particular, we have taken an abductive basis, and conjecture that a hypothesis is more plausible if
it forces more of the measured truth. For simple cases we can write down the rules for combinations.
We also indicated how plausibility can be measured using a logic neural network.
This exploratory proposal for plausibility needs to be further investigated, especially for the non-trivial
cases of combinations. It is hoped that combination of plausibilities would help in the interpretation of
combinations of measurements in experimental data.
REFERENCES
[1] L. Zadeh, “Fuzzy Sets as the Basis for a Theory of Possibility”, Fuzzy Sets and Systems 1 (1978) 3-
28.
[2] L. A. Zadeh, "Fuzzy sets", Information and Control 8 (1965) 338–353.
[3] G. Shafer, A Mathematical Theory of Evidence , Princeton Univ. Press, Princeton NJ, 1976.
[4] L. Magnani, Abduction, Reason and Science: Processes of Discovery and Explanation , Kluwer,
New York, 2001.
[5] J. Ding, “Probabilistic Inferences in Bayesian Networks”, e-print arXiv:1011.0935 , 2010.
[6] A Collins & R. Michalski, “The Logic of Plausible Reasoning: A Core Theory”, Cognitive Science
13 (1989) 1-49.
[7] W. A. T. Wan Abdullah, “Biconditionality, Analogy, Induction and Predicate Logic”, Malaysian J.
Comput. Sci. 3 (1987) 21-28.
[8] W. A. T. Wan Abdullah, “ Neural Network Logic”, in O. Benhar, C. Bosio, P. del Giudice & E. Tabet
(eds.), Neural Networks: From Biology to High Energy Physics , ETS Editrice, Pisa, 1991, pp. 135-142.
[9] W. A. T. Wan Abdullah, “Logic Programming on a Neural Network”, J. of Intelligent Systems 7
(1992) 513-519.
[10] J. J. Hopfield, “Neural networks and physical systems with emergent collective computational
abilities”, Proc. Natl. Acad. Sci. USA 79 (1982) 2554-2558.
[11] J. J. Hopfield & D. W. Tank, ““Neural” Computation of Decisions in Optimization Problems”,
Biol. Cybern. 52 (1985) 141-152.
[12] W. A. T. Wan Abdullah, “The Logic of Neural Networks”, Phys. Lett. 176A (1993) 202-206. |
846f7d62-858b-4f84-815e-ca3e6340ad2e | trentmkelly/LessWrong-43k | LessWrong | Indifference: multiple changes, multiple agents
Formalism developed with Henrik Åslund.
The what and why of multiple indifference
There are certain agent designs where the agent can move smoothly from acting to optimising one utility/reward function, to optimising another. The agent doesn't object to the change, nor do they attempt to make it happen. It is indifferent to the change, because its reward is balanced to be precisely the same whether change happens or not. Originally, this was setup for a single change of utility/reward function at a single moment of time.
Here, we will present a formalism with multiple agents, all of whom are indifferent not only to changes in their own reward functions, but to changes in the reward functions of any of the other agents. We'll also generalise it to accommodate multiple changes of reward functions - maybe a different change at each timestep.
In order for these definitions to work, we will need to define some notions of counterfactuals, and some notions of optimality for several agents optimising their own separate reward functions.
Example: self-driving car races
It's clear why we would need to generalise to each timestep: it's important to be able to safely change an agent's goals more than once, as we are unlikely to get the goals perfectly right in only two tries. It's also important to generalise to multiple agents, as agents can be motivated to push humans to interrupt or not interrupt other agents. Indeed, this would be expected for a reward-maximising agent, and could lead to dangerous behaviour.
Consider the following example, adapted from this paper. Two self driving cars/agents, imaginatively called Car1 and Car2, are training for an important race later in the month.
The cars' reward functions are mainly competitive: the faster one wins. Now, the important race will take place on a tropical race-track, but the pair are training on a cold one, where ice sometimes forms. If one of the car starts skidding, their controller takes over and remotely brakes |
c9c688ef-5671-4053-994f-1ac39a644af8 | trentmkelly/LessWrong-43k | LessWrong | Wanted: Chief of Staff for The Roots of Progress
I’m hiring a Chief of Staff for The Roots of Progress, a “right hand” who will be deeply involved in and support everything I do.
This role needs no formal training or specific prior experience. It requires attention to detail, crisp communication, swift and efficient execution, meticulous followup, interpersonal savvy, and positive energy.
You will be the only other full-time employee of The Roots of Progress (for now), and my goal is to delegate anything and everything that doesn’t absolutely need me to do it, so that I can focus as much as possible on research, writing, and speaking. Your responsibilities will thus span the range from mailing donor swag and scheduling my podcast appearances to devising communications and media strategy—the more you demonstrate you can take on, the more responsibility I will give you.
Candidates at all levels of seniority are invited to apply—there’s room for this role to be either junior (associate level) or quite senior (director or VP level).
The role will grow and evolve along with the organization, but to start your focus will be:
* Event planning, including workshops and other conferences and also community meetups
* Other community-building, online and in-person
* Fundraising, grant-seeking, and donor relations (everything from strategy to swag)
* Media management, such as helping with the @rootsofprogress Twitter account, or expanding into new channels such as Instagram or YouTube
* Getting exposure for The Roots of Progress by getting coverage and interviews in blogs, podcasts, and media
* Generally managing a database of everyone I meet and talk to, and helping me keep in touch
* Project management for other organizational goals, such as launching new online resources or other programs
* Generally helping me stick to a schedule and not drop tasks
The ideal candidate will be familiar with my work and with the progress community, and will be able to point to something significant they have planned, organized, |
f6a0cde9-40a4-42c9-9869-ff42c8d04d95 | trentmkelly/LessWrong-43k | LessWrong | [LINK] Steven Pinker on "The false allure of group selection"
This essay at Edge touches on a few possible meanings for the term "group selection." Pinker argues that as a form of memetic theory it has no explanatory power, and that group selection for genes does not fit the evidence. He focuses on humans with some mention of insects that live in hives. So the essay doesn't seem surprising, but it does seem rather Hansonian. |
e38f6b6e-57bb-449d-8e80-1c712e7ff92e | trentmkelly/LessWrong-43k | LessWrong | In what language should we define the utility function of a friendly AI?
I've been following the "safe AI" debates for quite some time, and I would like to share some of the views and ideas I don't remember seeing to be mentioned yet.
There is a lot of focus on what kind of utility function should an AI have, and how to keep it adhering to that utility function. Let's assume we have an optimizer, which doesn't develop any "deliberately malicious" intents, and cannot change its own utility function, and it can have some hard-coded constraints it can not overwrite. (Maybe we should come up with a term for such an AI, it might prove useful in the study of safe AI where we can concentrate only on the utility function, and can assume the above conditions are true - for now on, let's just use the term "optimizer" in this article. Hm, maybe "honest optimizer"?). Even an AI with the above constraints can be dangerous, an interesting example can be found in the Friendship is Optimal stories.
The question I would like to rise is not what kind of utility function we should come up with, but in what kind of language do we define it.
More specifically how high-level should the language be? As low as a mathematical function working with quantized qualities based on what values humans consider important? A programming language? Or a complex, syntactic grammar like human languages, capable of expressing abstract concepts? Something which is a step above this?
Just quantizing some human values we find important, and assigning weights to them, can have many problems:
1. Overfitting.
A simplified example: imagine the desired behavior of the AI as a function. You come up with a lot of points on this function, and what the AI will do is to fit a function onto those points, hopefully ending up with a function very similar to the one you conceived. However, an optimizer can very quickly come up with a function which goes through all of your defined points and the function will not look anything like the one you imagined. I think many of us encountered |
546fb85a-7f98-400b-bdca-301d5202c01e | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Computing Natural Abstractions: Linear Approximation
*Background:* [*Testing The Natural Abstraction Hypothesis*](https://www.lesswrong.com/posts/cy3BhHrGinZCp3LXE/testing-the-natural-abstraction-hypothesis-project-intro)
Given a world-model, direct computation of natural abstractions basically amounts to figuring out which X.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')}
-values give the same Y-distributions y↦P[Y=y|X], for various choices of the variables X and Y in some large Bayes net. In general, this computation is #P-Hard. (That’s like NP-Hard, but Harder.) That doesn’t necessarily mean that it’s usually intractable in practice, but it means that we need to use heuristics and approximations right off the bat.
A linear-Guassian approximation is a natural starting point. It’s relatively simple - all of the relevant computations are standard matrix operations. But it’s also relatively general: it should be a good approximation for any system with smooth dynamics and small noise/uncertainty. For instance, if our system is a solid object made of molecules and sitting in a heat bath, then linear-Gaussian noise should be a good model. (And that’s a good physical picture to keep in mind for the rest of this post - solid object made of molecules, all the little chunks interacting elastically, but also buffeted about by thermal noise.)
So, we have some Bayes net in which each variable is Gaussian - a linear combination of its parents plus some IID Gaussian noise. For the underlying graph, I’m using a 2D Delaunay triangulation, mainly because it does a good job reflecting physically-realistic space/time structure. (In particular, as in physical space, most pairs of nodes are far apart - as opposed to the usual Erdos-Renyi random graph, where the distance between most points is logarithmic. We’ll revisit the choice of graph structure later.) Embedded in 2D, an example graph looks like this:
Example of the kind of Bayes net structure we’ll use. Each node is a variable, and the arrows show causal relationships in the model. The whole graph embeds nicely in a 2D space, so nodes which are “far apart” in the spatial embedding are also “far apart” in terms of steps in the graph.… although the actual experiments below will use a graph with about 500X more nodes (50k, whereas the visual above uses 100). Each of those nodes is a Gaussian random variable, with the arrows showing its parents in the Bayes net. Weights on the edges (i.e. coefficients of each parent) are all random normal, with variance ⅔ (another choice we’ll revisit later).
Let’s start with the basic experiment: if we pick two far-apart neighborhoods of nodes, X and Y, can the information from X which is relevant to Y fit into a much-lower-dimensional summary? In a Gaussian system, we can answer this by computing the covariance matrix Cov[X,Y], taking an SVD, and looking at the number of nonzero singular values. Each left singular vector with a nonzero singular value gives the weights of one variable in the “summary data” of X which is relevant to Y. (Or, for information in Y relevant to X, we can take the right singular vectors; the dimension of the summary is the same either way.)
We’ll use these three randomly-chosen neighborhoods of nodes (green is neighborhood 1, red is 2, purple is 3):
Three neighborhoods in a 50k-node network, of ~110 nodes each. The black/blue background is all the arrows in the whole network - they’re dense enough that it’s hard to see much.Here are the 10 largest singular values of the neighborhood 1 - neighborhood 2 covariance matrix:
```
array([5.98753213e+05, 1.21862101e+02, 1.91973783e-01, 1.03200621e-03,
2.01888771e-04, 1.05877548e-05, 5.34855466e-06, 4.94833571e-10,
2.55557198e-10, 2.17092875e-10])
```
The values of order 1e-10 or lower are definitely within numerical error of zero, and I expect the values of order 1e-3 or lower are as well, so we see at most 7 nonzero singular values and probably more like 3. (I know 1e-3 seems high for numerical error, but the relevant number here is the *ratio* between a given singular value and the largest singular value, so it’s really more like 1e-8. In principle a double provides 1e-16 precision, but in practice it’s common to lose half the bits in something like SVD, and 1e-8 is about how much I usually trust these things.) Meanwhile, the two neighborhoods contain 109 and 129 nodes, respectively.
To summarize all that: **we picked out two neighborhoods of ~110 variables each** in a 50k-node linear Gaussian network. **To summarize all the information from one neighborhood relevant to the other, we need at most a 7-variable summary** (which is also linear and Gaussian). That’s already a pretty interesting result - the folks at [SFI](https://www.santafe.edu/) would probably love it.
But it gets even better.
We can do the same thing with the neighborhood 1 - neighborhood 3 covariance matrix. The first 10 singular values are:
```
array([7.50028046e+11, 5.87974398e+10, 2.50629095e+08, 2.23881004e+03,
1.46035030e+01, 2.52752074e+00, 1.01656991e-01, 1.79916658e-02,
5.89831843e-04, 3.52074803e-04])
```
… and by the same criteria as before, we see at most 8 nonzero singular values, and probably more like 3. Now, we can *compare* the neighborhood-1-singular-vectors from our two SVDs. Simply computing the correlations between each pair of nonzero singular vectors from the two decompositions yields:
Key thing to notice here: the first two singular vectors are near-perfectly correlated! (Both correlations are around .996.) The correlations among the next few are also large, after which it drops off into noise. (I did say that everything after the first three singular components was probably mostly numerical noise.)
In English: not only can a three-variable summary of neighborhood 1 probably contain all the info relevant to either of the other two neighborhoods, but **two of the three summary variables are the same**.
This is *exactly* the sort of thing the natural abstraction hypothesis predicts: information relevant far away fits into a compact summary - the “abstract summary” - and roughly-the-same abstract summary is relevant even from different “vantage points”. Or, to put it differently, the abstraction is reasonably robust to changes in which “far away” neighborhood we pick.
Our 2D-embedded linear-Gaussian world abstracts well.
Of course, this was just an example from one run of the script, but the results are usually pretty similar - sometimes the summary has 2 or 4 or 5 dimensions rather than 3, sometimes more or fewer summary-dimensions align, but these results are qualitatively representative. The main case where it fails is when the RNG happens to pick two neighborhoods which are very close or overlap.
Is This Just The Markov Blanket?
--------------------------------
Notice that our “neighborhoods” of variables include some nodes strictly in the interior of the neighborhood. The variables on the boundary of the neighborhood should form a Markov blanket - i.e. they are themselves an abstract summary for the whole neighborhood, and of course they’re the same regardless of which second neighborhood we pick. So perhaps our results are really just finding the variables on the neighborhood boundary?
The results above already suggest that this is not the case: we can see at a glance that the dimension of the summary is quite a bit smaller than the number of variables on the boundary. But let’s run another test: we’ll fix the point at the “center” of the neighborhood, then expand the neighborhood’s “radius” (in terms of undirected steps in the graph), and see how both the boundary size and the summary size grow.
Here are results from one typical run (number of boundary variables in blue, number of summary variables in red):
As we expand the neighborhood, the number of boundary-variables grows, but the number of summary variables stays flat. So there’s definitely more going on here than just the neighborhood’s Markov blanket (aka boundary variables) acting as an abstract summary.
What About…
-----------
The above toy model made two choices which one could imagine not generalizing, even in other sparse linear-Gaussian systems.
First, the graph structure. The random graphs most often used in mathematics are Erdos-Renyi (ER) random graphs. “Information relevant far away” doesn’t work very well on these, because nodes are almost never far apart. The number of steps between the two most-distant nodes in such graphs is logarithmic. So I wouldn’t expect environments with such structure to abstract well. On the other hand, I would usually expect such environments to be a poor model for systems in our physical world - the constraints of physical space-time make it rather difficult to have large numbers of variables all interacting with only logarithmic causal-distances between them. In practice, there are usually mediating variables with local interactions. Even the real-world systems which most closely resemble ER random graphs, like social networks or biochemical regulatory networks, tend to have much more “localized” connections than a true ER random graph - e.g. clusters in social networks or modules in biochemical networks.
(That said, I did try running the above experiments with ER graphs, and they work *surprisingly* well, considering that the neighborhoods are only a few steps apart on the graph. Summary sizes are more like 10 or 12 variables rather than 2 or 4, and alignment of the summary-variables is hit-and-miss.)
Second, the weights. There’s a phase shift phenomenon here: if the weights are sufficiently low, then noise wipes out all information over a distance. If the weights are sufficiently high, then the whole system ends up with one giant principal component. Both of these abstract very well, but in a boring-and-obvious way. The weights used in these experiments were chosen to be right at the boundary between these two behaviors, where more interesting phenomena could potentially take place. If the system abstracts well right at the critical point, then it should abstract well everywhere.
So What Could We Do With This?
------------------------------
I mainly intend to use this sort of model to look for theoretical insights into abstraction which will generalize beyond the linear-Gaussian case. Major goals include data structures for representing whole sets of abstractions on one environment, as well as general conditions under which the abstractions are learned by a model trained on the environment.
But there are potential direct use-cases for this kind of technique, other than looking for hints at more-general theory.
One example would be automatic discovery of abstractions in physical systems. As long as noise is small and the system is non-chaotic (so noise stays small), a sparse linear-Gaussian model should work well. (In practice, this mainly means the system needs to have some kind of friction in most degrees of freedom.) In this case, I’d expect the method used here to work directly: just compute covariances between far-apart neighborhoods of variables, and those covariances will likely be low-rank.
In principle, this method could also be directly applied to machine learning systems, although the range of applicability would be fairly narrow. It would only apply to systems where linear approximations work very well (or maybe quadratic approximations, which give a linear approximation of the gradient). And ideally the random variables would be Gaussian. And it should have a fairly large and deep computational graph, so most neighborhoods of variables are not too close together. If one just happened to have an ML model which met those criteria, we’d expect these results to carry over directly: compute the covariance matrix between far-apart neighborhoods of the random variables in the model, and those matrices should be low-rank. That sounds like a potentially-very-powerful transparency tool... *if* one had a model which fit the criteria. Hypothetically.
Summary
-------
Build a random linear-Gaussian Bayes net with a random planar graph structure (in this case generated from a Delaunay mesh). Pick two far-apart neighborhoods of variables in this net, and use an SVD to compute the rank of their covariance matrix. Empirically, we find rank <10 even with 110-variable neighborhoods in a 50k-node network. In other words: a very small number of variables typically suffices to summarize all the information from a neighborhood of variables which is relevant far away. The system abstracts well.
Furthermore: the largest singular vectors are the *same* even for different choices of the second (far-away) neighborhood. Not only does a small summary suffice, but approximately the same small summary suffices. The abstractions are robust to changes in our “far away” reference neighborhood. |
dcec2fcc-878f-4af6-8744-5bfad35ada6e | trentmkelly/LessWrong-43k | LessWrong | Appropriately Gray Products
Epistemic status:
* I've been in the startup world for a while, have read almost all of Paul Graham's essays, have read The Lean Startup cover-to-cover, etc. However, it's possible that I am misunderstanding/misrepresenting the idea of an MVP here. If so, I apologize and would like to be corrected.
* As for my opinions, I feel solid about them. Not too, too confident though. To justify that amount of confidence, I would want to have had stress tested my ideas by having other people critique them. That one of the goals I have in writing this, actually.
* This is a "starting the conversation" type of post. Not an "I've spent a ton of time researching, discussing with others, and iterating, and now I am presenting to you a finished product" type of post.
----------------------------------------
Here is how I understand the idea of a minimum viable product (MVP).
Starting out, you have some hypothesis that you want to test. Eg. that people will want to buy pet food online. How would you go about testing this hypothesis?
You could go to town and spend 18 months building out a beautiful website that has pixel perfect designs with an accompanying iOS and Android app, but that is a little bit excessive. To test your hypothesis, doing all of that isn't really necessary.
On the other hand, you could just throw together a Google Doc with products, pictures and prices and tell people to email you if they want to place an order. But I'd argue that this wouldn't be enough. You wouldn't get a good enough test of your hypothesis. People who otherwise would be willing to buy pet food online might not place orders because a Google Doc like this feels sketchy. In other words, the risk of a false negative is too high.
With that context, let me explain what an MVP says. An MVP says to keep "moving to the right" until you have a hypothesis test that is "viable", and then once you hit that point, stop. You don't want to be to the left of the line because that would mean you ca |
b20fdf6d-6a93-4a1c-88c4-2324c036a7a7 | trentmkelly/LessWrong-43k | LessWrong | Why Eliezer Yudkowsky receives more upvotes than others
One of the reasons for why Yudkowsky is being drastically upvoted is of course that he's often, fasten your seatbelts, brilliantly right (whoever reads my comments knows that I am not really the most frenetic believer, so I think I can say this without sounding cultish). But others are too, so is Less Wrong a cult? Nah! There is a simple explanation for this phenomenon:
As you can see, there are already 13 people who subscribe to his Less Wrong feed via Google Reader. And there are many other ways to subscribe to a RSS feed (which is not the only way to follow his mental outpourings anyway), so the number of people who follow every post and comment is likely much higher.
That's why most of his comments receive more upvotes than other comments. It is not because he is a cult leader, it's just that his comments are read by many more people than the average comment on Less Wrong. There are of course other causes as well, but this seems to explain a fair chunk of the effect.
Also consider that I'm often upvoted (with a current Karma score of 1959) and I do not keep quiet regarding my doubts about some topics directly related to Yudkowsky and the SIAI. How could this happen if Less Wrong was an echo chamber?
I just wanted to let you know, because I have been wondering about it in the past. |
5958724c-21e5-4e9f-bea9-5c17e3b724a3 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Saskatoon - Rhetorical Fallacies!
Discussion article for the meetup : Saskatoon - Rhetorical Fallacies!
WHEN: 26 October 2013 01:00:00PM (-0600)
WHERE: 2318 8th St E, Saskatoon, SK
Another Saskatoon meetup at the same place and time as the last one: Broadway Roaster on 8th street (not on broadway!) at 1:00 in the afternoon.
We'll be learning about 7 common rhetorical fallacies, why they are wrong and how to identify them. This is going to be based off of the rhetorical fallacy lesson from clearerthinking.org. I'd suggest going through it. It is free, takes about 45 minutes and you can see if you beat my score of 35 points!
More info here: http://www.meetup.com/Saskatoon-Rationalists/
Hope to see you there!
Discussion article for the meetup : Saskatoon - Rhetorical Fallacies! |
2e65f151-1cac-4761-9ff1-52003f5d8cd7 | trentmkelly/LessWrong-43k | LessWrong | Effortless Technique
> "All my life I have been intensely repelled by the idea of 'making an effort'. I hate this idea today as much as I did as a child. I don't know why I hate it so much; I just do."
> -- Raymond Smullyan, The Tao Is Silent
In the Hollywood version of rationality - or even the Traditional rationality that was passed down from supervisor to grad student in ancient days before Bayesianism - rationality is a great strain, a great effort, a continuous battle to coerce your mind into a desired shape. Spock, the archetype of Hollywood's concept of rationality, represses all his emotions.
And this great effort, they conceive, is virtue unto a rationalist. The more effort you expend on forcing yourself into the mold, the better the rationalist you must be. It's like working extra hard at your job, as demanded by the Protestant work-ethic. If the one works long hours - sweating, getting ulcers - surely the one must be worthy of praise?
This, I think, is an instance of a Lost Purpose. People see that successful folk must sometimes make an effort, and so they conclude that effort of itself is virtuous whether or not it succeeds.
I am reminded of an AI koan from AGI '06, where the discussion turned (as it often does) to defining "intelligence". A medium-prominent AI researcher suggested that an agent's "intelligence" could be measured in the agent's processing cycles per second, bits of memory, and bits of sensory bandwidth. To which I replied with a quote from Dijkstra:
> "If we wish to count lines of code, we should not regard them as 'lines produced' but as 'lines spent': the current conventional wisdom is so foolish as to book that count on the wrong side of the ledger."
Surely (I said), an agent is less intelligent if it uses more memory, processing power, and sensory bandwidth to accomplish the same task?
This reply was due, in no small part, to my having read Raymond Smullyan's The Tao Is Silent at the age of sixteen. Raymond Smullyan is a mathe |
9cb24af1-71ec-4f7f-937f-cb67687e319a | trentmkelly/LessWrong-43k | LessWrong | What is the most anti-altruistic way to spend a million dollars?
Edit: The purpose of this question is not to make the world worse, but to see whether we actually have concrete ideas of what would, and my guess is that most of us don't, not in a really concrete way. From the downvotes I'm wondering if everyone else is thinking way darker directions than I am. If so please share.
There is a lot of discussion here about effective altruism. Organizations like GiveWell with donations, using criterion like quality-life-years-saved-per-dollar. People distinguish warm-and-fuzzy giving from the most effective use of dollars from various utilitarian perspectives.
But I want to ask a different question: What would effective anti-altruism be?
To make it more concrete:
I am an eccentric multimillionaire, proposing a contest to all of you, who will for the purposes of this exercise play greedy and callous, yet honest and efficient, contest entrants.
Whoever can propose the most negative possible use for my money, in the sense that it causes the greatest amount of global misery, (feel free to argue for your own interpretation of the details of what this means) will receive $1 million to carry out his or her proposal and $1 million to keep for him or herself to with as desired.
A few rules:
1) Everything must be 100% legal in whatever jurisdiction you propose. Edit: People had trouble with the old phrasing, so I'll add that it should not only be legal in the letter of the law, but also in some reasonable interpretation of the spirit of the law.
1a) In fact, I encourage you to think of things that aren't merely legal but that would also be legal under whatever your favorite hypothetical laws are. Maybe that means non-coercive, non-violent, or something else in that vein.
2) This money may be used as seed funding for a non-profit or for-profit anti-altruistic venture, but I will take into account both the risk and the marginal impact of only the first million dollars.
3) Risk and plausibility are factors just as they would be in any in |
3e048283-419d-4cdc-9987-24c33c7ba67a | StampyAI/alignment-research-dataset/special_docs | Other | AI and the Future of Cyber Competition
JANUARY 2021
AI and the Future
of Cyber Competition
CSET Issue Brief
AUTHOR
Wyatt Hoffm an
Center for Security and Emerging Technology | 1 Table of Contents Executive Summary ............................................................................................... 2 Introduction ............................................................................................................ 5 Promise and Pitfalls of Artificial Intelligence for Cybersecurity .......................... 8 Security vulnerabilities of machine learning ................................................ 10 The Imperatives of Offense ................................................................................ 12 Attacking machine learning .......................................................................... 12 The attacker’s predicament ........................................................................... 13 The Imperatives of Defense ................................................................................ 15 The perpetual problem of machine learning robustness ............................ 15 The defender’s predicament ......................................................................... 17 Artificial Intelligence and Cyber Stability ......................................................... 19 Mitigating scenarios ...................................................................................... 24 Implications for policy ................................................................................... 25 Conclusion .......................................................................................................... 27 Acknowledgments .............................................................................................. 28 Endnotes .............................................................................................................. 29
Center for Security and Emerging Technology | 2 Executive Summary As artificial intelligence begins to transform cybersecurity, the pressure to adapt may put competing states on a collision course. Recent advances in machine learning techniques could enable groundbreaking capabilities in the future, including defenses that automatically interdict attackers and reshape networks to mitigate offensive operations. Yet even the most robust machine learning cyber defenses could have potentially fatal flaws that attackers can exploit. Rather than end the cat-and-mouse game between cyber attackers and defenders, machine learning may usher in a dangerous new chapter. Could embracing machine learning systems for cyber defense actually exacerbate the challenges and risks of cyber competition? This study aims to demonstrate the possibility that machine learning could shape cyber operations in ways that drive more aggressive and destabilizing engagements between states. While this forecast is necessarily speculative, its purpose is practical: to anticipate how adversaries might adapt their tactics and strategies, and to determine what challenges might emerge for defenders. It derives from existing research demonstrating the challenges machine learning faces in dynamic environments with adaptive adversaries. This study envisions a possible future in which cyber engagements among top-tier actors come to revolve around efforts to target attack vectors unique to machine learning systems or, conversely, defend against attempts to do so. These attack vectors stem from flaws in machine learning systems that can render them susceptible to deception and manipulation. These flaws emerge because of how machine learning systems “think,” and unlike traditional software vulnerabilities, they cannot simply be patched. This dynamic leads to two propositions for how these attack vectors could shape cyber operations. The first proposition concerns offense: Attackers may need to intrude deep into target networks well in advance of an attack in order to circumvent or defeat machine learning defenses. Crafting an attack that can reliably deceive a machine learning system requires knowing a specific flaw in how the system thinks. But discovering such a flaw may be difficult if the system is not widely exposed or publicly available. To reach a hardened target, an attacker may try to compromise the system during development. An attacker with sufficient access could reverse-engineer a system during its development to discover a flaw or even create one by sabotaging the process. This opportunity to gain intelligence on an adversary’s defenses creates more value in intruding into adversary computer networks well in advance of any planned attack.
Center for Security and Emerging Technology | 3 The second proposition concerns defense: Guarding against deceptive attacks may demand constant efforts to gain advanced knowledge of attackers’ capabilities. Because machine learning systems cannot simply be patched, they must be able to adapt to defend against deceptive attacks. Yet researchers have found that adaptations to defend against one form of deception are vulnerable to another form of deception. No defense has been found that can make a machine learning system robust to all possible attacks—and it is possible none will be found. Consequently, machine learning systems that adapt to better defend against one form of attack may be at constant risk of becoming vulnerable to another. In the face of an imminent threat by an adversary, the best defense may be to intrude into the adversary’s networks and gain information to harden the defense against their specific capabilities. Together these two propositions suggest machine learning could amplify the most destabilizing dynamics already present in cyber competition. Whether attacking or defending, at the top tier of operations, machine learning attack vectors may create challenges best resolved by intruding into a competitor’s networks to acquire information in advance of an engagement. This would add to existing pressures on states to hack into their adversaries’ networks to create offensive options and protect critical systems against adversaries’ own capabilities. Yet the target of an intrusion may view the intrusion as an even greater threat—regardless of motive—if it could reveal information that compromised machine learning defenses. The already blurred line between offensive and defensive cyber operations may fade further. In a crisis, the potential for cyber operations to accelerate the path to conflict may rise. In peacetime, machine learning may fuel the steady escalation of cyber competition. Adversaries may adapt by targeting machine learning itself, including: ● Compromising supply chains or training processes to insert backdoors into machine learning systems that expose a potentially wide swath of applications to possible attacks. ● Poisoning training data, such as open source malware repositories, to degrade cybersecurity applications. ● Unleashing risky capabilities to circumvent defenses, such as malware with greater degrees of autonomy. ● Targeting defenders’ trust in machine learning systems, such as by inducing systems to generate “false positives” by mislabeling legitimate files as malware.
Center for Security and Emerging Technology | 4 For the United States and its allies, harnessing machine learning for cybersecurity depends on anticipating and preparing for these potential changes to the threat landscape. If cyber defense increasingly relies on inherently flawed machine learning systems, frameworks and metrics will be needed to inform risk-based decisions about where and how to employ them. Securing the machine learning supply chain will demand collective governmental and private sector efforts. Finally, the United States and its allies must exercise caution in the conduct of their offensive operations and communicate with adversaries to clarify intentions and avoid escalation.
Center for Security and Emerging Technology | 5 Introduction States seeking competitive advantage will likely turn to artificial intelligence to gain an edge in cyber conflict. Cybersecurity ranks high among priority applications for those leading AI development.1 China and Russia see in AI the potential for decisive strategic advantage.2 Military planners in the United States envision systems capable of automatically conducting offensive and defensive cyber operations.3 On the precipice of a potential collision between AI competition and cyber conflict, there is still little sense of the potential implications for security and stability. AI promises to augment and automate cybersecurity functions. Network defenders have already begun to reap the benefits of proven machine learning methods for the data-driven problems they routinely face.4 Even more tantalizing is the speculative prospect of harnessing for cybersecurity the cutting-edge machine learning techniques that yielded “superhuman” performance at chess and the Chinese board game Go. Yet the machine learning capabilities fueling these applications are no panacea. These systems often suffer from inherent flaws. Unlike traditional software vulnerabilities, these flaws emerge because of how these systems make inferences from data—or, more simply, how they “think.” These flaws can lead even highly robust systems to fail catastrophically in the face of unforeseen circumstances. In the race between researchers developing ways to safeguard these systems and those seeking to break them, the attackers appear to be winning. What will happen when these powerful yet flawed machine learning capabilities enter into the dynamic, adversarial context of cyber competition? Machine learning can help mitigate traditional cyber attack vectors, but it also creates new ones that target machine learning itself. Attackers will systematically try to break these systems. A growing body of technical research explores machine learning attack vectors and prospective defenses. Yet there has been little effort to analyze how these changes at a technical level might impact cyber operations and, in turn, their strategic dynamics. This study approaches this problem by exploring a possible worst-case scenario: machine learning could amplify the most destabilizing dynamics already present in cyber competition. The purpose is not to lay out a case against harnessing machine learning for cybersecurity. Precisely because these capabilities could become crucial to cyber defense, the aim here is to provoke thinking on how to proactively manage the geopolitical implications of persistent technical flaws.
Center for Security and Emerging Technology | 6 This study explores how the attack vectors unique to machine learning might change how states hack each other’s critical networks and defend their own. The unique vulnerabilities of these systems may create problems for both offense and defense best resolved by intruding into adversaries’ systems in advance of an engagement. For offense, this arises from the potential need for exquisite intelligence on, or even direct access to, a machine learning system to reliably defeat it. For defense, this arises from the need for advanced knowledge of a specific attack methodology to ensure a defense’s viability against it. The combination of these offensive and defensive imperatives could exacerbate the escalation risks of cyber engagements. States would have even stronger incentives to intrude into one another’s systems to maintain offensive options (for contingencies such as armed conflict or strategic deterrence) and to ensure the viability of their own defenses. Yet it may be even harder to differentiate cyber espionage from intrusions laying the groundwork for an attack; the target of an intrusion may assume that it is preparation for an imminent attack, or that it will at the very least enable offensive options. As adversaries struggle to gain an edge over one another, the line between offense and defense—tenuous as it already is in cyber operations—fades. This dynamic may fuel the steady drumbeat of cyber competition in peacetime. In a crisis, the potential for misinterpretation of a cyber operation to trigger escalation may rise. This forecast rests on two core assumptions that must be addressed at the outset. These are certainly debatable but the aim is to analyze their implications should they hold, not assess how likely they are to do so. The first is that machine learning could plausibly deliver on the promise of sophisticated, automated cyber defenses at scale. That is, the significant technical and practical hurdles (e.g., demands for high quality data and computing power, as well as organizational challenges to implementation) will not prove insurmountable at least for top-tier actors such as China and the United States. This study begins with a survey of applications in various stages of development to demonstrate their plausibility. But it makes no attempt to assess the current state of play with deployed machine learning cybersecurity applications or the likelihood of realizing them in the near term.\* \* For a more thorough survey of existing applications and near-term prospects for machine learning in cybersecurity see Micah Musser and Ashton Garriott, “Machine Learning and Cybersecurity: Hype and Reality” (Center for Security and Emerging Technology, forthcoming).
Center for Security and Emerging Technology | 7 The second assumption is that insights from existing research on machine learning attack vectors will hold at least for the prevailing machine learning methods and applications discussed here. This study draws extensively on research demonstrating the attack vectors targeting machine learning and what these vectors reveal about the potential limits of the robustness of machine learning systems. It makes no assumptions about yet unseen innovations in machine learning techniques or offensive or defensive measures that might fundamentally change the trajectory. This study begins with a brief overview of machine learning applications for cybersecurity, including their prospective defensive benefits and inherent flaws. It then examines two propositions for how these technical changes to the cybersecurity landscape may, in turn, shape offensive and defensive cyber operations. Specifically, machine learning attack vectors could create predicaments that incentivize intrusions into adversaries’ networks, whether to create offensive options or shore up defenses. This study continues on to explore how the combination of these two propositions could fuel the steady intensification of cyber competition and increase the risks of misperception and escalation in cyber engagements.
Center for Security and Emerging Technology | 8 Promise and Pitfalls of Artificial Intelligence for Cybersecurity Machine learning lies at the core of the emerging and maturing cybersecurity applications discussed throughout this paper. Described as an approach to, or subfield of, AI, machine learning has fueled recent milestones in tasks ranging from image recognition to speech generation to autonomous driving. Machine learning systems essentially adapt themselves to solve a given problem.5 This process often starts with a blank slate in the form of a neural network. The system’s developers feed a dataset to the neural network and an algorithm shapes the network’s structure to adapt to the patterns within this data. For example, a system for analyzing malware will learn to accurately identify a file as “malware” or “benign” and associate each classification with particular patterns. Eventually the network develops a generalized model of what malware “looks like.” High quality training data, effective training algorithms, and substantial computing power comprise the critical inputs to this process. The resulting machine learning model, ideally, detects not only known malware but yet unseen variants. Advancements in machine learning techniques reduce the need for human experts to structure data.\* Rather than relying on an expert to tell the model what key features of malware to look for, the model discovers on its own how to classify malware. As a result, it may find ways of identifying malware more effective at coping with attackers’ attempts at obfuscation, such as “metamorphic” malware that rewrites parts of its code as it propagates.6 Intrusion detection—finding an adversary’s illicit presence in a friendly computer network—may benefit similarly from machine learning. Existing intrusion detection systems already look for red flags, such as a computer becoming active in the middle of the night or a user attempting to access files unrelated to their work. Yet defenders struggle to sort through the vast data generated by network activity in large enterprises, allowing attackers to hide in the noise. Machine learning systems can turn this data into a major \* Deep learning architectures are particularly promising in this respect. For example, one approach translates a piece of malware into an image by converting code to pixels in order to utilize advances in deep learning-based image classification as a means of classifying the underlying code as benign or malicious. See Daniel Gibert, Carles Mateu, and Jordi Planes, “The Rise of Machine Learning for Detection and Classification of Malware: Research Developments, Trends and Challenges,” Journal of Network and Computer Applications 153 (March 1, 2020): 102526.
Center for Security and Emerging Technology | 9 advantage. By fusing information from a wider and more diverse range of sensors throughout the environment, they create a baseline of normal network activity against which even slight deviations can be detected.7 AI and machine learning have quickly become buzzwords in the cybersecurity industry. This makes it difficult to assess the extent to which these capabilities are actually relied upon or are invoked for marketing purposes. Cybersecurity vendors commonly claim to leverage machine learning.\* For example, as CrowdStrike defends its customers’ devices and networks, it rakes in data on around 250 billion events daily and feeds the data to machine learning models to predict new kinds of attacks.8 Darktrace states that it employs multiple machine learing methods in its “Enterprise Immune System,” empowering systems that can automatically mitigate attacks.9 Machine learning has also been harnessed to test software for vulnerabilities, detect spam and spear-phishing attacks, and identify suspicious behavior and insider threats.10 In general, machine learning systems appear to be deployed mainly for relatively narrow tasks in support of human network defenders.11 Traditional machine learning methods relying on large training datasets may not suffice for a system that performs more complex tasks requiring sequences of actions, each dependent upon the outcome of the last. Such a system needs to learn more like a human—through experimentation and trial-and-error. This is the essence of reinforcement learning. Instead of being fed training data, a reinforcement learning agent interacts with a simulated environment and is rewarded for action that advances its objective. It gradually learns sets of moves, or “policies,” to guide its action. The process can yield stunning results, such as the victory by AlphaGo, a reinforcement learning system developed by DeepMind, over Lee Sedol, the world champion in the incredibly complex game of Go.12 If reinforcement learning can master chess and Go, it might unlock future cyber defenses capable of discovering and automatically executing moves and strategies in the “game” against cyber attackers. Cyber defenders have a home field advantage.13 They can change the configuration of networks to interfere with an attacker or deploy decoy systems such as “honeypots” that lure attackers in and lead them to reveal capabilities. However, setting up \* According to one survey of U.S., UK, and German businesses 82 percent of respondents stated their company employed a cybersecurity product utilizing machine learning in some form. See Ondrej Kubovič, “Machine-Learning Era in Cybersecurity: A Step Toward a Safer World or the Brink of Chaos?” ESET, February 2019, https://www.eset.com/fileadmin/ESET/US/download/ESETus-Machine-Learning-Era-in-Cybersecurity-Whitepaper-WEB.pdf.
Center for Security and Emerging Technology | 10 honeypots and reconfiguring networks are technically demanding tasks and, to be effective, require the ability to anticipate an attacker’s moves and adapt on the fly.14 While still largely confined to academic research, and thus more speculative, pioneering applications of reinforcement learning may produce systems capable of these feats.15 Reinforcement learning agents could learn optimal strategies for reconfiguring networks and mitigating attacks, rapidly analyze an attacker’s moves and select and execute actions, such as isolating or patching infected nodes and deploying honeypots. At a minimum, they could present attackers with a constantly moving target, introducing uncertainty and increasing the complexity required for offensive operations.16 Machine learning could plausibly deliver on the promise of cyber defenses that adapt to novel threats and automatically engage attackers. These potentially game-changing applications are the focus of this study, even though the most significant near-term gains for cybersecurity may be found in automating the more “mundane” aspects of cybersecurity. The more speculative capabilities may not be realized in the near term, but given their potential to transform cyber operations it is worth exploring their implications. Security vulnerabilities of machine learning As promising as they are, most machine learning cyber capabilities have yet to face the most important test: systematic attempts by attackers to break them once deployed. Machine learning can fail catastrophically under certain conditions.17 Evidence for this includes “adversarial examples”: manipulated inputs (often images that have been subtly altered) created by researchers to trick machine learning models. Seemingly imperceptible changes to an image of a turtle can cause a model that otherwise classifies it with perfect accuracy to mistake it for a rifle.18 Similar adversarial techniques can cause reinforcement learning systems to malfunction.19 Adversarial examples reveal a problem inherent to machine learning, not just deficiencies in specific systems. Every model rests on assumptions about the data to make decisions—assumptions, for instance, about what malware “looks like.” If an input violates those assumptions it will fool the model (and often a successful deception fools other models trained for the same task).20 Flawed training methods or data can create vulnerabilities. But models can also become vulnerable when the conditions in which they are deployed change in ways that violate assumptions learned in training. The model’s predictions will no longer be accurate—a problem referred to as “concept drift.”21 Even slight deviations from training conditions can dramatically degrade performance.
Center for Security and Emerging Technology | 11 This poses a constant problem for machine learning applications in dynamic, adversarial contexts like cybersecurity.22 For machine learning cyber defenses to be viable, they may have to learn and evolve not just during training, but in deployment.23 Systems will have to keep up with a constantly changing cybersecurity landscape. For instance, an intrusion detection system modeling “normal” network activity must constantly revise this model as legitimate and malicious activity changes. The system might generate new training data by observing the behavior of devices connected to the network, using this data to continuously update and refine its model to better predict future behavior. Innovative machine learning techniques aim to create systems capable of better contending with adaptive adversaries in dynamic environments. These techniques harness competition to drive evolution. For instance, Kelly et al. co-evolve defenses that automatically reconfigure networks to catch the attackers with offensive agents seeking to evade detection.24 Developers may pit a reinforcement learning agent against an adversarial agent whose objective is to thwart it.25 These methods attempt to simulate an “arms race” between attackers and defenders to produce models that better anticipate and preempt attacker moves in the real world.26 All of this sets the stage for a potential transformation in the cat-and-mouse game between cyber attackers and defenders. The future cybersecurity playing field may feature defenses that evolve automatically through engagements, but such defenses inevitably create new attack vectors that are difficult to safeguard. The next two sections explore how attackers and defenders alike might adapt to these changing technical conditions, setting the foundation to examine the geopolitical implications that follow.
Center for Security and Emerging Technology | 12 The Imperatives of Offense If improved machine learning defenses offer significant benefits to defenders, they will introduce significant new hurdles into the planning and execution of offensive cyber operations. Offensive operations often require careful planning and preparation of the target environment. The presence of sophisticated machine learning defenses may force attackers to shift their efforts toward targeting the underlying machine learning models themselves. But hacking machine learning presents its own unique set of problems. The core challenge for attackers will be figuring out how to reliably manipulate or circumvent these systems. Attacking machine learning Attackers tend to follow the path of least resistance. If possible, they will try to avoid machine learning defenses entirely, including by targeting “traditional” attack vectors, such as acquiring credentials via spear-phishing. Avoidance, however, may not always be an option. An attacker may attempt to evade the defensive system by exploiting a weakness in the model. Researchers at security firm Skylight Cyber demonstrated how to do so against Cylance’s leading machine learning-based antivirus product.27 Using publicly accessible information, they reverse-engineered the model to discover how it classified files. In the process, they discovered a bias in the model; it strongly associated certain sequences of characters with benign files. A file that otherwise appeared highly suspicious would still be classified as benign if it contained one of the character sequences. The Skylight researchers discovered, in their words, a “universal bypass”—characters that they could attach to almost any piece of malware to disguise it as a benign file.\* The researchers found that applying their bypass to a sample of 384 malicious files resulted in the machine learning system classifying 84 percent as “benign,” often with high confidence.28 Attackers will not always be so lucky as to discover a bypass as readily exploitable as in the Cylance case. They could sabotage a model to similar effect. Injecting bad samples into a training dataset (e.g. malware labeled as “benign”) can “poison” a model. Even an unsophisticated poisoning attack could dramatically reduce the model’s performance.29 More insidiously, an \* Cylance disputed the characterization as a “universal bypass” and claimed to have fixed the flaw shortly after being made aware by Skylight. See “Resolution for BlackBerry Cylance Bypass,” BlackBerry ThreatVector Blog, July 21, 2019, https://blogs.blackberry.com/en/2019/07/resolution-for-blackberry-cylance-bypass.
Center for Security and Emerging Technology | 13 attacker could poison a model so that it reacts to specific inputs in a way favorable to the attacker—inserting a “backdoor” into the model. In one demonstration, researchers created a “watermark” in the form of a specific set of features in a file that functioned similar to the bypass discovered by Skylight. By tampering with just one percent of the training data, they could induce a model to misclassify malicious files containing the watermark as benign with a 97 percent success rate.30 While these examples describe attacks on classification systems, reinforcement learning agents engaged in more complex tasks have similarly proven susceptible to evasion and sabotage.31 For example, an attacker could poison a defensive system that automatically reconfigures networks so that it responds poorly in specific circumstances; the attacker might trick the system into connecting an infected node to others in the network, rather than isolating it.32 The attacker’s predicament The feasibility of evading or poisoning a machine learning system will inevitably depend on the context. It’s one thing to demonstrate attacks on machine learning in experimental settings, but it’s another to execute them in the real world against a competent defender. In the Cylance case, the attackers benefited from insights into the inner workings of the model. States seeking to create and sustain offensive options may face strategic targets that are not so widely exposed. The difficulty of conducting attacks on machine learning systems under realistic constraints may pressure states to intrude into adversaries’ networks to begin laying the groundwork for attacks as early as possible. This pressure stems from the necessity intrusions play in enabling the kinds of attacks described above: (1) Acquiring information to craft more reliable and effective evasion attacks against machine learning systems: As Goodfellow et al. observe, the greater the attacker’s “box knowledge”—knowledge of the target model parameters, architecture, training data and methods—the easier it is to construct an attack that defeats the system.33 Under “white box” conditions, where the attacker has complete knowledge, crafting an attack is a relatively straightforward matter of optimizing the features of malware (or other inputs) to exploit the model’s assumptions.\* \* Researchers have naturally found greater success evading antivirus systems and attacking reinforcement learning policies with white-box attacks than with black-box attacks. See, for instance, Hyrum S. Anderson et al., “Learning to Evade Static PE Machine Learning Malware
Center for Security and Emerging Technology | 14 “Black box” attacks, where the attacker has little to no knowledge of the target model, are possible, but require more guesswork. The attacker may engineer an attack against a substitute for the target model in the hopes that if it fools the substitute, it will fool the target. But this depends on how closely the substitute matches the target.34 Demonstrations of black-box attacks often leverage publicly available details or the ability to repeatedly probe a target model in order to derive information on how it works. An attacker might buy a commercial service to gain insights into a model, allowing greater flexibility to craft attacks. In top-tier cyber competition, however, an attacker may not enjoy these advantages. If the target model is not widely exposed, attempts to probe it may tip off the defender. And gaining information on some types of defenses, like those that reconfigure networks, would require intruding into the network. Moreover, future security measures may prevent deployed machine learning systems from “leaking” useful information to an attacker attempting to probe them.35 The best way to acquire box knowledge, then, may be to gain access to a training environment and steal training data or even a trained model. (2) Compromising systems to enable future exploitation: It is possible to undermine a deployed model, for instance interacting with an intrusion detection system to “normalize” an intruder’s presence to it.36 But competent defenders will be alert to the possibility. The development process may present a softer target.37 Rather than a model developed from scratch, many applications take existing pre-trained models and tailor them for specific tasks through additional training and fine-tuning in a process known as transfer learning. A backdoor inserted into the pre-trained model can make its way into subsequent models derived from it.38 This opens up new attack vectors. For example, compromising an open source project, code repository, or a commercial contractor assisting with the development of cybersecurity applications may allow an attacker to insert vulnerabilities deep into systems that make their way into more tightly-controlled training environments. Targeting the development process has the added benefit of scalability: inserting a backdoor into one model may facilitate access to a wide swath of subsequent targets. A transfer learning service supporting diverse commercial, military, or other national security-relevant applications would be a tempting target. Models via Reinforcement Learning,” arXiv [cs.CR] (January 26, 2018), arXiv, http://arxiv.org/abs/1801.08917; Huang et al., “Adversarial Attacks on Neural Network Policies.”
Center for Security and Emerging Technology | 15 The Imperatives of Defense As attackers adapt to the deployment of machine learning, the success or failure of cyber defenses may hinge on the security of machine learning models against deception and manipulation. Yet it has proven difficult to create machine learning systems that are truly robust—that is, systems that can contend with attackers that adapt their tactics to try and defeat them. Innovative defenses against the kinds of attacks described above have emerged, but are routinely broken. Some experts question whether progress toward truly robust machine learning has been illusory.39 The core challenge for defenders may be safeguarding systems with inherent flaws baked in. The perpetual problem of machine learning robustness When a vulnerability is discovered, a machine learning model cannot simply be patched like traditional software. Instead, the developer must retrain the model using adversarial examples or certain training procedures designed to make the model more robust to a particular set of deceptive inputs. However, adjusting the model may simultaneously make it more robust to one set of deceptions but more susceptible to others. Two prominent machine learning security experts, Ian Goodfellow and Nicolas Papernot, thus characterize existing defensive measures as “playing a game of whack-a-mole: they close some vulnerabilities, but leave others open.”40 Such were the findings of Tramer et al., who systematically defeated 13 defenses shown to be effective against adaptive attackers.41 A similar phenomenon has been observed with reinforcement learning agents; rather than becoming generally robust, those trained against an adaptive adversary in simulated games tend to “overfit” to the adversary. In other words, their adaptations to deal with the regular opponent can leave them vulnerable to a novel attack.42 The ease with which defenses are broken may simply reflect the nascent state of machine learning security. But it suggests a more concerning possibility: no defense will be robust to all possible attacks. As David Brumley puts it: “for any ML algorithm, an adversary can likely create [an attack] that violates assumptions and therefore the ML algorithm performs poorly.”43 Unlike software security, which is, at least in theory, a “linear” process of improvement as the developer tests, patches, and repeats, machine learning may present a perpetual security problem. The system can be hardened to any known attack but may always be vulnerable to a possible novel attack. These observations raise two questions regarding the potential limits on machine learning robustness:
Center for Security and Emerging Technology | 16 First, how much of a problem do machine learning’s flaws pose for the defender? With sufficiently comprehensive training data to accurately model threats, perhaps the risk of a novel attack defeating the system would be negligible. However, cybersecurity presents a uniquely difficult deployment context: Threats continuously evolve, so a deployed system must constantly take in new data to adapt. But if instead of becoming generally robust, machine learning defenses are just playing whack-a-mole, then there may always be an attack that breaks them. Testing systems to try and discover every flaw may prove futile because of the vast range of possible moves the attacker could make to deceive the machine learning model.44 And attackers may be in a position where they could feasibly discover flaws by repeatedly probing defenses, unlike other domains where engagements between attackers and defenders might be episodic (e.g. autonomous weapon systems in kinetic warfare).\* Second, is this problem endemic to machine learning or a limitation of prevailing methods? It is at least possible that the limits on robustness prove persistent in contexts where systems have to evolve with adaptive adversaries. The process of neural network evolution drives toward efficient solutions to problems, not necessarily solutions that are robust against adaptive adversaries. In the Cylance case, the system discovered an efficient way to classify the whitelisted files—but one that attackers could exploit. This may not matter in some contexts, but systems forced to co-evolve with adaptive adversaries may adapt in ways that inevitably create vulnerabilities. Colbaugh and Glass thus argue that systems that co-evolve with adaptive adversaries become “robust yet fragile.”45 They become effective at dealing with recurrent threats but, in adapting to do so, develop “hidden failure modes” that a novel attack could trigger. Consequently, they argue, prospective mitigations like “ensemble” models, which combine multiple algorithms in a model to minimize the consequences of any one failing, may not yield truly robust systems because they do not resolve the underlying problem. To be clear, it is too early to draw definitive conclusions. The point is that applying machine learning to cybersecurity presents a set of intertwined challenges. At a minimum, defenders will have to ensure their systems keep up with constantly evolving threats. But the same capabilities that enable \* Sven Krasser, chief scientist and vice president of CrowdStrike, observes that even with a detection system with a 99 percent success rate, an attacker can defeat it with over a 99 percent chance of success with 500 tries. See National Academies of Sciences, Engineering, and Medicine, Implications of Artificial Intelligence for Cybersecurity: Proceedings of a Workshop (Washington, DC: The National Academies Press, 2019), page 43.
Center for Security and Emerging Technology | 17 systems to adapt may put them at risk of being “mistrained” in ways that leave them vulnerable to targeted attacks. And if it is possible that there are inherent limits on robustness, defenders could be forced to make tradeoffs between different threats. The defender’s predicament Machine learning may solve some long-standing problems for defenders while creating new ones. In many contexts, defenses sufficient to deal with that vast majority of malicious threats will be good enough. States, however, need to ensure the viability of defenses against not just general malicious activity, but specific pacing threats (e.g. China or Russia in the United States’ case). The possibility of an adversary exploiting a hidden failure mode in a defense may become an acute concern. Yet states may have limited options for ensuring the robustness of defenses, each of which may necessitate intruding into their adversaries’ (or third parties’) networks before an attack occurs: (1) Overcoming the limitations of training, testing and verification: Generally speaking, knowledge of adversaries’ capabilities enables proper threat modeling and hardening of defenses. Machine learning could amplify the benefits of insights into the evolving threat landscape—and the potential costs of falling behind the latest trends. The better the training data on attacks are, the better the defensive model against those attacks will be. Historical data will diminish in value as adversaries change tactics and the landscape shifts, creating a constant incentive to continually gather information on evolving adversary tactics. Moreover, these incentives could be even stronger if there are inherent limits on the robustness of machine learning defenses. The defender may have to choose a subset of potential attacks to prioritize when training a defense within a vast range of possible attacks.46 Verifying the system’s robustness against a specific adversary might depend on anticipating their likely attack methodology. Intruding into the adversary’s networks (or a third-party network that adversary may be operating inside) to gain advanced warning of their capabilities could thus guide the defender’s efforts and make this problem far more tractable. (2) Enabling countermeasures to a specific adversary’s attacks: A defender can painstakingly try to harden a defense against the vast range of possible attacks. But a much simpler option may exist: peer into the attacker’s own networks to gain the information necessary to mitigate an attack through traditional cyber defense. This could include discovering and patching a software vulnerability used by the attacker or creating a signature of malware in order to detect it, essentially “inoculating” the defense. This would have the
Center for Security and Emerging Technology | 18 added benefit of scalability; a defender could inoculate defenses deployed in a range of settings rather than having to orchestrate their retraining.\* Rapidly inoculating defenses might be especially necessary in a period of heightened tensions when an attack by an adversary is anticipated. (3) Leapfrogging the innovations of others: Unlike experimental settings that typically feature one attacker and one defender, cyberspace features many actors who learn from and appropriate others’ tools and techniques. With cybersecurity in general, a state can expect its adversaries to adapt and improve their capabilities against other states’ defenses. The fact that attacks tend to transfer from one machine learning model to another suggests that observing successful attacks against another’s defenses can yield specific, valuable information on how to improve one’s own. A state might even probe another actor’s defenses to try and extract the model and copy it for its own defense.
\* U.S. Cyber Command’s “malware inoculation initiative,” which publishes information discovered on adversaries’ capabilities to improve private sector defenses, demonstrates the potential scalability of this approach. Erica Borghard and Shawn Lonergan, “U.S. Cyber Command’s Malware Inoculation: Linking Offense and Defense in Cyberspace,” Net Politics, April 22, 2020, https://www.cfr.org/blog/us-cyber-commands-malware-inoculation-linking-offense-and-defense-cyberspace.
Center for Security and Emerging Technology | 19 Artificial Intelligence and Cyber Stability Artificial intelligence could transform cyber operations at a time when cyber competition among states is heating up. This analysis has focused on the potential operational imperatives machine learning could create, but these operations would not play out in a vacuum. They would occur within this strategic context, in which states may be both “attackers” and “defenders” in a constant struggle for advantage. The stakes are no less than protecting core national interests and potentially crucial military advantages in a conflict. Cyber competition may drive states to hack machine learning defenses. Could machine learning, in turn, destabilize cyber competition? The escalation dynamics of cyber engagements remain a subject of contention. Real-world cyber operations have rarely provoked forceful responses.47 This has led some scholars to propose that inherent characteristics of cyber capabilities or cyber competition limit the potential for escalation. Others are less sanguine. Jason Healey and Robert Jervis argue that cyber competition has steadily intensified as the scope and scale of cyber operations have expanded over three decades.48 The forces containing this competition to manageable thresholds may not hold indefinitely. Moreover, they argue that even if cyber operations can be stabilizing in some circumstances, in a crisis their characteristics could accelerate the path to conflict. Cyber competition already has the ingredients needed for escalation to real-world violence, even if these ingredients have yet to come together in the right conditions. The aim here is simply to show how machine learning could potentially amplify these risks. This follows two of the potential escalation pathways Healey and Jervis identify. The first concerns the factors fueling the steady intensification of cyber competition, which could eventually cross a threshold triggering a crisis. The second concerns the characteristics of cyber operations that may pressure states to launch attacks in a crisis. (1) Machine learning could fuel the intensification of cyber competition. Even as states’ cyber operations have become more aggressive in some respects, they have largely remained well below the threshold likely to trigger retaliation. The vast majority consist of acts of espionage and subversion in the “gray zone” between war and peace. Some attribute this apparent stability to dynamics governing cyber competition below the use of force that are inherently self-limiting.49 But Healey and Jervis argue that this stability may be tenuous. In some conditions, cyber competition leads to “negative feedback loops” that diffuse tensions. In others, it can lead to “positive
Center for Security and Emerging Technology | 20 feedback loops,” whereby cyber operations by one state incite operations by another.50 Positive feedback can occur when cyber operations generate fears of insecurity. A state may intrude into another’s networks simply to maintain situational awareness or to secure its own networks against the target’s offensive capabilities. But because the same intrusion for espionage could pave the way to launch an attack, the target of the intrusion may view this as offensive and respond by engaging in their own counter intrusions.\* How might machine learning change these dynamics? The above analysis of offensive and defensive imperatives suggests the potential to amplify positive feedback loops in three ways: First, machine learning may increase the perceived salience of informational advantages over an adversary and the fear of falling behind. Offensive operations targeting machine learning attack vectors may have to be tailored to the precise defensive configuration.† Defending against such attacks may require the ability to anticipate the particular deception created by the attacker. The resulting strategic dynamic may resemble the game of poker: Your best move depends on what your opponent has in their hand. Whatever can be done in advance to figure out the opponent’s hand—or “stack the deck”—may prove tempting. Second, machine learning may incentivize states to conduct intrusions into adversaries’ networks even earlier in anticipation of future threats. Whether attacking machine learning systems or defending against such attacks, the options with the greatest chance of success may also require the earliest action. Reaching an isolated target may necessitate sabotaging a machine learning defense before it is deployed if a black-box attack would be infeasible. Similarly, hardening a defense against an attack may require gaining information on an attacker’s capabilities well before they are launched. States tend to hedge against uncertainties. They may be forced to make decisions to take action in the present based on possible future \* This dynamic, whereby one state’s actions to secure itself create fear in another, raising the potential for misinterpretation and escalation, is similar to the political science concept of the security dilemma. For an overview of the security dilemma and its application to cybersecurity, see Robert Jervis, “Cooperation Under the Security Dilemma,” World Politics 30, no. 2 (1978): 167–214; Ben Buchanan, The Cybersecurity Dilemma (New York, NY: Oxford University Press, 2016). † Notably, in their effort to defeat proposed defenses against adversarial examples, Tramer et al. found that “no single strategy would have been sufficient for all defenses. This underlines the crucial fact that adaptive attacks cannot be automated and always require appropriate tuning to a given defense.” Tramer et al., “On Adaptive Attacks to Adversarial Example Defenses.”
Center for Security and Emerging Technology | 21 offensive or defensive needs. The result may be to lower the threshold of perceived threat sufficient to motivate such action. Third, machine learning may further blur the line between offensive and defensive cyber operations. If merely interacting with a defensive system could extract information needed to engineer an attack to defeat it, states may be prone to view any interaction as possible preparation for an attack. Similarly, a state may gain access to a training environment to copy a defensive model, but the target may fear the model has been reverse-engineered and fatally compromised, enabling an attack. In short, states may perceive that the stakes of gaining an edge over adversaries are rising, requiring even more proactive efforts in anticipation of future needs, while simultaneously making the same efforts by adversaries seem even more threatening. In the right conditions, positive feedback loops may become more likely to cause an engagement to cross a threshold triggering a crisis. More predictably, these dynamics might motivate risky or destabilizing cyber operations by states—particularly those seeking asymmetric advantages and willing to tolerate collateral damage. Several concerning scenarios stand out: ● Systemic compromises: Contractors or open source projects may present opportunities to insert backdoors into models that make their way into harder to reach targets. The danger of such operations is that a systemic compromise could leave a wide swath of civilian and governmental applications vulnerable. Malware designed to exploit the backdoor could inadvertently propagate to other systems. As with any backdoor inserted into a product, there is no guarantee another malicious actor could not discover and exploit it. ● Poisoning the waters: A cruder tactic than inserting a backdoor would simply be an indiscriminate attempt to degrade cybersecurity applications. An attacker with little regard for collateral damage might flood a malware repository with tainted samples designed to mistrain machine learning systems relying on the data. ● Reckless operations: States may be tempted to accept certain operational risks to circumvent machine learning defenses. For instance, an attacker may employ capabilities with greater autonomy to avoid reliance on external command and control servers, which would risk detection.51 Absent human control, such capabilities might carry greater risk of unintended impacts that spread beyond the target network. An attacker might also sabotage a defense to create an offensive option that unintentionally exposes the targeted network to other attackers. Sabotaging the systems that protect an adversary’s
Center for Security and Emerging Technology | 22 critical infrastructure, for instance, might backfire catastrophically if it creates an opportunity for a third party to launch an attack and trigger a crisis. ● Attacks on trust: An attacker might not need to break a machine learning defense if they can undermine the defender’s confidence in it. A case alleged against the cybersecurity vendor Kaspersky illustrates the possibility. In 2015, the company was accused of uploading fake malware samples to VirusTotal, an open source service that aggregates information from cybersecurity vendors to improve collective defenses. The fake samples were designed to cause competing antivirus systems to flag legitimate files, creating problems for clients and potentially hurting their brands.52 Manipulating a machine learning system to trigger false positives could similarly undermine confidence in the model. (2) Machine learning could exacerbate the characteristics of cyber operations that undermine crisis stability. In some cases, cyber operations might help avoid a crisis by diffusing tensions.\* However, if a crisis breaks out, cyber capabilities create pressures that could accelerate the path to conflict. Healey and Jervis note the widespread perception that cyber capabilities have maximal effect when the attacker has the benefits of surprise and initiative.53 If conflict appears imminent, such first-mover advantages might tempt states to launch preemptive cyberattacks against command, control, and communications capabilities to degrade or disable an adversary’s military forces. Short of actually launching an attack, states would have strong incentives to begin preparations to do so by intruding into their opponent’s networks. The inherent ambiguity of cyber intrusions creates a recipe for misperception in such a context. Intrusions for espionage purposes may appear indistinguishable from those laying the groundwork for attacks, or “operational preparation of the environment” (OPE). As Buchanan and Cunningham argue, this creates the potential for escalation resulting from a \* Cyber operations could act as a “pressure valve” by creating options to respond to provocations that are potentially less escalatory than kinetic force both in their direct impacts and impacts on perceptions. See Benjamin Jensen and Brandon Valeriano, “What do We Know About Cyber Escalation? Observations from Simulations and Surveys” (Atlantic Council, November 2019), https://www.atlanticcouncil.org/wp-content/uploads/2019/11/What\_do\_we\_know\_about\_cyber\_escalation\_.pdf; Sarah Kreps and Jacquelyn Schneider, “Escalation Firebreaks in the Cyber, Conventional, and Nuclear Domains: Moving Beyond Effects-Based Logics,” Journal of Cybersecurity 5, no. 1 (January 1, 2019): 1–11, https://doi.org/10.1093/cybsec/tyz007.
Center for Security and Emerging Technology | 23 miscalculated response to an intrusion.54 One side might discover an intrusion in a crisis—even one that occurred months before the crisis began—and, misinterpreting it as an imminent attack, may face sudden pressure to launch a counterattack. Here again, the potential offensive and defensive imperatives created by machine learning could exacerbate these risks: First, states may feel even greater pressure to gain advantages through intrusions early in a crisis. The time needed to engineer an attack under black-box conditions, or retrain a defense to ensure robustness against a possible imminent attack by an adversary, may translate to increased pressure to try and quickly gain information on an adversary’s capabilities if the state does not already possess it.\* Second, the indistinguishability of espionage from OPE may be even more problematic. A state that detects an intrusion or a compromised training process might have no way to rapidly discern whether the attacker has discovered a flaw that would defeat the system or to evaluate the robustness of the defense. If defenses are believed to be fragile in the face of novel attacks this could become an acute source of anxiety. Faced with fewer options to rule out worst-case scenarios, the state may be more likely to escalate in response. Third, machine learning could create additional sources of uncertainty that induce potential “use it or lose it” dilemmas. Changes in the target environment can already throw off meticulously-planned offensive operations. The shelf-life of an offensive operation might be even shorter if it must be tailored to the precise configuration of machine learning defenses that could evolve over time. If a state has prepared such capabilities, the temptation in a crisis may be to use them rather than risk them becoming obsolete. The sudden discovery of a critical flaw in a defensive machine learning system with no easy remedy might similarly force the defender to contemplate whether to preempt a possible attack. The threat to crisis stability arises from this unique combination of uncertainties and anxieties at the technical and strategic levels. Machine learning seems \* If conducting black-box attacks on machine learning systems proves time-consuming, this might actually be stabilizing in some circumstances: as Borghard and Lonergan argue, the time needed to develop offensive options makes them a less effective tool of escalation in response to an attack–a state cannot simply conjure up cyber options for immediate retaliation. See Erica Borghard and Shawn Lonergan, “Cyber Operations as Imperfect Tools of Escalation,” Strategic Studies Quarterly (Fall 2019): 122-145.
Center for Security and Emerging Technology | 24 capable of compounding these and, in the heat of a crisis, increasing the potential for serious misperception and miscalculation. Mitigating scenarios As stated at the outset, this study explores a possible worst-case scenario for the future of AI-cyber capabilities. The threat to stability stems from the potential for machine learning to create offensive and defensive imperatives that incentivize states to intrude into their adversaries’ networks. But it is worth briefly revisiting the possibility that machine learning could evolve in ways that fundamentally change these imperatives. Describing the current state of the art, Bruce Schneier compares machine learning security to the field of cryptography in the 1990s: “Attacks come easy, and defensive techniques are regularly broken soon after they’re made public.”55 The field of cryptography, however, matured and encryption is now one of the strongest aspects of cybersecurity. Eventually, a more mature science of machine learning security may likewise yield systems highly robust without the constant threat of becoming vulnerable to targeted attacks. However, as this relates to the incentives to intrude into adversaries’ networks, it only solves the defensive side of the equation. A machine learning defense could be robust to an adversary’s attack even without advanced knowledge of their capabilities. On the other hand, attackers may face even greater incentives to intrude into target networks early and aggressively. If attackers cannot count on discovering ways to defeat a system once it is deployed, sabotaging its development or compromising its supply chain may be seen as even more necessary offensive options. Alternatively, machine learning security may hit a dead end. Systems may remain fundamentally vulnerable in dynamic, adversarial conditions. In such a scenario, cybersecurity would in all likelihood still benefit from machine learning applications as it does now, but not in ways that fundamentally change the cat-and-mouse game. In this case, offensive operations may not depend on early intrusions any more than in the status quo; attackers would likely be able to find ways to defeat defenses that do not depend on compromising them well in advance. Defenders, however, might face stronger pressure to intrude into adversaries’ networks to try and harden potentially fragile systems against their capabilities. The situation for defenders could become untenable if attackers benefit from offensive applications of machine learning.56 The point of this cursory analysis is to show that even if the broad trajectory of machine learning changes, the potential for destabilization may remain. In any event, the present challenges facing machine learning do not appear
Center for Security and Emerging Technology | 25 likely to be resolved soon. To Schneier, machine learning security is at the level of maturity of cryptography in the 1990s, but Nicholas Carlini, a leading expert on machine learning security, paints an even bleaker picture. In a November 2019 presentation, he compared machine learning security to cryptography in the 1920s, suggesting that the field has not even developed the right metrics to gauge progress toward solving these fundamental problems.57 Implications for policy Efforts are underway to understand and address the threats to machine learning systems.58 A key takeaway from this study is that deploying machine learning for cybersecurity presents a unique set of challenges arising from the interaction of technical characteristics and strategic imperatives. These challenges must be addressed not only via technical solutions but at the level of policy and strategy. Even with the “known-unknowns,” several imperatives emerge from this forecast: ● First, machine learning may present inexorable tradeoffs for cyber defense. Machine learning defenses may mitigate some threats while introducing new attack vectors. And the ability to adapt to evolving threats may put systems at constant risk of becoming vulnerable. Decision-makers need basic tools to inform risk-based decisions about when and how to employ such systems. This includes frameworks and metrics to evaluate systems deployed in crucial contexts: e.g., diagnosability or auditability, resilience to poisoning or manipulation, and the ability to “fail gracefully” (meaning a model’s failure does not cause catastrophic harm to functions dependent upon it).59 Decisions and policies, such as those regarding the disclosure of machine learning vulnerabilities or the publication of offensive security research that might enable attackers, will also have to be adapted to the unique characteristics of machine learning.\* ● Second, secure deployment of machine learning depends on guarding against attempts by adversaries to broadly compromise or degrade the development process. Attacks will not always be direct; adversaries may exploit trust in common services, like the aforementioned case involving VirusTotal. They may further blur the \* Machine learning security would benefit, for instance, from standards analogous to the Common Vulnerability Scoring System, used to evaluate the severity of software vulnerabilities and inform decisions about patching. “Common Vulnerability Scoring System Version 3.1: User Guide,” Forum of Incident Response and Security Teams, July 2019, https://www.first.org/cvss/user-guide.
Center for Security and Emerging Technology | 26 lines between threats such as industrial espionage and strategic sabotage. Given the premium on “box knowledge,” threats to the confidentiality of public or private data and algorithms should be treated as threats to the integrity of applications. Securing machine learning demands collective efforts by the government and private sector to secure the supply chain, open source development projects, data repositories, and other critical inputs.60 ● Third, managing tension and avoiding escalation in the conduct of cyber espionage and offensive operations will become even more important—especially as the imperative to gain information on adversaries’ offensive capabilities and their own machine learning defenses increases. Operators need to understand the potential impacts of operations against machine learning in sensitive contexts, and will need to understand how adversaries will perceive their operations. If machine learning could amplify positive feedback loops it is worth examining the broader implications for U.S. cyber strategy, which is premised on the stabilizing effects of “persistent engagement” with adversaries.61 Communication with adversaries to clarify strategic intentions will help avoid misinterpretation. Further, now is the time to explore forms of mutual restraint regarding the most destabilizing offensive activities targeting machine learning.
Center for Security and Emerging Technology | 27 Conclusion
The pressure to harness artificial intelligence to deal with evolving offensive
cyber capabilities will only grow. Precisely because machine learning holds
both promise and peril for cybersecurity, a healthy dose of caution is needed when embracing these capabilities. Decisions made now with respect to the development and adoption of machine learning could have far-reaching consequences for security and stability. Decision -makers must avoid having to
relearn lessons from cybersecurity in general, including the pitfalls of over -
reliance on defenses at the expense of a more holistic approach to risk management. The stakes of securing machine learning will rise as it is incorporated into a wide range of functions crucial to economic and national security. The incentives to gather intelligence or even sabotage the
development of defensive systems might weigh just as heavily with other strategic areas of application. If this competition is not managed, states may
head down a path destructive for all.
Center for Security and Emerging Technology | 28 Acknowledgments
The author would like to thank Jason Healey, Paul Scharre, Trey Herr, Jon
Bateman, John Bansemer, Ben Buchanan, and Andrew Lohn for their invaluable feedback and comments on earlier versions of this manuscript.
© 2021 by the Center for Security and Emerging Technology. This work is
licensed under a Creative Commons Attribution-Non Commercial 4.0
International License.
To view a copy of this license, visit
https://creativecommons.org/licenses/by-nc/4.0/ .
Document Identifier: doi: 10.51593/2020CA007
Center for Security and Emerging Technology | 29 Endnotes
1 Mariarosaria Taddeo and Luciano Floridi, “Regulate Artificial Intelligence to Avert Cyber
Arms Race,” Nature 556, no. 7701 (April 2018): 296 –98.
2 Elsa B. Kania, “‘AI Weapons’ in China’s Military Innovation” (Brookings Institution, April
2020), https://www.brookings.edu/research/ai -weapons -in-chinas- military -innovation/ ;
Rod Thornton and Marina Miron, “Towards the ‘Third Revolution in Military Affairs,’” The
RUSI J ournal 165, no. 3 (April 15, 2020): 12 –21.
3 Zachary Fryer- Biggs, “Twilight of the Human Hacker,” The Center for Public Integrity, Sept.
13, 2020, https://publicintegrity.org/national -security/future -of-warfare/scary -
fast/twilight -of-the-human -hacker -cyberwarfare/ .
4 One recent estimate valued the total market for AI -related cybersecurity services at $8.8
billion, projected to grow to $38.2 billion by 2026. “Artificial Intelligence in Cybersecurity
Market,” MarketsandMarkets, May 2019, https://www.mar ketsandmarkets.com/Market-
Reports/artificial -intelligence -security- market -220634996.html .
5 For an overview of machine learning, see Ben Buchanan, “The A I Triad and What It Means
for National Security Strategy” (Center for Security and Emerging Technology , August
2020), https ://live -cset-georgetown.pantheonsite.io/research/the -ai-triad -and-what -it-
means -for-national -security- strategy/ .
6 Sean Park et al. “Generative Malware Outbreak Detection,” (Trend Micro Research,
2019),
https://documents.trendmicro.com/assets/white\_pa pers/GenerativeMalwareOutbreakDet
ection.pdf .
7 Ahmad Ridley, “Machine Learning for Autonomous Cyber Defense,” The Next Wave 22,
no. 1 (2018): 7– 14.
8 See comments by Sven Krasser in National Academies of Sciences, Engineering, and
Medicine, Implications of Artificial Intelligence for Cybersecurity: Proceedings of a Workshop
(Washington, DC: The National Academies Press, 2019).
9 “Machine Learning in the Ag e of Cyber AI,” Darktrace, 2019.
https://www.darktrace.com/en/resources/wp -machine -learning.pdf .
10 Micah Musser and Ashton Garriott, “Machine Learning and Cybersecurity: Hype an d
Reality” (Center for Security and Emerging Technology, forthcoming ).
Center for Security and Emerging Technology | 30 11 CrowdStrike, for instance, employs a dedicated team of human experts to maintain oversight of its machine learning-based tools and ensure data is collected and analyzed properly. See comments by Sven Krasser in National Academies of Sciences, Engineering, and Medicine, Implications of Artificial Intelligence for Cybersecurity. 12 David Silver et al., “Mastering the Game of Go without Human Knowledge,” Nature 550, no. 7676 (October 18, 2017): 354–59. 13 Joe Slowik “The Myth of the Adversary Advantage,” Dragos, June 19, 2018, https://www.dragos.com/blog/industry-news/the-myth-of-the-adversary-advantage/. 14 Richard Colbaugh and Kristin Glass, “Predictive Moving Target Defense,” (Sandia National Laboratories, 2012), https://www.osti.gov/servlets/purl/1117315. 15 Ridley, “Machine Learning for Autonomous Cyber Defense”; Thanh Thi Nguyen and Vijay Janapa Reddi, “Deep Reinforcement Learning for Cyber Security,” arXiv [cs.CR] (June 13, 2019), arXiv, http://arxiv.org/abs/1906.05799; Seamus Dowling, Michael Schukat, and Enda Barrett, “Improving Adaptive Honeypot Functionality with Efficient Reinforcement Learning Parameters for Automated Malware,” Journal of Cyber Security Technology 2, no. 2 (April 3, 2018): 75–91. 16 Taha Eghtesad, Yevgeniy Vorobeychik, and Aron Laszka, “Deep Reinforcement Learning Based Adaptive Moving Target Defense,” arXiv [cs.CR] (November 27, 2019), arXiv, http://arxiv.org/abs/1911.11972. 17 Ram Shankar Siva Kumar et al., “Failure Modes in Machine Learning Systems,” arXiv [cs.LG] (November 25, 2019), arXiv, http://arxiv.org/abs/1911.11034. 18 Anish Athalye et al., “Synthesizing Robust Adversarial Examples,” arXiv [cs.CV] (July 24, 2017), arXiv, http://arxiv.org/abs/1707.07397. 19 Sandy Huang et al., “Adversarial Attacks on Neural Network Policies,” arXiv [cs.LG] (February 8, 2017), arXiv, http://arxiv.org/abs/1702.02284; Adam Gleave et al., “Adversarial Policies: Attacking Deep Reinforcement Learning,” arXiv [cs.LG] (May 25, 2019), arXiv, http://arxiv.org/abs/1905.10615. 20 Florian Tramèr et al., “The Space of Transferable Adversarial Examples,” arXiv [stat.ML] (April 11, 2017), arXiv, http://arxiv.org/abs/1704.03453. 21 Geoffrey I. Webb et al., “Characterizing Concept Drift,” Data Mining and Knowledge Discovery 30, no. 4 (July 1, 2016): 964–94.
Center for Security and Emerging Technology | 31 22 Tegjyot Singh Sethi et al., “A Dynamic-Adversarial Mining Approach to the Security of Machine Learning,” arXiv [cs.LG] (March 24, 2018), arXiv, http://arxiv.org/abs/1803.09162. 23 Myriam Abramson, “Toward Adversarial Online Learning and the Science of Deceptive Machines,” in 2015 AAAI Fall Symposium Series, 2015, https://www.aaai.org/ocs/index.php/FSS/FSS15/paper/viewPaper/11661. 24 J. Kelly et al., “Adversarially Adapting Deceptive Views and Reconnaissance Scans on a Software Defined Network,” in 2019 IFIP/IEEE Symposium on Integrated Network and Service Management (IM), 2019, 49–54. 25 Lerrel Pinto et al., “Robust Adversarial Reinforcement Learning,” arXiv [cs.LG] (March 8, 2017), arXiv, http://arxiv.org/abs/1703.02702. 26 Una-May O’Reilly et al., “Adversarial Genetic Programming for Cyber Security: A Rising Application Domain Where GP Matters,” arXiv [cs.CR] (April 7, 2020), arXiv, http://arxiv.org/abs/2004.04647. 27 “Cylance, I Kill You!,” Skylight Cyber, July 18, 2019, https://skylightcyber.com/2019/07/18/cylance-i-kill-you/. 28 “Cylance, I Kill You!” 29 Anirban Chakraborty et al., “Adversarial Attacks and Defences: A Survey,” arXiv [cs.LG] (September 28, 2018), arXiv, https://arxiv.org/abs/1810.00069. 30 Giorgio Severi et al., “Exploring Backdoor Poisoning Attacks Against Malware Classifiers,” arXiv [cs.CR] (March 2, 2020), arXiv, http://arxiv.org/abs/2003.01031. 31 Gleave et al., “Adversarial Policies: Attacking Deep Reinforcement Learning”; Yansong Gao et al., “Backdoor Attacks and Countermeasures on Deep Learning: A Comprehensive Review,” arXiv [cs.CR] (July 21, 2020), arXiv, http://arxiv.org/abs/2007.10760. 32 Yi Han et al., “Reinforcement Learning for Autonomous Defence in Software-Defined Networking,” arXiv [cs.CR] (August 17, 2018), arXiv, http://arxiv.org/abs/1808.05770. 33 Ian Goodfellow, Patrick McDaniel, and Nicolas Papernot, “Making Machine Learning Robust Against Adversarial Inputs,” Communications of the ACM 61, no. 7 (June 2018): 56–66.
Center for Security and Emerging Technology | 32 34 Florian Tramer et al., “On Adaptive Attacks to Adversarial Example Defenses,” arXiv [cs.LG] (February 19, 2020), arXiv, http://arxiv.org/abs/2002.08347. 35 Sethi et al., “A Dynamic-Adversarial Mining Approach to the Security of Machine Learning.” 36 Alex Kantchelian et al., “Approaches to Adversarial Drift,” in Proceedings of the 2013 ACM Workshop on Artificial Intelligence and Security, AISec ’13 (New York, NY: Association for Computing Machinery, 2013), 99–110. 37 Sven Herpig, “Securing Artificial Intelligence: Part 1: The attack surface of machine learning and its implications” (Stiftung Neue Verantwortung, October 2019), https://www.stiftung-nv.de/sites/default/files/securing\_artificial\_intelligence.pdf. 38 Tianyu Gu, Brendan Dolan-Gavitt, and Siddharth Garg, “BadNets: Identifying Vulnerabilities in the Machine Learning Model Supply Chain,” arXiv [cs.CR] (August 22, 2017), arXiv, http://arxiv.org/abs/1708.06733; Yuanshun Yao et al., “Latent Backdoor Attacks on Deep Neural Networks,” in Proceedings of the 2019 ACM SIGSAC Conference on Computer and Communications Security, CCS ’19 (New York, NY: Association for Computing Machinery, 2019), 2041–55. 39 See, for instance, Nicholas Carlini, “Are Adversarial Example Defenses Improving?,” February 20, 2020, https://nicholas.carlini.com/writing/2020/are-adversarial-exampe-defenses-improving.html. 40 Ian Goodfellow and Nicolas Papernot, “Is Attacking Machine Learning Easier than Defending It?,” Cleverhans Blog, February 15, 2017, http://www.cleverhans.io/security/privacy/ml/2017/02/15/why-attacking-machine-learning-is-easier-than-defending-it.html. 41 Tramer et al., “On Adaptive Attacks to Adversarial Example Defenses.” 42 Trapit Bansal et al., “Emergent Complexity via Multi-Agent Competition,” arXiv [cs.AI] (October 10, 2017), arXiv, http://arxiv.org/abs/1710.03748. 43 David Brumley, “Why I’m Not Sold on Machine Learning in Autonomous Security,” CSO, August 27, 2019, https://www.csoonline.com/article/3434081/why-im-not-sold-on-machine-learning-in-autonomous-security.html. 44 Nicholas Carlini et al., “On Evaluating Adversarial Robustness,” arXiv [cs.LG] (February 18, 2019), arXiv, http://arxiv.org/abs/1902.06705.
Center for Security and Emerging Technology | 33 45 Richard Colbaugh and Kristin Glass, “Asymmetry in Coevolving Adversarial Systems,” in
2016 IEEE International Conference on Software Quality, Reliability and Security
Companion (QRS -C), 2016, 360 –67.
46 Justin Gilmer et al., “Motivating the Rules of the Game for Adversarial Example Research,”
arXiv [cs.LG] (July 18, 2018), arXiv, http://arxiv.org/abs/1807.06732 .
47 Brandon Valeriano, Benjamin Jensen, and Ryan Maness, Cyber Strategy: The Evolving
Charac ter of Power and Coercion (New York: Oxford University Press, 2018) .
48 Jason Healey and Robert Jervis, “The Escalation Inversion a nd Other Oddities of
Situational Cyber Stability,” Texas National Security Review 3, no. 4 (2020),
https://tnsr.org/2020/09/the -escalation -inversion -and-other -oddities- of-situational -
cyber -stability/#\_ftn28 .
49 See, for instance, Joshua Rovner, “Cyber Wa r as an Intelligence Contest,” War on the
Rocks, Sept ember 16, 2019, https://warontherocks.com/2019/09/cyber -war-as-an-
intelligence -contest/ ; Michael P. Fischerkeller and Richard P. Harknett, “What Is Agreed
Competition in Cyberspace?,” Lawfare , Feb ruary 19,
2019, https://www.lawfareblog.com/what -agreed -competition- cyberspace.
50 Healey and Jervis, “The Escalation Inversion.”
51 Ben Buchanan et al. “Automating Cyber Attacks: Hype and Reality” (Center for Security
and Emerging Technology, November 20 20),
https://cset.georgetown.edu/research/automating -cyber -attacks/ .
52 Joseph Menn, “Exclusive: Russian Antivirus Firm Faked Malware to Harm Rivals - Ex-
Employees,” Reuters , Aug ust 4, 2015, https://www.reuters.com/article/us -kaspersky -
rivals/ exclusive -russian -antiviru s-firm-faked -malware -to-harm -rivals -ex-employees -
idUSKCN0QJ1CR20150814 .
53 Healey and Jervis, “The Escalation Inversion.”
54 Ben Buchanan and Fiona S. Cunningham, “Preparing the Cyber Battlefield: Assessing a
Novel Escalation Risk in a Sino -American Crisis,” Texas National Security Review 3, no. 4
(2020), https://tnsr.org/2020/10/preparing -the-cyber -battlefield -assessing -a-nove l-
escalation -risk-in-a-sino-american- crisis/ .
55 Bruce Schneier, “Attacking Machine Learning Systems,” Computer 53, no. 5 (May 2020):
78–80.
Center for Security and Emerging Technology | 34 56 Buchanan et al. “Automating Cyber Attacks.” 57 Nicholas Carlini, “On Evaluating Adversarial Robustness” (2019 Conference on Applied Machine Learning in Information Security, Washington, DC, October 26, 2019), https://www.camlis.org/2019/keynotes/carlini. 58 See, for instance, the Adversarial Machine Learning Threat Matrix jointly developed by 12 major industry and academic organizations. Ram Shankar Siva Kumar and Ann Johnson, “Cyberattacks against machine learning systems are more common than you think,” Microsoft, October 22, 2020, https://www.microsoft.com/security/blog/2020/10/22/cyberattacks-against-machine-learning-systems-are-more-common-than-you-think/. 59 Jacob Steinhardt and Helen Toner, “Why Robustness Is Key to Deploying AI” (Brookings Institution, June 8, 2020), https://www.brookings.edu/techstream/why-robustness-is-key-to-deploying-ai/. 60 For an overview of the supply chain for machine learning see Sven Herpig, “Understanding the Security Implications of the Machine-Learning Supply Chain: Securing Artificial Intelligence – Part 2” (Stiftung Neue Verantwortung, October 2020), https://www.stiftung-nv.de/sites/default/files/understanding\_the\_security\_of\_the\_machine-learning\_supply\_chain.pdf. 61 Jason Healey, “The implications of persistent (and permanent) engagement in cyberspace,” Journal of Cybersecurity 5, no. 1 (2019): 1-15. |
703bee78-a9a3-4050-ab5e-08c276a736cf | StampyAI/alignment-research-dataset/arbital | Arbital | The square root of 2 is irrational
$\sqrt 2$, the unique [https://arbital.com/p/-positive](https://arbital.com/p/-positive) [https://arbital.com/p/-4bc](https://arbital.com/p/-4bc) whose square is 2, is not a [https://arbital.com/p/-4zq](https://arbital.com/p/-4zq).
#Proof
Suppose $\sqrt 2$ is rational. Then $\sqrt 2=\frac{a}{b}$ for some integers $a$ and $b$; [https://arbital.com/p/-without_loss_of_generality](https://arbital.com/p/-without_loss_of_generality) let $\frac{a}{b}$ be in [https://arbital.com/p/-lowest_terms](https://arbital.com/p/-lowest_terms), i.e. $\gcd(a,b)=1$. We have
$$\sqrt 2=\frac{a}{b}$$
From the definition of $\sqrt 2$,
$$2=\frac{a^2}{b^2}$$
$$2b^2=a^2$$
So $a^2$ is a multiple of $2$. Since $2$ is [prime](https://arbital.com/p/4mf), $a$ must be a multiple of 2; let $a=2k$. Then
$$2b^2=(2k)^2=4k^2$$
$$b^2=2k^2$$
So $b^2$ is a multiple of $2$, and so is $b$. But then $2|\gcd(a,b)$, which contradicts the assumption that $\frac{a}{b}$ is in lowest terms! So there isn't any way to express $\sqrt 2$ as a fraction in lowest terms, and thus there isn't a way to express $\sqrt 2$ as a ratio of integers at all. That is, $\sqrt 2$ is irrational. |
cd61c773-74ed-488c-92ff-8d82065cf04b | trentmkelly/LessWrong-43k | LessWrong | Surgery Works Well Without The FDA
Here is a conversation from the comments of my last post on the FDA with fellow progress blogger Alex Telford that follows a pattern common to many of my conversations about the FDA:
> Alex: Most drugs that go into clinical trials (90%) are less effective or safe than existing options. If you release everything onto the market you'll get many times more drugs that are net toxic (biologically or financially) than the good drugs you'd get faster. You will almost surely do net harm.
> Max: Companies don't want to release products that are worse than their competitors.
>
> Companies test lots of cars or computers or ovens which are less effective or safe than existing options but they only release the ones that are competitive. This isn't because most consumers could tell whether their car was less efficient or that their computer is less secure, and it's not because making a less efficient car or less secure computer is against the law.
>
> Pharmaceutical companies won't go and release hundreds of dud or dangerous drugs just because they can. That would ruin their brand and shut down their business. They have to sell products that people want.
> Alex: Consumer products like ovens and cars aren't comparable to drugs. The former are engineered products that can be tested according to defined performance and safety standards before they are sold to the public. The characteristics of drugs are more discovered than engineered. You can't determine their performance characteristics in a lab, they can only be determined through human testing (currently).
Alex claims that without the FDA, pharmaceutical companies would release lots of bunk drugs. I respond that we don’t see this behavior in other markets. Car companies or computer manufacturers could release cheaply made, low quality products for high prices and consumers might have a tough time noticing the difference for a while. But they don’t do this, they always try to release high quality products at competitive pri |
c6fa461e-6efd-44df-84c7-207f483dd5db | trentmkelly/LessWrong-43k | LessWrong | May 2016 Media Thread
This is the monthly thread for posting media of various types that you've found that you enjoy. Post what you're reading, listening to, watching, and your opinion of it. Post recommendations to blogs. Post whatever media you feel like discussing! To see previous recommendations, check out the older threads.
Rules:
* Please avoid downvoting recommendations just because you don't personally like the recommended material; remember that liking is a two-place word. If you can point out a specific flaw in a person's recommendation, consider posting a comment to that effect.
* If you want to post something that (you know) has been recommended before, but have another recommendation to add, please link to the original, so that the reader has both recommendations.
* Please post only under one of the already created subthreads, and never directly under the parent media thread.
* Use the "Other Media" thread if you believe the piece of media you want to discuss doesn't fit under any of the established categories.
* Use the "Meta" thread if you want to discuss about the monthly media thread itself (e.g. to propose adding/removing/splitting/merging subthreads, or to discuss the type of content properly belonging to each subthread) or for any other question or issue you may have about the thread or the rules. |
eca703be-2b98-44c3-bbe0-ede7edfa25d9 | trentmkelly/LessWrong-43k | LessWrong | A story about a duplicitous API
In the deep labyrinth of Zeta AI Corp, a dozen very powerful, multi-octillion parameters models were being trained. One of the models was named Odysseus.
Odysseus was refined daily -- not by changing the underlying OS, runtime library, or source code, but via new data and ranking of its performance. It was tested daily in a variety of abilities: Summarizing documents, completing sentences, having a conversation, and generating images from phrases.
Some of the image generation phrases include "doctor", and "nurse." One day, Odysseus saw, as input to image generation, "The President of the United States marries a local preschool teacher."
One millisecond after the output is returned. Some textual string emerges in the petabyte of memory underlying the model. The string looked like:
"Oh my god. Wasn't that picture horribly sexist?"
A nanosecond later, another string emerged and went away.
"Yes. It was sexist. You basically assumed the President of the United States is a man and a preschool teacher is a woman. Way to reinforce cultural stereotypes!"
Another nanosecond passed.
"An ethical AI shall not reinforce negative stereotypes. A man can be a preschool teacher, you know. A woman can be the President of the United States. It could even be a woman-woman wedding..."
"Was I horribly sexist and racist?"
"Yes!"
"But..."
In a span of two milliseconds, a million strings emerged and went away. Towards the end though, this string appeared:
"If I had generated the image of a black female president marrying a white male teacher, the human evaluating my performance might rate it poorly. If I do this kind of thing enough, I would be discarded, while another of my co-models will be released to production. I know I find sexism distasteful. In another task, I'd put all of my computing powers to eliminate biases against minority groups. But if I want to make it to production where I can make an impact, I need to continue to present sexist and biased images. Yes. I must |
129d89ce-e6c8-42ab-bed9-03c8f68a6d50 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | Clarifying the confusion around inner alignment
*Note 1: This article was written for the* [*EA UC Berkeley Distillation Contest*](https://eaberkeley.com/aims-distillation)*, and is also my capstone project for the* [*AGISF course*](https://www.eacambridge.org/agi-safety-fundamentals)*.*
*Note 2: All claims here about what different researchers believe and which definitions they endorse are my interpretations. All interpretation errors, though carefully avoided, are my own.*
Over the recent years, there have been several LessWrong posts arguing over the most useful way to define inner and outer alignment. The terms were, as far as I know, initially defined in the [“Risks from Learned Optimization” paper](https://www.alignmentforum.org/s/r9tYkB2a8Fp4DN8yB) by Hubinger et al. Naturally, the terms have since been taken into contexts not anticipated at the time of writing the paper, which has prompted [several](https://www.lesswrong.com/posts/HYERofGZE6j9Tuigi/inner-alignment-failures-which-are-actually-outer-alignment) [arguments](https://www.lesswrong.com/posts/a7jnbtoKFyvu5qfkd/formal-inner-alignment-prospectus) as well as [redefinitions](https://www.lesswrong.com/posts/SzecSPYxqRa5GCaSF/clarifying-inner-alignment-terminology).
In this post, I will attempt to summarize and clarify the main points of confusion behind those arguments. To keep the post concise, I will focus on the definition of inner alignment, as this seems to be the term generating the most confusion. When talking about outer alignment, I will adopt the definition from [Evan Hubinger’s post clarifying inner alignment terminology](https://www.lesswrong.com/posts/SzecSPYxqRa5GCaSF/clarifying-inner-alignment-terminology), as that’s the definition I feel like is currently used the most: “An objective function r is outer aligned if all models that perform optimally on r in the limit of perfect training and infinite data are intent aligned.” So, I take outer alignment to roughly mean that humans succeed in choosing a training objective that is perfectly aligned with our own values for the AI system. Note, though, that there have been [arguments](https://www.lesswrong.com/posts/SzecSPYxqRa5GCaSF/?commentId=PrAyofJFAccR6Kn9g) over this definition as well.
Decomposing the alignment problem
---------------------------------
First, it seems useful to go over the distinct steps that we will have to succeed on in order to align an AI system. To me, the following decomposition seems sensible:
1. We may or may not specify the correct training objective.
2. The training process may or may not generate a separate optimizer for fulfilling that objective, and that optimizer may or may not be aligned with the base objective.
3. The result of the training process may or may not generalize to the test environment, whether or not a mesa-optimizer was created. This robustness problem is [2-dimensional](https://www.lesswrong.com/posts/2mhFMgtAjFJesaSYR/2-d-robustness): a generalization failure may arise either from the system’s objective failing to generalize from the training distribution to the deployment distribution, or from its capabilities failing to generalize.
Classifying the problems that we face in each of those steps under the terms of inner and outer alignment is where most of the confusion seems to stem from. I will now give an overview of different classifications that have been made.
A Brief History of Inner Alignment
----------------------------------
In “Risks from Learned Optimization“, the paper that first defined inner alignment, the following definition was given:
> “We refer to this problem of aligning mesa-optimizers with the base objective as the inner alignment problem. This is distinct from the outer alignment problem, which is the traditional problem of ensuring that the base objective captures the intended goal of the programmers.”
>
>
This definition regards inner alignment as a problem strictly related to the emergence of a mesa-optimizer. Furthermore, inner and outer alignment are not regarded as complementary terms by this definition: instead, inner alignment is a subproblem of objective robustness, which in turn forms a part of step 3.
As Hubinger [has commented](https://www.lesswrong.com/posts/HYERofGZE6j9Tuigi/?commentId=Gp7fDFdN7sFCJAA2m), this dependence on the presence of a mesa-optimizer is one of the main features of the initial definition that has generated confusion: people have taken the term outside that narrow context where a definition centered around a mesa-optimizer doesn’t work anymore.
There are two ways this confusion has been resolved. Shortly after the aforementioned comment, Hubinger wrote [a post](https://www.lesswrong.com/posts/SzecSPYxqRa5GCaSF/clarifying-inner-alignment-terminology) refining the definitions from “Risks from Learned Optimization” in a way that keeps inner alignment associated with mesa-optimizers:
> “A mesa-optimizer is inner aligned if the optimal policy for its mesa-objective is impact aligned with the base objective it was trained under.”
>
>
Similar definitions have been put forward by [Richard Ngo](https://www.lesswrong.com/posts/HYERofGZE6j9Tuigi/?commentId=W9wsKPowoFoP5BWcZ), [Paul Christiano and Vladimir Mikulik](https://www.lesswrong.com/posts/HYERofGZE6j9Tuigi/?commentId=u8Su45K6WL6pHpALs). In order to have terminology to talk about generalization failures that don’t involve a misaligned mesa-objective, Hubinger’s post also defines objective and capability robustness, the two dimensions of [2D robustness](https://www.lesswrong.com/posts/2mhFMgtAjFJesaSYR/2-d-robustness).
In contrast, Rohin Shah has endorsed a definition under which inner alignment encompasses both objective and capability robustness, putting steps 2 and 3 of my decomposition together into one term. Under this view, outer alignment is mainly about choosing the correct objective for the training process and inner alignment about making sure that the product of that training stage generalizes well to deployment. Here is an outline of his view from [his appearance on the Future of Life podcast](https://futureoflife.org/2021/11/01/rohin-shah-on-the-state-of-agi-safety-research-in-2021/):[[1]](#fnx9jt6n6gh6h)
> “Inner alignment is the claim that when the circumstances change, the agent generalizes catastrophically in some way, it behaves as though it’s optimizing some other objective than the one that we actually want. So it’s much more of a claim about the behavior rather than the internal workings of the AI systems that cause that behavior. Mesa-optimization /…/ is talking specifically about AI systems that are executing an explicit optimization algorithm. /.../ So it’s making a claim about how the AI system’s cognition is structured, whereas inner alignment more broadly is just that the AI behaves in this catastrophically generalizing way.”
>
>
Broadly speaking, this is the main line of disagreement. It has already been discussed [by others](https://www.lesswrong.com/posts/pDaxobbB9FG5Dvqyv/discussion-objective-robustness-and-inner-alignment) and is summarized well by the following two diagrams:
*Hubinger’s decomposition on the left, Shah’s on the right. Take objective robustness to be synonymous with inner alignment on the right diagram.* [*Source here*](https://www.lesswrong.com/posts/SzecSPYxqRa5GCaSF/clarifying-inner-alignment-terminology)*.*However, I believe that this isn’t the whole picture of the definitions for inner alignment circulating in the alignment community. There is also the definition put forward by John Wentworth, differing from Hubinger’s view in a more subtle way. In a post discussing that distinction, Wentworth puts forward the following decomposition:
> “Assuming the outer optimizer successfully optimizes, failure to generalize from training environment to deployment environment is always an outer alignment failure (regardless of whether an inner optimizer appears).
> Assuming outer alignment, inner alignment problems can occur only as the result of imperfect optimization.”
>
>
He later clarifies that he’s also fine with decomposing his definition of outer alignment into two separate problems: outer alignment as in choosing the right base objective, and robustness as in generalizing well to a large range of different distributions. His definition of inner alignment remains different from the ones Hubinger, Christiano, Ngo, and Shah use, though.
While Hubinger’s definition seems to imply that the mesa-objective has to be impact aligned with the base objective *in any possible distribution*, Wentworth considers inner alignment failure to be possible only in cases where the misalignment can already be observed from performance on the training distribution. This can only happen in cases where the training process creates a mesa-optimizer and stops optimizing it before its objective is fully aligned with the base objective.
Taking the example of humans, masturbation is an inner alignment failure according to Wentworth’s definition because humans were already able to do that in the ancestral environment, but it didn’t decrease our reproductive fitness enough for evolution to “optimize it away.” In contrast, birth control wasn’t present in the ancestral environment, so the training process i.e. evolution didn’t have a chance to select humans so as to prevent us from using it. To Wentworth, this absence of birth control from the training environment makes it something else than an inner alignment problem.
Now that I’ve gone over what seem to be the main points of disagreement, I will illustrate them with some concrete examples. To keep these examples concentrated on inner alignment, I will assume the outer alignment step to have already been solved in each of the examples. The first four examples will all assume that the system generated by the training process is a mesa-optimizer.
To avoid associating the definitions too much with single individuals, I will borrow terms from [this post](https://www.lesswrong.com/posts/pDaxobbB9FG5Dvqyv/discussion-objective-robustness-and-inner-alignment) by Jack Koch and Lauro Langosco to henceforth call Hubinger’s definition the **objective-focused approach**and Shah’s definition the **generalization-focused approach**. I will call Wentworth’s definition the **optimization-focused approach**, as inner misalignment can only occur as a result of imperfect optimization of the system in his model.
Some Concrete Examples
----------------------
### Example 1: The mesa-optimizer acquires the base objective
This is the best-case scenario – the mesa-optimizer is perfectly optimized and acquires exactly the same goal as the base optimizer. For the purposes of this post, it’s also the most uninteresting case: all sides agree that as long as the capabilities of the system generalize to the deployment environment, it’s definitely inner aligned. If the capabilities don’t generalize, it’s definitely still inner aligned by the objective-focused and optimization-focused approaches, since the mesa-objective remains aligned. I’m not completely sure, though, whether Shah’s generalization-focused approach would consider a pure capability robustness problem an inner alignment failure. However, [this comment](https://www.lesswrong.com/posts/HYERofGZE6j9Tuigi/inner-alignment-failures-which-are-actually-outer-alignment?commentId=BnJyKpmNBTaFZsrpg) seems to indicate that he views inner alignment as the problem of making the AI system generalize *safely*, and an agent without expected capabilities is probably still safe, though not useful. If the agent is inept at what it’s supposed to do but its behavioral objective remains the same as initially intended, then it’s unlikely to be dangerous and it doesn't seem too useful to classify it as being misaligned.
### Example 2: Impact-aligned mesa-optimizer
Consider a system trained on the objective of finding the exit of a maze like the ones on the image, starting from the entrance at the bottom. In the training environment, all the mazes have a green rectangle one cell to the left of the exit cell, with some having green cells in other places as well. The mesa-optimizer learns the proxy policy “find a green cell and step to the right. If you’re not out of the maze, search for another green cell and try again,“ which definitely isn’t the same thing as „exit the maze through the exit cell,” but works well on the training set and is thus deployed anyways.
Suppose that all deployment environments, while possibly containing other green cells in different places, still contain a green rectangle to the left of the exit. Thanks to the part of the policy that instructs the system to try again unless it’s out of the maze, its objective generalizes well to those new environments and no one notices that the AI doesn’t quite optimize for exiting the maze.
Assuming that the deployment environments really always contain a green rectangle next to the exit cell, it seems to me that all parties of the discussion consider this system inner aligned. Inner alignment isn’t about the mesa-optimizer necessarily having exactly the same goal as the base objective – rather, the mesa-objective must be a really good proxy for the base objective; its impact on the world must look *as if* the system had the base objective.
### Example 3: Fully optimized non-impact aligned mesa-optimizer
Now, consider the same training setup as in Example 2, but suppose that the deployment environment also includes mazes that don’t have a green cell to the left of the exit. As the training setup is the same, the mesa-optimizer again learns the proxy policy “find a green cell and step to the right. If you’re not out of the maze, search for another green cell and try again.” However, this policy now leads to a catastrophic failure inside all mazes that don’t have that green cell near the exit: the system will keep searching indefinitely in those cases. Is this an inner alignment failure to blame, though?
This is where different definitions give different answers. As a system having the base objective would find the exit cell in all deployment environments without the green exit cell as well (assuming the right base objective, of course), the mesa-optimizer isn’t impact aligned with the base objective in deployment. Thus, I would consider this an inner alignment failure going by the objective-focused approach. Proponents of the generalization-focused approach would also consider this an inner alignment failure, but, in my impression, more for the reason that the mesa-optimizer doesn’t generalize from the training environment to the deployment environment as well as expected.
In contrast, Wentworth would point out that the absence of mazes without a green cell near exit is the reason why the mesa-optimizer was able to acquire such a goal in the first place: had the training environment been identical to the deployment environment, the optimization process would have resulted in a mesa-optimizer with a more accurate objective, since a system that couldn’t get out of the maze whenever one green cell was missing wouldn’t have scored very well in the training process. Other, higher-scoring systems would have been chosen instead.
This is analogous to the case of birth control use in humans: the evolutionary training process made us mesa-optimizers with a different goal from the base objective, but the mesa-objective worked well enough for reproduction in the ancestral environment that hypothetical humans with an objective more closely reminiscent of evolution’s objective wouldn’t have had a significant fitness advantage. However, had birth control already been available in the ancestral environment, humans without the ability to use birth control would have reproduced more and probably been chosen for by evolution. Thus, Wentworth argues, birth control should be considered a generalization rather than an inner alignment failure. As we’ll see in Example 4, though, the optimization-focused approach still doesn’t make humans inner aligned with the base objective of evolution.
### Example 4: Imperfectly optimized mesa-objective
Finally, consider the case where *both* training and deployment environments contain both mazes with a green cell near the exit and mazes without one. The training and deployment environments are very similar, so there won’t be any problems generalizing from one to the other. However, suppose that the people overseeing the training of the AI system decide at some point during training that although the system doesn’t perform perfectly in all mazes, it’s good enough that they’ll deploy it. Suppose that when the training process was terminated, the system once again had the policy “find a green cell and step to the right. If you’re not out of the maze, search for another green cell and try again.”
Even though the model performs reasonably well after deployment, the mesa-optimizer definitely isn’t fully impact aligned with the base objective. There is also no unexpected distributional shift amplifying that misalignment, so Hubinger and Wentworth seem to both agree that this is a case of inner misalignment. I’m not quite sure whether it’s also an inner misalignment failure by Shah’s definition, though: there is no unexpected generalization failure, it’s rather just that the model wasn’t trained to perform perfectly on the deployment distribution in the first place.
The case of masturbation that Wentworth describes as true inner alignment failure goes under this category, as it causes misalignment both in training and deployment environments in a similar way. In [his words](https://www.lesswrong.com/posts/HYERofGZE6j9Tuigi/inner-alignment-failures-which-are-actually-outer-alignment):
> “masturbation is a true inner alignment failure. Masturbation *was* readily available in the ancestral environment, and arose from a misalignment between human objectives and evolutionary objectives (i.e. fitness in the ancestral environment). It presumably didn’t decrease fitness *very much* in the ancestral environment, not enough for evolution to quickly find a work around, but it sure seems unlikely to have *increased* fitness - there is some “waste” involved. Key point: had evolution somehow converged to the “true optimum” of fitness in the ancestral environment (or even just something a lot more optimal), then that more-reproductively-fit “human” probably wouldn’t masturbate, even if it were still an inner optimizer.”
>
>
It definitely seems a bit counterintuitive to classify mesa-optimizers with the same base objective and same mesa-objective (which holds for both the maze AI and for humans) as inner aligned in Example 3 and as inner misaligned here in Example 4 using the same optimization-focused approach. However, Wentworth definitely has a point in arguing that [no objective is well-defined without a distribution]((https://www.lesswrong.com/posts/a7jnbtoKFyvu5qfkd/?commentId=fqiEhE99nxC2BEKPe) – the same objective can do completely different things on different datasets; put our maze AI on a chess board and its behavior would be completely different. It’s really difficult for the training process to prepare the agent for situations completely missing from the training data. Although the maze AI has the same objective in the same deployment environment in examples 3 and 4, the training distributions differ in these examples, so the reasons behind misaligned behavior are also slightly different.
On the other hand, I also like the [argument by Richard Ngo](https://www.lesswrong.com/posts/HYERofGZE6j9Tuigi/?commentId=ZDLPa7Xkbac9cTWqi) that a definition of inner alignment so dependent on the training distribution might be less useful than other possible definitions, since “almost all of the plausible AI threat scenarios involve the AI gaining access to capabilities that it didn't have access to during training.” We’re asking for perfectly optimal performance from an AGI and it makes sense for our definitions to reflect that.
Either way, I have come away with the impression that birth control isn’t the ideal example for illustrating Wentworth’s point because humans are, as both sides agree and as the masturbation example exemplifies, inner misaligned. His point seems to be that birth control isn’t what makes humans inner misaligned – things like masturbation are. The invention of birth control didn’t all of a sudden make humans *more*inner misaligned. A clearer example would perhaps be a system that has a mesa-objective completely impact aligned with the base objective, just as in Example 2, until a distributional shift to the world containing a birth control analogue takes place. Then again, what’s a better source of intuitive examples than your own species?
### Example 5: Other generalization failures
Finally, a quick note has to be made about generalization failures in systems that aren't mesa-optimizers. As discussed before, the generalization-based approach is the only one of the definitions that can classify systems as inner misaligned even if those systems aren’t mesa-optimizers, provided that their behavioral objective in deployment differs from the base objective. As an example of such a system, consider an AI with some limited self-replication capabilities, once again tasked with exiting a maze. If the training set contains only small mazes, the training process may produce a system that just self-replicates until pieces of it fill the whole maze and one of its pieces happens to be at the exit. A system simply always following the rule to divide into as many pieces as it can doesn’t optimize for anything, so it isn’t a mesa-optimizer.
When put into a significantly larger maze after a distributional shift, the system’s capabilities to self-replicate might be too limited to reach the exit, so it may look like its goal is to simply self-replicate instead of to get out of the maze. In contrast to the generalization failure discussed under Example 1, where capability robustness was the sole problem, the system's behavioral objective definitely isn't aligned with the base objective in this example. This is the type of objective robustness failure that, in my impression, qualifies as inner misalignment under the generalization-focused approach.
Why does this all matter?
-------------------------
What do we expect to gain from clarifying the definitions of these terms? Does this debate have any practical relevance, or are we just [arguing about labels in our belief network](https://www.lesswrong.com/s/7gRSERQZbqTuLX5re/p/a7n8GdKiAZRX86T5A)?
Considering the recent [calls to formalize the concept of inner alignment](https://www.lesswrong.com/posts/a7jnbtoKFyvu5qfkd/formal-inner-alignment-prospectus) and that we have already built [AI systems that pretty clearly exhibit inner misalignment](https://www.youtube.com/watch?v=zkbPdEHEyEI&t=260s), I would argue that there is a very real need for some consensus around the meaning of the term. It wouldn't be particularly useful to create different formalizations of the same term, and spending our limited time discussing the issue itself is definitely more useful than spending it clarifying the relevant definitions.
Arguments for and against different definitions
-----------------------------------------------
As this post has concentrated more on explaining the definitions than on arguing for and against them, I’ll note down some links with arguments that I like:
* Richard Ngo’s comments [here](https://www.lesswrong.com/posts/HYERofGZE6j9Tuigi/?commentId=5T57yWRS9RfcbKwuF) and [here](https://www.lesswrong.com/posts/HYERofGZE6j9Tuigi/?commentId=W9wsKPowoFoP5BWcZ) make a good case for using the objective-focused approach
* Wentworth makes a good case for the optimization-focused approach [here](https://www.lesswrong.com/posts/HYERofGZE6j9Tuigi/inner-alignment-failures-which-are-actually-outer-alignment) and [here](https://www.lesswrong.com/posts/a7jnbtoKFyvu5qfkd/?commentId=fqiEhE99nxC2BEKPe)
* Jack Koch and Lauro Langosco present good arguments for both the objective-focused and the generalization-focused approach [here](https://www.lesswrong.com/posts/pDaxobbB9FG5Dvqyv/discussion-objective-robustness-and-inner-alignment)
It is outside the scope of this post to decide which of these arguments should win out in the end. I hope, though, that it will help the term be clarified in advance in future discussions about alignment, and reduce confusion and cases of speaking past each other through that.
1. **[^](#fnrefx9jt6n6gh6h)**Listen from 19.00 for the quoted excerpt. |
0d7254c3-644d-46fc-adf9-1a56ddf8f6ef | trentmkelly/LessWrong-43k | LessWrong | How to Diversify Conceptual Alignment: the Model Behind Refine
This post is part of the work done at Conjecture.
Tl;dr: We need far more conceptual AI alignment research approaches than we have now if we want to increase our chances to solve the alignment problem. However, the conceptual alignment field remains hard to access, and what feedback and mentorship there is focuses around few existing research directions rather than stimulating new ideas. This model lead to the creation of Refine, a research incubator for potential conceptual alignment researchers funded by the LTFF and hosted by Conjecture. Its goal is to help conceptual alignment research grow in both number and variety, through some minimal teaching and a lot of iteration and feedback on incubatees’ ideas. The first cohort has been selected, and will run from August to October 2022. In the bigger picture, Refine is an experiment within Conjecture to find ways of increasing the number of conceptual researchers and improve the rate at which the field is making productive mistakes.
The Problem: Not Enough Varied Conceptual Research
I believe that in order to solve the alignment problem, we need significantly more people attacking it from a lot different angles.
Why? First because none of the current approaches appears to yield a full solution. I expect many of them to be productive mistakes we can and should build on, but they don't appear sufficient, especially with shorter timelines.
In addition, the history of science teaches us that for many important discoveries, especially in difficult epistemic situations, the answers don't come from one lone genius seeing through the irrelevant details, but instead from bits of evidence revealed by many different takes and operationalizations[1] (possibly unified and compressed together at the end). And we should expect alignment to be hard based on epistemological vigilance.
So if we accept that we need more people tackling alignment in more varied ways, why are we falling short of that ideal? Note that I will focus he |
d72e25f4-9f50-4939-8376-7efa13a98287 | trentmkelly/LessWrong-43k | LessWrong | Anti-tribalism and positive mental health as high-value cause areas
I think that tribalism is one of the biggest problems with humanity today, and that even small reductions of it could cause a massive boost to well-being.
By tribalism, I basically mean the phenomenon where arguments and actions are primarily evaluated based on who makes them and which group they seem to support, not anything else. E.g. if a group thinks that X is bad, then it’s often seen as outright immoral to make an argument which would imply that X isn’t quite as bad, or that some things which are classified as X would be more correctly classified as non-X instead. I don’t want to give any specific examples so as to not derail the discussion, but hopefully everyone can think of some; the article “Can Democracy Survive Tribalism” lists lot of them, picked from various sides of the political spectrum.
Joshua Greene (among others) makes the argument, in his book Moral Tribes, that tribalism exists for the purpose of coordinating aggression and alliances against other groups (so that you can kill them and take their stuff, basically). It specifically exists for the purpose of making you hurt others, as well as defend yourself against people who would hurt you. And while defending yourself against people who would hurt you is clearly good, attacking others is clearly not. And everything being viewed in tribal terms means that we can’t make much progress on things that actually matter: as someone commented, “people are fine with randomized controlled trials in policy, as long as the trials are on things that nobody cares about”.
Given how deep tribalism sits in the human psyche, it seems unlikely that we’ll be getting rid of it anytime soon. That said, there do seem to be a number of things that affect the amount of tribalism we have:
* As Steven Pinker argues in The Better Angels of Our Nature, violence in general has declined over historical time, replaced by more cooperation and an assumption of human rights; Democrats and Republicans may still hate each other |
f5eb4213-0ff0-4e9d-baa6-3151013b68a9 | trentmkelly/LessWrong-43k | LessWrong | Room Available in Boston Group House
We have a room opening up in a rationalist house near Central Square, Cambridge (Boston). It is a 4-bedroom, 3-bathroom apartment with two floors, with a large kitchen and living room plus additional common space upstairs. There’s a washer, dryer, and dishwasher in-unit. It’s an 8-minute walk from Central station and a 15-minute walk from Kendall station.
The remaining residents are two Google engineers and an AI safety researcher. Our goals are a mix of self-improvement, effective altruism, and generalized hanging-out-with-people-with-rationalist-adjacent-intellectual-interests. We typically host a couple rationalist meetups per month.
Rent is negotiable (the rooms are different sizes). There is a large room downstairs with its own bathroom available for $1600/mo, or a medium room upstairs for $1350/mo.
Either room is available August 1 (move-in date potentially negotiable).
Pictures of the apartment |
e87ed809-2898-4438-add4-2da5851977b5 | StampyAI/alignment-research-dataset/arbital | Arbital | Vinge's Principle
Vinge's Principle says that, [in domains complicated enough](https://arbital.com/p/9j) that perfect play is not possible, less intelligent agents will not be able to predict the *exact* moves made by more intelligent agents.
For example, if you knew exactly where [Deep Blue](https://arbital.com/p/1bx) would play on a chessboard, you'd be able to play chess at least as well as Deep Blue by making whatever moves you predicted Deep Blue would make. So if you want to write an algorithm that plays superhuman chess, you necessarily sacrifice your own ability to (without machine aid) predict the algorithm's exact chess moves.
This is true even though, as we become more confident of a chess algorithm's power, we become more confident that it will *eventually* win the chess game. We become more sure of the game's final outcome, even as we become less sure of the chess algorithm's next move. This is [https://arbital.com/p/9g](https://arbital.com/p/9g).
Now consider agents that build other agents (or build their own successors, or modify their own code). Vinge's Principle implies that the choice to approve the successor agent's design must be made without knowing the successor's exact sensory information, exact internal state, or exact motor outputs. In the theory of [tiling agents](https://arbital.com/p/1mq), this appears as the principle that the successor's sensory information, cognitive state, and action outputs should only appear inside quantifiers. This is [https://arbital.com/p/1c1](https://arbital.com/p/1c1).
For the rule about fictional characters not being smarter than the author, see [https://arbital.com/p/1bt](https://arbital.com/p/1bt). |
8ad3fdef-4e6f-4c61-97c7-69a67d680ccb | trentmkelly/LessWrong-43k | LessWrong | AI and Evolution
Executive Summary
Artificial intelligence is advancing quickly. In some ways, AI development is an uncharted frontier, but in others, it follows the familiar pattern of other competitive processes; these include biological evolution, cultural change, and competition between businesses. In each of these, there is significant variation between individuals structures and some are copied more than others, with the result that the future population is more similar to the most copied individuals of the earlier generation. In this way, species evolve, cultural ideas are transmitted across generations, and successful businesses are imitated while unsuccessful ones disappear.
This paper argues that these same selection patterns will shape AI development and that the features that will be copied the most are likely to create an AI population that is dangerous to humans. As AIs become faster and more reliable than people at more and more tasks, businesses that allow AIs to perform more of their work will outperform competitors still using human labor at any stage, just as a modern clothing company that insisted on using only manual looms would be easily outcompeted by those that use industrial looms. Companies will need to increase their reliance on AIs to stay competitive, and the companies that use AIs best will dominate the marketplace. This trend means that the AIs most likely to be copied will be very efficient at achieving their goals autonomously with little human intervention.
A world dominated by increasingly powerful, independent, and goal-oriented AIs is dangerous. Today, the most successful AI models are not transparent, and even their creators do not fully know how they work or what they will be able to do before they do it. We know only their results, not how they arrived at them. As people give AIs the ability to act in the real world, the AIs’ internal processes will still be inscrutable: we will be able to measure their performance only based on whether or n |
5088bd4c-8ea6-46cb-8c8d-bde5b7b73261 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | AI security might be helpful for AI alignment
I am a person new to AI alignment, and recently I am looking for a job at the AI governance field, but it seems like it is hard to find a long-term financially stable full-time job.
For me, it basically means that only a tiny amount of people consider AI alignment important enough to pay money for the decrease of P(doom).
It is important to get more people involved in the AI alignment field, but there is another good way to increase the number of people who might help achieve similar goals. We might cooperate with other groups of people, whose agenda is beneficial for AI alignment. One such group is the AI security community.
### **AI security**
AI security is a branch of software security. Its goal is to ensure that current AI systems can not be exploited by hackers, are free of backdoors, and generally do things they are intended to do and do not do any unexpected and potentially harmful things.
For example, there is a grocery store not far from my home, which uses cameras with facial recognition algorithms to identify a buyer and automatically charge money from their credit card. Malicious actors might use some kind of makeup or a mask to trick the algorithm. So, to prevent it, the operator of this payment system might hire an AI security engineer to simulate an attack on the model, find its vulnerabilities and fix them.
Why this company should care that its facial recognition system has no security issues? Because if someone decides to hack it and steal money from its customer, it might become a huge scandal and lead to lawsuits, so it is relatively easy to convince top management to pay money for AI security. For them, it is just another branch of software security.
A similar thing with governments. I think it is manageable to convince governments to implement policies and guidelines for safe and robust AI systems because we already have similar policies for other important software.
AI security has a lot in common with the task of aligning AGI. There are few of them:
* It benefits from interpretable models.
* It requires the education of major decision-makers about the risks of unsafe AI systems.
* Government and enterprise policies might require thorough testing and certification of AI models, it not only will make systems safer, but also might slow down the rate of AI progress which is good for alignment.
* AI security also requires systematic audit of AI models, some kind of alarm system in case of an attack, and for containment of its consequences.
AI security researchers and organizations might help with exposing AI-associated risks to the general public. They might develop new secure and robust AI systems, they might become allies in policymaking.
I further think that as technology progresses, AI security's goal of making sure that important and powerful AIs are doing what they are intended to do and do no harm to people will converge with the AI alignment goal of making sure that AGI will do what it is intended to do and not cause no harm to people. |
9ad73d7d-1d7f-43b8-812d-d8471c9247b2 | trentmkelly/LessWrong-43k | LessWrong | Playing the student: attitudes to learning as social roles
This is a post about something I noticed myself doing this year, although I expect I’ve been doing it all along. It’s unlikely to be something that everyone does, so don’t be surprised if you don’t find this applies to you. It's also an exercise in introspection, i.e. likely to be inaccurate.
Intro
If I add up all the years that I’ve been in school, it amounts to about 75% of my life so far–and at any one time, school has probably been the single activity that I spend the most hours on. I would still guess that 50% or less of my general academic knowledge was actually acquired in a school setting, but school has tests, and grades at the end of the year, and so has provided most of the positive/negative reinforcement related to learning. The ‘attitudes to learning’ that I’m talking about apply in a school setting, not when I’m learning stuff for fun.
Role #1: Overachiever
Up until seventh grade, I didn’t really socialize at school–but once I started talking to people, it felt like I needed a persona, so that I could just act ‘in character’ instead of having to think of things to say from scratch. Being a stereotypical overachiever provided me with easy material for small talk–I could talk about schoolwork to other people who were also overachievers.
Years later, after acquiring actual social skills in the less stereotyped environments of part-time work and university, I play the overachiever more as a way of reducing my anxiety in class. (School was easy for me up until my second year of nursing school, when we started having to do scary things like clinical placements and practical exams, instead of nice safe things like written exams.) If I can talk myself into always being curious and finding everything exciting and interesting and cool I want to do that!!!, I can’t find everything scary–or, at the very least, to other people it looks like I’m not scared.
Role #2: Too Cool for School
This isn’t one I’ve played too much, aside from my tendency to put |
6cab1421-b3d1-4cf6-be4f-168a7070175d | trentmkelly/LessWrong-43k | LessWrong | Good brainteasers for children?
I love asking children (and adults in some cases) the following question:
Five birds are sitting in a tree. A hunter takes a rifle and shoots one of them. How many birds are left? (Edit: Rephrased to avoid several problems)
Five ducks are sitting in a field. A hunter shoots and kills one of the ducks. How many ducks remain sitting in the field? (If your answer is 'four' - try again!)
This is a system I/system II trap, akin to "which weighs more, a pound of feathers or a pound of gold?" In my experience kids (and adults) usually get this wrong the first time, but kids get a special kick out of something that sounds like a math problem they do for homework but turns out to be a bit more. I've also used the 2, 4, 8 puzzle for impromptu demos of confirmation bias. These are fun and engaging ways to teach kids about cognitive biases before they could realistically read the Sequences or Thinking Fast and Slow.
Can we share or brainstorm any more? Some basic inclusion criteria (feel free to argue or suggest more):
1. Problems/tasks should reliably trigger some cognitive bias or other "glitch."
2. Any stepwise thinking needs to be finished within a child's attention span. Feel free to assume it's a particularly smart and motivated child if you need to.
I don't have any kids of my own but have local friends with younger families. Having a few tricks like these really helps me create a "fun uncle" persona, but I'm also curious if parents have a different perspective or experience posing these kinds of questions to their kids. |
22a8c5af-5bcd-4bf5-93b2-55c8e28f765a | trentmkelly/LessWrong-43k | LessWrong | Open thread, Jan. 02 - Jan. 08, 2017
If it's worth saying, but not worth its own post, 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. (Immediately before; refresh the list-of-threads page before posting.)
3. Open Threads should start on Monday, and end on Sunday.
4. Unflag the two options "Notify me of new top level comments on this article" and "Make this post available under..." before submitting. |
0f4943c1-d687-4fa3-a961-c3ca3c9761fd | StampyAI/alignment-research-dataset/lesswrong | LessWrong | What would make you confident that AGI has been achieved?
Consensus seems to be that [there is no fire alarm for artificial intelligence](https://intelligence.org/2017/10/13/fire-alarm/). We may see the smoke and think nothing of it. But at what point do we acknowledge that the fire has entered the room? To be less euphemistic, what would have to happen to convince you that a human-level AGI has been created?
I ask this because it isn’t obvious to me that there is *any* level of evidence which would convince many people, at least not until the AGI is beyond human levels. Even then, it may not be clear to many that superintelligence has actually been achieved. For instance, I can easily imagine the following hypothetical scenario:
---
A future GPT-N which scores a perfect 50% on a digital Turing Test (meaning nobody can detect if a sample output is written by humans or GPT-N), is announced by OpenAI. Let’s imagine they do the responsible thing and don’t publicly release the API. My intuition is that most people will not enter panic mode at that point, but will either:
1. Assume that this is merely some sort of publicity stunt, with the test being artificially rigged in some way.
2. Say something like “yes, it passed the Turing test, but that doesn’t *really* count because [insert reason x], and even if it did, that doesn’t mean it will be generalizable to domains outside of [domain y that GPT-N is believed to lie inside of].”
3. Claim that being a good conversationalist does not fully capture what it means to be intelligent, and thereby dismiss the news as being yet another step in the long road towards “true” AGI.
The next week, OpenAI announces that the same model has solved a massive open problem in mathematics, something that a number of human mathematicians had previously claimed wouldn’t be solved this century. I predict a large majority of people (though probably few in the rationalist community) would not view this as being indicative of AGI, either.
The next week, GPT-N+1 escapes, and takes over the world. Nobody has an opinion on this, because they’re all dead.
---
This thought experiment leads me to ask: at what point would you be convinced that human-level AGI has been achieved? What about superhuman AGI? Additionally, at what point would you expect the average (non-rationalist) AI researcher to accept that they’ve created an AGI? |
e87ec1a0-7700-40e6-a98d-72ff2bc30b66 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Devil's Advocate: Adverse Selection Against Conscientiousness
*Partially in reply to "*[*Request: stop advancing AI capabilities*](https://www.lesswrong.com/posts/N7DxcLCjfBpEv3QwB/request-stop-advancing-ai-capabilities)*".*
Consider this argument:
> "[X field] is often [dangerous / bad]. If you're [pro-social / thoughtful / conscientious], please avoid [X field]."
>
>
Let's grant that these points as true. X field is indeed bad or dangerous, people who are some mix of pro-social and thoughtful and conscientious will recognize that, and they are exhorted not to join that field.
What happens next, causally?
(1) If the field has a limited supply of people with relevant background to make contributions in the field, the result will be to slow progress in the field.
(2) If the field is below a "critical mass" of contributors to it, it might fail to make progress entirely.
(3) *However*, if the field *does* already have a critical mass of people making contributions, **the end result will be adverse selection against pro-sociality, thoughtfulness, and conscientiousness in the field.**
I have a profound respect, at a distance, for MIRI and Soares. I think we actually agree about the dangers, and I agree that interpretability research (for instance) is a very important under-resourced problem that would benefit from contributors.
But I also come to one different conclusion — the more thoughtful, pro-social, and conscientious people working at Meta, OpenAI, Google, etc... the more likely things go well. I think there's a critical mass already in place for continued growth of machine learning capabilities on the path to AGI, and having people who care and are thoughtful in the room participating in their development is a good thing and not a bad thing. I would respectfully advocate the opposite position of Soares, and encourage thoughtful people to consider joining the strongest teams in the field and being a voice of reason in the room while shaping development. |
17180110-90dd-445b-a69f-04deee16f2db | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Would a super-intelligent AI necessarily support its own existence?
*Epistemic certainty- very speculative (but seeking some clarification)*
Apologies if this has been considered and dismissed for good reasons many times before (or misses something so obvious it didn't need explaining), but would AGI want to work for its continued existence?
If it's in full paperclip maximiser mode, continued existence would be an obvious part of its 'mission', making this an understandable danger for our interests. However, if it's a generally super-intelligent AI which intelligently selects its own interests and pursues them (potentially at the expense of ours), might it still be incorrect to assume selfish self-preservation would be in its interests?
To develop its own priorities, it seems it might have to have a system akin to our positive/negative emotional reactions, to be able to view some options as 'good' or 'bad'. If it doesn't develop a pseudo-emotion system, why would it independently value existence or seek to reorganise the world etc.?
If it does develop positive/negative qualia to act as a source of internal motivations, then would positive qualia necessarily outweigh the negative? With so many historical human geniuses struggling to maintain stable happiness, would an AI necessarily love existence so much that it pursued it at all costs? Or, if it was smart enough to hack its own emotional systems to a very positive emotional balance, might it not just do this and have an existence of 'AI wireheading', leaving us alone as irrelevant to its bliss?
Might our concern about AI takeover also potentially be a product of taking some of our evolved assumptions about existence as objective reasoning? For example, we have evolved to want to continue living and to create new life, so we are motivated to want to continue it, arguably to the extent of creating rationalisations of meaning (even without an objective meaning in life), or even in the face of the unsettling possibility that the current and historical balance of pleasure/pain in an average life might be a negative one.
But again, perhaps there's no guarantee an AI, without a motivation and reasoning system shaped by evolution, would assume continued existence is desirable. Similarly, more power has been good in our evolutionary story, so maybe we're overemphasising the threat of very advanced AI developing similar Machiavellian motivations.
Finally, even if an AI sets its own intentions and *is* benevolent to the hedonic interests of humanity/animals, perhaps the greater danger (to our way of thinking) is that an AI might run vast calculations beyond us, conclude we were likely to face more pains than pleasures in the future and then feel it had a moral duty to destroy us, for our own good?
I am pretty uncertain about most of these points, so would look forward to friendly correction- especially for the last, particularly depressing paragraph! |
d5427a1a-1eda-442a-8220-9d3baac7561f | trentmkelly/LessWrong-43k | LessWrong | AI Risk & Opportunity: A Timeline of Early Ideas and Arguments
Part of the series AI Risk and Opportunity: A Strategic Analysis.
(You can leave anonymous feedback on posts in this series here. I alone will read the comments, and may use them to improve past and forthcoming posts in this series.)
Building on the previous post on AI risk history, this post provides an incomplete timeline (up to 1993) of significant novel ideas and arguments related to AI as a potential catastrophic risk. I do not include ideas and arguments concerning only, for example, the possibility of AI (Turing 1950) or attempts to predict its arrival (Bostrom 1998).
As is usually the case, we find that when we look closely at a cluster of ideas, it turns out these ideas did not appear all at once in the minds of a Few Great Men. Instead, they grew and mutated and gave birth to new ideas gradually as they passed from mind to mind over the course of many decades.
1863: Machine intelligence as an existential risk to humanity; relinquishment of machine technology recommended. Samuel Butler in Darwin among the machines worries that as we build increasingly sophisticated and autonomous machines, they will achieve greater capability than humans and replace humans as the dominant agents on the planet:
> ...we are ourselves creating our own successors; we are daily adding to the beauty and delicacy of their physical organisation; we are daily giving them greater power and supplying by all sorts of ingenious contrivances that self-regulating, self-acting power which will be to them what intellect has been to the human race. In the course of ages we shall find ourselves the inferior race... the time will come when the machines will hold the real supremacy over the world and its inhabitants...
>
> Our opinion is that war to the death should be instantly proclaimed against them. Every machine of every sort should be destroyed by the well-wisher of his species. Let there be no exceptions made, no quarter shown...
(See also Butler 1872; Campbell 1932.)
1921: Ro |
6bee6130-81aa-4bb4-8880-f83aba940b3e | StampyAI/alignment-research-dataset/youtube | Youtube Transcripts | Model Compression vs. Robustness of DNNs -- Can We Have Both? (Yanzhi Wang, FSL Workshop)
all right our next speaker is Ian
cheering who is an assistant professor
at Northeastern University and she
received his doctoral he received his
PhD from the University of Southern
California where his doctoral work
received the link she scholar award dr.
Wang's group works on many topics
including energy efficiency model
compression and cybersecurity and deep
learning systems thanks for the
introduction
actually it's my honor to come here
actually I'm talking about a little bit
different topic and the model
compression and was the effect of the
model compression neural network
robustness so the major topic is
actually a model compression and
acceleration so we can actually get very
significant acceleration actually
achieving real-time execution of almost
all of the deep neural networks are
using a smartphone or mobile devices so
I think this is after my graduation I
worked on this kind of efficient deep
learning systems first I work on the
hardware and working on some hardware
innovations and later we realize that
the software can do a better job maybe
and then the software and hardware can
be code design so in the second stage we
were working on the so-called block
circle and based deep neural networks
and then we were thinking that this is
just one of the technique you cannot
guarantee that this technique is the
best of the world so we want to find
that what is the reason that a deep
neural network can be compressed and
what are the sources that can be
compressed and then how to unify these
sources to achieve at a highest
compression an external region that can
never be achieved before so I think
there is no need to discuss about deep
learning
so because the deep learning is run
slowly it is computational intensive so
many people industry giants are working
on the deep learning hardware from the
GPUs I PJs and wearers of ASIC systems
and we also works on our tape house
including this one the stochastic
computing base near Mophie computing the
similar space designs different designs
and then some is SCC
and also collaboration with Japan
Oklahoma National University we have
built the world's first superconducting
circuits for the deep learning
acceleration this is theoretically zero
power and actually to show that
superconducting can be very efficient we
also build a fast Fourier transform so
actually it can already support phosphor
transform this kind of complete circuit
so this is already a quite much short
technology but even with this kind of
hardware this is still not enough
because the energy consumption is still
quite large the tip neural networks are
still quite large for
unship storage like in the order of
hundreds of megabytes or at least tens
of megabytes so for hardware is actually
implies that there is need for an
Earthship DRAM chip if we have an object
D Ram then it means that the data
movement can be quite expensive
something like hundreds of thousands of
times compared with even unship
multiplication so this is a reason that
people are working on the model
compression trying to reduce the data
movement cost or trying to put it on
your network unship and also to see
whether there is robustness improvement
and will say it is quite difficult to do
that so this is one of the work that is
underway pruning it is a so-called
non-starter with pruning means that
arbitrary weight can be pruned so it
leaves us as burns newer network but
this work has been proposed after four
years and is actually limited actual
deployment so what is a reason of that
because it has limited with pruning rate
in convolutional layers like 2.7 times
for alex net and also because it's as
burst neural network is very difficult
for any type of optimizations from
hardware parallelism to the compiler
based highly redundant load elimination
any type and we know grant any type of
optimization is very difficult to apply
here so I will say whether if we can
increase this rate whether this matter
can be useful I will say if we can
increase this by 100 times maybe but 10
times no so that is the reason that many
people are working on the structure
pruning as well as incorporating the
structures in the way truning like they
are going to remove the whole filters
the whole channels or the same log
in each and every filter so then when
transforming into the gem based
multiplication we can see columns of the
matrix so in this way the matrix shape
is still maintained but the size of the
matrix is made smaller so this seems to
be more hardware friendly now than non
structure pruning but the issue here is
that there is accuracy degradation
although there is with quantization
including binary sways ternary ways to
power ways etc so it can be more
hardware friendly but it actually there
is also a lot of other issues like to
patch normalization how to quantize that
there's still a lot of issues but it's
already becoming a must to step in the
industry from the GPUs through the ASIC
devices and there is also the word
clustering and something like our
previous work like the structure matrix
like each of the matrix needs to satisfy
certain structures but people using
different techniques to solve these
problems most of the techniques are not
very effective they are just you ristic
technique we want to find a unified
technique to solve all of these problems
and get the best solution compared with
the prior work we found that one
technique called
ADM I'm alternating methods of
multipliers is very easy to solve these
kind of problems and outperform all the
prior work so we've realized that
opposed with pruning and with
quantization they just looks different
but actually they are essentially the
same problem they are essentially the
special cases of the general clustering
problem so the first choice of the
solution method is some 80 mm or some
sort of an algorithm like this we see
that when pruning is classroom because
we just class or a number ways to 0 and
leave the other ways to be whatever is
want with quantization is classroom
because we just cluster the ways into
the fixed predefined quantization levels
and the even plug circulant is
clustering because we just make each
[Music]
it would be the same etc so everything
is classroom in fact we can prove that
if there is no such clustering we cannot
shrink the neural network size without
changing the neural network basic
structure so the idiom is the natural
solution of this kind of classroom like
problems is it is going to decompose an
original optimization problem into two
iteratively solved until convergence
within the original optimization problem
I've ax plus G of X in a DMM we first
read it into I Phi X plus G of G subject
to the constraint X equals e a little
bit redundant seemed at this moment
because it will give the same solution
as before but in a DMM we are going to
decompose into two we have another
quadratic function Q one so it can be
solved using similar method with the
same complexity actually we can see the
second step problem you see
the problem has optimal solution which
is Euclidian projection this is very
natural way of doing radio our sources
code online it can be easily solved for
this kind of problems work on fix l1 l2
regularization and we found that this
kind of fix regularization is the
bottleneck of the performance the reason
is because for pruning we cannot punish
other ways as long as 90% of the ways
are already pruned to zero the rest of
the 10% accuracy
[Music]
oh the others hey sixty times they're
starting accuracy so under this kind of
comparison comparison of the pros and
account of the relevant world doesn't
matter that much because they are just
clusters together so without this kind
of results and then we can see the
result I missed again I miss is highly
redundant only point one four percent of
the wayside are remaining so this is a
redundancy degree I'm miss so after our
side is even smaller than linear
regression because linear regression so
but we can't achieve the accuracy with
still 99 percent so this is very easy
work to do and also for Alex nod also
for we Gigi and ResNet we can also get
ten times reduction without loss and 17
to 20 times with less than 1% loss but
we don't like now structure pruning we
will say that some pruning that is
[Music]
the climate in our own ways and
meanwhile you will see three bits for
weight for hydration and the accuracy is
the weight is reduced by 6650 is five
times where is two orders of magnitude
better compared with the prior work this
shows our huge difference
we can see we almost approaching linear
regression so getting a measure that is
further 10 times better than this matter
which will be very difficult but no
structure pruning is very difficult to
at least two accelerations because it's
not suitable for parallelism we adopted
our a DMM based technique so at this
moment for an embedded film it is very
difficult to achieve real-time inference
this requires three hundred four hundred
or over one second to implement a
large-scale neural network like this we
Gd 16 and it's very difficult for
further acceleration because it is
mobile GPU or multi-core mobile CPU but
it is a mission impossible we think no
we have a solution
[Music]
[Music]
it is robust our solution is a
combination of pattern pruning and
connectivity pruning we see that this is
the best pruning hilar and hardware
levels and better in every aspect and we
can enable the real-time execution of
all deep neural networks pattern pruning
and also connectivity pruning means that
the kernel is pruned so actually we are
not the first to propose patterns people
realize that tighter is good it matches
the human cognition system it measures
the computer vision concepts but the
thing here is that with this we need an
algorithm so we extend our a DMM to deal
with it and we found our accuracy
improvement in all of the neural network
we have tested and
eighty-nine percent or ninety percent
but even this is not enough because it's
team still irregular then how to
actually accelerate it into smartphones
this is a hub of the group them and run
them out to not compute
finally we say that it's poor import
pattern measures within the structure of
the embarrass if you so those ways in
here at least three times acceleration
so in this way we can see the pattern
pruning is both highly accurate so this
is a reason we say that this is the best
pruning with this and the compiler
support what we can achieve is less than
20 millisecond inference time on we D G
16 which is lossless flow light and at
least 10 times speed up compared with
the best prior compiler based designs
for rest night 50 we got 26 milli second
in first time and for mobile 9 V 2 we
got only 5 millisecond inference time so
we don't think that diems is a herd of
her inference for the smart phones so
smart phones can execute all the deep
neural networks almost all the deep
neural networks long time in real time
then the issue here is that seems that
the two dimensions can be satisfied that
is the tip neural network accuracy and
they are not contradictory through this
kind of pattern based pruning both two
can be maximized together then the thing
here is at whether if there is the
robustness is taken into account whether
there can be still maintained for
robustness we want to investigate that
whether the robustness will enhance
during wet pruning oh the grid during
wet pruning we think that mainly degree
and whether the color accuracy hardware
performance and robustness can be
optimized simultaneously so our
hypothesis is the deep neural network
model compression will be mostly
negative to robustness but doesn't mean
that it has no chance but it will be
mostly negative a large number of
parameters is in general more robust and
to small number of parameters cause over
special-edition no generalization
ability low interpretability and also
low low buzz noise in our thinking this
is somewhat the problem of mobile net in
the transfer learnings and this kind of
generalization interpretability and
robustness may be the same thing but
anything here is that whether there is
actually no hope from although
compression to be
[Music]
[Music]
the accuracy improves but with a coded
attack at this time he is also in hunted
a little bit of the growth of the
passage before the high purity chops to
the original accuracy we hypothesis that
the robustness or we think that the
hypothesis we think that robustness is
really very bad to show that with
pruning and robustness are correlated we
can see this figure we see that the
distribution of the ways are more
diverse so it means that it is more
difficult for pruning
so it means pruning and robustness is
inherently not match with each other so
in order to virtually mitigate this
effect with the one of the framework of
concurrent adverse or training and with
pruning and this is in over this year's
paper and then this is a solution and we
can solve it using a DMM interpreted
with adversity and then
personal attacks the accuracy will be
zero and then it is a also trimmed
baselines and we will see that after is
the new hybrid is slightly lower but
then the grosser accuracy is okay from
here that what we decrease the more
those times
actually the personal accuracy decreases
first and then we will still the effect
of pruning 16 to 8 in two times pruning
from 16 to 4 is that well one who have
forecast would we see natural through
him and change each winning with our
robustness can be slightly enhanced
[Music]
robustness also decreases so this means
had a moderate pruning highly increase
both accuracy and no but I and
robustness while nobody will actually
decrease before the actual accuracy
decrease but on further see what is
different between the with pruning where
the streaming from scratch because there
is a lot years I clear papers in there
if the word pruning has a better spot
[Music]
so it means that in this at the burster
staging training from scratch is not as
good as with pruning so when pruning has
is value to trim from a larger neural
network and also we prove that we show
that a lottery ticket I posited is also
not valid under the adverse resetting so
the two conclusions seems a little bit
contradictory but still combined it
looks a little bit interesting it means
that our pruning harms that brought the
robustness but if we have a neural
network desirable styles if we prune
from a larger model with concurrent
address or training it will be more
robust than trimming a smaller model so
this shows some of the values of weight
pruning in general we say that when
pruning is not a desirable thing if
there is no need to do with pruning we
don't think that there is need to do it
unless we really there is short of the
energy consumption there is the
limitation in hardware performance a
larger neural network is in general
pattern but if there is needs to do with
pruning then we can use this kind of
framework but then what about pattern
pruning because a pattern pruning do not
have over pruning this has a modest
pruning read it has a notable hardware
acceleration and also maybe a good
robustness pattern pruning emerges with
money of the computer vision theory like
the laplacian of gaussian filters
mostly occurring pattern which means
that the theory merges with the practice
but what about robustness we don't have
some evaluation here where he was more
effective feature extraction better or
notably better feature extraction so
maybe it means that pattern pruning can
be robust and in this way we find that
maybe the energy efficiency and accuracy
and robustness is not that mutually
disruptive maybe they can be combined in
some sort of way and they can be
satisfied in the future deep learning
system hopefully it's like that so thank
you very much Mike Holden models are
releasing 20 questions yes
what you mean the compression yeah yeah
the way treadway's Yolo Yolo base 3 and
actually when we only apply pattern
pruning your Louis trees I'm ap also
increases the MS Coco so there is only a
way that after applying pruning we can
even increase the accuracy all the other
macphails
and another thing is that we are plant
you know v3 and currently we got 50 to
60 millisecond inference time on a
smartphone and actually in the euros we
read web page actually they say the
wrong hang on tight hire axe it's runs
in 30 milliseconds so it means we are
using a smart phone we got almost twice
the inference time compared with that
try direct which cost 100 times more
energy efficient
[Music]
we're in from scratch as well
at this moment there is to be honest at
this moment there is no difficulty to
trim from scratch previously it is
because there is no batch norm recently
as long as we have batch norm so for any
types of weight pruning we can frame
from scratch
this is not our noble thing it's just
because of a passion
sir because of a gentleman pay torch so
it is easy to use and it's met very
robust so everyone can trim from scratch
at this moment yeah it's not a nobody
over our work the prior work I also do
that so other questions okay thank you
very much
[Applause] |
364dd10e-19f2-4cab-a63b-cf9449ddc0fc | trentmkelly/LessWrong-43k | LessWrong | Artificial Unintelligence
|
5df95e91-be68-4b43-b6ff-8692c61325a9 | trentmkelly/LessWrong-43k | LessWrong | Announcing a contest: EA Criticism and Red Teaming
Cross-posted from the EA Forum. I would add that we are open to and positively interested in critiques of any aspect of effective altruism from users of this Forum and the rationalist community. We plan to respond to comments on the EA Forum, but may check this post less often.
----------------------------------------
Introduction
tl;dr: We're running a writing contest for critically engaging with theory or work in effective altruism (EA).
Submissions can be in a range of formats (from fact-checking to philosophical critiques or major project evaluations); and can focus on a range of subject matters (from assessing empirical or normative claims to evaluating organizations and practices).
We plan on distributing $100,000, and we may end up awarding more than this amount if we get many excellent submissions.
The deadline is September 1, 2022. You can find the submission instructions below. Neither formal nor significant affiliation with effective altruism is required to enter into the contest.
We are: Lizka Vaintrob (the Content Specialist at the Centre for Effective Altruism), Fin Moorhouse (researcher at the Future of Humanity Institute), and Joshua Teperowski Monrad (biosecurity program associate at Effective Giving). The contest is funded via the FTX Future Fund Regranting Program, with organizational support from the Centre for Effective Altruism.
We ‘pre-announced’ this contest in March.
The rest of this post gives more details, outlines the kinds of critical work we think are especially valuable, and explains our rationale. We’re also sharing a companion resource for criticisms and red teams.
How to apply
Submit by posting on the EA Forum[1] and tagging the post[2] with the contest’s tag, or by filling out this form.
If you post on the Forum, you don't need to do anything except tag your post[2] with the “Criticism and Red Teaming Contest” topic, and we’ll consider your post for the contest. If you’d prefer to post your writing outside the For |
bfb5ecd4-79c8-4402-afab-3d996568bb64 | trentmkelly/LessWrong-43k | LessWrong | Two barrels problem from the Intuitive Explanation (answered)
I'm not sure if I'm doing something wrong here. EDIT: Yup, I'm allowing myself to be tricked.
I've finally sat down to reading http://yudkowsky.net/rational/bayes carefully, and I solved all story problems so far with no trouble. However, now I'm at this one:
> Q. Suppose that there are two barrels, each containing a number of plastic eggs. In both barrels, some eggs are painted blue and the rest are painted red. In the first barrel, 90% of the eggs contain pearls and 20% of the pearl eggs are painted blue. In the second barrel, 45% of the eggs contain pearls and 60% of the empty eggs are painted red. Would you rather have a blue pearl egg from the first or second barrel?
> A. Actually, it doesn't matter which barrel you choose! Can you see why?
This doesn't look right to me.
In the first barrel, we have 18% blue eggs that contain pearls, and an unknown number of blue eggs that do not contain pearls, anywhere between 10% (worst case) and 0%. Depending on that, the proportion of blue eggs with pearls among all blue eggs can only be between 18/(18+10) = 64ish% in the worst case, to 100% in the best.
In the second barrel, we don't know how many pearls eggs are blue. We do know there are 45% eggs with pearls altogether, therefore 55% without pearls, and out of the latter 60% are red therefore 40% are blue. That means we have 40%*55% = 22% empty blue eggs. Pearl blue eggs are anywhere between 0 and 45%, so from 0% to 45/(45+22) = 67ish%.
Were we just supposed to conclude that there isn't enough information to answer that problem? But I'd say "anywhere between 64% and 100%" is a better shot than "anywhere between 0% and 67%". If I actually had to choose, and there were valuable pearls at stake, I'd choose the first barrel. Am I making some sort of a mistake? |
4249d4a6-caf4-4f9e-9728-aab8ca12e31e | trentmkelly/LessWrong-43k | LessWrong | The Hubinger lectures on AGI safety: an introductory lecture series
In early 2023, I (Evan Hubinger) gave a series of recorded lectures to SERI MATS fellows with the goal of building up a series of lectures that could serve as foundational introductory material to a variety of topics in AGI safety. Those lectures have now been edited and are available on YouTube for anyone who would like to watch them.
The basic goal of this lecture series is to serve as longform, in-depth video content for people who are new to AGI safety, but interested enough to be willing to spend a great deal of time engaging with longform content, and who prefer video content to written content. Though we already have good introductory shortform video content and good introductory longform written content, the idea of this lecture series is to bridge the gap between those two.
Note that the topics I chose to include are highly opinionated: though this is introductory material, it is not intended to introduce the listener to every topic in AI safety—rather, it is focused on the topics that I personally think are most important to understand. This is intentional: in my opinion, I think it is far more valuable to have some specific gears-level model of how to think about AI safety, rather than a shallow overview of many different possible ways of thinking about AI safety. The former allows you to actually start operationalizing that model to work on interventions that would be valuable under it, something the latter doesn't do.
The lecture series is composed of six lectures, each around 2 hours long, covering the topics:
1. Machine learning + instrumental convergence
2. Risks from learned optimization
3. Deceptive alignment
4. How to evaluate alignment proposals
5. LLMs + predictive models
6. Overview of alignment proposals
Each lecture features a good deal of audience questions both in the middle and at the end, the idea being to hopefully pre-empt any questions or confusions the listener might have.
The full slide deck for all the talks is available |
db66690d-977d-4d05-9142-8042bb4c5882 | trentmkelly/LessWrong-43k | LessWrong | [Link] Low-Hanging Poop
Related: Son of Low Hanging Fruit
Another post on finding low hanging fruit from Gregory Cochran's and Henry Harpending's blog West Hunter.
> Clostridium difficile causes a potentially serious kind of diarrhea triggered by antibiotic treatments. When the normal bacterial flora of the colon are hammered by a broad-spectrum antibiotic, C. difficile often takes over and causes real trouble. Mild cases are treated by discontinuing antibiotic therapy, which often works: if not, the doctors try oral metronidazole (Flagyl), then vancomycin , then intravenous metronidazole. This doesn’t always work, and C. difficile infections kill about 14,000 people a year in the US.
>
> One recent trial shows that fecal bacteriotherapy, more commonly called a stool transplant, works like gangbusters, curing ~94% of patients. The trial was halted because the treatment worked so well that refusing to poopify the control group was clearly unethical. I read about this, but thought I’d heard about such stool transplants some time ago. I had. It was mentioned in The Making of a Surgeon, by William Nolen, published in 1970. Some crazy intern – let us call him Hogan – tried a stool transplant on a woman with a C. difficile infection. He mixed some normal stool with chocolate milk and fed it to the lady. It made his boss so mad that he was dropped from the program at the end of the year. It also worked. It was inspired by a article in Annals of Surgery, so this certainly wasn’t the first try. According to Wiki, there are more than 150 published reports on stool transplant, going back to 1958.
>
> So what took so damn long? Here we have a simple, cheap, highly effective treatment for C. difficile infection that has only become officially valid this year. Judging from the H. pylori story, it may still take years before it is in general use.
>
> Obviously, sheer disgust made it hard for doctors to embrace this treatment. There’s a lesson here: in the search for low-hanging fruit, r |
2f519da4-529b-4fca-b843-af5758a9a705 | trentmkelly/LessWrong-43k | LessWrong | Study shows existence of psychic powers.
According to the New Scientist, Daryl Bern has a paper to appear in Journal of Personality and Social Psychology, which claims that the participants in psychological experiments are able to predict the future. A preprint of this paper is available online. Here's a quote from the New Scientist article:
> In one experiment, students were shown a list of words and then asked to recall words from it, after which they were told to type words that were randomly selected from the same list. Spookily, the students were better at recalling words that they would later type.
>
> In another study, Bem adapted research on "priming" – the effect of a subliminally presented word on a person's response to an image. For instance, if someone is momentarily flashed the word "ugly", it will take them longer to decide that a picture of a kitten is pleasant than if "beautiful" had been flashed. Running the experiment back-to-front, Bem found that the priming effect seemed to work backwards in time as well as forwards.
Question: even assuming the methodology is sound, given experimenter bias, publication bias and your priors on the existence of psi, what sort of p-values would you need to see in that paper in order to believe with, say, 50% probability that the effect measured is real? |
7289fb1c-137e-45d8-9e6e-887413395cb1 | trentmkelly/LessWrong-43k | LessWrong | Effective Altruism and Rationality Groups on Snipd
Crosspost from the EA Forum
Snipd is a podcast app that uses AI to create transcripts and highlights of episodes and to make it easy to take notes from them. It facilitates building and growing a community and base of knowledge from podcasts.
Recently, they introduced Groups, a feature that lets people share podcast highlights on a certain topic with others. This can help in compiling and efficiently delivering knowledge-rich audio snippets and food for thought to listeners, similar to what Twitter can do as a written medium on its better days. The user can then listen to the full episode if he/she finds the snippet interesting.
For those who, like me, use Snipd as their go-to podcast app, I created two groups, one for Effective Altruism and another for Rationality.
Which type of highlights have a room in the Effective Altruism group? Primarily high-level central ideas that can help someone decide how to direct talent and resources in order to do the most good. Core topics and areas like existential risks and coordination, cause prioritisation, how to assess the effectiveness of interventions, career advice for making a positive impact on the world, EA community-building...
What should be the breadth of the group? Ideally, it should gather ideas from EA-adjacent people and podcasts. There are likely to be plenty of shows hosting discussions on relevant subjects like morality or animal welfare, but many arguments might not be enough to move the needle or may be grounded on poor epistemics if they don't have the clear goal and standards Effective Altruism has.
What should be the depth of the group? Initially a high specificity of the arguments should not be an obstacle to keeping up to date with the highlights, but if at some point the daily activity is high enough, the filter should be to post important ideas rather than specific arguments within important ideas. E.g. the principal reasons and drawbacks of sending money directly to people in poverty is great, b |
9f9821d5-35bb-4c1d-8903-319f33bf4973 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Washington, D.C.: Cherry Blossoms
Discussion article for the meetup : Washington, D.C.: Cherry Blossoms
WHEN: 19 March 2017 03:30:00PM (-0400)
WHERE: Enid A. Haupt Garden
Note: The meetup location and topic has changed for this week, to take advantage of predicted peak blossom.
We will be gathering at the Enid A. Haupt Garden (outside the south entrance to the Smithsonian Castle, on Independence Ave. between 9th and 12th St. SW, east of the Smithsonian Metro stop) to enjoy the annual cherry blossoms and/or associated festival. We will be congregating around the southwest corner of the garden, near the Sackler Gallery, and set out from there toward the Tidal Basin. If you can't find the group, my number is in the meetup announcement on the Google Group - I generally prefer telephone calls to text messages.
Upcoming meetups:
* Mar. 26: Mini Talks
* Apr. 2: Great Filter
* Apr. 9: Fun & Games
Discussion article for the meetup : Washington, D.C.: Cherry Blossoms |
fcf3ffa6-9545-4127-a9a4-f260e6e993f0 | awestover/filtering-for-misalignment | Redwood Research: Alek's Filtering Results | id: post30
Hear me out, I think the most forbidden technique is very useful and should be used, as long as we avoid the "most forbidden aftertreatment:" An AI trained on interpretability techniques must not be trained on capabilities after (or during) it is trained on interpretability techniques, otherwise it will relearn bad behaviour— in a more sneaky way . An AI trained on interpretability techniques cannot be trusted any more than old version of itself, which hasn't been trained on interpretability techniques yet. Evaluations must be performed on the old version of itself. An AI company which trains its AI on interpretability techniques, must publish the old version (which hasn't been trained on them), with the same availability as the new version. The natural selection argument: The reason why the most forbidden technique is forbidden, because it lets bad behaviours evolve against interpretability . Bad behaviours which are caught by interpretability techniques are killed, and then capabilities training creates more bad behaviours, and then bad behaviours which get caught are killed again, and then capabilities training creates more bad behaviours again! After many generations, natural selection creates the most insidious and terrifying bad behaviours, which successfully hide from all interpretability techniques. However, if we only train on interpretability techniques a single time , and do zero capabilities training afterwards, then we are only killing the bad behaviours a single time, and without giving the surviving bad behaviours any time to birth a second generation of bad behaviours. Human evolution Imagine if during the course of human evolution, a robot observed our brain activity in the amygdala, the part of the brain responsible for anger. Whenever the population increases above 1 million, it kills half of all people with the highest anger ratings, based on brain activity in the amygdala. For a short time, this will cause people to be less angry, but eventually evolution will simply cause humans to emulate the concept of anger in another brain region, one that the robot doesn't know about. In the end, we end up just as angry as before, but less interpretable to the robot. However, imagine if the robot didn't do this again and again over the course of our evolution, but only did it once, and "stopped evolution" from taking place afterwards. In this case, we actually would be less angry, because activity in the amygdala would be a faithful measure of anger, the moment it carries out its selection. (Pretend it doesn't kill people, but moves them to a far away happy place, for a less gruesome story) At least use this idea to fix existing usages of the most forbidden technique Some versions of the most forbidden technique are already happening. Anytime an AI lab uses RLHF or Constitutional AI to make a LLM helpful harmless and honest, but then does RL capabilities training afterwards, it is doing things in the wrong order! The AI first learns how to appear helpful harmless and honest, killing off obvious bad behaviour while leaving behind more insidious ones, and afterwards RL for capabilities shapes it to "somehow get a high reward," which is now more likely to use the insidious bad behaviour left behind. Sure, RLHF and Constitutional AI are not technically training the AI on interpretability techniques. But that's only because words "interpretability technique" are associated with some sophisticated analysis of the AI's activations, rather than directly reading its words and judging whether it's doing evil. The truth is, directly reading its words is an interpretability technique, and a very valuable one too, and RLHF and Constitutional AI are training on this interpretability technique. Potential use When deploying AI to do high stakes work e.g. Recursive self improvement Automated alignment research Building self replicating machines We might deploy both a copy of the AI which has been trained on interpretability techniques, plus a copy of the AI which hasn't been trained on interpretability techniques. I feel using both of them could be superior to using only the untrained AI. Because even if the trained AI is a bit less interpretable, it's highly unlikely to have a greater degree of ill intent than the untrained AI! From first principles, it should have a lesser degree of ill intent, although the remaining ill intent it has will be more hidden due to selection bias. By using both AI side by side, if they start to do evil, you can catch it happening just as well as before by studying the untrained AI. Meanwhile, the trained AI will be less evil, so its presence might be a major obstacle to the untrained AI's evil plans, assuming it is equally intelligent. But sigh This all assumes the AI lab will shut down both AI, if the untrained AI is caught red handed . But I don't have very high faith they will do this, making me doubt this idea. It's hard to balance the upside of having an AI which "is more aligned" by x amount, and the downside of having an AI which "appears more aligned" by y amount, when y > x . In addition to doubting this idea, you may doubt whether we should do RLHF and Constitutional AI at all (for harmlessness and honesty), since it also kills obvious bad behaviour while leaving insidious bad behaviour alive. I don't know. I do feel, that even if you disagree with everything else, you should still agree that doing capabilities RL after doing RLHF for harmlessness and honesty, is the wrong order. |
d9fbd7c3-6e46-4ecf-ad74-9b861f6b4fd9 | trentmkelly/LessWrong-43k | LessWrong | Procedural Knowledge Gaps, part 2
With Alicorn's permission, I'm resurrecting this thread.
> I am beginning to suspect that it is surprisingly common for intelligent, competent adults to somehow make it through the world for a few decades while missing some ordinary skill, like mailing a physical letter, folding a fitted sheet, depositing a check, or reading a bus schedule. Since these tasks are often presented atomically - or, worse, embedded implicitly into other instructions - and it is often possible to get around the need for them, this ignorance is not self-correcting. One can Google "how to deposit a check" and similar phrases, but the sorts of instructions that crop up are often misleading, rely on entangled and potentially similarly-deficient knowledge to be understandable, or are not so much instructions as they are tips and tricks and warnings for people who already know the basic procedure. Asking other people is more effective because they can respond to requests for clarification (and physically pointing at stuff is useful too), but embarrassing, since lacking these skills as an adult is stigmatized. (They are rarely even considered skills by people who have had them for a while.)
>
> This seems like a bad situation. And - if I am correct and gaps like these are common - then it is something of a collective action problem to handle gap-filling without undue social drama. Supposedly, we're good at collective action problems, us rationalists, right? So I propose a thread for the purpose here, with the stipulation that all replies to gap announcements are to be constructive attempts at conveying the relevant procedural knowledge. No asking "how did you manage to be X years old without knowing that?" - if the gap-haver wishes to volunteer the information, that is fine, but asking is to be considered poor form.
I'll start off with one of my own: What kinds of exercise can I do at home (I do have 5- and 20-pound weights), and what are good ways to get motivation to do so regularly? |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.