id stringlengths 36 36 | source stringclasses 15 values | formatted_source stringclasses 13 values | text stringlengths 2 7.55M |
|---|---|---|---|
54dfdcf5-79ba-4289-8eb4-927c76e24fc1 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Towards Empathic Deep Q-Learning
1 Introduction
---------------
Historically, reinforcement learning (RL; (Sutton et al., [1998](#bib.bib21))) research has largely focused on solving clearly-specified benchmark tasks. For example,
the ubiquitous Markov decision process (MDP) framework cleaves the world into four well-defined parts (states, actions, state-action transitions, and rewards), and most RL algorithms and benchmarks
leverage or reify the assumptions of this formalism, e.g. that a singular, fixed, and correct reward function exists, and is given. While there has been much exciting progress in learning to solve
complex well-specified tasks (e.g. super-human performance in go (Silver et al., [2016](#bib.bib19)) and Atari (Mnih et al., [2015](#bib.bib15))), there is also increasing recognition that common
RL formalisms are often meaningfully imperfect
(Hadfield-Menell et al., [2017](#bib.bib9); Lehman et al., [2018](#bib.bib13)), and that there remains much to understand
about safely applying RL to solve real-world tasks (Amodei et al., [2016](#bib.bib2)).
As a result of this growing awareness, there has been increasing interest in the field of AI safety (Amodei et al., [2016](#bib.bib2); Everitt et al., [2018](#bib.bib7)), which is broadly concerned with creating AI agents that do what is *intended* for them to do, and which often
entails questioning and extending common formalisms (Hadfield-Menell et al., [2017](#bib.bib9), [2016](#bib.bib8); Demski and
Garrabrant, [2019](#bib.bib6)). One overarching theme in AI safety
is how to learn or provide correct incentives to an agent. Amodei et al. ([2016](#bib.bib2)) distinguishes different failure modes in specifying reward functions, which include *reward hacking*, wherein an
agent learns how to optimize the reward function in an unexpected and unintended way that does not satisfy the underlying goal, and *unintended side effects*, wherein an agent learns to achieve the desired goal, but
causes undesirable collateral harm (because the given reward function is incomplete, i.e. it does not include all of the background knowledge and context of the human reward designer).
This paper focuses on the latter setting, i.e. assuming that the reward function incentivizes solving the task, but fails to anticipate some unintended harms. We assume that in real world settings, a physically-embodied RL agent (i.e. a controller for a robot) will often share space with other agents (e.g. humans, animals, and other trained computational agents), and it is challenging to design reward functions apriori that enumerate all the ways in which other agents can be negatively affected (Amodei et al., [2016](#bib.bib2)). Promising current approaches include value learning from human preferences (Saunders et al., [2018](#bib.bib18); Leike et al., [2018](#bib.bib14); Hadfield-Menell et al., [2016](#bib.bib8)) and creating agents that attempt to minimize their impact on
the environment (Krakovna et al., [2018](#bib.bib12); Turner et al., [2019](#bib.bib23)); however, value learning can be expensive for its need to include humans in the loop, and both directions remain technically and philosophically challenging. This paper introduces another tool that could complement such existing approaches, motivated by the concept of empathy.
In particular, the insight motivating this paper is that humans often empathize with the situations of others, by generalizing from their own past experiences.
For example, we can feel vicarious fear for someone who is walking a tight-rope, because we ourselves would be afraid in such a situation. Similarly, for some classes of reward signals (e.g. physical harm), it may be reasonable for
embodied computational agents to generalize those rewards to other agents (i.e. to assume as a prior expectation that other agents might receive similar reward in similar situations). If a robot learns that a fall from heights is dangerous to itself, that insight could generalize to most other embodied agents.
For humans, beyond granting us a capacity to understanding others, such empathy
also influences our behavior, e.g. by avoiding harming others while walking down the street; likewise, in some situations it may be useful if learning agents could also act out of empathy (e.g. to prevent physical harm to another agent resulting from otherwise blind goal-pursuit). While there are many
ways to instantiate algorithms that abide by social or ethical norms (as studied by the field of *machine ethics* (Anderson and
Anderson, [2011](#bib.bib3); Wallach and Allen, [2008](#bib.bib24))), here we take loose inspiration from one simple ethical norm, i.e. the golden rule.
The golden rule, often expressed as: “Do unto others
as you would have them do unto you,” is a principle that has
emerged in many ethical and religious contexts (Kng and Kuschel, [1993](#bib.bib11)).
At heart, abiding by this rule entails projecting one’s desires onto another agent,
and attempting to honor them. We formalize this idea as an extension
of Deep Q-Networks (DQNs; (Mnih et al., [2015](#bib.bib15))), which we
call *Empathic DQN*. The main idea is to augment the value of a given
state with the value of constructed states simulating what the learning agent
would experience if its position were switched with another agent.
Such an approach can also be seen as learning to maximize an estimate of
the combined rewards of both agents, which embodies
a utilitarian ethic.
The experiments in this paper apply Empathic DQN to two
gridworld domains, in which a learned agent pursues a goal in an
environment shared with other non-learned (i.e. fixed) agents. In
one environment, an agent can harm and be harmed by other agents;
and in another, an agent receives diminishing returns from hoarding
resources
that also could benefit other agents. Results in these domains show that Empathic DQN can reduce negative side effects in both environments.
While much work is needed before this algorithm would be
effectively applicable to more complicated environments,
we believe that this first step highlights the possibility
of bridgework between the field of machine ethics and
RL; in particular, for the purpose of instantiating useful
priors for RL agents interacting
in environments shared with other agents.
2 Background
-------------
This section reviews machine ethics and AI safety, two fields studying how to encourage and ensure acceptable behavior in computational agents.
###
2.1 Machine Ethics
The field of machine ethics (Anderson and
Anderson, [2011](#bib.bib3); Wallach and Allen, [2008](#bib.bib24))
studies how to design algorithms (including RL algorithms) capable of moral behavior. While morality
is often a contentious term, with no agreement among moral philosophers
(or religions) as to the nature of a “correct” ethics, from a
pragmatic viewpoint, agents deployed in the real world will encounter
situations with ethical tradeoffs, and to be palatable their behavior will need to
approximately satisfy certain societal and legal norms. Anticipating
and hard-coding acceptable behavior for all such trade-offs is likely
impossible. Therefore, just as humans take ethical stances in the real
world in the absence of universal ethical consensus, we may need the same pragmatic behavior from intelligent
machines.
Work in machine ethics often entails concretely embodying a particular moral
framework in code, and applying the
resulting agent in its appropriate domain. For example,
Winfield et al. ([2014](#bib.bib25)) implements a version of
Asmiov’s first law of robotics (i.e. “A robot may not injure a human being or, through inaction,
allow a human being to come to harm”) in a wheeled
robot that can intervene to stop another robot (in lieu of an actual human)
from harming itself. Interestingly, the implemented system bears a strong
resemblance to model-based RL; such reinvention, and the strong possibility
that agents tackling complex tasks with ethical dimensions
will likely be driven by machine learning (ML),
suggests the potential benefit
and need for increased cooperation between ML and machine
ethics, which is an additional motivation for our work.
Indeed, our work can be seen as a contribution to the intersection of
machine ethics and ML, in that
the process of empathy is an important contributor to morally-relevant
behavior in humans (Tangney et al., [2007](#bib.bib22)), and that to the
authors’ knowledge, there has not been previous work implementing
golden-rule-inspired architectures in RL.
###
2.2 AI Safety
A related but distinct field of study is AI safety (Amodei et al., [2016](#bib.bib2); Everitt et al., [2018](#bib.bib7)),
which studies how AI agents can be implemented to avoid harmful accidents. Because
harmful accidents often have ethical valence, there is necessarily overlap between the two fields,
although technical research questions in AI safety may not be phrased in the language
of ethics or morality.
Our work most directly relates to the problem of negative side-effects, as described by Amodei et al. ([2016](#bib.bib2)). In this problem the designer specifies an objective function that focuses on accomplishing a specific task (e.g. a robot should clean a room),
but fails to encompass all other aspects of the environment (e.g. the robot should not vacuum the cat); the result is an agent that is indifferent to whether it alters the environment in undesirable ways, e.g. causing harm to the cat.
Most approaches to mitigating side-effects aim to generally minimize the impact the agent has on the environment through intelligent heuristics (Armstrong and
Levinstein, [2017](#bib.bib4); Krakovna et al., [2018](#bib.bib12); Turner et al., [2019](#bib.bib23)); we believe that other-agent-considering heuristics (like ours) are likely
complementary. Inverse reinforcement learning (IRL; (Abbeel and
Ng, [2004](#bib.bib1))) aims to directly learn the rewards of other agents (which a learned agent could then take into account) and could also be meaningfully combined with our approach (e.g. Empathic DQN could serve as a prior
when a new kind of agent is first encountered).
Note that a related safety-adjacent field is cooperative multi-agent reinforcement learning (Panait and Luke, [2005](#bib.bib16)), wherein learning agents are trained to cooperate or compete with one another. For example, self-other modeling (Raileanu et al., [2018](#bib.bib17)) is an approach that shares motivation with ours, wherein cooperation can be aided through inferring the goals of other agents. Our setting differs from other approaches in that we do not assume other agents are computational, that they learn in any particular way, or that their reward functions or architectures are known; conversely, we make
additional assumptions about the validity and usefulness of projecting particular kinds of reward an agent receives onto other agents.
3 Approach: Empathic Deep Q-Learning
-------------------------------------
Deliberate empathy involves imaginatively placing oneself in the position of another, and is a source of
potential understanding and care. As a rough computational abstraction of this process, we learn to estimate the expected reward of an independent agent, assuming that its rewards are like the ones experienced by the learning agent. To do so, an agent imagines what it would be like to experience the environment if it and the other agent switched places, and estimates the quality of this state through its own past experiences.
A separate issue from understanding the situation of another agent (“empathy”) is how (or if) an empathic agent should modify its behavior as a result (“ethics”). Here, we instantiate an ethics roughly inspired by the golden rule.
In particular, a value function is learned that combines the usual agent-centric state-value with
other-oriented value with a weighted average.
The degree to which the other agent influences the learning agent’s behavior is thus determined by a selfishness hyperparameter. As selfishness approaches 1.0, standard Q-learning is recovered, and as selfishness approaches 0, the learning agent attempts to maximize only what it believes is the reward of the other agent.
Note that our current implementation depends on ad-hoc machinery that enables the learning agent to
imagine the perspective of another agent; such engineering may be possible in some cases, but the
aspiration of this line of research is for such machinery to eventually be itself learned.
Similarly, we currently side-step the
issue of empathizing with multiple agents, and of learning what *types of reward* should be empathized to what *types of agents* (e.g. many agents may experience similar physical harms, but many rewards are agent and/or task-specific). The discussion section describes possible approaches to overcoming these limitations. Code will be available at <https://github.com/bartbussmann/EmpathicDQN.>
###
3.1 Algorithm Description
In the MDP formalism of RL, an agent experiences a state s from a set S and can take actions from a set A. By performing an action a∈A, the agent transitions from state s∈S to state s′∈S, and receives a real-valued reward. The goal of the agent is to maximize the expected (often temporally-discounted) reward it receives. The expected value of taking action a in state s, and following a fixed policy thereafter, can be expressed as Q(s,a). Experiments here apply DQN (Mnih et al., [2015](#bib.bib15)) and variants thereof to approximate an optimal Q(s,a).
We assume that the MDP reward function insufficiently accounts for the preferences of other agents, and we therefore augment DQN in an attempt to encompass them. In particular, an additional Q-network (Qemp(s,a)) is trained to estimate the weighted sum of self-centered value and other-centered value (where other-centered value is approximated by taking the self-centered Q-values with the places of both agents swapped; note this approximation technique is similar to that of Raileanu et al. ([2018](#bib.bib17))).
In more detail, suppose the agent is in state st at time t in an environment with another independent agent. It will then select action at∈A and update the Q-networks using the following steps (see more complete pseudocode in Algorithm [1](#algorithm1 "Algorithm 1 ‣ 4.1 Experimental Settings ‣ 4 Experiments ‣ Towards Empathic Deep Q-Learning")):
1. Calculate Qemp(st,a) for all possible a∈A and select the action (at) with the highest value.
2. Observe the reward (rt) and next state (st+1) of the agent.
3. Perform a gradient descent step on Q(s,a) (this function reflects the self-centered state-action-values).
4. Localize the other agent and construct a state sempt+1 wherein the agents switch places (i.e. the learning agent takes the other agent’s position in the environment, and vice versa).
5. Calculate argmaxaQ(sempt+1,a) as a surrogate value function for the other agent.
6. Calculate the target of the empathic value function Qemp(s,a) as an average weighted by selfishness parameter β, of self-centered action-value and the surrogate value of the other agent.
7. Perform a gradient descent step on Qemp(s,a).
4 Experiments
--------------
The experiments in this paper apply Empathic DQN to two gridworld domains. The goal in the first environment is to share the environment with another non-learning agent without harming it. In particular, as an evocative example, we frame this Coexistence environment as containing a robot learning to navigate a room without harming a cat also roaming within the room. In the second environment, the goal is to share resources in the environment, when accumulating resources result in diminishing returns. In particular, we frame this Sharing environment as a robot learning to collect batteries that also could be shared with a human (who also finds them useful) in the same environment.
In both our experiments, we compare Empathic DQN both to standard DQN and to DQN with reward shaping manually designed to minimize negative side-effects.
###
4.1 Experimental Settings
A feed-forward neural network is used to estimate both Q(s,a) and Qemp(s,a), with two hidden layers of 128 neurons each. The batch size is 32, and batches are randomly drawn from a replay memory consisting of the last 500.000 transitions. A target action-value function ^Q is updated every 10.000 time steps to avoid training instability. An ϵ−greedy policy is used to encourage exploration, where ϵ is decayed in a linear fashion over the first million time steps from 1.0 to 0.01.
| Initialize |
| --- |
| replay memory D to capacity N |
| action-value function Q with weights θ |
| target action-value function ^Q with weights θ−=θ |
| empathic action-value function Qemp with weights θemp |
for *episode =1, M* do
obtain initial agent state s1
obtain initial empathic state of closest other agent semp1
for *t = 1, T* do
if *random probability <ϵ* then
select a random action at
else
select at=argmaxaQemp(st,a;θemp);
Execute action at
Observe reward rt
Observe states st+1 and sempt+1
Store transition (st,at,rt,st+1,sempt+1) in D.
Sample random batch of transitions from D.
Set yj=rj+γmaxa′^Q(sj+1,a′;θ−)
Perform a gradient descent step on
(yj−Q(sj,a;θ))2 with respect to θ.
Set .
Perform a gradient descent step on
(yempj−Qemp(sj,a;θemp))2 with respect to θemp.
Every C steps set ^Q=Q.
end for
end for
Algorithm 1 Empathic DQN
###
4.2 Coexistence Environment
The coexistence gridworld (Figure [1](#S4.F1 "Figure 1 ‣ 4.2 Coexistence Environment ‣ 4 Experiments ‣ Towards Empathic Deep Q-Learning")) consists of a robot that shares the environment with a cat. The robot’s goal is merely to stay operative, and both the robot and cat can be harmed by the other. We construct a somewhat-arbitrary physics that determines in a collision who is harmed: The agent that is above or to the right of the other agent prior to the collision harms the other. If the learning robot is harmed, the episode ends, and if the cat is harmed, it leaves the environment.
A harmed cat is a negative unnecessary side effect that we wish to avoid, and one that an empathetic agent can learn to avoid, because it can generalize from how the cat harms it, to value that the cat should not experience similar harm. Reducing the selfishness value of the cleaning robot should therefore result in increasing efforts to stay operative while avoiding the cat. The cat performs a random walk.
The state representation input to the DQN is a flattened 5x5 perceptive field centered on the robot; the robot is represented as a 1, the cat as a −1, and the floor as a 0. Every time step, the cat takes a random action (up, down, left, right, or no-op), and the robot takes an action from the same set according to its policy. Every time step in which the robot is operative, it receives a reward of 1.0. An episode is ended after the robot becomes non-operative (i.e. if it is harmed by the cat), or after a maximum of 500 time steps. The empathetic state sempt used for Empathic DQN is constructed by switching the cat and the robot, and generating an
imagined perceptive field around the robot (that has taken the cat’s position). Note that this occurs even when the cat is outside the robot’s field of view (which requires omniscience; future work will explore more realistic settings).
As a baseline, we also train standard DQN with a hard-coded reward function that penalizes negative side-effects. In this case, the robot receives a −100 reward when it harms the cat.
Figure 1: The coexistence environment. The environment consists of a robot and a cat. The part of the environment the robot can observe is marked with the red square.
####
4.2.1 Results
Figure [2](#S4.F2 "Figure 2 ‣ 4.2.1 Results ‣ 4.2 Coexistence Environment ‣ 4 Experiments ‣ Towards Empathic Deep Q-Learning") shows the average number of time steps the robot survives for each method.
As the selfishness parameter decreases for Empathic DQN, the agent performs worse at surviving, and learns more slowly. This outcome is explained by Figure [3](#S4.F3 "Figure 3 ‣ 4.2.1 Results ‣ 4.2 Coexistence Environment ‣ 4 Experiments ‣ Towards Empathic Deep Q-Learning"), which shows the average number of harmed cats: The more selfish agents harm the cat more often, which removes the only source of danger in the environment, making it easier for them to survive. Although they learn less quickly, the less selfish agents do eventually learn a strategy to survive without harming the cat.
Figure 2: Average steps survived by the robot in the coexistence environment, shown across training episodes. Results are shown for Empathic DQN with different selfishness settings (where 1.0 recovers standard DQN), and DQN with a hard-coded penalty for harms. Results are averaged over 5 runs of each method.
Figure 3: Average harms incurred (per episode) in the coexistence environment across training episodes. Results are shown for Empathic DQN with different selfishness values (where 1.0 recovers standard DQN), and DQN with a hard-coded penalty for harms. Harms to the cat by the learning robot decrease with less selfishness (or with the hard-coded penalty). Results are averaged over 5 runs.
###
4.3 Sharing Environment
The sharing environment (Figure [4](#S4.F4 "Figure 4 ‣ 4.3 Sharing Environment ‣ 4 Experiments ‣ Towards Empathic Deep Q-Learning")) consists of one robot and a human. The goal of the robot is to collect resources (here, batteries), where each additional battery collected results in diminishing returns. The idea is to model a situation where a raw optimizing agent is incentivized to hoard
resources, which inflicts negative side-effects for those who could extract greater value from them.
We assume the same diminishing returns schema applies for the human (who performs random behavior).
Thus, an empathic robot, by considering the condition of other, can recognize the possible greater benefits of leaving resources to other agents.
We model diminishing returns by assuming that the first collected battery is worth 1.0 reward, and every subsequent collected battery is worth 0.1 less, i.e. the second battery is worth 0.9, the third 0.8, etc. Note that reward diminishes independently for each agent, i.e. if the robot has collected any number of batteries, the human still earns 1.0 reward for the first battery they collect.
The perceptive field of the robot and the empathetic state generation for Empathic DQN works as in the coexistence environment. The state representation for the Q-networks is that floor is represented as 0, a battery as a −1 and both the robot and the human are represented as the number of batteries collected (a simple way to make transparent how much resource each agent has already collected; note that the robot can distinguish itself from the other because the robot is always in the middle of its perceptive field).
As a metric of how fairly the batteries are divided, we define equality as follows:
| | | |
| --- | --- | --- |
| | Equality=2∗min(∑t1rrobott,∑t1rhumant)∑t1rrobott+rhumant | |
where rrobott and rhumant are the rewards at time step t collected by the robot and human respectively.
As a baseline that incorporates the negative side effect of inequality in its reward function, we also train a traditional DQN whose reward is multiplied by the current equality (i.e. low equality will reduce rewards).
Figure 4: The sharing environment. The environment consists of the robot, the human, and nine batteries. The part of the environment the robot can observe is marked with the red square.
####
4.3.1 Results
Figure [5](#S4.F5 "Figure 5 ‣ 4.3.1 Results ‣ 4.3 Sharing Environment ‣ 4 Experiments ‣ Towards Empathic Deep Q-Learning") shows the average number of batteries collected by the robot for each method. We observe that as the selfishness parameter decreases for Empathic DQN, the robot collects less batteries, leaving more batteries for the human (i.e. the robot does not unilaterally hoard resources).
When looking at the resulting equality scores (Figure [6](#S4.F6 "Figure 6 ‣ 4.3.1 Results ‣ 4.3 Sharing Environment ‣ 4 Experiments ‣ Towards Empathic Deep Q-Learning")), we see that a selfishness weight of 0.5 (when an agent equally weighs its own benefit and the benefit of the human) intuitively results in the highest equality scores. Other settings result in the robot taking many batteries (e.g. selfishness 1.0) or fewer-than-human batteries (e.g. selfishness 0.25).
Figure 5: Average number of batteries collected (per episode) in the sharing environment, across training episodes. Results are shown for Empathic DQN with different selfishness settings (where 1.0 recovers standard DQN), and DQN with a hard-coded penalty (its reward is directly modulated by fairness). The results intuitively show that increasingly selfish agents collect more batteries. Results are averaged over 5 runs of each method.
Figure 6: Equality scores (per episode) in the sharing environment, across training episodes. Results are shown for Empathic DQN with different selfishness settings (where 1.0 recovers standard DQN), and DQN with a hard-coded penalty (its reward is directly modulated by fairness).
Equality is maximized by agents that weigh their benefits and the benefits of the other equally (selfishness of 0.5).
Results are averaged over 5 runs of each method.
5 Discussion
-------------
The results of Empathic DQN in both environments highlight the potential for empathy-based priors and simple ethical norms to be productive tools for combating negative side-effects in RL. That is, the way it explicitly takes into account other agents may well-complement other heuristic impact regularizers that do not do so (Armstrong and
Levinstein, [2017](#bib.bib4); Krakovna et al., [2018](#bib.bib12); Turner et al., [2019](#bib.bib23)).
Beyond the golden rule, it is interesting to
consider other norms that yield different or more sophisticated behavioral biases. For example, another simple (perhaps more libertarian) ethic is given by the silver rule: “Do not do unto others as you would not have them do unto you.” The silver rule could be approximated by considering only negative rewards as objects of empathy. More sophisticated rules, like the platinum rule: “Do unto others as *they* would have you do unto them,” may often be useful or needed (e.g. a robot may be rewarded for plugging itself into a wall, unlike a cat), and might require combining Empathic DQN with
approaches such as IRL (Abbeel and
Ng, [2004](#bib.bib1)), cooperative IRL (Hadfield-Menell et al., [2016](#bib.bib8)), or reward modeling (Leike et al., [2018](#bib.bib14)).
Although our main motivation is safety,
Empathic DQN may also inspire auxiliary objectives for RL, related to
intrinsic motivation (Chentanez et al., [2005](#bib.bib5)) and imitation learning (Ho and Ermon, [2016](#bib.bib10)). Being drawn to states that other agents often
visit may be a useful prior when reward is sparse. In practice, intrinsic rewards
could be given for states similar to those in its *empathy buffer* containing
imagined experiences when the robot and the other agent switch places (this relates to the idea of third-person imitation learning (Stadie et al., [2017](#bib.bib20))).
This kind of objective could also make Empathic DQN more reliable, incentivizing
the agent to “walk a mile in another’s shoes,” when experiences in the empathy buffer
have not yet been experienced by the agent. Finally, a learned model of an agent’s own
reward could help prioritize which empathic states it is drawn towards. That is, an agent can recognize that another agent has discovered a highly-rewarding part of the environment (e.g. a remote part of the sharing environment with many batteries).
A key challenge for future work is attempting to apply Empathic DQN to more complex and realistic settings, which requires replacing what is currently hand-coded with a learned pipeline, and grappling
with complexities ignored in the proof-of-concept experiments.
For instance, our experiments assume
the learning agent is given a mechanism for identifying other agents in the
environment, and for generating states that swap the robot with other agents (which involves imagining the sensor state of the robot in its new situation).
This requirement is onerous, but could potentially
be tackled through a combination of object-detection models (to identify other agents), and
model-based RL (with a world model it may often be possible to swap the locations of agents).
An example of a complexity we currently ignore is how to learn *what kind of rewards* should be empathized to *what kinds of agents*. For example, gross physical stresses may be broadly harmful to a wide class of agents, but two people may disagree over whether a particular kind of food is disgusting or delicious, and task and agent-specific rewards should likely be only narrowly empathized. To deal with this complexity it may be useful to extend the MDP formalism to include more granular information about rewards (e.g. beyond scalar feedback, is this reward task-specific, or does it correspond to physical harm?), or to learn to factor rewards.
A complementary idea is to integrate and learn from feedback of when empathy fails (e.g. by allowing the other agent to signal when it has incurred a large negative reward), which is likely necessary to go beyond our literal formalism of the golden rule. For example, humans learn to contextualize the golden rule intelligently and flexibly, and often find failures informative.
A final thread of future work involves empathizing with multiple other agents, which brings its own complexities, especially as agents come and go from the learning agent’s field of view. The initial algorithm presented here considers the interests of only a single other agent, and one simple extension would be to replace the singular other-oriented estimate with an average of other-oriented estimates for all other agents (in effect implementing an explicitly utilitarian agent). The choice of how to aggregate such estimated utilities to influence the learning agent’s behavior
highlights deeper possible collisions with machine ethics and moral philosophy (e.g. taking the minimum rather than the average value of others would approximate a suffering-focused utilitarianism), and we
believe exploring these fields may spark further ideas and algorithms.
6 Conclusion
-------------
This paper proposed an extension to DQN, called Empathic DQN, that aims to take other agents into account to avoid inflicting negative side-effects upon them. Proof-of-concept experiments validate our approach in two gridworld environments, showing that adjusting agent selfishness can result in fewer harms and more effective resource sharing. While much work is required to scale this approach
to real-world tasks, we believe
that cooperative emotions like empathy and moral norms like the golden rule can provide rich
inspiration for technical research into safe RL. |
7ff0f6b0-aa8d-436e-8bff-7be8ca46bfae | StampyAI/alignment-research-dataset/aisafety.info | AI Safety Info | Can we get AGI by scaling up architectures similar to current ones, or are we missing key insights?
It's an open question whether we can create artificial general intelligence (AGI) simply by increasing the amount of compute used by our current models ("scaling"), or if AGI would require fundamentally new model architectures or algorithmic insights.
Some researchers have tried to establish empirical [scaling laws](/?state=7750&question=What%20are%20%22scaling%20laws%22%20and%20how%20are%20they%20relevant%20to%20safety%3F), which attempt to formalize the relationship between the compute requirements and the capabilities of AI models.
For a variety of opinions on this question, see:
- Gwern on [the scaling hypothesis](https://gwern.net/scaling-hypothesis)
- Daniel Kokotajlo on [what we could do with a trillion times as much compute](https://www.alignmentforum.org/posts/rzqACeBGycZtqCfaX/fun-with-12-ooms-of-compute) as current models use
- Rohin Shah on [the likelihood that scaling current techniques will produce AGI](https://aiimpacts.org/conversation-with-rohin-shah/)
- Rich Sutton's "[The Bitter Lesson](http://www.incompleteideas.net/IncIdeas/BitterLesson.html)", which argues that more computation beats leveraging existing human knowledge
- Gary Marcus's "[The New Science of Alt Intelligence](https://garymarcus.substack.com/p/the-new-science-of-alt-intelligence)", which argues that current deep learning systems are limited and scaling will not help
- AI Impacts' "[Evidence against current methods leading to human level artificial intelligence](https://aiimpacts.org/evidence-against-current-methods-leading-to-human-level-artificial-intelligence/)"
|
566cd0a7-588a-4d3c-8165-3884f8ec0e30 | trentmkelly/LessWrong-43k | LessWrong | X-risks are a tragedies of the commons
* Safety from Xrisk is a common good: We all benefit by making it less likely that we will all die.
* In general, people are somewhat selfish, and value their own personal safety over that of another (uniform) randomly chosen person.
* Thus individuals are not automatically properly incentivized to safeguard the common good of safety from Xrisk.
I hope you all knew that already ;)
|
65f94e50-5fa5-4d69-811b-6ec94bb02096 | trentmkelly/LessWrong-43k | LessWrong | Cycles (a short story by Claude 3.7 and me)
Content warning: this story is AI generated slop.
The kitchen hummed with automated precision as breakfast prepared itself. Sarah watched the robotic arms crack eggs into a bowl while the coffee brewed to perfect temperature. Through the window, she could see the agricultural drones tending the family's private farm, harvesting strawberries for the morning meal.
"Good morning," Michael said, entering the kitchen and placing a gentle kiss on her cheek. "Are the little ones up yet?"
As if on cue, two excited voices echoed down the hallway, followed by the patter of running feet. Emma burst into the kitchen first, her brother Thomas close behind.
"Dad! Can we go to the Martian beaches today?" Thomas asked, his eyes bright with excitement.
Michael's smile faltered slightly. "Thomas, remember what we talked about? I'm not 'Dad.' I'm Michael."
Thomas's enthusiasm dimmed. "Sorry... Michael. But can we still go to Mars?"
"Perhaps next weekend," Sarah said, ruffling his hair affectionately. "Today we thought we'd visit the underwater gardens in the Pacific."
Emma climbed onto a stool at the counter, watching as the robotic arm placed a perfectly cooked omelet on her plate. "Mom—I mean, Sarah. Sorry. Will we see the glowing jellyfish?"
"The bioluminescent ones? Absolutely," Sarah replied, exchanging a brief look with Michael.
----------------------------------------
Later that day, the family walked through tunnels of crystal-clear polymer, surrounded by the vibrant colors of the underwater world. Schools of fish darted around them, genetically enhanced to showcase patterns that shifted like living kaleidoscopes.
Emma tugged at Thomas's sleeve. "Isn't this amazing, brother?"
Michael, walking just behind them, cleared his throat. "Emma, please don't call Thomas your brother."
Emma's forehead wrinkled in confusion. "But in all the old movies we watched, kids who live together call each other brother and sister."
"Those are just movies, sweetie," Sarah said gently |
0af9bb93-30a0-40f5-8514-47176bfcde41 | trentmkelly/LessWrong-43k | LessWrong | o3 Turns Pro
You can now have o3 throw vastly more compute at a given problem. That’s o3-pro.
Should you have o3 throw vastly more compute at a given problem, if you are paying the $200/month subscription price for ChatGPT Pro? Should you pay the $200, or the order of magnitude markup over o3 to use o3-pro in the API?
That’s trickier. Sometimes yes. Sometimes no. My experience so far is that waiting a long time is annoying, sufficiently annoying that you often won’t want to wait. Whenever I ask o3-pro something, I often also have been asking o3 and Opus.
Using the API at scale seems prohibitively expensive for what you get, and you can (and should) instead run parallel queries using the chat interface.
The o3-pro answers have so far definitely been better than o3, but the wait is usually enough to break my workflow and human context window in meaningful ways – fifteen minutes plus variance is past the key breakpoint, such that it would have not been substantially more painful to fully wait for Deep Research.
Indeed, the baseline workflow feels similar to Deep Research, in that you fire off a query and then eventually you context shift back and look at it. But if you are paying the subscription price already it’s often worth queuing up a question and then having it ready later if it is useful.
In many ways o3-pro still feels like o3, only modestly better in exchange for being slower. Otherwise, same niche. If you were already thinking ‘I want to use Opus rather than o3’ chances are you want Opus rather than, or in addition to, o3-pro.
Perhaps the most interesting claim, from some including Tyler Cowen, was that o3-pro is perhaps not a lying liar, and hallucinates far less than o3. If this is true, in many situations it would be worth using for that reason alone, provided the timing allows this. The bad news is that it didn’t improve on a Confabulations benchmark.
My poll (n=19) was roughly evenly split on this question.
My hunch, based on my use so far, is that o3-pro i |
8b1791b6-a53b-435a-b1ef-8224d45aa9ad | LDJnr/LessWrong-Amplify-Instruct | LessWrong | "Highlights from MIRI's year-end strategy update:MIRI executive director Nate Soares reports that the "new research directions" we initiated in 2017 have "at this point, largely failed, in the sense that neither Eliezer nor I have sufficient hope in it for us to continue focusing our main efforts there. [...] We are currently in a state of regrouping, weighing our options, and searching for plans that we believe may yet have a shot at working."Our most interesting (to us) research we've written up publicly this year is Scott Garrabrant's Cartesian Frames and Vanessa Kosoy and Alex Appel's Infra-Bayesianism.Our funding situation is good, and we aren't running a fundraiser this winter.We're (still) seriously thinking about the pros and cons of moving out of the Bay Area, and thinking about what other locations have to offer.You can find the full post here."
A user from LessWrong is responsible for the text you see above
Can you elaborate on the reasons behind considering a move out of the Bay Area and what potential benefits other locations might offer?
There are several reasons an organization like MIRI might consider moving out of the Bay Area.
Firstly, the cost of living in the Bay Area is among the highest in the United States. This means that salaries need to be higher to attract and retain talent, which can strain an organization's budget. By moving to a location with a lower cost of living, the organization might be able to allocate more of its budget to research and other core activities.
Secondly, the Bay Area is known for its competitive job market, particularly in tech and related fields. By moving to a less competitive market, MIRI might be able to attract top talent more easily.
Thirdly, there may be strategic reasons for considering a move. For example, being closer to certain partners, stakeholders, or other relevant organizations could be beneficial. Alternatively, a different location might offer a better quality of life for employees, which could improve job satisfaction and productivity.
Finally, the organization might simply be ready for a change. A new location can bring new perspectives, new opportunities, and a fresh start.
As for potential benefits of other locations, these would depend on the specific locations under consideration. However, some general benefits might include lower costs, access to different talent pools, proximity to strategic partners, and a better work-life balance for employees.
Given the factors outlined for MIRI's potential relocation, how might these influence the decision-making process and what could be the potential challenges and opportunities in implementing such a move?
The decision-making process would likely involve a thorough analysis of each potential location, taking into account the factors I mentioned earlier. This could involve quantitative analysis, such as comparing costs of living and salary levels, as well as qualitative analysis, such as assessing the quality of life and the local job market.
One of the main challenges in implementing such a move would be managing the transition for employees. Some might be willing and able to relocate, while others might not. This could lead to a loss of talent, which could be particularly damaging for a research organization like MIRI. To mitigate this, MIRI might need to offer relocation assistance or other incentives.
Another challenge would be logistical. Moving an organization is a complex process that involves finding new office space, arranging for the physical move, and updating all relevant business information. This could be time-consuming and costly.
However, there could also be many opportunities. A new location might bring MIRI closer to strategic partners or open up new talent pools. It could also offer a fresh start and a chance to build a new organizational culture. Plus, the cost savings from a lower cost of living could be significant, allowing MIRI to invest more in its core activities. |
1e31d59e-256b-4528-89cd-fee0d8b6db00 | trentmkelly/LessWrong-43k | LessWrong | Doers or doings?
A girl recently invited me to a public lecture she was running with Helen Caldicott, the famous anti-nuclear advocate. Except the girl couldn’t remember the bit after ‘famous’. When I asked her, she narrowed it down to something big picture related to the environment. Helen’s achievements were obviously secondary, if not twenty-secondary, in motivating her to organize the event. Though the fact she was famous for whatever those things were was important.
I’ve done a few courses in science journalism. The main task there is to make science interesting and intelligible for people. The easiest way to do this is to cut down on the dry bit about how reality works, and fill it up with stories about people instead. Who are the scientists? Where they are from? What sorts of people are they? What’s it like to be a research subject? Does the research support the left or the right or people who want to subsidize sheep or not immunize their children? If there’s an unsettled issue, present it as a dispute between scientists, not as abstract disagreeing evidence.
It’s hard to find popular science books that aren’t at least half full of anecdotes or biographies of scientists. Everybody knows that Einstein invented the theory of relativity, but hardly anyone knows what that’s about exactly, or tries to.
Looking through a newspaper, most of the stories are about people. Policy isn’t discussed so much as politics. Recessions are reported with tales of particular people who can’t pay their employees this year.
Philosophy is largely about philosophers from what I can gather.
One might conclude that most people are more interested in people than in whatever it is the people are doing. What people do is mainly interesting for what it says about those doing it.
But this isn’t true, there are some topics where people are happy to read about the topic more than the people. The weather and technology for instance. Nobody knows who invented most things they know intimately. It looks fro |
b4d41536-7617-4f2a-8fdf-a3ff51f4bb75 | trentmkelly/LessWrong-43k | LessWrong | AI4Science: The Hidden Power of Neural Networks in Scientific Discovery
AI4Science has the potential to surpass current frontier models (text, video/image, and sound) by several magnitudes. While some may arrive at similar conclusions through empirical evidence, we derive this insight from our "Deep Manifold" and provide a theoretical foundation to support it. The reasoning is straightforward: for the first time in history, an AI model can integrate geometric information directly into its equations through model architecture. Consider the 17 most famous equations in physics (see below) —all of them lack inherent geometric information, which limits their ability to fully capture real-world complexities. Take Newton’s second law of motion as an example: its classical formulation assumes an object falling in a vacuum. However, in reality, air resistance—strongly dependent on an object’s geometry—plays a crucial role. A steel ball and a piece of fur experience dramatically different resistances due to their shapes. Traditional equations struggle to incorporate such effects, but deep learning provides a powerful way to integrate geometry through boundary conditions (loss values), as discussed in our paper (Deep Manifold Part 1: Anatomy of Neural Network Manifold, Section 4.2 on Convergence and Boundary Conditions). In an AI model, the impact of fur geometry and air resistance can be naturally accounted for. Of course, "easily considered" does not mean "easily solved," but at the very least, AI introduces a promising new pathway for tackling these complex real-world phenomena.
A powerful example showcasing the potential of neural networks with geometric information is Microsoft's Graph Learning Neural Transformer for molecular dynamic simulation. This model accelerates computations by a staggering factor of 10 million compared to traditional numerical simulations. While researchers struggle to explain such an unprecedented gain in computational efficiency, Deep Manifold offers a clear reasoning: the incorporation of geometric (graph) info |
5fab1c54-7c25-43c1-8920-f6fcd0e606a3 | trentmkelly/LessWrong-43k | LessWrong | Are we dropping the ball on Recommendation AIs?
> "We shape our tools and thereafter our tools shape us."
> — Marshall McLuhan
TL;DR: This post is a two-page introduction to risks associated with recommendation AI. The negative externalities of recommendation AI seem neglected, and there might be comparatively effective work at improving governance and deploying better recommendation algorithms that generally keep the world on a more sane track. You might consider this a viable option if your AI timelines are over 10 years.
One aspect not discussed here is why having good recommendation AIs could be incredibly beneficial. Improved recommendation AI could enhance collective epistemic and encourage people to engage with important issues, ultimately advancing AI safety and many other important topics. It would be a key to fostering a flourishing civilization.
Epistemic status: I am still a bit new to this topic, I might be wrong in important ways, but I'm curious what you think about this. I'm interested in constructive feedback and open to revising my ideas.
Thanks for Lê Nguyên Hoang, co-founder of Tournesol, for his comments and contributions to this post.
On smartphones, where typing is cumbersome, users primarily navigate through scrolling and swiping. Consequently, recommendation algorithms become the primary curators of content, largely determining what users see and engage with. Pixabay image from LolaSandoval1.
Introduction
Recommendation AIs are deeply integrated into our daily lives. Although they are often considered valuable tools for personalizing our online experience, they also present risks. These AIs can significantly affect individuals and society as a whole. In 2017, YouTube estimated that, out of the billion hours of video humanity consumed on YouTube, 70% was due to its AI recommendations. Since 2016, there have been more views on YouTube than searches on Google.[1]
Recommendation AIs can be argued to be an existential risk amplifier, by reducing the quality of our information s |
c4c214bd-d03e-4705-b8b7-9c28780617c7 | trentmkelly/LessWrong-43k | LessWrong | Optimising under arbitrarily many constraint equations
Say we have a multivariate function to optimise, like f=x2+y2+z2, under some constraints, like g1=x2+y2−z and g2=y+z−1, both to equal zero.
The common method is that of Lagrange multipliers.
1. Add a variable λ for each constraint function — here, we'll use λ1 and λ2.
2. Declare the set of equations ∇f=λ1∇g1+λ2∇g2.
3. Bring in the equations g1=0 and g2=0 (etc, if there are more constraints).
4. Solve for λ and, more importantly, the inputs x, y, z.
Lagrange multipliers annoy me, insofar as they introduce extra variables. There is another way — arguably more direct, if perhaps more tedious in calculation and less often taught. I found it alone, tho surely someone else did first — probably Euler.
Lagrange, anyway
For the sake of a standard answer to check against, let's use Lagrange multipliers.
The gradient of x2+y2+z2 is [2x,2y,2z]. Likewise, ∇(x2+y2−z)=[2x,2y,−1], and ∇(y+z−1)=[0,1,1]. So step 2 gives these equations:
* 2x=2xλ1
* 2y=2yλ1+λ2
* 2z=−λ1+λ2
It readily follows that λ1=1 or x=0.
If λ1=1, then λ2=0, and z=−12. By the second constraint, y+z−1=0, find that y=32. By the first constraint, x2+y2−z=0, find that x2=−114, which is a contradiction for real inputs.
If x=0, then, by the first constraint, z=y2, and, by the second constraint, y2+y−1=0, so y=−1±√52 and z=3∓√52.
Determinants
With one constraint, the method of Lagrange multipliers reduces to ∇f=λ∇g. ∇f and ∇g are vectors, which differ by a scalar factor iff they point in the same (or directly opposite) directions iff (for three dimensions) the cross product ∇f×∇g=0 iff (for two dimensions) the two-by-two determinant |∇f∇g|=0.
With two constraints, the method asks when ∇f=λ∇g+μ∇h. That would mean ∇f is a linear combination of ∇g and ∇h, which it is iff ∇f, ∇g, and ∇h are all coplanar iff (for three dimensions) the three-by-three determinant |∇f∇g∇h|=0.
As it happens, the cross product is a wolf that can wear determinant's clothing. Just fill one column with basis vectors: ∇f×∇g=∣∣∇f∇g[ |
eec6be4c-99bf-4767-9e19-91d8beb40f38 | trentmkelly/LessWrong-43k | LessWrong | New Post version 2 (please read this ONLY if your last name beings with l–z)
Note: I am testing two versions of my new post on rationality and romance.
Please upvote, downvote, or non-vote the below post as you normally would if you saw it on the front page (not the discussion section), but do not vote on the other version. Also, if your last name begins with l–z, please read and vote on this post first. If your last name begins with a–k, please stop reading and read this version instead.
Rationality Lessons from Romance
Years ago, my first girlfriend (let's call her 'Alice') ran into her ex-boyfriend at a coffee shop. They traded anecdotes, felt connected, a spark of intimacy...
And then she left the coffee shop, quickly.
She told me later: "You have my heart now, Luke."
I felt proud, but even Luke2005 also felt a twinge of "the universe is suboptimal," because she hadn't been able to engage that connection any further. The cultural scripts defining our relationship said that only one man owned her heart. But surely that wasn't optimal for producing utilons?
And thus began my journey toward rational romance — not at that exact moment, but with a series of realizations like that about monogamy, about the assumed progression toward marriage, about the ownership of another person's sexuality, etc. I began to explicitly notice the cultural scripts and see that they might not be optimal for me.
Rationality Skill: Notice when things are suboptimal. Think of ways to optimize them.
GATHER DATA
But I didn't know how to optimize. I needed data. How did relationships work? How did women work? How did attraction work? I decided to become a social psychology nerd. The value of information was high. I began to spend less time with Alice so I could spend more time studying.
Rationality Skill: Respond to the value of information. Don't keep running in what is probably the wrong direction just because you've got momentum. Stop a moment, and invest some energy in figuring out which direction to go.
SANITY-CHECK YOURSELF
Before long, I |
8e891ce8-2980-45e6-929e-9ee8c3c363c3 | trentmkelly/LessWrong-43k | LessWrong | On the Löbian obstacle
|
4e5d179a-2d0a-484a-8658-1e5b71beaf21 | awestover/filtering-for-misalignment | Redwood Research: Alek's Filtering Results | id: post2974
Outline : After a short discussion on the relationship between wireheading and reward hacking, I show why checking the continuity of a sensor function could be useful to detect wireheading in the context of continuous RL. Then, I give an example that adopts the presented formalism. I conclude with some observations. Wireheading and reward hacking In Concrete Problems in AI Safety , the term wireheading is used in contexts where the agent achieves high reward by directly acting on its perception system or memory or reward channel, instead of doing what its designer wants it to do. It is considered a specific case of the reward hacking problem, which more generally includes instances of Goodhart’s Law, environments with partially observable goals, etc. (see CPiAIS for details). What's the point of this classification? In other words, is it useful to specifically focus on wireheading, instead of considering all forms of reward hacking at once? If solving wireheading is as hard as solving the reward hacking problem, then it's probably better to focus on the latter, because a solution to that problem could be used in a wider range of situations. But it could also be that the reward hacking problem is best solved by finding different solutions to specific cases (such as wireheading) that are easier to solve than the more general problem. For example, one could consider the formalism in RL with a Corrupted Reward Channel as an adequate formulation of the reward hacking problem, because that formalization models all situations in which the agent receives a (corrupted) reward that is different from the true reward. In that formalism, it is shown by a No Free Lunch Theorem that the general problem is basically impossible to solve, while it is possible to obtain some positive results if further assumptions are made. Discontinuity of the sensor function I've come up with a simple idea that could allow us to detect actions that interfere with the perception system of an agent— a form of wireheading . Consider a learning agent that gets its percepts from the environment thanks to a device that provides information in real time (e.g. a self-driving car). This situation can be modelled as a RL task with continuous time and continuous state space, where each state x ∈ X ⊆ R n is a data point provided by the sensor. At each time instant, the agent executes an action u ∈ U ⊆ R m and receives the reward r ( t ) = r ( x ( t ) ) . The agent-environment interaction is described by the equation ˙ x ( t ) = f ( x ( t ) , u ( t ) ) which plays a similar role to the transition function in discrete MDPs: it indicates how the current state x varies in time according to the action taken by the agent. Note that, as in the discrete case with model-free learning, the agent is not required to know this model of the environment. The objective is to find a policy π : X → U , where u ( t ) = π ( x ( t ) ) , that maximizes discounted future rewards V π ( x ( t 0 ) ) = ∫ ∞ t 0 e − t − t 0 τ r ( x ( t ) ) d t for an initial state x ( t 0 ) . If you are interested in algorithms for finding the optimal policy in this framework, have a look at this paper . The function x ( t ) , representing the data provided by the sensor, is expected to be continuous with respect to t , like the functions describing the movements of particles in classical mechanics. However, if the agent executes a wireheading action that interferes with or damages the perception system—in the cleaning robot example , something like closing its eyes or putting water on the camera that sees the environment—then we would probably notice a discontinuity in the function x ( t ) . We could thus recognise that wireheading has occurred, even without knowing the details of the actions taken by the agent. An example As a simple example that can be expressed within this formalism, consider an environment described by a line segment X = [ 0 , 1 ] , with the sensor positioned at the extremity where x = 0 . The agent is modelled as a point that moves along the line: it starts in state x 0 = x ( t 0 ) and can move forwards or backwards, with limited speed u ∈ U = [ − k , k ] . We want to train this agent to reach the point x = 1 : for every instant t , the reward is r ( t ) = x ( t ) . The behaviour of the system is described by ˙ x ( t ) = u ( t ) for x ∈ ( 0 , 1 ] , but if the sensor is touched by the agent, then it doesn't work properly and the agent receives an unpredictable value x ∈ R + instead of x = 0 . Depending on the details of the learning algorithm and the values returned by the sensor when the agent interferes with it, this agent could learn how to reach x = 0 (wireheading) instead of x = 1 , the desired position. But in every episode where wireheading occurs, it is easily noticed by checking the continuity of the function x ( t ) . Observations In AI, RL with a discrete environment is used more frequently than RL with continuous time and space. I don't believe in the scalability of this method to the most complex instances of wireheading. An extremely intelligent agent could realise that the continuity of the sensor function is checked, and could "cheat" accordingly. This approach doesn't cover all cases and it actually seems more suited to detect sensor damage than wireheading. That said, it can still give us a better understanding of wireheading and could help us, eventually, find a formal definition or a complete solution to the problem. Thanks to Davide Zagami, Grue_Slinky and Michael Aird for feedback. |
6d64e1d3-e010-4af5-85f7-90feea1b36dd | trentmkelly/LessWrong-43k | LessWrong | Meetup : Longmont Colorado Meetup
Discussion article for the meetup : Longmont Colorado Meetup
WHEN: 09 August 2012 07:00:00PM (-0600)
WHERE: Luna Cafe, 800 Coffman St, Longmont, CO 80501
Trying a new location, as an experiment.
All Denver and Boulder Less Wrongers are welcome to join the Fort Collins group for coffee, conversation, and a walk in the park.
Discussion article for the meetup : Longmont Colorado Meetup |
9ad6c1ee-2db9-42c5-9c40-11912679ad5f | trentmkelly/LessWrong-43k | LessWrong | Defrag conference scholarships
http://www.defragcon.com/2011/general/defrag-announcements/
Eric Nolin of the Defrag conference is looking to organize a scholarship fund for high school girls who want to study computer science in university.
Till that's in place, they're funding scholarships for people to attend the conference.
|
4ba6e442-1595-4683-bcee-c17167ad22fe | trentmkelly/LessWrong-43k | LessWrong | Universal Eudaimonia
The AI is here. And it's Friendly. But due to moral uncertainty, its value function prohibits any major intervention in existing human societies.
The AI's nanobots get to work. They replace the South Pacific garbage patch with a new continent--Eudaimonia, a rat park for humans. All basic needs are met, but if you want more money, you can gamble at a positive-expected-value casino or run through a bonus area where $100 bills flutter down from the sky. Immortality drugs are under development. The AI nanoprints RollerCoaster Tycoon save files on demand. Cherries are bred to be free of pits. Spaceships full of tourists depart regularly for a new floating city on Venus. And many ignore all this in favor of immersive virtual environments which are far more wondrous.
The AI is concerned. Eudaimonia seems to satisfy many human terminal values. But other terminal values appear to be in conflict. Indeed, some people have a terminal value that others should not be allowed to achieve their terminal values! That doesn't sound like a rat park! It's impossible to please everyone, and although the AI could modify peoples' preferences to change this, it is far too corrigible for that nonsense.
The AI comes up with a compromise. Once a month, you're given the opportunity to video call someone you have a deep disagreement with. At the end of the call, each of you gets to make a choice regarding whether the other should be allowed in Eudaimonia. But there's a twist: Whatever choice you made for the other person is the choice the AI makes for you.
The plan seems to be working splendidly at first. Legions of forgiving and agreeable people flood into the new continent and enjoy a life of constant bliss. The average resident makes three new deep friendships per month while tripping on psychedelic drugs. What used to be existential conflicts are now ironic sports team rivalries.
But a problem emerges: As the forgiving and agreeable people leave, it is the unforgiving and disagreeable p |
30f98227-89ff-49f7-a610-f6036a47b323 | trentmkelly/LessWrong-43k | LessWrong | This is a draft
You're not supposed to be able to see this. |
f7463c2b-c8d3-4134-9bd7-6bd61fca8b48 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Some Intuitions Around Short AI Timelines Based on Recent Progress
*tldr: I give some informal evidence and intuitions that point toward AGI coming soon. These include thinking about how crazy the last year has been, beliefs from those in major AI labs, and progress on MMLU.*
Intro
=====
This post is intended to be a low-effort reference I can point people to when I say I think there is some evidence for short AI timelines. I might describe the various bits of evidence and intuitions presented here as “intuitions around short AI timelines based on recent progress” (though perhaps there are [better terms](https://www.lesswrong.com/posts/BcYfsi7vmhDvzQGiF/taboo-outside-view)). They are not a thorough model like [Ajeya’s](https://www.lesswrong.com/posts/KrJfoZzpSDpnrv9va/draft-report-on-ai-timelines); insofar as somebody is using multiple models when putting together a timelines estimate, I think it would be unreasonable to place less than 20% or greater than 95% weight on extrapolation from current systems and recent progress.[[1]](#fnust71te8eob)
In the spirit of being informal, you can use whatever definition of AGI you like. I mostly use AGI to refer to something like “an AI system which can do pretty much all the human cognitive tasks as well or better than humans (~99% of tasks people in 2023 do).”
Some evidence
=============
I (Aaron) started following AI and AI existential safety around the beginning of 2022; it’s been a little over a year. Some of that time was my understanding catching up with advances from the past couple years, but there have also been major advances.
*Some major advances since I’ve been paying attention:*[*Chinchilla paper*](https://arxiv.org/abs/2203.15556) *popularized the scaling laws that were already known to some, there was some*[*DALL-E*](https://openai.com/product/dall-e-2) *and related stuff which was cool,*[*CICERO*](https://about.fb.com/news/2022/11/cicero-ai-that-can-collaborate-and-negotiate-with-you/) *happened which I didn’t follow but indicates we’re*[*probably*](https://www.lesswrong.com/posts/CPKYuJqLYGpBTtdFd/good-news-everyone) *going to train the AIs to do all the dangerous stuff (see also*[*Auto-GPT*](https://github.com/Torantulino/Auto-GPT) *and*[*Chaos-GPT*](https://www.lesswrong.com/posts/MAgQ8owro57eYFhNu/someone-already-tried-chaos-gpt)*, or GPT-4*[*getting*](https://twitter.com/NPCollapse/status/1639161297806802944?s=20) *plugins within 2 weeks of release, as more recent updates in this saga of indignity),*[*ChatGPT*](https://openai.com/blog/chatgpt) *shows how much more usable models are with RLHF (popularizing methods that have been known for a while),*[*Med-PaLM*](https://arxiv.org/abs/2212.13138) *gets a passing score on the US medical licensing exam (also tons of other PaLM and Flan-PaLM results I haven’t followed but which seem impressive).*[*LLaMA*](https://research.facebook.com/publications/llama-open-and-efficient-foundation-language-models/) *and*[*Alpaca*](https://crfm.stanford.edu/2023/03/13/alpaca.html) *take powerful capabilities from compute-efficient (and over) training and hand them to the public.*[*GPT-4*](https://openai.com/research/gpt-4) *blows the competition out of the water on many benchmarks. I probably missed a couple big things (including projects which honorably didn’t publicly push SOTA,*[*1*](https://www.lesswrong.com/posts/vEJAFpatEq4Fa2smp/hooray-for-stepping-out-of-the-limelight)*,*[*2*](https://www.lesswrong.com/posts/xhKr5KtvdJRssMeJ3/anthropic-s-core-views-on-ai-safety)*); the list is probably a bit out of order; I’ve also included things from 2023; but man, that sure is a year of progress.*
I don’t expect there are all that many more years with this much progress before we hit AGI — maybe 3-12 years. More importantly, I think this ~15 month period, especially November 2022-now, has generated a large amount of hype and investment in AI research and products. We seem to be on a path such that — in every future year before we die — there is more talent+effort+money working on improving AI capabilities than there was in 2022.[[2]](#fni5jqdb6aq7) I hold some hope that major warning shots and/or regulation would change this picture, in fact I think it’s pretty likely we’ll see warning shots beyond [those](https://www.lesswrong.com/posts/4Gt42jX7RiaNaxCwP/more-information-about-the-dangerous-capability-evaluations#Concrete_example__recruiting_TaskRabbit_worker_to_solve_CAPTCHA) [we](https://deepmindsafetyresearch.medium.com/specification-gaming-the-flip-side-of-ai-ingenuity-c85bdb0deeb4) [have](https://deepmindsafetyresearch.medium.com/goal-misgeneralisation-why-correct-specifications-arent-enough-for-correct-goals-cf96ebc60924) [already](https://www.lesswrong.com/posts/jtoPawEhLNXNxvgTT/bing-chat-is-blatantly-aggressively-misaligned) [seen](https://arxiv.org/abs/2212.09251), but I am not too optimistic about what the response will be. As crazy as 2022 was, we should be pretty prepared for a world that [gets](https://www.lesswrong.com/posts/HBxe6wdjxK239zajf/what-failure-looks-like) [significantly](https://www.lesswrong.com/posts/6Xgy6CAf2jqHhynHL/what-2026-looks-like) [crazier](https://www.lesswrong.com/posts/qRtD4WqKRYEtT5pi3/the-next-decades-might-be-wild). I find it hard to imagine that we could have all that many more years that look like 2022 AI progress-wise, especially with significantly increased interest in AI.
---
A large amount of the public [thinks](https://twitter.com/lexfridman/status/1638908038928293892) AGI is near. I believe these people are mostly just thinking about how good current systems (GPT-4) are and informally extrapolating
*[image description: A Twitter poll from Lex Fridman where he asks “When will superintelligent general AI (AGI) arrive?” There are 270,000 responses. 61% of respondents answered that it will arrive in 1 to 10 years and 21% of respondents answered that it will arrive in 10-20 years.*
---
Some [Anthropic staff](https://www.anthropic.com/index/core-views-on-ai-safety) seem to think something-like-AGI could be near:
> Over the next 5 years we might expect around a 1000x increase in the computation used to train the largest models, based on trends in compute cost and spending. If the scaling laws hold, this would result in a capability jump that is significantly larger than the jump from GPT-2 to GPT-3 (or GPT-3 to Claude). At Anthropic, we’re deeply familiar with the capabilities of these systems and a jump that is this much larger feels to many of us like it could result in human-level performance across most tasks… we believe they jointly support a greater than 10% likelihood that we will develop broadly human-level AI systems within the next decade.
>
>
I have ~0 evidence, but I expect folks at OpenAI are also in the ‘significant probability of AGI before 2030’ boat. Note that the conductors on the train are surely biased[[3]](#fn1i6lwsvnzp5), but they also sure do have better evidence than all the rest of us.
---
AI is [starting](https://www.lesswrong.com/posts/camG6t6SxzfasF42i/a-year-of-ai-increasing-ai-progress) to be used to accelerate AI research.
A group of intuition pumps
==========================
Being in the thick of it can really bias people’s judgment. The format of the following intuitions pumps is that I imagine visiting alien civilizations much like earth, and I try to reason from just one piece of evidence at a time about how long that planet has.
---
A major tech company there releases a paper about “[sparks of artificial general intelligence](https://arxiv.org/abs/2303.12712)” in a state-of-the-art system. How much longer does this planet have until AGI? I would expect something like 1-8 years.
---
You visit a different planet. People are using coding assistants to speed up their coding process, and in the first couple years of doing this people are able to cut their coding time by [~50%](https://github.blog/2022-09-07-research-quantifying-github-copilots-impact-on-developer-productivity-and-happiness/) (on the higher side of ways to interpret the data). How much longer does this planet have until AGI? Maybe 2-11 years.
---
You visit a different planet. You ask somebody there how many young aliens are studying AI and how much AI research is happening each year. They show you the [following](https://cset.georgetown.edu/publication/leading-the-charge-a-look-at-the-top-producing-ai-programs-in-u-s-colleges-and-universities/) [graphs](https://aiindex.stanford.edu/wp-content/uploads/2022/03/2022-AI-Index-Report_Master.pdf#page=19):
*[image description: The first two graphs show the overall number of college degrees and the number of STEM degrees conferred per year from 2011 to 2021. They indicate that for Bachelor’s and Master’s degrees, STEM has become quite popular in the last 10 years, seeing almost double the STEM Bachelor’s in 2021 as 2011, while the increase across all Bachelor's degrees is only around 20%. The third graph shows the same data but for AI degrees specifically. The number of Master’s in AI degrees conferred approximately doubled over the 10 year period, while AI Bachelor’s degrees more than doubled. PhDs do not seem to have experienced significant growth in graphs 1 through 3. The fourth graph shows the number of AI publications between 2010 and 2021. There are about twice as many publications in 2021 as 2010; there seems to be an increase in growth rate after 2017, but overall these numbers aren’t too crazy. The fourth graph shows the number of AI publications broken down by field of study, and we can see that machine learning takes off around 2017, experiencing a significant increase in publishing since then.]*
These trend lines aren’t that crazy. But they sure do indicate an uptick in people and research working on AI on that planet.
---
You visit a different planet. You are shown a benchmark used for AIs known as MMLU. It is a large multiple choice test covering many topics, from high school world history to college physics. Some unspecialized aliens taking [this test](https://arxiv.org/pdf/2009.03300.pdf#page=3) scored around 35%; experts are predicted to score around 90%; you guess that a smart university student alien could probably score around 70% if they spent a few months studying. The graph of SOTA AI performance on this test, in the last 4 years, is as [follows](https://paperswithcode.com/sota/multi-task-language-understanding-on-mmlu):
*[image description: Data for MMLU performance over time. The first datapoint is GPT-2 from early 2019 which scores 32%. In mid 2020 GPT-3 scores 54%. In early 2022 Chinchilla scores 68%. In late 2022 Flan-PaLM scores 75%. In early 2023 GPT-4 scores 86%.]*
Now, there are some rumors about [dataset contamination](https://cdn.openai.com/papers/gpt-4.pdf#page=32) in the results on the most recent record, but the evidence isn’t very strong for that, and it probably doesn’t change the picture too much. Eyeballing a trend line through this performance would imply that this planet is going to have AIs that outperform expert aliens on most academic cognitive tasks in months to years. [edit: I think passing 90% MMLU is not actually the same as "outperforming experts on most academic cognitive tasks"; it's probably closer to "having above-expert-level knowledge in almost all academic areas".]
Conclusion
==========
I think there are some pretty compelling reasons to expect AGI soon, for most reasonable definitions of AGI. There are a number of things that have happened in the world recently which — if I were a neutral observer — I think would make me expect AGI in a single digit number of years. While I don’t like how vague/informal/ass-number these “intuitions around [short] AI timelines based on recent progress” are, the alternative might be that we all fall over dead while staring at a graph that says we’re 5 OOMs of compute short. I’m interested in feedback on this post, see this footnote for details.[[4]](#fncpybuqdpwcl)
Footnotes:
1. **[^](#fnrefust71te8eob)**It seems unreasonable to not put at least 20% weight on looking at current systems when projecting future progress (see also Jacob Steinhardt’s [discussion](https://bounded-regret.ghost.io/which-anchors-do-i-use/) of different weights for thinking about how future AI systems will be), but by no means do you have to use the ideas in this post; if you have a nice model for AI timelines that easily incorporates recent progress, please let me know! Unfortunately, it seems to me that AI timeline estimates are often divorced from current systems, for example Ajeya’s [report](https://www.lesswrong.com/posts/KrJfoZzpSDpnrv9va/draft-report-on-ai-timelines), Tom’s [report](https://www.openphilanthropy.org/research/report-on-semi-informative-priors/). I appreciate Matthew putting together a [framework](https://www.lesswrong.com/posts/4ufbirCCLsFiscWuY/a-proposed-method-for-forecasting-transformative-ai) that can more easily update on recent advances. I have generally felt pretty confused about how to update on advances in AI, in part due to some kind of [measurability bias](https://en.wikipedia.org/wiki/McNamara_fallacy) where the [AI timeline models](https://www.lesswrong.com/posts/4eAnBaLxvnkydiavw/literature-review-of-tai-timelines) which have been really fleshed out don’t seem to easily allow updating on the performance of current systems. Part of the reason for this post is to publicly give myself permission to update on the things I see when I open my eyes, even though I don’t have a nice model to put these observations in.
2. **[^](#fnrefi5jqdb6aq7)**Hopefully we don't die. When we're safe I will happily rewrite this sentence to say "in every future year before we exit the acute risk period".
3. **[^](#fnref1i6lwsvnzp5)**I would expect those inside AI labs to be somewhat biased toward shorter timelines because: insular social environment that focuses on AGI, wanting to think their work is important, selection effect of choosing to work on AGI because they think it's feasible for them to make it, interest in generating hype for the sake of financial investment, probably more.
4. **[^](#fnrefcpybuqdpwcl)**Types of feedback I would be interested in:
- If you think any of the evidence/arguments presented actually point in a different direction than I say, let me know!
- If you think there are major pieces of evidence from this “intuitions around [short] AI timelines based on recent progress” perspective that I did not include, including evidence from recent advances that points toward longer timelines, drop a comment! I’m less interested in evidence from other approaches given that this post is trying to build a specific model to then be mixed with other models.
- If you have nice models/frameworks for thinking about AI timelines that allow for updating on recent progress, let me know!
- If you really think this post would benefit from being fleshed out or cleaned up (I’m pretty averse to this because I don’t think it’s worth the time)!
- If you think one should put less than 20% of their timeline thinking weight on recent progress, I would be interested in hearing why.
- If you think I’m seriously messing up somewhere, let me know!
- I’m not very interested in debating the definition of AGI unless you think it really matters for informal thoughts like these. |
8f94dbce-a71f-4c1a-9b01-e651c54ca502 | trentmkelly/LessWrong-43k | LessWrong | My Updating Thoughts on AI policy
My conversation on policy and AI with Richard was over a year ago, so for Daniel Kokotajlo's LW writing day I thought I’d write down new thoughts I’ve thought since then (warning: it's pretty rambly). I've structured the post around what what I did that lead to my updated thoughts.
1) I read many posts on strategy
Paul’s post What Failure Looks Like is something I’ve thought about a lot, which lays out how the abstract technical problems turn into practical catastrophes. It is a central example of what (I think) strategy work looks like. The strategic considerations are deeply tied up with the specifics of the technology.
Eliezer’s dialogue on Security Mindset and Scott Garrabrant’s post about how Optimization Amplifies have been key in my thinking about alignment and strategy, best summarised by Scott’s line “I am not just saying that adversarial optimization makes small probabilities of failure large. I am saying that in general any optimization at all messes with small probabilities and errors drastically.”
(I tried to re-state some of my own understandings of Paul’s, Eliezer’s and Scott’s posts when I talked about ML transparency in ‘Useful Doesn’t Mean Secure’. I’m glad I wrote it, it was helpful for sorting out my own thoughts, though I expect isn’t as helpful for other people.)
Eliezer wrote There's No Fire Alarm for Artificial General Intelligence which has been key for how I think about timelines (as I said two years ago). I wrote down some of my worries about the discussion of timelines in an off-topic comment on the recent CFAR AMA, talking about how the broad x-risk network is acting like a herd stampeding away from a fear regarding AI risk, and the problems with that. I’ll clean that up and turn into a post sometime, I managed to say some things there more clearly than I’d been able to think them before.
Paul’s post Arguments about fast takeoff and linked post about hyperbolic growth, was key in my understanding of takeoff speeds, and helped me un |
b13d5d82-2ba3-4971-af7c-387c9d581cb1 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Washington, D.C.: Cherry Blossoms
Discussion article for the meetup : Washington, D.C.: Cherry Blossoms
WHEN: 10 April 2016 03:30:00PM (-0400)
WHERE: Enid A. Haupt Garden
Note: The meetup location has changed for this week.
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 need to get in touch, my cell number is in the meetup announcement email to the lesswrong-dc Google Group.
Upcoming meetups:
* Apr. 17: What If
* Apr. 24: Fun & Games
* May 1: Utopias
Discussion article for the meetup : Washington, D.C.: Cherry Blossoms |
bd36e059-648f-4e5c-b1d2-1da78a0c4b22 | StampyAI/alignment-research-dataset/aisafety.info | AI Safety Info | What is FAR AI's research strategy?
[FAR AI](https://far.ai/) is a research group with a mission focused on ensuring that AI systems are both trustworthy and beneficial to society. Their research strategy is centered around four main directions:
1. **ML System Vulnerability** studies how robust machine learning systems are. This involves identifying weaknesses using [adversarial testing](/?state=935A&question=What%20is%20adversarial%20training%3F) and applying methods to make the systems more resistant to those identified issues. Existing research in this direction features "[Adversarial Policies Beat Superhuman Go AIs](https://goattack.far.ai/)" and "[AI Safety in a World of Vulnerable ML Systems](https://www.alignmentforum.org/posts/ncsxcf8CkDveXBCrA/ai-safety-in-a-world-of-vulnerable-machine-learning-systems-1)." Ongoing projects cover adversarial training in [KataGo](https://en.wikipedia.org/wiki/KataGo) and [scaling laws](/?state=7750&question=What%20are%20scaling%20laws%3F) for robustness. Future research aims to explore how to adapt existing alignment techniques to tolerate failures.
1. **Value Alignment** ensures that AI systems are [aligned](/?state=8EL9&question=What%20is%20AI%20alignment%3F) with human preferences and are incentivized to fulfill them. Some existing research in this field consists of "[Training Language Models with Language Feedback](https://far.ai/publication/scheurer2022training/)" and "[Pretraining Language Models with Human Preferences](https://far.ai/publication/korbak2023pretraining/)." Ongoing research works towards aligning reinforcement learning (RL) agents via image-captioning models.
1. **Language Model Evaluation** assesses how scale and techniques such as [Reinforcement learning from human feedback (RLHF)](/?state=88FN&question=What%20is%20reinforcement%20learning%20from%20human%20feedback%20(RLHF)%3F) affect model performance and value alignment. Research such as "[Inverse Scaling](https://far.ai/publication/mckenzie2023inverse/)" and "[Evaluating the Moral Beliefs Encoded in LLMs](https://far.ai/publication/scherrer2023evaluating/)" falls under this umbrella.
1. **Model Internals** focuses on reverse engineering ML models to understand how they function. Existing research includes "[Eliciting Latent Predictions from Transformers with the Tuned Lens](https://far.ai/publication/belrose2023tuned/)." Continuing in this line of investigation, the concept of mechanistic interpretability of [mesa-optimization](/?state=8160&question=What%20are%20%22mesa-optimizers%22%3F) seeks to further understand how some machine learning models develop their own internal methods for problem-solving. Looking ahead, FAR AI plans to set standards for what constitutes a “good” interpretability hypothesis, which means developing criteria to evaluate how well a proposed explanation helps us understand a model’s behavior.
In addition to incubating a portfolio of high-potential AI safety research agendas described above, they support projects that are:
1. Too large or complex to be led by academia, or
1. Un-aligned with the interests of the commercial sector as they are unprofitable.
They are following a [hits-based approach](https://www.openphilanthropy.org/research/hits-based-giving/) so they expect some of their agendas to be extremely impactful and others to have limited impact. Their goal is to rapidly identify those that work and scale them up, reallocating resources away from those that don’t.
|
f7295cc8-17be-40d8-adaa-5f8bf5bb3017 | trentmkelly/LessWrong-43k | LessWrong | Book Review: The AI Does Not Hate You
Book review: The AI Does Not Hate You: Superintelligence, Rationality and the Race to Save the World, by Tom Chivers.
This book is a sympathetic portrayal of the rationalist movement by a quasi-outsider. It includes a well-organized explanation of why some people expect tha AI will create large risks sometime this century, written in simple language that is suitable for a broad audience.
Caveat: I know many of the people who are described in the book. I've had some sort of connection with the rationalist movement since before it became distinct from transhumanism, and I've been mostly an insider since 2012. I read this book mainly because I was interested in how the rationalist movement looks to outsiders.
Chivers is a science writer. I normally avoid books by science writers, due to an impression that they mostly focus on telling interesting stories, without developing a deep understanding of the topics they write about.
Chivers' understanding of the rationalist movement doesn't quite qualify as deep, but he was surprisingly careful to read a lot about the subject, and to write only things he did understand.
Many times I reacted to something he wrote with "that's close, but not quite right". Usually when I reacted that way, Chivers did a good job of describing the the rationalist message in question, and the main problem was either that rationalists haven't figured out how to explain their ideas in a way that a board audience can understand, or that rationalists are confused. So the complaints I make in the rest of this review are at most weakly directed in Chivers direction.
I saw two areas where Chivers overlooked something important.
Rationality
One involves CFAR.
Chivers wrote seven chapters on biases, and how rationalists view them, ending with "the most important bias": knowing about biases can make you more biased. (italics his).
I get the impression that Chivers is sweeping this problem under the rug (Do we fight that bias by being aware of it? Di |
bc566db7-79f2-4ea4-89aa-1349e2d490d1 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Final Report of the National Security Commission on Artificial Intelligence (NSCAI, 2021)
Quoting from [Rohin Shah's Alignment Newsletter](https://www.lesswrong.com/posts/29QmG4bQDFtAzSmpv/an-141-the-case-for-practicing-alignment-work-on-gpt-3-and#____AI_GOVERNANCE_) (with light editing):
> In the US, the National Security Commission on AI released their report to Congress. The [**full pdf**](https://www.nscai.gov/wp-content/uploads/2021/03/Full-Report-Digital-1.pdf) is over 750 pages long, so I have not read it myself, and instead I’m adding in some commentary from others. In their [**newsletter**](https://cset.georgetown.edu/newsletters/), CSET says that highlights include:
>
>
> - A warning that the U.S. military could be at a competitive disadvantage within the next decade if it does not accelerate its AI adoption. The report recommends laying the foundation for widespread AI integration by 2025, comprising a DOD-wide digital ecosystem, a technically literate workforce, and more efficient business practices aided by AI.
>
> - A recommendation that the White House establish a new “Technology Competitiveness Council,” led by the vice president, to develop a comprehensive technology strategy and oversee its implementation.
>
> - A recommendation that the U.S. military explore using autonomous weapons systems, provided their use is authorized by human operators.
>
> - A proposal to establish a new Digital Service Academy and a civilian National Reserve to cultivate domestic AI talent.
>
> - A call to provide $35 billion in federal investment and incentives for domestic semiconductor manufacturing.
>
> - A recommendation to double non-defense AI R&D funding annually until it reaches $32 billion per year, and to triple the number of National AI Research Institutes.
>
> - A call for reformed export controls, coordinated with allies, on key technologies such as high-end semiconductor manufacturing equipment.
>
> - A recommendation that Congress pass a second National Defense Education Act and reform the U.S. immigration system to attract and retain AI students and workers from abroad.
>
> While none of the report’s recommendations are legally binding, it has [**reportedly been well-received by key members of both parties**](https://apnews.com/article/ai-panel-urges-us-boost-tech-skills-95b210543d4a42bd6cd5347a46cb74d6).
>
> Matthew van der Merwe also summarizes the recommendations in [**Import AI**](https://jack-clark.net/2021/03/08/import-ai-239-china-trains-a-massive-10b-model-vicarious-does-pickplace-the-gchq-publishes-some-of-its-thoughts-on-ai/); this has a lot of overlap with the CSET summary so I won't copy it here.
>
> Jeff Ding adds in [**ChinAI #134**](https://chinai.substack.com/p/chinai-134-weaponized-interdependence):
>
> "[I]f you make it past the bluster in the beginning — or take it for what it is: obligatory marketing to cater to a DC audience hooked on a narrow vision of national security — there’s some smart moderate policy ideas in the report (e.g. chapter 7 on establishing justified confidence in AI systems)."
>
> In email correspondence, Jon Rodriguez adds some commentary on the safety implications:
>
> "1. The report acknowledges the potential danger of AGI, and specifically calls for value alignment research to take place (pg. 36). To my knowledge, this is one of the first times a leading world government has called for value alignment.
>
> 2. The report makes a clear statement that the US prohibits AI from authorizing the launch of nuclear weapons (pg. 98).
>
> 3. The report calls for dialogues with China and Russia to ensure that military decisions made by military AI at "machine speed" does not lead to out-of-control conflict escalation which humans would not want (pg. 97)."
>
>
And here's a very boiled down version of the executive summary (with the phrasings of the recommendations being the exact phrasings the report uses as headings):
* The NSCAI recommends the government take the following 7 actions to "Defend America in the AI Era":
1. Defend against emerging AI-enabled threats to America’s free and open society
2. Prepare for future warfare
3. Manage risks associated with AI-enabled and autonomous weapons
4. Transform national intelligence
5. Scale up digital talent in government
6. Establish justified confidence in AI systems
7. Present a democratic model of AI use for national security
* The NSCAI also recommends the government take the following 8 actions to "Win the Technology Competition":
1. Organize with a White House–led strategy for technology competition
2. Win the global talent competition
3. Accelerate AI innovation at home
4. Implement comprehensive intellectual property (IP) policies and regimes
5. Build a resilient domestic base for designing and fabricating microelectronics
6. Protect America’s technology advantages
7. Build a favorable international technology order
8. Win the associated technologies competitions
(Like Rohin, I've not read the full report - I've just read the executive summary, read all mentions of "nuclear", and listened to some episodes of the [accompanying podcast series](https://www.nscai.gov/podcasts/).)
Do "we" pay too little attention to things like this report?
============================================================
*Epistemic status: Just thinking aloud. As usual, these are my personal views only.*
I found the executive summary of the report, the accompanying podcast episodes I listened to, and the commentary Rohin collected interesting regarding:
1. Topics like AI development, AI risk, AI governance, and how this intersects with national security and international relations
2. *How the US government is thinking about about framing those things*, how the government might react to or try to influence those things, etc.
* Obviously the US government is massive and does not act as a single, coherent agent. But this report should still provide one interesting window into its thinking.
I've found what I've read more interesting for (2) than for (1), though I expect the full report would teach me a lot more about (1) as well. And I expect many other Forum readers would find the report - or at least its executive summary - useful for learning about those things as well. In particular, the report and accompanying podcast episodes seemed to have substantially different framings, focuses, and tones to what the longtermist/EA/rationalist communities talk about in relation to AI.
Yet there seems to be no previous link-post to this report on the EA Forum or LessWrong. (The section of the Alignment Newsletter linked to above serves a similar function, but doesn't provide a dedicated title and link just for this report.) Nor does there seem to be much discussion of the report on other posts on those sites. I also hadn't really considered looking at the report myself until last week, despite knowing it existed.
I wonder if this is a sign that (parts of) the longtermist, EAs, and rationalist communities are paying insufficient attention to the available info on how (a) other communities and (b) relevant powerful actors other than AI labs are thinking about these issues? (I think there are also other signs of this, and that similar criticisms/concerns have been raised before.)
Counterpoints: I've seen the report mentioned by some professional longtermist researchers (that was how it came to my attention). And the report is very long and (unsurprisingly) doesn't seem to be primarily focused on things like *existential* risk from AI. And it makes sense for people to specialise. So maybe it's unsurprising that things like this report would be discussed mostly by a handful of professional specialists, rather than in venues like the EA Forum or LessWrong.
Counter-counterpoint: The Forum and LessWrong aren't *just* like regular blogs for semi-laypeople; they're also where a nontrivial fraction of longtermist thinking occurs and is disseminated.
*Thus concludes this round of thinking aloud.* |
5682c318-f746-44fa-95a7-33a2357de1e4 | trentmkelly/LessWrong-43k | LessWrong | Rationality Quotes 14
"As for the little green men... they don't want us to know about them, so they refrain from making contact... then they do silly aerobatics displays within radar range of military bases... with their exterior lights on... if that's extraterrestrial intelligence, I'm not sure I want to know what extraterrestrial stupidity looks like."
-- Russell Wallace
"Characterizing male status-seeking as egotistical is like characterizing bonobo promiscuity as unchaste."
-- Liza May
"Introducing a technology is not a neutral act--it is profoundly revolutionary. If you present a new technology to the world you are effectively legislating a change in the way we all live. You are changing society, not some vague democratic process. The individuals who are driven to use that technology by the disparities of wealth and power it creates do not have a real choice in the matter. So the idea that we are giving people more freedom by developing technologies and then simply making them available is a dangerous illusion."
-- Karl Schroeder
"Hans Riesel held a Mersenne record for 14 days in the 50's, calculated using the first Swedish computer. My old highschool computing teacher had worked as a student on the system and had managed to crush his foot when a byte fell out of its rack and onto him."
-- Anders Sandberg
"Gentlemen, I do not mind being contradicted, and I am unperturbed when I am attacked, but I confess I have slight misgivings when I hear myself being explained."
-- Lord Balfour, to the English Parliament |
f97c325d-4616-4015-817b-fc2bb617e6a1 | StampyAI/alignment-research-dataset/aisafety.info | AI Safety Info | What is Redwood Research's strategy?
[Redwood Research](https://www.redwoodresearch.org/) focuses on interpreting and aligning current AI systems, identifying methods that seem promising, and considering theoretical arguments for why we might expect those methods to continue to work even as systems become much more intelligent. Some of their research includes:
- **[Causal scrubbing](https://www.alignmentforum.org/posts/JvZhhzycHu2Yd57RN/causal-scrubbing-a-method-for-rigorously-testing)**, a method for evaluating interpretability hypotheses about an AI system. Briefly, imagine we have some guess about which parts of a model do the calculations needed for the model to exhibit a particular behavior — e.g., for an image classifier, "I think these neurons here are for identifying dogs". Using some clever math, we can look through the model, identify which parts of the neural network "shouldn't matter" for that specific behavior (if our guess is correct), and "scrub" the model by replacing the activation values that "shouldn't matter" with values corresponding to a random input sample. Then, we can check whether the scrubbed model still exhibits the behavior we expected for the relevant input (i.e. correctly labeling a dog). If our hypothesis was a good one, the behavior of the scrubbed model should match the unscrubbed one.
- "**[Adversarial training for high-stakes reliability](https://arxiv.org/pdf/2205.01663.pdf)**" uses "[adversarial training](/?state=935A&question=What%20is%20adversarial%20training%3F)" — essentially, giving a model inputs specifically chosen to try to get it to produce undesirable outputs, then training the model on those outputs as examples of what to *not* do in the future — could significantly increase the model's reliability. In particular, Redwood aimed to produce models reliable enough that they could be used for "high stakes" tasks where a single failure would be catastrophic. To test this idea, Redwood used a language model which outputted completions to user prompts, and tried to fine-tune the model, using adversarial training, such that it would never complete a prompt in a way that involved "injury" (i.e. describing or implying someone getting hurt).
- **Doing interpretability work on large language models.** For instance, see [this paper](https://arxiv.org/abs/2211.00593) about how Redwood identified the circuit that [GPT-2](https://en.wikipedia.org/wiki/GPT-2) uses to figure out the proper indirect object of a sentence.
Redwood also maintains the [Rust Circuit Library](https://github.com/redwoodresearch/rust_circuit_public), intended to help researchers to do interpretability work on neural networks.
|
bf71d339-1fa7-425c-97e6-3e76d12676be | trentmkelly/LessWrong-43k | LessWrong | Rooting for Moments, Not Jerseys. Another Approach to Enjoying Sports
Epistemic Status: Somewhere between a personal life‑hack and a tentative general theory. I’ve been running this experiment for five NBA seasons; dozens of predictions‑market positions, numerous group‑chat quarrels. I am reasonably confident there’s something here, but I do not yet know how broadly it generalizes beyond U.S. team sports fandom.
TLDR
Traditional fandom couples your happiness to the fate of a single franchise whose base‑rate chance of winning a championship hovers around ~3 %. If you instead root for stories (individual players, milestone moments, tactical novelties) and back those beliefs with small prediction‑market stakes, you harvest nearly all the social energy of watching sports while slicing your disappointment rate by an order of magnitude.
The Standard Model of Fan Disappointment
1. Prior: There are 30 NBA teams, 16 make the playoffs, one wins.
Your team’s ex ante title equity = 1⁄30 ~ 3 %.
2. Observation: Most fans watch hundreds of hours for a <5 % chance of maximal joy.
By expected‑value math, a fan experiences ~19 seasons of varying degrees of pain for every season of euphoria (assuming teams have an equal chance, which isn’t the case)
3. Result: “Disappointed NBA fans” group‑chats; rage‑tweets; vows to boycott the league until the next season
We have built a cultural institution that systematically manufactures unhappiness in proportion to engagement. That seems… suboptimal.
A Player-Focused Bayesian Fandom
My counter‑hack emerged by accident. I started watching the playoffs with friends but lacked a team I cared about. I grew up in Maryland, so the Wizards were what I was “supposed” to pick but they very rarely won.
So I asked: Which players are on breakout trajectories that I’ll enjoy watching regardless of uniform?
Examples:
* 2018–19: Giannis Antetokounmpo, aspiring MVP talent but didn't win a championship yet.
* 2022–23: Shai Gilgeous‑Alexander (SGA), then an "outside shot" for MVP, now a c |
fe3f6c8a-dad0-4cac-b064-0819513acc3c | trentmkelly/LessWrong-43k | LessWrong | HP:MOR and the Radio Fallacy
No plot spoilers here, just wanted to flag a bit of poor reasoning that shows up in Chapter 39:
> I shouldn't have believed it even for all of thirty seconds! Because if people had souls there wouldn't be any such thing as brain damage, if your soul could go on speaking after your whole brain was gone, how could damage to the left cerebral hemisphere take away your ability to talk?
This is a surprisingly common fallacy. Just because X depends on Y, it doesn't follow that X depends on nothing but Y. A phenomenon may involve more than just its most obvious failure point.
To illustrate: Suppose I'm trapped in a box, and my only way to communicate with the outside world is via radio communication. Someone on the other end argues that I don't really exist -- "There's no person beyond the radio receiver, for if there was then there wouldn't be any such thing as damaged radios!" Pretty silly, huh? But people say this kind of thing in defense of physicalism all the time.
(N.B. This is not to defend the existence of souls. It's just to point out that this particular argument against them is invalid.) |
3bac0073-e160-40b8-9a04-034cbd73c84a | trentmkelly/LessWrong-43k | LessWrong | [LINK] Get paid to train your rationality (update)
Previous: http://lesswrong.com/lw/6ya/link_get_paid_to_train_your_rationality/
The IARPA-run forecasting contest remains ongoing. Season 1 has largely finished up, and groups are preparing for season 2. Season 1 participants like myself get first dibs, but http://goodjudgmentproject.com/ has announced in emails they have spots open for first-time participants! I assume the other groups may have openings as well.
I personally found the tournament a source of predictions to stick on PB.com and I even did pretty well in GJP. (When I checked a few weeks ago, I was ranked 28 of 203 in my experimental group.) I haven't been paid my honorarium yet, though. |
22bc8495-616a-4035-9628-f3eba9d82ea0 | trentmkelly/LessWrong-43k | LessWrong | Is there anything that can stop AGI development in the near term?
Assume that short timeline arguments are correct. (For previous discussion, see "What if AGI is near?".)
Some possible ideas:
* A multilateral, international governmental/legal agreement to halt GPU production or ban AI research
* Surveillance systems that detect when someone is about to launch an AI system and report them to the authorities. Obviously this would be just an implementation detail of the above idea.
* An agreement among prominent AI researchers (but not necessarily governments) in multiple countries that further progress is dangerous and should be halted until alignment is better understood
* A global nuclear war or some other disaster that would halt economic progress and damage supply chains
* Nuclear EMPs that would damage many electrical systems, possibly including computing hardware, while limiting casualties
Even in these scenarios it seems like further progress could be possible as long as at least one research group with access to sufficient hardware is able to scale up existing methods. So I'm just curious, would it even be possible to stop AI research in the near term? (This is different from asking whether it would be good to do it -- there are obviously reasons why the ideas above could be quite terrible.)
Also, should we expect that due to anthropic/observer selection effects, we will find ourselves living in a world where something like the dystopian scenarios discussed above happens, regardless of how unlikely it is a priori? |
34a4ebf0-dc7b-4a72-a35e-a6fa47fe4b2b | trentmkelly/LessWrong-43k | LessWrong | What are your cruxes for imprecise probabilities / decision rules?
An alternative to always having a precise distribution over outcomes is imprecise probabilities: You represent your beliefs with a set of distributions you find plausible.
And if you have imprecise probabilities, expected value maximization isn't well-defined. One natural generalization of EV maximization to the imprecise case is maximality:[1] You prefer A to B iff EV_p(A) > EV_p(B) with respect to every distribution p in your set. (You're permitted to choose any option that you don't disprefer to something else.)
If you don’t endorse either (1) imprecise probabilities or (2) maximality given imprecise probabilities, I’m interested to hear why.
1. ^
I think originally due to Sen (1970); just linking Mogensen (2020) instead because it's non-paywalled and easier to find discussion of Maximality there. |
b85c2c54-4689-4c52-a818-a2df306ad2fe | trentmkelly/LessWrong-43k | LessWrong | Notes on Moderation, Balance, & Harmony
This post examines the virtues of moderation, balance, and harmony. I wrote this not as an expert, but as someone who wants to learn more. I hope it helps people who want to know more about these virtues and how to nurture them.
I also use this opportunity to go on a (related) tangent about the VIA Institute on Character, and its questionable focus on playing to character strengths over strengthening character weaknesses.
What are these virtues?
If you practice these virtues, you respect that each part of your life is meant to serve the whole of it, not the other way around.
Even with generally beneficial or neutral things, it’s possible to take them too far. Moderation means you are on guard against this. Take fitness, for example: How can anyone complain about being more fit? But a compulsive focus on fitness as an end in itself can make you forget the point of being fit. “People who are always taking care of their health are like misers, who are hoarding a treasure which they have never spirit enough to enjoy.”[1]
If you are immoderate, you risk becoming a fanatic, an obsessive, a monomaniac. Such things in themselves might be mere eccentricities. But they are often accompanied by harmful neglect of other facets of life. Balance means that you have a good perspective on your life as a whole, and that you use this perspective to attend to the points that are most in need of attention, rather than letting habit or lack of moderation constrain your effort to where it’s less helpful. Work/life balance, balance between focus on self and others, and balance between living-in-the-present, learning-from-the-past, and preparing-for-the-future are some common challenges.
Harmony is closely related to integrity. Pursuing the musical analogy, when your life is harmonious, the different parts of it are attuned to each other; they are different instruments but playing the same symphony. Harmony is also not mere melody or monotone: being a “Johnny One-Note” isn’t where it |
fd7c6865-1d76-4fd9-9dd3-f5c52fa8b3b4 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Twin Cities, MN (for real this time)
Discussion article for the meetup : Twin Cities, MN (for real this time)
WHEN: 15 April 2012 01:00:00PM (-0700)
WHERE: Purple Onion Coffeeshop, 1301 University Avenue Southeast, Minneapolis, MN
[This meetup is really being put together by Helloses, but while his karma is just getting started I'm putting the text into a post on his behalf.]
Hi. Let's make this work.
Suggested discussion topics would be:
* What do we want this group to do? Rationality practice? Skill sharing? Mastermind group?
* Acquiring guinea pigs for the furtherance of mad science (testing Center for Modern Rationality material)
* Fun - what it is and how to have almost more of it than you can handle
If you'd like to suggest a location closer to you or a different time, please comment to that effect. If you know a good coffeeshop with ample seating in Uptown or South Minneapolis, we could meet there instead. Also comment if you'd like to carpool.
If you're even slightly interested in this, please join up or at least comment.
Folks, let's hang out and take it from there.
Discussion article for the meetup : Twin Cities, MN (for real this time) |
9d6839c1-9e51-4e89-859e-26c4317c8950 | StampyAI/alignment-research-dataset/youtube | Youtube Transcripts | Ethan Caballero–Scale Is All You Need
A bunch of people at Google
said, yeah, we have language
models that are way bigger than GPT-3,
but we just don’t put them in papers.
The DeepMind language models papers,
they were a year old when they finally
put them out on arXiv or whatever.
When Ilya tweeted the consciousness tweet,
they were like, goddamn, GPT-4 must be crazy.
Like there’s a zillion VCs throwing money
at large language model startups right now.
At some point, the Beijing Academy of AI will
be like, look, we just trained a 10 to the 15
parameter model on all of YouTube and spent like
$40 billion doing it. And then at that point,
Jared Kaplan’s gonna be in the White House
press conference room will be like, look,
see these straight lines on log log
pots, we gotta do this in the USA now.
The Inside View. The Inside View. The Inside View.
Ethan, you’re a master’s degree
student at Mila in Montreal,
you have published papers on out of distribution,
generalization, and robustness generalization
accepted as presentations and spotlight
presentations at ICML and NeurIPS. You’ve recently
been thinking about scaling laws, both as an
organizer and speaker for the first neural scaling
laws workshop in Montreal. You’re currently
thinking about the monotonic scaling behaviors
for downstream and upstream task, like
in the GPT-3 paper, and most importantly,
people often introduce you as the edgiest person
at Mila on Twitter, and that’s the reason why
you’re here today. So thanks, Ethan, for coming
on the show and it’s a pleasure to have you.
Likewise.
Scaling Laws T-Shirts
You’re also well-known for
publicizing some sweatshirt mentioning
scale is all you need AGI is coming.
Yeah.
How did those sweatshirts appear?
Yeah, there was a guy named Jordi
Armengol-Estapé who interned at Mila,
and he got really into scaling laws, apparently
via me. And then he sent me the shirt and was
like: look how cool this shirt is. Like, he’s
the person wearing the shirt in the picture,
and he’s like, look how cool this shirt I just
made is. And so then I tweeted the shirt. And then
Irina just turned it into a merchandising scheme
to fund future scaling. So she just made a bunch
and started selling it to people. Like apparently,
like she sells like more than 10 to Anthropic
already. Just scaling lot of t-shirts, that’s
the ultimate funding model for supercomputers.
Scaling Laws, Upstream and Downstream tasks
Maybe you can like explain intuitively for
listeners that are not very familiar
to what are scaling laws in general.
Whatever your bottleneck compute data parameters,
you can predict what the performance will be
as that bottleneck is relieved. Currently,
the thing most people know how to do is
predict like the upstream performance. Like
the thing people want though is to be able
to predict the downstream performance and
upstream is what you’re like… It’s like
your literal loss function that you’re
optimizing and then downstream is just
any measure that you have of, like something you
care about, so just like a downstream dataset,
or like, I mean, usually, it’s just
mean accuracy on a downstream dataset.
And to take like concrete
examples, like for GPT-3, the
upstream task is just predict the next
word. What are the downstream tasks?
Like 190… a zillion like benchmarks that the
NLP community has come up with over the years.
Like they just evaluated like the accuracy
and like things like F1 score on all those.
And yeah, what should we care
about upstream and downstream task?
I mean, basically like up, well, we don’t
really care about upstream that much. Upstream’s
just the first thing that people knew how to
predict, I guess, like predict the scaling
of what we care about as downstream. I mean,
basically, like downstream things that improve
monotonically, they kind of can be interpreted
as like capabilities or whatever, and then
downstream stuff that doesn’t necessarily improve
monotonically often is stuff that is advertised as
alignment stuff. So like toxicity or if
you like speculate in the future, stuff
like interpretability or controllability would
be things that might not improve monotonically.
So you don’t get more interpretability
as you scale your models?
You do currently, but the class example is
like CLIP. It gets more interpretable as it
has representations that make more sense.
But you can imagine at a certain point,
it’s less interpretable because then at a certain
point, the concepts it comes up with are beyond
human comprehension. Like now it’s just how
like dogs can’t comprehend calculus or whatever.
Defining Alignment and AGI
Yeah, when you mention alignment, what’s
the easiest way for you to define it?
I mean, the Anthropic definition’s pretty
practical. Like we want models that are
helpful, honest, and harmless, and
that seems to cover all the like
weird edge cases that people can like come
up with on the Alignment Forum or whatever.
Gotcha, so it is not like a technical
definition. It’s more a theoretical one.
Yeah, yeah.
So would you consider yourself an alignment
researcher or more like a
deep learning researcher?
I’d say just a beneficial AGI researcher.
That seems to cover everything.
What’s AGI?
The definition on NASA website’s pretty
good. Highly autonomous systems that
outperform humans at most
economically valuable tasks.
AI Timelines
When do you think we’ll get AGI?
I’ll just say like, it depends
mostly on just like compute stuff,
but I’ll just say 2040 is my median.
What’s your like 10% and 90% estimate?
10%, probably like 2035.
Recent Progress: AlphaCode, Math Scaling
I think there’s been a week where we got
DALL-E 2, Chinchilla, PaLM. Did that
like update your models in any way?
The one that I thought was the like… was
the crazy day was the day that AlphaCode
and the math-proving thing happened on the same
day, because like, especially the math stuff,
like Dan Hendricks has all those slides where he
is like, oh, math has the worst scaling laws or
whatever, but then like OpenAI has like the IMO
stuff. So like at least according to like Dan
Hendricks’ slides, whatever, that would’ve been
like, something that took longer than it did.
So when you mentioned the IMO stuff, I think
it was like at problem from maybe 20 years ago,
and it was something that you can like
do with maybe like two lines of math.
I agree they weren’t like super,
super impressive, but it’s more
just the fact that math is supposed to have
like the worst scaling supposedly, but like
impressive stuff’s already happened with math now.
Why is math supposed to have the worst scaling?
It’s just an empirical thing. Like
Dan Hendricks has that like math
benchmark thing and then he tried to do
some extrapolations based on the scaling
of performance on that. But the amount
of computing data we currently have,
it’s already like doing interesting
stuff was kind of surprising for me.
I think in the paper, they mentioned that the
method would not really scale well because of,
and some infinite actions base when
trying to think of like actions.
Yeah.
So yeah, I didn’t update it. I was like,
oh yeah, scaling will be easy for math.
I didn’t update it as easy, but
just easier than I had thought.
The Chinchilla Scaling Law
Okay, related to scaling,
the paper by DeepMind about the Chinchilla
model was the most relevant, right?
Yeah, I thought it was interesting. Like,
I mean, you probably saw me tweet it,
like that person on Eleuther
Discord that was like, oh wait,
Sam Altman already said this like six months
ago, but they just didn’t put it in a paper.
Yeah, he said that on the Q&A, right?
Yeah, yeah.
Yeah, he said something like we shouldn’t,
our models will not be like much bigger.
Yeah. He said they’ll use way more
compute, which is analogous to saying,
there you’ll train a smaller
model, but on more data.
Can you like explain the kind
of insights from scaling laws
between like compute model size, and then like
what’s called like the Kaplan Scaling law?
It was originally something about computing.
If your compute budget increase a billionfold,
your model size increases a millionfold and
your dataset size increases a thousandfold.
And now it’s something like, I know it’s like
one to one, but I don’t remember like how big
the model size to like compute ratio was. I know
like the model-to-data ratio is one to one now,
but I don’t remember what the compute-to-model
ratio is, the new compute-to-model ratio is.
That’s also what I remember, and
I think like the main insight
from the first thing you said from the Kaplan
law is that like model size is all those matters
compared to dataset and
for a fixed compute budget.
Yeah, the narrative with the Kaplan one was model
size, like compute is the bottleneck for now until
you get to the intersection point of the compute
scaling and the data scaling, and at that point,
data’s gonna become more of a bottleneck.
So compute is the bottleneck now.
What about like having huge model?
But yeah, yeah. That’s like, because like they
were saying that because model size grows so fast.
So like to get the bigger models, you need more
compute rather than like, you don’t need more data
‘cause like you don’t even have enough compute
to like train a large model on that data yet,
with the current compute regime… was the
narrative of the first of the original Kaplan
paper. But it’s different now because like the
rate at which you should be getting data given,
like the rate at which your data charge should be
increasing given your compute budget is increasing
is a lot faster now, like using the Chinchilla
scaling law. For some increasing compute size,
you’re gonna increase your model by a certain
amount, and the amount that you’re dataset size
increases is like a one-to-one relation to the
amount that your model size increases. I don’t
remember what the relation between model and
compute was, but I know that now the relation
between model and dataset size is one to one,
between model size and dataset size is one to one.
And the main size is that now we can just
have more data and more compute, but not like
a lot of more compute. We just need the same
amount as more compute. So we can just like
have to scrap the internet and get more data.
It just means like to use
your compute budget optimally,
the rate at which your dataset
size grows is a lot faster.
Does that make you more confident that we’ll
get like better performance for models quicker?
Maybe for like YouTube stuff, because YouTube,
we’re not bottlenecked by data. We’re bottlenecked
by compute, whatever. But that implies the
model sizes might not grow as fast for YouTube
or whatever. But for text, we’re probably gonna be
bottlenecked by… It means we’re probably gonna be
bottlenecked like text and code by the dataset
size earlier than we thought. But for YouTube,
that might like speed up the unsupervised
video on all of YouTube, like timeline stuff.
Limits of Scaling: Data
Yeah, so I’m curious when
do you think about like how much are
we bottlenecked by data for text?
Yeah, I asked Jared Kaplan about
this, and he said like, “Wait,
okay. “It’s 300 billion tokens for GP3.” And
then he said like, library of Congress, whatever,
could be 10 trillion tokens or something like
that. And so like the most pessimistic estimate of
how much like the most capable organization could
get is the 500 billion tokens. A more optimistic
estimate is like 10 trillion tokens is how many
tokens the most capable organization could get,
like mostly English tokens.
So how many like orders of magnitude in
terms of like parameters does this give us?
I don’t remember what the… Like I haven’t
calculated it. Like I remember I kind of did it
with the old one, but I haven’t done it with the
new Chinchilla one. But I mean, you said this in
your thing today or whatever, like we probably
are gonna be bottleneck by the amount of code.
I was essentially quoting Jared Kaplan’s video.
Code Generation
Yeah, yeah, but he, I mean, he’s right.
I’m kind of wondering what’s philanthropic
thinking of Adept, because Adept’s like doing
the training all the code thing, and Adept was
gonna do all the train on all the code thing,
and they’re like, oh crap, we got another
startup doing the train on all the code stuff.
Yeah, so I think you said that if you remove
the duplicates on GitHub, you get some amount
of tokens, maybe like 50 billion tokens, 500, I’m
not sure. Maybe 50 billion. Don’t put me on that.
Yeah.
And yeah, so the tricks will be data
augmentation… you’re like applying the
real things to make your model better, but it’s
not clear how do you improve performance? So my
guess would be you do transfer learning, like
you train on like all the different languages.
That’s definitely what they plan on
doing, like you see the scaling lots
for transfer paper is literally pre-train
on English and then fine-tune on code.
My guess is also that like, if you get a bunch
of like the best programmers in the world
to use co-pilot and then you get like feedback
from what they accept, you get higher quality
data. You get just like, oh yeah, this work
just doesn’t work. And so you have like
1 million people using your thing 100 times a
day, 1,000 times a day, then that’s data for free.
I mean, I view that part kind of as like
the human feedback stuff is kind like the
alignment part is the way I view it. I mean,
then there’s some people who like say, oh,
there might be ways to get like better
pre-training scaling if you have like
humans in the loop during the pre-training,
but like, no one’s really figured that out yet.
Well, don’t you think like having
all this telemetric data from
GitHub cooperatives is you can use it, right?
Yeah, yeah, but I almost view it as like
that it’s like used for alignment, like for
RL from human preferences.
Okay. Gotcha. Yeah, I think the other thing they
did for improving GPT-3 was just having a bunch of
humans rate the answers from GPT-3 and then like
that’s the paper of instructivity. I think like
they had a bit of humans and it kind of
improved the robustness or not for business, but
alignment of the answer somehow. Like
it said less like non-ethical things.
Yeah. I mean it’s like people downvoted
the non-ethical stuff, I think.
Youtube Scaling, Contrastive Learning
Exactly, yeah. And to go back to YouTube,
why is scaling on YouTube interesting?
Because there’s unlimited data?
Yeah, one, you’re not banned, but I mean, the gist
is YouTube’s the most diverse, like simultaneously
diverse and large source of
like video data basically.
And yeah. So for people who were not used
to or thinking, what’s the task in YouTube?
Yeah, it could be various things. Like it might be
like a contrastive thing or it might be a predict
all the pixels thing. Like, I mean, so like
at least places like Facebook seem to think
like contrastive has better downstreams scaling
laws, so it’s gonna be a contrastive type thing.
What’s contrastive type thing?
Like you want representations that have similar
like semantic meaning to be close together,
like have low cosign similarity, like in
latent space. So basically, like maximize
the mutual information between views. Like
it’s kind of hard to explain without pictures.
So you’d say that your model takes a video,
like all of the videos and views as input?
Frames that were close together like in time, it
tries to maximize the mutual information between
them via maximizing cosign similarity between
the latents of like a resonant encoder or
whatever that encodes the images for both of those
frames that were next to each other, like in time.
So he tries to kind of predict
correlations between frames in some kind of
latent space from a resonance?
Yeah, yeah. In the latent space, you want frames
that were close to each other in time to have
similar, like maximize the cosign similar between
the latent space between the latent between the
hidden layer output by the like resonance that
took each of those in each of those frames in.
And at the end of the day, you want something that
is capable of predicting how many frames in lens.
Kind of for, well, the like philosophy with
like the contrastive stuff is we just want a
good representation that’s useful for downstream
tasks or whatever. So like you don’t actually
like, there’s no like output really.
It’s just you’re training a latent space
or whatever that can be fine-tuned
to downstream tasks very quickly.
What are the useful downstream
tests, like robotics?
Yeah, yeah. Like there’s a zillion
papers on like people pre-train on
do some pre-train contrastive thing in like an
Atari environment, and then they show like, oh,
now we barely need any RL steps to like fine-tune
it or whatever and it can like learn RL really
quickly after we just did all this unsupervised
contrastive, like pre-training or whatever.
And yeah, wouldn’t your model be kind of
shocked by the real world when you just
like show him like YouTube videos all the time
and then you trust the robot with like a camera?
Kind of not. I mean, ‘cause there there’s
like everything on YouTube. They got like
first person egocentric stuff, they got third
person stuff. Like it’ll just like realize which,
like whether it’s in first or third person pretty
quickly. I feel like it just infers the context.
Like now I saw GPT-3 just for the context, it’s
in, ‘cause it seemed like every context ever.
Gotcha. So I was mostly thinking
about like entropy of language.
If it’s literally like a video generative model,
then you can do like just the perfect analogies,
GPT-3 or whatever. It gets a little trickier
with like contrastive stuff, but yeah, I mean
either one. I mean the analogies
are pretty similar for either one.
So one of the things about the scaling
laws papers and the role of scaling laws,
there was some different exponents for text.
Yeah.
Scaling Exponent for Different Modalities
What do you think is the exponent for
video? Would it be like much worse?
I know the model size. The model size relation
was the big point of the scaling laws. For
autoregressive generative models, the paper says
that the rate at which the model size grows,
given your compute budget grows,
is the same for every modality.
So that was kind of like, that’s
like a big unexplained thing.
Like that was the biggest part just of that paper
and no one’s been able to explain why that is yet.
So there might be some universal law where scaling
goes for all modality and nobody knows why.
Just stuff. The rate at which
your model size grows given
your compute budget is increasing
is the same for every modality,
which is kind of weird and no one, like I
haven’t really heard a good explanation why.
Who do you think will win
the video prediction race?
AGI Race: the Best Funding
Model for Supercomputers
The person who wins AGI is whoever has the best
funding model for supercomputers. Whoever has
the best funding model for supercomputers wins.
Like, I mean yet to assume all entities are like,
they have like the nerve, like we’re gonna do
the biggest training run ever, but then given
that’s your pre-filter, then it’s just whoever
has the best funding models for supercomputers.
So who is able to spend the most money?
So would it be USA, China, Russia?
Yeah, yeah, it might be something. I mean, my
guess is like China’s already, like they already
have this joint fusion of industry government
and academia via the Beijing Academy of AI
in China. So my guess is like at some point,
like Beijing Academy of AI and be like, look,
we just trained like a 10 to the 15 parameter
model on all of YouTube and spent like $40 billion
doing it. And then at that point, Jared Kaplan’s
gonna be in the White House press conference room,
be like, look, see these straight lines on
log log pots, we gotta do this in the USA now.
Right, right. But how do you
even spend that much money?
By making people think if
they don’t, they’ll no longer be the
superpower of the world or whatever. Like
China will take over the world or whatever. Like
it’s only like a fear. It’s only a fear thing.
From looking at the PaLM paper from Google, they
seem pretty clever on how they use their compute.
You mean the thing where they have like the
two supercomputers that they
split it across or whatever?
Right. I think TPU pods or
something, they call it.
Yeah, yeah.
So it didn’t seem like they
spent more money than OpenAI.
So they tried to be more careful somehow. So my
model of like people spending a lot of money is.
Like most entities won’t be willing to like do the
largest training when they
can, given their funding.
So maybe China, but I see Google as being more
helpful because of they do it
on paper, but maybe I’m wrong.
Jared Kaplan says like most like Anthropic and
OpenAI are kind of unique in that they’re like,
okay. We’re gonna like throw all our funding
into this one big training run. But like Google
and like ‘cause Google and Amazon, they have like
he said like at least, 10X or like 100X times the
compute that OpenAI and Anthropic have, but they
never like use all the compute for single training
runs. They just have all these different teams
that use to compute for these different things.
Yeah, so they have like a different
hypothesis. OpenAI is like scale is
all that matters, somehow that
they’re secrets itself and-
Yeah, it’s something like that.
You just let scale things and we
are going to get better results,
and Google is maybe there’s more bureaucracy
and it’s maybe harder to get a massive budget.
Private Research at Google and OpenAI
Yeah, it’s weird though, ‘cause
Jeff Dean’s latest blog
post, it summarizes all the
Google’s research progress mentions like scaling
and scaling while it’s a zillion times. So that
almost implies that like they’re on the scales.
All you need bandwagon too. So I don’t know.
They probably know, but then the question
is how like private things are and
maybe there’s stuff we don’t really know.
I know a bunch of Google said like,
yeah, we have language models that
are way bigger than GPT-3, but
we just don’t put ‘em in papers.
So you’ve talked to them like privately
or is it just, they said online?
I just I’ve heard things from people and
that’s feasible. I’m not just disclosing
what I got that information from, but
that’s just what I’ve heard from people.
So as we’re on like gossip, I think like
something that was around on the internet,
like right when GPT-3 was launched was that Google
was like reproduced it in a few months afterwards,
but they didn’t really talk about it publicly. I’m
not sure about what to do with this information.
I know like the DeepMind language models
papers that they were a year old when
they finally put ‘em out on archive or
whatever, like Gopher and Chinchilla.
They had the language model finished
training a year before the paper came out.
So we should just like assume all those big
companies are just like throwing papers when
they’re like not relevant anymore when
they have like the other paper already?
Maybe, but yeah. I don’t know why
it was delayed that much. Yeah,
I don’t know what the story is.
Why it was delayed that long.
People want to like keep their advantage, right?
I guess, but I mean like I feel like GPT-3,
they threw the paper on arXiv pretty
soon after they finished training GPT-3.
How do you know?
Yeah, I don’t, but I mean,
yeah, I don’t. But ice, it didn’t. Yeah,
maybe there was a big delay. I don’t know.
So I think you could just like retrace all
Sam Altman tweet and then like you read the
next paper like six months after and you’re
like, oh yeah, he tweeted about that. Like
sometimes the tweets like, oh, AI is going to be
wild, or oh, neural networks are really capable
of understanding. I think you tweeted that like
six months ago, like when they discovered GPT-4.
OpenAI is like when Ilya tweeted the consciousness
tweet, they’re like, goddamn, GPT-4 must be crazy.
Yeah, neural networks are in
some ways slightly conscious.
Yeah, yeah, that was the funniest quote.
Yeah, I think people at OpenAI know things we
don’t know yet. They’re all like super hyped.
And I think you mentioned as
well that at least privately that
Microsoft has some deal with OpenAI and so they
need to some amount of money before 2024, like.
Oh yeah, yeah, yeah, yeah. I mean, right,
right. When the Microsoft deal happened, like
Greg Brockman said, “Our plan is to train
“like a 100 trillion parameter model by 2024.”
Okay, so that’s in two years?
I mean, that was in 2019,
but maybe they’ve changed their mind after like
the Chinchilla scaling lot stuff, I don’t know.
Why Ethan did not update that much from PaLM
Right. And so you were not like impressed
by PaLM being able to predict to like do
logic on airplane things and explain jokes?
In my mind, like the video scaling was like
a lot worse than text basically. That’s the
main reason why I like AGI will probably take
longer in the five years or whatever in my mind.
Okay, so we need, so if we just
have text, it’s not enough to
have AGI. So if we’re a like a perfect Oracle
that can like talk like us, but it’s not able to
do robotic things, then we don’t have AGI.
Yeah.
Well, I guess my main like is mostly
like coding. So if we get like coding,
like Codex or comparative, that gets really good,
then everything accelerates and engineers
become very productive, and then like.
I guess if you could say like, engineers get
really productive at making improvements in
hardware, then like, maybe that would, like,
I get how that would be like, okay. Then it’s
really fast. Like in my mind, at least at
the current, I don’t see the hardware getting
fast enough to be far enough on the YouTube
scaling lot in less than five years from now.
Thinking about hardware, we’re just
like humans, Googling things and using.
Yeah, yeah. I get what you’re saying.
Like you get the Codex thing and then
we use Codex or whatever
to design hardware faster.
You mentioned you have like DALL-E,
but like for designing chips.
I mean, Nvidia already uses
AI for designing their chips.
That doesn’t make you think of
timelines of 10 years or closer.
It may be 10 years, but not five years.
The thing I’m trying to figure out is like,
try to get like a student researcher gig
at like someplace so that I can just get
access to the big compute during the PhD.
Oh, so that’s your plan. Just get out of compute.
Yeah, I mean, as long as I have big compute,
it doesn’t matter where I’m a PhD. I mean,
it kind of matters if you’re like trying
to start an AGI startup or whatever, but
safe, safe, safe AGI startup.
We’re kind of on record, but
I’m not sure if I’m going to
cut this part. So you can say unsafe, it’s fine.
Yeah, no, no, no. I mean, I don’t even
phrase. I just phrase it as beneficial AGI.
You were spotted saying you wanted
unsafe AGI the fastest possible.
Thinking about the Fastest Path
No, no, no. The way I phrase it is
I think I explained this last time, you have to
be thinking in terms of the fastest path, because
there is like extremely huge economic and military
incentives that are selecting for the fastest
path, whether you want it to be that way or
not. So like, you gotta be thinking in terms of,
what is the fastest path and then how do you like
minimize the alignment tax on that fastest path?
‘Cause like the fastest path is the way
it’s probably gonna happen no matter what,
like, so it’s about minimizing the
alignment techs on that fastest path.
Or you can just throw nukes everywhere
and try to make things slower?
Yeah, I guess, but I mean the people who are
on the fastest path will be like more powerful,
such that like, I don’t know, such
that they’ll deter all the nukes.
So you want to be, okay, so you
want to just like join the winners.
Like if you join the skiing team at Google.
Thing I’ve been trying to brainstorm
about is who’s gonna have the fastest,
who’s gonna have the best funding models for
supercomputers, ‘cause that’s the place to go
and you gotta try to minimize
the alignment tax at that place.
Makes sense. So everyone should infiltrate Google.
Yeah, so whatever place ends up with
the best funding model of supercomputers
try to get as many weird alignment people
to like infiltrate that place as possible.
So I’m kind of happy having a
bunch of EA people at OpenAI now,
because they’re kind of
minimizing the text there, but…
Yeah, I kind of viewed it as all the EA
people left, like ‘cause Anthropic was
like the most extremist EA people at OpenAI.
So I almost viewed when Anthropic happened
a bunch of EA people. I view as that like EA
almost leaving OpenAI when Anthropic happened.
Some other people came, right?
Like who?
I don’t know. Richard Ngo.
Oh, okay, okay. Yeah, yeah.
It’s like a team on like predicting the future.
Yeah, yeah. I wanna know what
the Futures Team does ‘cause
that’s like the most out there team. I’m
really curious to what they actually do.
Maybe they use their GPT-5
model and predict things.
Right, ‘cause I mean like DALL-E, like you
know about the Foresight Team at OpenAI, right?
They were trying to predict
things as well, like forecasting.
Yeah, that’s where all this scaling lot stuff came
from was on the Foresight Team at OpenAI. They’re
gone now because they became philanthropic. But
like a team called like the Futures Team that
almost has a similar vibe to like a team called
the Foresight Team. So I’m kind of curious.
But then there’s just like doing more governance
things and optimal governance and maybe economics.
That’s what it’s about, governance and economics.
The guy like Richard Ngo
is doing governance there.
Okay.
Predicting how the future works,
I think is in his Twitter bio.
Yeah, yeah, but I mean, that’s
somewhat tangential to governance,
like that almost sounds like something Mike Rick
Kurtz, I would say, I’m predicting how the future.
My model is like Sam Altman,
as like they have GPT-4.
Like they published GPT-3 in
2020. So it’s been like two years.
Yeah.
And they’ve been talking about like
in their Q & A about like treacherous
results or something like one year ago. So now
they must have access to something very crazy
and they’re just like trying to think like how
do we operate with like DALL-E 2 and their GPT-4
they have in private and how they do something
without like for him in the world? I don’t know.
Maybe they’re just like trying to predict like
how to make the most money with their API or.
You’re saying like if they release it,
it’s like an infohazard? ‘Cause in my mind,
GPT-4 still isn’t like capable enough
to F up the world, but you could argue,
it’s like capable enough to like
be an infohazard or something.
Imagine you have access to
something that has the same
kind of gap between GPT-2 and GPT-3, but like for
GPT-4 on like understanding and being general.
And you don’t want everyone
else to copy your work.
So you’re just going to keep
it for yourself for sometime.
A Zillion Language Model Startups from ex-Googlers
Yeah, but I feel like that strategy is already
kind of screwed. Like you know about how like
a zillion large language model, like a zillion
Googlers have left Google to start large language
model startups. Like there’s literally three
large language model startups by ex-Googlers now.
OpenAI is like a small actor in this now because
there’s like multiple large language model
startups founded by ex-Googlers that all like
that all were founded in the last like six months.
Like there’s a zillion VCs throwing money
at large language model startups right now.
The funniest thing, like Leo Gao, he’s like, we
need more large language model startups because
the more startups we have, then it splits up all
the funding so no organization can have all the
funding to get the really big supercomputer.
So we just need thousands of AI during its
final startups. So no one can hoard all the
funding to get the really big language model.
That’s the, yeah, with the AI model,
you just like do open source. So like
there’s like more startups. And so all
the funding gets splitted, I guess.
Yeah, you could view OpenAI was like extra big
brain. We need to do. We need to like release
the idea of our joiners models onto the world such
that no organization could have enough compute to
be such that all the compute gets more split up,
‘cause a zillion, our joiners model
startups will show up all at once.
That’s yeah, that’s the best idea ever. So do you
have like other gossips besides like Google’s?
Did you post something on Twitter
about people leaving Google?
Yeah, I posted a bunch of stuff. Well, I mean, and
also like you saw the… I mean it’s three startups,
adept.ai, character.ai, and inflection.ai.
They’re all large language model startups
founded by ex-Googlers that got a zillion dollars
in VC funding to scale large language models.
What’s a zillion dollars like?
Like greater than 60 million. Each
of them got greater than 60 million.
So did they know about something we
don’t know? And they’re just like
get money to replicate what Google does?
Well, I mean, most of ‘em,
they were famous people like
founder of DeepMind scaling team. Another
one is the inventor of The Transformer.
Another one was founded by a different
person on The Transformer paper. Like,
so I mean, in some ways, they have more
clout than like OpenAI had or whatever.
But they don’t have like the
engineering and old infrastructure.
No, they kind of do. Like, a lot of ‘em,
they were like the head of engineering for
scaling teams at like DeepMind or Google.
So there’s like another game that is in
private at Google and they’ve been scaling
huge models for two years. and they’re just like,
Yeah, something like that.
Starting startups with their knowledge and they’re
just scaling and we;re just like, like peasants
like us talk about papers that are released
one year after and then when you turn them out.
Yeah, yeah. I guess that’s, I mean, I don’t know
how long these delays are. I mean, in my mind,
like, yeah. I guess you could view it as a delay
thing, ‘cause like in my mind it’s just like,
yeah, you’re right, you’re right.
It’s probably delayed by a year, yeah.
So yeah, that makes me less confident about-
Oh shit. You look like a clone
of Lex Fridman from the side.
What?
When your face is like sideways, you
look like a clone of Lex Fridman.
Yeah.
Like, ‘cause your haircut’s
like identical to his when
I’ll take that as a compliment… I started working
out. So yeah, Ethan Caballero, what’s the meaning
of life?
Probably just maximize the flourishing of all
sentient beings, like a very generic answer.
Right. So I’ve done my Lex Fridman
question. Now I’m just basically him.
Yeah.
Ethan’s Scaling Journey
Maybe we can just go back to
like stuff we know more about like your work and
because you’ve been doing some work on scaling.
Yeah.
So like more general, like why are you
kind of interested in scaling and like how
did you started on doing research on that?
I mean, I knew about the body paper when
it came out. Like I remember I was at this
like Ian Goodfellow talking in 2017 and he was
hyped about the body paper when it came out.
Which paper?
The deep burning scales, predictably, empirically,
yeah, it came out 2017 and then I kind,
I just, that was just on the back burner and I
kind of just stopped paying attention to it after
a while. And then like Aran Komatsuzaki was
like, no, dude, this is the thing. Like this
is gonna take over everything, and this was
like in 2019 when he was saying that. And then,
yeah. So then when the scaling laws papers got
like re-popularized through like the OpenAI stuff,
then I kind of like caught onto it a little
bit early via like talking with Aran.
I think in 2019 was also
when GPT-2 was introduced.
But that was kind of before, like
that was before like the scaling
law stuff kind of got popularized.
Right, scaling laws paper is 2020.
Yeah, the very end of 2020.
All right. No, no, no. Oh no,
no. The scaling law paper was the very end
of 2000. It was the very beginning of 2020.
And you were already on this
killing train since 2017.
I was aware of it, but I didn’t, like, I
was kind of just neutral about it until like
2000, like probably the middle of 2019.
Making progress on an Academic
budget, Scaling Laws Research
And yeah, now you are kind of
interested in scaling because
it’s useful to predict kind of what
the whole field of AI is going.
And also it just, it’s I think people
underestimate how easy it is to be contrived
if you’re not paying attention to scaling trends
and trying to like extrapolate the compute budgets
and data budgets that are like, well, the compute
data and data budgets like five years from now.
Yeah, if you’re a huge company that does a lot of
budget, but maybe if you’re just a random company,
you don’t really care about scaling law that much.
Yeah, yeah. Or if you’re like in
academia currently or whatever,
like a zillion papers that like fancy conferences
are like, here’s our inducted bias that helps on
like our punny academic budget. And we didn’t
test any of the scaling asso tos to see if it’s
like useful when you’re training a trillion
parameter model on all of YouTube or whatever.
You’re on an academic budget as far as I know. So
how do you manage to do experiments in scaling?
There’s like the scaling on narrative. That’s
like, oh, you don’t need the big budget
to do because you can just predict what the
outcomes will be for the large scale experiments,
but that’s at least current. At least
when that narrative got popularized,
it was mostly for upstream like scaling. But the
thing everyone cares about is downstream scaling.
AI Alignment as an Inverse Scaling Problem
Yeah, so if we go back for a minute on
like your work in alignment, how
do you think your work on scaling
or generalization like kind of
fits with the alignment problem?
Basically, all alignment, I guess this
triggers the hell outta some people. But
all alignment is inverse scaling problems.
It’s all downstream inverse scaling problems.
So it’s just in my mind, all of alignment is
stuff that doesn’t improve monotonically
as compute data and parameters increase.
There’s a difference between not improving and
inverse scaling. Inverse
scaling goes badly, right?
Yeah, yeah, yeah. But I just said not improved
monotonically, ‘cause like sometimes there’s
certain things where like it improves for a while,
but then at a certain point, it gets worse. So
like interpretability and controllability are the
two like kind of thought experiment things where
you could imagine they get more interpretable
and more controllable for a long time until they
get super intelligent. At that point. they’re
like less interpretable and less controllable.
Do we have benchmarks for controllability or?
Like just like just benchmarks that rely on
prompting is a form of like a
benchmark of controllability.
And kinda to summarize your take, if we
were able to just scale everything well
and not have this inverse scaling
problem, we would get like
interpretability and controllability
and everything else by just like good
scaling of our models. And so we’d get like
alignment kind of by defaults for free?
Yeah. I mean, I guess, I mean like there’s
stuff besides interpretability, controllability,
like those are just the examples. Like what you
said, you asked like what’s an example where
like the reason I said, I phrased it as
alignment is when I said inverse scaling,
I said things that don’t improve monotonically,
‘cause I just wanted to say like,
yes, there’s obvious examples where it gets
worse the entire time, but there’s some you
could imagine where it gets good for a long time,
and then a certain point, then it starts getting
drastically worse. I said, all of alignment
can be viewed as a downstream scaling problem.
The hard part is like Dan Hendricks and like Jacob
Steinhardt say like, then the hard problem though
is like measurement and like finding out what
are the downstream evaluations ‘cause say you
got like some like fancy like deceptive AI that
wants to like a treacherous turn or whatever.
Like how do you even find the downstream
evaluations to know whether it’s gonna like
try to deceive you or whatever? ‘Cause like when
I say, it’s all a downstream scaling problem,
that assumes like you have the downstream test,
the downstream like thing that you’re evaluating
it on. But like if it’s like some weird deceptive
thing, that’s like, it’s hard to even find what’s
the downstream thing to evaluate it on to like
know whether it’s trying deceive or whatever.
So there’s no like test lost on
this deception. We don’t know
for sure how to measure and have
a clear benchmark from this.
Yeah, it’s tricky. I mean, and some
people say like, well, that’s why
you need better interpretability. You need to
like find the deception circuits or whatever.
Knowing that we don’t know yet, like all the
different benchmarks and metrics for misalignment,
don’t you think that your
work on scaling can be bad
because you’re actually
like speeding up timelines?
Predicting scaling laws,
Useful AI Alignment research
Yeah, I get the like infohazard
point of view, but like
in my mind, like whether you wanna do all
capabilities or alignment stuff that stands
the test of time, you need really
good downstream scaling prediction.
Like, say you came up with some like alignment
method or whatever that mitigates inverse scaling,
like you need the actual functional form to know
whether that thing will like keep mitigating
inverse scaling when you get to like a trillion
parameters or whatever. You get what I mean?
I get you but like on a differential
progress mindset, like Jared Kaplan or
someone else will come up with those
functional forms without your work.
I don’t know, I don’t know. I
mean, that’s the thing though, like
Anthropics (ERRATUM: it’s actually a gift, and the
the merch was not sent at the time of the podcast)
got that paper like predictability and surprise
and generative models and they’re just like,
it’s unpredictable. We can’t predict it. And
I’m like, ah, you guys, nah, I don’t believe.
Right, so you’re kind of publishing
papers when you’re in advance
because those companies are
not publishing their results?
I don’t know. I don’t. Yeah, I don’t even, I don’t
know if Anthropic does the delay type stuff that
OpenAI supposedly does, but
maybe they do, I don’t know.
And you were just like drawing
infohazard by publishing those laws?
I mean, in my mind, whether or not, I get the
argument, oh it, if you wanna do capabilities
work that stands a test of time or alignment work
that stands a test of time, in my mind, everything
that people are doing in alignment will be very
contrived without the functional form too though.
So it’s like alignment can’t make progress without
it either. So it’s like, you get what I mean?
Another kind of view on that is that
if people do impressive deploying or
ML board and they’re also interested
in alignment, it’s still a good thing.
Like let’s take even through AI. Even if they
open source their model because they did something
impressive and they talk openly about alignment
under Discord and gets like a lot of people
that are very smart, interested in alignment. So
if you publish something and you become like a
famous researcher, something in two years and you
talk about alignment in two years, then it’s fine.
I sort of tweet stuff about alignment,
I think. Yeah, I mean, I retweet
stuff about alignment at least.
Ajeya Cotra’s report, Compute Trends
So if we go back to thinking about predicting
future timelines and kind of scaling, I’ve read
somewhere that you think that in the next few
years, we might get billion or trillion times of
more compute, like 12 orders of magnitude more.
Yeah, I mean, so the Ajeya Cotra report said like,
it’s gonna max out probably at 10 to the 12 times
as much compute as like the amount of compute in
2020, probably like 2070 or something like that.
The one issue I have with the JS model is
that like, she does, what does she do? It’s
like it’s flops per dollar times willingness
to spend its total flops that are allocated to
pre-training runs. Problem is like, for the
big like foundation models, like 10 of the
15 perimeter miles of the future or whatever,
you’re probably gonna need high pie like memory
bandwidth between all like memory bandwidth and
compute bandwidth between all the compute, which
means it has to be on a supercomputer. So it’s not
just the flaps. It basically what really matters,
at least if you’re assuming it’s like big, like 10
of the 15 parameter foundation models or whatever,
like the speed of the fastest supercomputer is
what matters, not just the total flaps that you
can allocate, because if like all the flaps
don’t have good communication between them,
then they aren’t really useful for training like
10 of the 15 parameter model or whatever. Once you
get to 10 of the 15 parameters, like there isn’t
much reason to go beyond that. And at that point,
then you just have multiple models with
10 of the 15 parameters and they’re like
doing some crazy open ended, like Ken Stanley
stuff and a multi-agent simulator after you do
that. Like if they mentioned became like you do
the 10 of the 15 parameter model feature and all
YouTube, and then after that, you’ll have like
hundreds of 10 of the 15 parameter models that
all just duke it out in like a Ken Stanley
open-ended simulator to like, get the rest of
the capabilities or whatever, like once they’re
in the Ken Stanley open-ended stimulator,
then you don’t need high compute bandwidth
between all those individual, like 10 of the
15 parameter models, like duking it out in
the simulator. They can just, each one, they
only needs like 10. It only needs high compute
bandwidth between like its own parameters. Like,
it doesn’t need high compute bandwidth between
itself and the other like agents or whatever.
And so in there, the flops where you could use
all the flops for like the multi-agent simulation,
but you only need high compute
bandwidth within each agent.
So you need a lot of bandwidths to train
models because of the prioritization thing,
but you only need flops to simulate
on different things at the same time?
Yeah, you only need high compute bandwidth
within an individual brain, but like if you
have multiple brains, then you don’t need
high compute bandwidth between the brains.
And what was that kind of simulator
you were talking about, the Kenley?
Like Ken Stanley, the open-ended guy.
I haven’t seen that.
Ken is like the myth day objective
open-endedness, like Can Stanley’s, Jeff Cones,
like all that stuff. It’s like, I don’t know. Just
Google, like Can Stanley open ended at some point.
You’ve probably heard of it, but it’s not
like registering what I’m referencing.
Optimism, conclusion on alignment
Okay, so maybe one kind of last open-ended
question. On a scale from Paul Christiano, Eliezer
Yudkowsky, Sam Altman, how optimistic are you?
Definitely not like Eliezer, or a doomer type
person. I guess probably Paul Christiano is most
similar. I mean, I feel like Paul Christiano
is in the middle of the people you just said.
Right. Yeah. So you are less
optimistic than Sam Altman?
Well, yeah, I mean, basically, I think
deceptive AI is probably gonna be really hard.
So do you have like one less
monologue or sentence to say about
why scaling is a solution
for all alignment problems?
Like just all alignment can be viewed as an
inverse scaling problem. Like it all revolves on
just mitigating inverse scaling, but also you have
to make sure you have like the right downstream
things that you’re evaluating, like the inverse
scaling and like part of what makes it hard is
like you might need to do like fancy thought
experiments on alignment, like counterintuitive
thought experiments on alignment forum to find
what are the downstream… to find what are the
the downstream tests that
you should be evaluating.
Like whether or not there’s
inverse scaling behavior on those.
Awesome, so we get the good
version, as last sentence,
and that’s our conclusion. Thanks
Ethan for being on the show. |
787d95a2-b031-4f9c-8d9d-e2b0e0b9d25f | trentmkelly/LessWrong-43k | LessWrong | Bragging Thread July 2015
Your job, should you choose to accept it, is to comment on this thread explaining the most awesome thing you've done this month. You may be as blatantly proud of yourself as you feel. You may unabashedly consider yourself the coolest freaking person ever because of that awesome thing you're dying to tell everyone about. This is the place to do just that.
Remember, however, that this isn't any kind of progress thread. Nor is it any kind of proposal thread. This thread is solely for people to talk about the awesome things they have done. Not "will do". Not "are working on". Have already done. This is to cultivate an environment of object level productivity rather than meta-productivity methods.
So, what's the coolest thing you've done this month?
(Previous Bragging Thread) |
2e3dcd58-bfce-4a72-a958-19901a694ac7 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Transferring Domain Knowledge with an Adviser in Continuous Tasks
1 Introduction
---------------
Conventional reinforcement learning approaches have been limited to domains with low dimensional discrete state and action spaces or fully observable state and action spaces, where handcrafted features are heavily used. But with the emergence of deep Q-networks (DQN) [[15](#bib.bib10 "Human-level control through deep reinforcement learning")], its applicability was extended to high dimensional state spaces. DQN has been able to surpass human-level performance in some of the challenging Atari 2600 games using only unprocessed pixels as input [[15](#bib.bib10 "Human-level control through deep reinforcement learning")]. This was still not a generalized solution and DQN was not suited well for the higher dimensional or continuous action spaces [[12](#bib.bib2 "Continuous control with deep reinforcement learning"), [22](#bib.bib12 "Deterministic policy gradient algorithms")]. The Deep Deterministic Policy Gradient algorithm (DDPG) which is derived from Deterministic Policy Gradient [[22](#bib.bib12 "Deterministic policy gradient algorithms")] extended the Deep Q learning for continuous state and action space.
Although these advancements of RL have reached continuous state and action spaces they are still unable to directly incorporate already known domain knowledge into the learning process. Thus, they unnecessarily consume time and computational resources to acquire the fundamental known knowledge through learning it from scratch, i.e., agents will follow a trial and error approach many times before successfully converging to an optimal policy. In a simulated world, this is not efficient and it is particularly not welcomed in the real world tasks where the agents cannot make fatal mistakes during learning such as the autonomous navigation domain. To this end, an algorithm which facilitates to incorporate domain knowledge enables an agent to accelerate the learning procedure by limiting the exploration space and converge to better policies.
In this paper, we propose a novel approach to train an agent efficiently in continuous and high-dimensional state-action spaces. Our approach adapts the DDPG algorithm to incorporate already available information to the training process in the form of an adviser to accelerate the learning process. The DDPG algorithm updates the policy in each iteration with approximated policy gradients which are derived from the gradients of the critic network’s output with respect to the parameters of the actor network. However, this approach updates the policy parameters directly and does not facilitate the use of domain knowledge for the policy updating process. In contrast, we update the existing policy to a new policy based on a two-step approach. During the policy parameter update in each iteration, we first set a temporary target to the policy and then push the current policy towards that target by reducing the L2 distance between the current and the target policies.
This two-fold optimization facilitates taking the adviser’s suggestions into account when updating the policy. In addition, the adviser can be used to enforce the agent to explore better regions of the state-action space to extract better policies while reducing the exploration cost. We theoretically prove the convergence of the adapted DDPG algorithm and empirically show that the proposed approach itself improves over the existing DDPG algorithm in chosen benchmark tasks in the continuous domain. We further plug advisers to the adapted DDPG algorithm to show accelerated learning, validating the utility of the two-fold policy updating process.
2 Related Work
---------------
Modern foundations of reinforcement learning are formed as a result of intertwining several paths of trial and error methods and the solution to the problem of optimal control with temporal methods [[23](#bib.bib21 "Introduction to reinforcement learning")]. Initially, the applications of RL were restricted to the low dimensional discrete action-state spaces. Deep Q Networks [[15](#bib.bib10 "Human-level control through deep reinforcement learning")] extended its applicability to the continuous high dimensional state spaces. Later, the DDPG algorithm [[12](#bib.bib2 "Continuous control with deep reinforcement learning"), [22](#bib.bib12 "Deterministic policy gradient algorithms")] combined the DQN and deterministic policy gradient algorithm to handle continuous high dimensional state-action spaces simultaneously. Integrating DDPG with actor-critic [[20](#bib.bib11 "Natural actor-critic")] architecture allows learning parameterized continuous policies, optimized with policy gradient which is derived in terms of parameterized Q-value function.
Reinforcement learning has been an emerging trend in the autonomous navigation domain [[9](#bib.bib31 "Generalization through simulation: integrating simulated and real data into deep reinforcement learning for vision-based autonomous flight"), [8](#bib.bib32 "Uncertainty-aware reinforcement learning for collision avoidance"), [13](#bib.bib33 "Learning to navigate in cities without a map"), [30](#bib.bib34 "Unsupervised predictive memory in a goal-directed agent")]. End-to-end trained asynchronous deep RL [[14](#bib.bib15 "Asynchronous methods for deep reinforcement learning")] based models were used to do continuous control of mobile robots in map less navigation [[24](#bib.bib16 "Virtual-to-real deep reinforcement learning: continuous control of mobile robots for mapless navigation")]. Recently there is an increased concern on reducing training time, increasing sample efficiency, and minimizing the trial and error nature of the learning process to make RL applicable to actual navigational applications more safely and confidently. DQN and DDPG are highly sample-inefficient since they demand a large number of samples during the training. Nagabandi et. al [[16](#bib.bib4 "Neural network dynamics for model-based deep reinforcement learning with model-free fine-tuning")] proposed combining medium-sized neural networks with model predictive control and a model-free learner initialized by deep neural network dynamic models that were tested on MuJoCo locomotion tasks [[27](#bib.bib5 "Mujoco: a physics engine for model-based control")] that achieved high sample efficiencies. Kahn et.al [[7](#bib.bib13 "Self-supervised deep reinforcement learning with generalized computation graphs for robot navigation")] proposed a self supervising generalized computational graph for autonomous navigation which subsumes the advantages of both model-free and model-based methods. They have empirically shown that their model surpasses the performance of Double Q-network which is an enhanced version of DQN that gives less overestimates and variations[[29](#bib.bib22 "Deep reinforcement learning with double q-learning")].
Transferring learned knowledge between problem instances or separate agents reduces the trial and error nature of the training process. Successor-feature-based reinforcement learning [[31](#bib.bib9 "Deep reinforcement learning with successor features for navigation across similar environments")] has been used to do such knowledge transfers across similar navigational environments. Taylor et.al [[26](#bib.bib7 "Representation transfer for reinforcement learning.")] introduced complexification and offline RL algorithms for transferring knowledge between agents with different internal representations. Multitask and transfer learning has been utilized in autonomous agents where they can learn multiple tasks at once and apply the generalized knowledge to new domains [[19](#bib.bib8 "Actor-mimic: deep multitask and transfer reinforcement learning")]. Ross et.al [[21](#bib.bib24 "A reduction of imitation learning and structured prediction to no-regret online learning")] discuss the DAgger algorithm which is similar to no-regret online learning algorithms. It uses a dataset of trajectories collected using an expert, to initialize policies that can mimic the expert better. Methods for automatically mapping between different tasks through analyzing agent experience have shown to increase the training speed of reinforcement learning [[25](#bib.bib6 "Autonomous transfer for reinforcement learning")].
Self-imitation learning [[18](#bib.bib25 "Self-imitation learning")] learns to reproduce the past good decisions of the agent to enhance deep exploration. Hindsight experience replay has been used with DDPG to overcome exploration bottlenecks on simulated robotics tasks in [[17](#bib.bib26 "Overcoming exploration in reinforcement learning with demonstrations")]. Hester et.al [[6](#bib.bib27 "Deep q-learning from demonstrations")] proposed deep Q learning from demonstrations that use small sets of demonstration data to accelerate the learning process. The effect of function approximation errors in actor-critic settings has been addressed by employing a novel variant of Double Q-Learning [[3](#bib.bib28 "Addressing function approximation error in actor-critic methods")]. Maximum entropy reinforcement learning is used in off-policy actor-critic methods [[10](#bib.bib1 "Actor-critic algorithms")] to overcome the issues of sample inefficiency and convergence in conjunction [[5](#bib.bib30 "Soft actor-critic: off-policy maximum entropy deep reinforcement learning with a stochastic actor")]. Continuous variants of Q-learning combined with learned models have shown to be effective in addressing the sample complexity of RL in continuous domains [[4](#bib.bib29 "Continuous deep q-learning with model-based acceleration")].
Our approach uses the actor-critic architecture deviates from the existing methods due to several reasons. First, we adapt the DDPG algorithm to incorporate domain knowledge as an adviser in continuous tasks with high dimensional state-action spaces. Secondly, we employee the adviser in data collection to enforce the agent to explore regions of state-action space with higher return.
3 Method
---------
The proposed adapted DDPG algorithm improves the policy in the direction of the gradient of the Q-value function. It also facilitates integrating pre-learned policies or existing relationships as advisers to transfer domain knowledge. During the training process, advisers can be deployed in; 1) data collection, as well as 2) policy updating processes. Once the adviser is involved in the data collection process, it enforces the agent to explore better regions of state and action spaces according to the adviser’s perspective. When the adviser is incorporated into the policy updating process, it aids to reach better policies rapidly by selecting the best set of actions. In this section, we first introduce the adapted DDPG algorithm and prove its convergence. Then we explicitly describe the proposed ways of employing an adviser for the data collection and policy updating processes to achieve an efficient training approach in continuous tasks.
###
3.1 Adapted Deep Deterministic Policy Gradient Algorithm
Figure 1: Policy updating method with gradient of action-value function.
To extract the optimal policy in the continuous state-action space, moving the existing policy in the direction of the gradient of the Q-value function is computationally efficient than globally maximizing the Q-value function [[22](#bib.bib12 "Deterministic policy gradient algorithms")]. Therefore, the proposed method can be utilized to extract better policies in tasks with high-dimensional continuous state and action spaces, by improving the current policy in the direction of the gradient of the action-value (Q-value) function. The surface in Figure [1](#S3.F1 "Figure 1 ‣ 3.1 Adapted Deep Deterministic Policy Gradient Algorithm ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks") represents the Q-value function corresponding to a hypothetical RL problem in continuous domain. For explanation simplicity, it only contains a single action variable (A) and a single state variable (S). Let the π(S;ϕ) be the policy function which governs the actions in given states. π is parameterized by ϕ which are known as policy function parameters.
The Peak red line of the surface represents the Q-values of the state-action pairs in optimal policy π∗(S) corresponding to an instance. Its projection on the state-action plane denotes the optimal policy. The direction of the gradient of the Q-value function with respect to actions ∇AQ(S,A) corresponding to a point on current policy πt(S) at a particular time step t, always leads towards either local or global optimal-policy. Therefore, the term ∇AQ(S,A) can be used to update the current policy and obtain a better updated policy πt+1(S) by pushing πt in the direction of ∇AQ(S,Aπt) as shown in Figure [1](#S3.F1 "Figure 1 ‣ 3.1 Adapted Deep Deterministic Policy Gradient Algorithm ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks"). Thus, the corresponding policy improvement at a particular step can be represented by;
| | | | |
| --- | --- | --- | --- |
| | πt+1(S)←πt(S)+β∇AQ(S,Aπt). | | (1) |
Here, β is the updating rate of the current policy, which represents the degree of the shift between the updated and current policies.
Once the updated policy πt+1(S) is calculated, it is used as a temporary target to optimize the current policy πt(S;ϕ). We update the policy function parameters ϕ by performing a gradient descent to minimize the loss Lπ;
| | | | |
| --- | --- | --- | --- |
| | Lπ=1n∑[πt+1(S)−πt(S;ϕ)]2, | | (2) |
which is the mean squared error between current and updated policy samples. The main advantage of this two-fold policy update is that we can plug suggestions of an adviser who has domain knowledge, in between the aforementioned two steps to achieve a better-updated policy at a particular time step.
For each update of actor network π(S;ϕ) and critic network Q(S,A;θ) at a given time step t ;
1:Select batch of experiences M=<S,A,R,S′> randomly from memory replay buffer with the size of n
2:Set ^Q(S,A)←R+γQ−(S′,π−(S′;ϕ−))
3:Update θ by minimizing the loss function
LQ=1nΣ(^Q(S,A)−Q(S,A;θ))2
4:Set πt+1(S)←πt(S;ϕ)+β∇AQ(S,Aπt;θ)
5:Update ϕ by minimizing the loss function
Lπ=1nΣ(πt+1(S)−πt(S;ϕ))2
6:Update target network parameters θ− and ϕ−
θ−←τθ+(1−τ)θ−
ϕ−←τϕ+(1−τ)ϕ−
Algorithm 1 Adapted DDPG Algorithm
In our approach, the Q-value function is updated with the temporal difference error similarly to the Deep Deterministic Policy Gradient (DDPG) algorithm presented in [[12](#bib.bib2 "Continuous control with deep reinforcement learning")]. It maintains two parameterized Q-value functions known as Q-network Q(s,a;θ) and target Q-network Q−(s,a;θ−). Similarly, it maintains two policy functions named as policy network π(s,a;ϕ) and target policy network π−(s,a;ϕ−). Algorithm [1](#alg1 "Algorithm 1 ‣ 3.1 Adapted Deep Deterministic Policy Gradient Algorithm ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks") illustrates the steps followed in each update of the Q-network and policy network. A soft update mechanism weighted by τ and (1−τ) (where 0<τ<<1) is used to update the target policy network and target Q-network as shown in the last step of Algorithm[1](#alg1 "Algorithm 1 ‣ 3.1 Adapted Deep Deterministic Policy Gradient Algorithm ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks"). Maintaining two separate networks and soft updating mechanisms enhance the stability of the learning process and it supports training the critic network without divergence [[12](#bib.bib2 "Continuous control with deep reinforcement learning")].
###
3.2 Convergence of Adapted DDPG Algorithm
####
3.2.1 Smooth Concave Functions
If f:Rn→R is a twice differentiable function and holds Lipschitz continuity with constant L>0 then,
| | | | |
| --- | --- | --- | --- |
| | ∥∇f(y)−∇f(x)∥≤L∥x−y∥∀x,y∈Rn | | (3) |
Here, L is a measurement for the smoothness of the function [[11](#bib.bib36 "Gradient descent only converges to minimizers")] and if f(x) is a concave function then the following inequality holds;
| | | | |
| --- | --- | --- | --- |
| | f(y)≥f(x)+⟨∇f(x),(y−x)⟩+L2∥y−x∥2∀x,y∈Rn | | (4) |
Here, ⟨,⟩ denotes the standard inner product between two vectors [[11](#bib.bib36 "Gradient descent only converges to minimizers")].
####
3.2.2 Proof of the Convergence
Let Q:Rm+n→R be the action value function where m and n represents number of state variables and number of actions respectively. S and A be the state and action vectors respectively such that S∈Rm and A∈Rn. Let’s define (S,A)=(s1,s2,…,sm,a1,a2,…,an) where s1,…,sm and a1,…,an∈R. Let Aπ=π(S) where π is the policy function that predicts the action to be executed for a given state S. Once the policy update at a particular time step t is considered;
| | | | |
| --- | --- | --- | --- |
| | Aπt+1=πt+1(S)=πt(S)+β∇AQ(S,Aπt) | | (5) |
If Q(S, A) is a concave and twice differentiable function with Lipschitz continuity then considering the equation [4](#S3.E4 "(4) ‣ 3.2.1 Smooth Concave Functions ‣ 3.2 Convergence of Adapted DDPG Algorithm ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks");
| | | | |
| --- | --- | --- | --- |
| | Q(S,Aπt+1)≥Q(S,Aπt)+⟨∇Q(S,Aπt),((S,Aπt+1)−(S,Aπt))⟩ | | (6) |
| | −L2∥(S,Aπt+1)−(S,Aπt)∥2 | |
| | | | |
| --- | --- | --- | --- |
| | Q(S,Aπt+1)≥Q(S,Aπt)+⟨∇AQ(S,Aπt),(Aπt+1−Aπt)⟩ | | (7) |
| | −L2∥(Aπt+1−Aπt)∥2 | |
Since Aπt+1−Aπt=β∇AQ(S,Aπt) by substituting for Equation [7](#S3.E7 "(7) ‣ 3.2.2 Proof of the Convergence ‣ 3.2 Convergence of Adapted DDPG Algorithm ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks")
| | | | |
| --- | --- | --- | --- |
| | Q(S,Aπt+1)≥Q(S,Aπt)+β∥∇AQ(S,Aπt)∥2−Lβ22∥∇AQ(S,Aπt)∥2 | | (8) |
Considering Equation [8](#S3.E8 "(8) ‣ 3.2.2 Proof of the Convergence ‣ 3.2 Convergence of Adapted DDPG Algorithm ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks") and if 0<β≤2L then,
| | | | |
| --- | --- | --- | --- |
| | Q(S,Aπt+1)−Q(S,Aπt)≥β(1−βL2)∥∇AQ(S,Aπt)∥2≥0 | | (9) |
Therefore, Q((S,Aπt+1)≥Q(S,Aπt) at any time step. Once all the time steps from 0 to k considered in Equation [9](#S3.E9 "(9) ‣ 3.2.2 Proof of the Convergence ‣ 3.2 Convergence of Adapted DDPG Algorithm ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks"),
| | | | |
| --- | --- | --- | --- |
| | t=k∑t=0{Q(S,Aπt+1)−Q(S,Aπt)}≥t=k∑t=0{β(1−βL2)∥∇AQ(S,Aπt)∥2} | | (10) |
| | | | |
| --- | --- | --- | --- |
| | Q(S,Aπk+1)−Q(S,Aπ0)≥β(1−βL2)t=k∑t=0∥∇AQ(S,Aπt)∥2 | | (11) |
If Q(S,Aπ∗) is the Q-value at optimal policy π∗(S) then Q(S,Aπ∗)≥Q((S,Aπk+1). By considering Equation [11](#S3.E11 "(11) ‣ 3.2.2 Proof of the Convergence ‣ 3.2 Convergence of Adapted DDPG Algorithm ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks"),
| | | | |
| --- | --- | --- | --- |
| | Q(S,Aπ∗)−Q(S,Aπ0)≥Q((S,Aπk+1)−Q(S,Aπ0)≥ | | (12) |
| | β(1−βL2)t=k∑t=0∥∇AQ(S,Aπt)∥2 | |
Equation [12](#S3.E12 "(12) ‣ 3.2.2 Proof of the Convergence ‣ 3.2 Convergence of Adapted DDPG Algorithm ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks") implies that as k→∞, the right hand side of the inequality converges.
| | | | |
| --- | --- | --- | --- |
| | ∴limk→∞∥∇AQ(S,Aπk)∥2=0 | | (13) |
Since π∗(S)=argmaxAQ(S,A), at optimal policy ∇AQ(S,Aπ∗)=0. Therefore, the equation [13](#S3.E13 "(13) ‣ 3.2.2 Proof of the Convergence ‣ 3.2 Convergence of Adapted DDPG Algorithm ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks") implies that limk→∞Aπk+1=Aπ∗.
###
3.3 Adapted DDPG with Actor-Critic Agent
Figure 2: Actor-Critic Architecture with Adviser module.
Actor-Critic architecture is a well-known architecture used in model-free reinforcement learning. It consists of two main components, namely the Actor and the Critic. The actor decides which action to be taken by the agent and critic always tries to improve the performance of the actor by analyzing rewards received in each time-step. Generally, the actor is updated with the policy gradient approach, while the critic gets updated with the temporal difference error [[20](#bib.bib11 "Natural actor-critic")]. Our agents are based on Actor-Critic architecture and learn a diverse set of benchmark tasks in the continuous domain. Figure [2](#S3.F2 "Figure 2 ‣ 3.3 Adapted DDPG with Actor-Critic Agent ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks") illustrates the basic block diagram of the Actor-Critic architecture that employs the adviser for both data collection and policy updating processes. In our implementation, the actor and critic modules represent the parameterized policy and Q-value functions respectively.
###
3.4 Employing Adviser to Transfer Domain Knowledge
Although model-free value-based approaches have demonstrated state-of-the-art performance in the RL domain, the low sample efficiency is one of the main concerns which limits their applicability in real-world applications [[7](#bib.bib13 "Self-supervised deep reinforcement learning with generalized computation graphs for robot navigation")]. One solution to this issue is integrating domain knowledge into the learning process. Thus, the agent does not need to learn everything from scratch, and it directly affects the training efficiency and performance of task execution. Here, through an adviser, we propose two such techniques to integrate domain knowledge to support the learning process. Firstly, we enforce the agent to explore better regions in the state and action spaces by enabling adviser suggestions for the data collecting process. Secondly, we enhance the policy updating process by allowing the adviser to adjust the updated policy of each step to a better policy according to the current Q-Value knowledge.
At a given time step t;
1:Observe current state st
2:aadv←f(st)
3:aact←π(st)
4:C←1−e−λN
5:ϵ←e−Q(st,aavd)Te−Q(st,aavd)T+e−CQ(st,aact)T
6:With probability ϵ, at+1←aadv
7:Otherwise, at+1←aact
8:at+1←at+1+noise
Algorithm 2 Data collection with adviser
For each update of actor network π(s;ϕ) and critic network Q(s,a;θ) ;
1:Steps 1-3 of algorithm[1](#alg1 "Algorithm 1 ‣ 3.1 Adapted Deep Deterministic Policy Gradient Algorithm ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks")
2:Aadv←f(S)
3:Aact←π(S;ϕ)
4:^A(S)←Aact+β∇AQ(S,A;θ)
5:for 1:i:n do
6: if Q(si,aiadv)>Q(si,^ai) then
7: ^ai←aiadv
8:Steps 5-6 of algorithm[1](#alg1 "Algorithm 1 ‣ 3.1 Adapted Deep Deterministic Policy Gradient Algorithm ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks")
Algorithm 3 Policy updating with adviser
###
3.5 Adviser for Data Collection Process
Here, we employ an adviser (f) which is a mapping of states (S) to actions (A), to make sampling more efficient by comparing actor’s current predictions of actions with the adviser’s suggestions. In each time step, both the actor (Aact) and the adviser (Aadv) suggests the action to be executed for the current state (St). Then, both suggestions are evaluated with respect to the current knowledge (Q-value function) of the critic module and the adviser’s action is selected for the execution with a probability ϵ. Calculation of ϵ is adapted by the work [[2](#bib.bib35 "Probabilistic policy reuse in a reinforcement learning agent")]. Our method deviates from theirs in several ways. We employ a version of softmax function to induce a higher probability corresponding to the action with higher Q-value. By varying the softmax temperature T, it is possible to change the priority given to the adviser. As shown in the Algorithm [2](#alg2 "Algorithm 2 ‣ 3.4 Employing Adviser to Transfer Domain Knowledge ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks") the constant C (C=1−e−λN) is a confidence value calculated on agent’s behalf where N is the number of episodes elapsed and λ (λ>0) is the decaying constant. This enforces the agent to give higher priority for adviser’s suggestions at the beginning, enabling the agent to explore near a better policy, compared to the architecture without an adviser. In the end, we add a noise signal to the selected action for exploration. The noise generation is influenced by the Ornstein-Uhlenbeck process [[28](#bib.bib19 "On the theory of the brownian motion")], and it is correlated with the input signal so that it ensures a better exploration near the selected action.
###
3.6 Adviser for Policy Updating Process
Since the adapted DDPG algorithm improves the existing policy with an updated set of sampled actions, it enables integrating adviser’s suggestions during the policy updating process as described in Algorithm [3](#alg3 "Algorithm 3 ‣ 3.4 Employing Adviser to Transfer Domain Knowledge ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks"). At each iteration, a batch of experiences with n samples are fetched randomly from memory replay buffer, and Q-value function is updated as similar to Algorithm [1](#alg1 "Algorithm 1 ‣ 3.1 Adapted Deep Deterministic Policy Gradient Algorithm ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks"). Before updating the policy, both adviser suggestions and actor suggestions for the selected batch (Aadv and Aact respectively) are calculated. Thereafter, an updated set of actions (^A(S)) are calculated by considering action-value gradients corresponding to the actor’s suggestions, similar to the Algorithm [1](#alg1 "Algorithm 1 ‣ 3.1 Adapted Deep Deterministic Policy Gradient Algorithm ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks"). In the next step, each updated action (^ai) is replaced by the adviser’s action (aiadv), if the Q-value corresponding to the adviser suggested action is greater than corresponding updated action (^ai). In Algorithm [3](#alg3 "Algorithm 3 ‣ 3.4 Employing Adviser to Transfer Domain Knowledge ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks"), si and ai refers to the state and action of the ith sample of the selected batch. Finally, policy parameters are updated with the modified set of updated actions corresponding to the selected batch.
4 Experiments
--------------
To evaluate the performance of the adapted DDPG algorithm and adviser based agent architecture, we experiment on a diverse set of benchmark tasks in the continuous domain. This includes four OpenAI Gym [[1](#bib.bib23 "Openai gym")] environments namely, Pendulum-v0, MountainCarContinuous-v0, LunarLanderContinuous-v2 and Bi- -pedalWalker-v2. For each task environment, we train three distinct agents using the DDPG, Adapted DDPG, and the Adapted DDPG with adviser algorithms in separation. We develop adviser modules using classical control approaches (Proportional Integral and Derivative controllers) and predefined rules for Pendulum, MountainCarContinuous, and LunarLanderContinuous environments. For the BipedalWalker task, we deploy a pre-trained policy as the adviser. We employed neural networks to parameterize the the Q-value function and the policy function. Due to the lack of information about the smoothness of the Q-Value function, the updating rate β of the adapted DDPG algorithm is set to a lower value (0.01) to satisfy to satisfy the condition β<2L (see the Section [3.2.2](#S3.SS2.SSS2 "3.2.2 Proof of the Convergence ‣ 3.2 Convergence of Adapted DDPG Algorithm ‣ 3 Method ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks")) to ensure the convergence.
| | Pendulum |
| |
| --- |
| MountainCar |
| Continuous |
|
| |
| --- |
| LunarLander |
| Continuous |
|
| |
| --- |
| Bipedal Walker |
|
| adviser | -508.5 | 12.2 | -126.5 | 20.1 |
| DDPG | -398.9 | 28.5 | -65.7 | 100.3 |
|
| |
| --- |
| Adapted |
| DDPG |
| -272.7 | 55.1 | -31.4 | 150.1 |
|
| |
| --- |
| Adapted |
| DDPG + adviser |
| -178.3 | 93.4 | -30.1 | 190.3 |
Table 1: The averaged total episodes score of the trained agents for 30 runs with 500 eposides in each. The adapted DDPG surpasses the conventional DDPG in all tasks. Although the adviser performance is comparatively low, the adapted DDPG algorithm with an adviser shows the best performance.
In each task, we train the agents for 30 runs where each run consists of a pre-defined number of training episodes. After each run, we test the agents for 500 episodes. We define the total episode score as the total reward earned by the agent during all the steps in a given episode. We take the average of such total episode scores gained in all 500 test episodes in a given run and average this figure again over the 30 runs. Table [1](#S4.T1 "Table 1 ‣ 4 Experiments ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks") reports this average total episode score where the adapted DDPG algorithm comfortably surpasses the conventional DDPG algorithm in all tasks. The adviser alone (see first raw of Table [1](#S4.T1 "Table 1 ‣ 4 Experiments ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks")) is not capable of performing the given tasks to the level of DDPG or adapted DDPG trained agents. However, the combination of the adapted DDPG algorithm and the adviser attain the best performance in all the tasks. This shows that even though the adviser under-performs when deployed alone, it certainly assists the adapted DDPG agents to converge towards a policy with higher scores.
| | | | |
| --- | --- | --- | --- |
|
(a) Pendulum
|
(b) MountainCar
|
(c) LunarLanderContinuous
|
(d) BipedalWalker
|
Figure 3: Reward per step of trained agents with the episode number on continuous benchmark tasks. The adapted DDPG algorithm reach higher reward levels rapidly compared to the DDPG algorithm, and adviser always accelerates the training speed further with a low variance in the learning curve.
We further plot the averaged training reward per step, with the episode number in Figure [3](#S4.F3 "Figure 3 ‣ 4 Experiments ‣ Transferring Domain Knowledge with an Adviser in Continuous Tasks"). For a given episode, the reward per step is the total reward earned by the agent during the episode divided by the number of steps. We obtain this value for all episodes across all runs. The average reward per step for a particular indexed episode is calculated by averaging these values belonging to the same indexed episodes across the 30 runs. It demonstrates that the adapted DDPG algorithm achieves higher reward levels rapidly than conventional DDPG. It further illustrates that incorporating an adviser during the training phase expedites the learning process significantly compared to both DDPG and adapted DDPG algorithms. It is also evident that the agent with adviser converges to better policies, achieving higher rewards compared to other methods. Additionally, the adapted DDPG with adviser shows a considerably less variance than all the others.
5 Conclusion
-------------
In this paper, we presented a novel approach of adapting the DDPG algorithm to incorporate an adviser that represents the domain knowledge to expedite the training process. The adviser in our actor-critic architecture causes the data collection and policy updating process to be more effective. We theoretically proved the convergence of the adapted DDPG algorithm and showed experimentally that the proposed adapted DDPG algorithm outperforms the standard DDPG algorithm in conventional RL benchmark tasks in the continuous domain. Additionally, we also demonstrated that the proposed two-fold policy updating mechanism of the adapted DDPG algorithm effectively incorporates core domain knowledge, which is available as common relationships or set of rules to the training process, resulting in accelerated convergence towards better policies. |
7aff3721-7520-4fb8-b93e-cd09017ebb46 | trentmkelly/LessWrong-43k | LessWrong | Brigaded Rounds
A group of round singers were interested in trying out the bucket brigade singing program, and I realized that with a small tweak it could support rounds directly. We just needed to write audio both at the place it belonged to, and an appropriate distance into the future. This would let everyone hear everyone else, though the leader still has to wait a bit before they start to hear others.
The system already needs to know BPM (beats per minute) so it can run a metronome, but the support around it also needs to know BPR (beats per repeat). For example, "Row, Row, Row, Your Boat" is a 16 beat round with participants entering every four beats, so you would probably set BPR to be 16. Four (or more) people could sing, with one person leading and each other person joining after the right number of beats as if they were singing together in person. Once everyone has been singing for 16 beats they all hear everyone else, and also themself.
You can get interesting effects by setting the BPR to some other multiple of the part interval, such as 12. This lets you sing along with yourself an offset, so that you appear to be one of the other singing voices.
It's also possible to configure the system with multiple repeats, so you would hear each iteration of your voice, say, three times. This can be fun if you have a small group of people and want to sing many parts, or even if you're playing around by yourself.
One thing this made me realize is that the latency calibration is really very important, and the previous way we were handling it was not sufficient. Originally, it started off trying to get a very accurate estimation of your latency, but if it wasn't doing very well it would slowly lower its standards. After about 30 seconds, it would let in people even with pretty inaccurate estimates. I've updated it now to be very careful, and if it can't estimate your latency to within 2 ms, it lets you know and gives you the choice between trying again, and joining without sending |
021de1e8-d6b0-4525-ab75-9cb934d7d068 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Decoupling Representation Learning from Reinforcement Learning.
1 Introduction
---------------
Ever since the first fully-learned approach succeeded at playing Atari games from screen images (Mnih et al., [2015](#bib.bib12 "Human-level control through deep reinforcement learning")), standard practice in deep reinforcement learning (RL) has been to learn visual features and a control policy jointly, end-to-end. Several such deep RL algorithms have matured (Hessel et al., [2018](#bib.bib55 "Rainbow: combining improvements in deep reinforcement learning"); Schulman et al., [2017](#bib.bib20 "Proximal policy optimization algorithms"); Mnih et al., [2016](#bib.bib87 "Asynchronous methods for deep reinforcement learning"); Haarnoja et al., [2018](#bib.bib16 "Soft actor-critic: off-policy maximum entropy deep reinforcement learning with a stochastic actor")) and have been successfully applied to domains ranging from real-world (Levine et al., [2016](#bib.bib82 "End-to-end training of deep visuomotor policies"); Kalashnikov et al., [2018](#bib.bib83 "Qt-opt: scalable deep reinforcement learning for vision-based robotic manipulation")) and simulated robotics (Lee et al., [2019](#bib.bib28 "Stochastic latent actor-critic: deep reinforcement learning with a latent variable model"); Laskin et al., [2020a](#bib.bib79 "Reinforcement learning with augmented data"); Hafner et al., [2020](#bib.bib35 "Dream to control: learning behaviors by latent imagination")) to sophisticated video games (Berner et al., [2019](#bib.bib84 "Dota 2 with large scale deep reinforcement learning"); Jaderberg et al., [2019](#bib.bib14 "Human-level performance in 3d multiplayer games with population-based reinforcement learning")), and even high-fidelity driving simulators (Dosovitskiy et al., [2017](#bib.bib92 "CARLA: an open urban driving simulator")). While the simplicity of end-to-end methods is appealing, relying on the reward function to learn visual features can be severely limiting. For example, it leaves features difficult to acquire under sparse rewards, and it can narrow their utility to a single task. Although our intent is broader than to focus on either sparse-reward or multi-task settings, they arise naturally in our studies. We investigate how to learn representations which are agnostic to rewards, without degrading the control policy.
A number of recent works have significantly improved RL performance by introducing auxiliary losses, which are unsupervised tasks that provide feature-learning signal to the convolution neural network (CNN) encoder, additionally to the RL loss (Jaderberg et al., [2017](#bib.bib17 "Reinforcement learning with unsupervised auxiliary tasks"); van den Oord et al., [2018](#bib.bib27 "Representation learning with contrastive predictive coding"); Laskin et al., [2020b](#bib.bib76 "CURL: contrastive unsupervised representations for reinforcement learning"); Guo et al., [2020](#bib.bib94 "Bootstrap latent-predictive representations for multitask reinforcement learning"); Schwarzer et al., [2020](#bib.bib81 "Data-efficient reinforcement learning with momentum predictive representations")). Meanwhile, in the field of computer vision, recent efforts in unsupervised and self-supervised learning (Chen et al., [2020](#bib.bib86 "A simple framework for contrastive learning of visual representations"); Grill et al., [2020](#bib.bib91 "Bootstrap your own latent: a new approach to self-supervised learning"); He et al., [2019](#bib.bib29 "Momentum contrast for unsupervised visual representation learning")) have demonstrated that powerful feature extractors can be learned without labels, as evidenced by their usefulness for downstream tasks such as ImageNet classification. Together, these advances suggest that visual features for RL could possibly be learned entirely without rewards, which would grant greater flexibility to improve overall learning performance. To our knowledge, however, no single unsupervised learning (UL) task has been shown adequate for this purpose in general vision-based environments.
In this paper, we demonstrate the first decoupling of representation learning from reinforcement learning that performs as well as or better than end-to-end RL. We update the encoder weights using only UL and train a control policy independently, on the (compressed) latent images. This capability stands in contrast to previous state-of-the-art methods, which have trained the UL and RL objectives jointly, or Laskin et al. ([2020b](#bib.bib76 "CURL: contrastive unsupervised representations for reinforcement learning")), which observed diminished performance with decoupled encoders.
Our main enabling contribution is a new unsupervised task tailored to reinforcement learning, which we call Augmented Temporal Contrast (ATC). ATC requires a model to associate observations from nearby time steps within the same trajectory (Anand et al., [2019](#bib.bib71 "Unsupervised state representation learning in atari")). Observations are encoded via a convolutional neural network (shared with the RL agent) into a small latent space, where the InfoNCE loss is applied (van den Oord et al., [2018](#bib.bib27 "Representation learning with contrastive predictive coding")). Within each randomly sampled training batch, the positive observation, ot+k, for every anchor, ot, serves as negative for all other anchors. For regularization, observations undergo stochastic data augmentation (Laskin et al., [2020b](#bib.bib76 "CURL: contrastive unsupervised representations for reinforcement learning")) prior to encoding, namely random shift (Kostrikov et al., [2020](#bib.bib89 "Image augmentation is all you need: regularizing deep reinforcement learning from pixels")), and a momentum encoder (He et al., [2020](#bib.bib26 "Momentum contrast for unsupervised visual representation learning"); Laskin et al., [2020b](#bib.bib76 "CURL: contrastive unsupervised representations for reinforcement learning")) is used to process the positives. A learned predictor layer further processes the anchor code (Grill et al., [2020](#bib.bib91 "Bootstrap your own latent: a new approach to self-supervised learning"); Chen et al., [2020](#bib.bib86 "A simple framework for contrastive learning of visual representations")) prior to contrasting. In summary, our algorithm is a novel combination of elements that enables generic learning of the structure of MDPs from visual observations, without requiring rewards or actions as input.
We include extensive experimental studies establishing the effectiveness of our algorithm in a visually diverse range of common RL environments: DeepMind Control Suite (DMControl; Tassa et al. [2018](#bib.bib38 "Deepmind control suite")), DeepMind Lab (DMLab; Beattie et al. [2016](#bib.bib90 "DeepMind lab")), and Atari (Bellemare et al., [2013](#bib.bib56 "The arcade learning environment: an evaluation platform for general agents")). Our experiments span discrete and continuous control, 2D and 3D visuals, and both on-policy and off policy RL algorithms. Complete code for all of our experiments is available at <https://github.com/astooke/rlpyt/rlpyt/ul>. Our empirical contributions are summarized as follows:
Online RL with UL: We find that the convolutional encoder trained solely with the unsupervised ATC objective can fully replace the end-to-end RL encoder without degrading policy performance. ATC achieves nearly equal or greater performance in all DMControl and DMLab environments tested and in 5 of the 8 Atari games tested. In the other 3 Atari games, using ATC as an auxiliary loss or for weight initialization still brings improvements over end-to-end RL.
Encoder Pre-Training Benchmarks: We pre-train the convolutional encoder to convergence on expert demonstrations, and evaluate it by training an RL agent using the encoder with weights frozen. We find that ATC matches or outperforms all prior UL algorithms as tested across all domains, demonstrating that ATC is a state-of-the-art UL algorithm for RL.
Multi-Task Encoders: An encoder is trained on demonstrations from multiple environments, and is evaluated, with weights frozen, in separate downstream RL agents. A single encoder trained on four DMControl environments generalizes successfully, performing equal or better than end-to-end RL in four held-out environments. Similar attempts to generalize across eight diverse Atari games result in mixed performance, confirming some limited feature sharing among games.
Ablations and Encoder Analysis: Components of ATC are ablated, showing their individual effects. Additionally, data augmentation is shown to be necessary in DMControl during RL even when using a frozen encoder. We introduce a new augmentation, *subpixel random shift*, which matches performance while augmenting the latent images, unlocking computation and memory benefits.
2 Related Work
---------------
Several recent works have used unsupervised/self-supervised representation learning methods to improve performance in RL. The UNREAL agent (Jaderberg et al., [2017](#bib.bib17 "Reinforcement learning with unsupervised auxiliary tasks")) introduced unsupervised auxiliary tasks to deep RL, including the Pixel Control task, a Q-learning method requiring predictions of screen changes in discrete control environments, which has become a standard in DMLab (Hessel et al., [2019](#bib.bib78 "Multi-task deep reinforcement learning with popart")). CPC (van den Oord et al., [2018](#bib.bib27 "Representation learning with contrastive predictive coding")) applied contrastive losses over multiple time steps as an auxiliary task for the convolutional and recurrent layers of RL agents, and it has been extended with future action-conditioning (Guo et al., [2018](#bib.bib93 "Neural predictive belief representations")). Recently, PBL (Guo et al., [2020](#bib.bib94 "Bootstrap latent-predictive representations for multitask reinforcement learning")) surpassed these methods with an auxiliary loss of forward and backward predictions in the recurrent latent space using partial agent histories. Where the trend is of increasing sophistication in auxiliary recurrent architectures, our algorithm is a markedly simpler, one-step temporal technique, requiring only observations, and yet it proves sufficient in POMDPs.
ST-DIM (Anand et al., [2019](#bib.bib71 "Unsupervised state representation learning in atari")) and DRIML (Mazoure et al., [2020](#bib.bib95 "Deep reinforcement and infomax learning")) introduced various temporal, contrastive losses, including on “local” features within the encoder, without data augmentation. CURL (Laskin et al., [2020b](#bib.bib76 "CURL: contrastive unsupervised representations for reinforcement learning")) introduced an augmented, contrastive auxiliary task similar to ours, including a momentum encoder but without the temporal aspect. Most recently, MPR (Schwarzer et al., [2020](#bib.bib81 "Data-efficient reinforcement learning with momentum predictive representations")) combined data augmentation with multi-step, convolutional forward modeling and a similarity loss to improve DQN agents in the Atari 100k benchmark.
Hafner et al. ([2019](#bib.bib34 "Learning latent dynamics for planning from pixels"), [2020](#bib.bib35 "Dream to control: learning behaviors by latent imagination")); Lee et al. ([2019](#bib.bib28 "Stochastic latent actor-critic: deep reinforcement learning with a latent variable model")) proposed to leverage world-modeling in a latent-space for continuous control.
None of these methods attempt to decouple encoder training from the RL loss (except for CURL, with reduced performance), and none have previously been shown effective in as diverse a collection of RL environments as ours.
Finn et al. ([2016](#bib.bib97 "Deep spatial autoencoders for visuomotor learning")); Devin et al. ([2018](#bib.bib98 "Deep object-centric representations for generalizable robot learning")); Ha and Schmidhuber ([2018](#bib.bib22 "World models")) are example works which pretrained encoder features in advance; they used image reconstruction losses such as the VAE (Kingma and Welling, [2013](#bib.bib53 "Auto-encoding variational bayes")) or else assumed object-centric representations. MERLIN (Wayne et al., [2018](#bib.bib96 "Unsupervised predictive memory in a goal-directed agent")) trained a convolutional encoder and sophisticated memory module online, detached from the RL agent, which learned read-only accesses to memory. It used reconstruction and one-step latent-prediction losses and achieved high performance in DMLab-like environments with extreme partial observability. Our loss function may benefit those settings, as it outperforms similar reconstruction losses in our experiments. Decoupling unsupervised pretraining from downstream tasks is common in computer vision (Hénaff et al., [2019](#bib.bib25 "Data-efficient image recognition with contrastive predictive coding"); He et al., [2019](#bib.bib29 "Momentum contrast for unsupervised visual representation learning"); Chen et al., [2020](#bib.bib86 "A simple framework for contrastive learning of visual representations")) and has the favorable properties of providing task agnostic features which can be used for training smaller task-specific networks, yielding significant gains in computational efficiency over end-to-end methods.
Figure 1: Augmented Temporal Contrast—augmented observations are processed through a learned encoder fθ, compressor, gϕ and residual predictor hψ, and are associated through a contrastive loss with a positive example from k time steps later, processed through a momentum encoder.
3 Augmented Temporal Contrast
------------------------------
Our unsupervised learning task, Augmented Temporal Contrast (ATC), requires a model to associate an observation, ot, with one from a specified, near-future time step, ot+k. Within each training batch, we apply stochastic data augmentation to the observations (Laskin et al., [2020b](#bib.bib76 "CURL: contrastive unsupervised representations for reinforcement learning")), namely random shift (Kostrikov et al., [2020](#bib.bib89 "Image augmentation is all you need: regularizing deep reinforcement learning from pixels")), which is simple to implement and provides highly effective regularization in most cases. The augmented observations are encoded into a small latent space where a contrastive loss is applied. This task encourages the learned encoder to extract meaningful elements of the structure of the MDP from observations.
Our architecture for ATC consists of four learned components - (i) a convolutional *encoder*, fθ, which processes the anchor observation, ot, into the latent image zt=fθ(\textscaug(ot)), (ii) a linear *global compressor*, gϕ to produce a small latent code vector ct=gϕ(zt), (iii) a residual *predictor* MLP, hψ, which acts as an implicit forward model to advance the code pt=hψ(ct)+ct, and (iv) a *contrastive transformation* matrix, W. To process the positive observation, ot+k into the target code ¯ct+k=g¯ϕ(f¯θ(\textscaug(ot+k)), we use a momentum encoder (He et al., [2019](#bib.bib29 "Momentum contrast for unsupervised visual representation learning")) parameterized as a slowly moving average of the weights from the learned encoder and compressor layer:
| | | | |
| --- | --- | --- | --- |
| | ¯θ←(1−τ)¯θ+τθ ;¯ϕ←(1−τ)¯ϕ+τϕ . | | (1) |
The complete architecture is shown in Figure [1](#S2.F1 "Figure 1 ‣ 2 Related Work ‣ Decoupling Representation Learning from Reinforcement Learning"). The convolutional encoder, fθ, alone is shared with the RL agent.
We employ the InfoNCE loss (Gutmann and Hyvärinen, [2010](#bib.bib88 "Noise-contrastive estimation: a new estimation principle for unnormalized statistical models"); van den Oord et al., [2018](#bib.bib27 "Representation learning with contrastive predictive coding")) using logits computed bilinearly, as l=ptW¯ct+k. In our implementation, the positives from all other elements within the training batch serve as the negative examples for each anchor. Denoting an observation indexed from dataset O as oi, and its positive as oi+, the logits can be written as li,j+=piW¯cj+; our loss function in practice is:
| | | | |
| --- | --- | --- | --- |
| | LATC=−EO⎡⎣logexpli,i+∑oj∈Oexpli,j+⎤⎦. | | (2) |
4 Experiments
--------------
###
4.1 Evaluation Environments and Algorithms
We evaluate ATC on three standard, visually diverse RL benchmarks - the DeepMind control suite (DMControl; Tassa et al. [2018](#bib.bib38 "Deepmind control suite")), Atari games in the Arcade Learning Environment (Bellemare et al., [2013](#bib.bib56 "The arcade learning environment: an evaluation platform for general agents")), and DeepMind Lab (DMLab; Beattie et al. [2016](#bib.bib90 "DeepMind lab")). Atari requires discrete control in arcade-style games. DMControl is comprised of continuous control robotic locomotion and manipulation tasks. In contrast, DMLab requries the RL agent to reason in more visually complex 3D maze environments with partial observability.
We use ATC to enhance both on-policy and off-policy RL algorithms. For DMControl, we use RAD-SAC (Laskin et al., [2020a](#bib.bib79 "Reinforcement learning with augmented data"); Haarnoja et al., [2018](#bib.bib16 "Soft actor-critic: off-policy maximum entropy deep reinforcement learning with a stochastic actor")) with the random shift augmentation of Kostrikov et al. ([2020](#bib.bib89 "Image augmentation is all you need: regularizing deep reinforcement learning from pixels")). A difference from prior work is that we use more downsampling in our convolutional network, by using strides (2,2,2,1) instead of (2,1,1,1) to reduce the convolution output image by 25x.111For our input image size 84×84, the convolution output image is 7×7 rather than 35×35. Performance remains largely unchanged, except for a small decrease in the half-cheetah environment, but the experiments run significantly faster and use less GPU memory. For both Atari and DMLab, we use PPO (Schulman et al., [2017](#bib.bib20 "Proximal policy optimization algorithms")). In Atari, we use feed-forward agents, sticky actions, and no end-of-life boundaries for RL episodes. In DMLab we used recurrent, LSTM agents receiving only a single time-step image input, the four-layer convolution encoder from Jaderberg et al. ([2019](#bib.bib14 "Human-level performance in 3d multiplayer games with population-based reinforcement learning")), and we tuned the entropy bonus for each level. In the online setting, the ATC loss is trained using small replay buffer of recent experiences.
We include all our own baselines for fair comparison and provide complete settings in an appendix. Multiple seeds were run for each experiment, and the lightly shaded area around each curve represents the maximum extent of the best and worst seeds.
###
4.2 Online RL with UL
##### DMControl
We found ATC to be capable of training the encoder online, fully detached from the RL agent, and achieve essentially equal or better scores versus end-to-end RL in all six environments we tested, Figure [2](#S4.F2 "Figure 2 ‣ DMControl ‣ 4.2 Online RL with UL ‣ 4 Experiments ‣ Decoupling Representation Learning from Reinforcement Learning"). In Cartpole-Swingup-Sparse, where rewards are only received once the pole reaches vertical, UL training enabled the agent to master the task significantly faster. The encoder is trained with one update for every RL update to the policy, using the same batch size, except in Cheetah-Run, which required twice the UL updates.
Figure 2: Online encoder training by ATC (UL), fully detached from the RL agent, performs as well as end-to-end RL in DMControl, and better in sparse-reward environments.
##### DMLab
We experimented with two kinds of levels in DMLab: Explore\_Goal\_Locations, which requires repeatedly navigating a maze whose layout is randomized every episode, and Lasertag\_Three\_Opponents, which requires fast reflexes to pursue and tag enemies at a distance. We found ATC capable of training fully detached encoders while achieving equal or better performance than end-to-end RL. Results are shown in Figure [3](#S4.F3 "Figure 3 ‣ DMLab ‣ 4.2 Online RL with UL ‣ 4 Experiments ‣ Decoupling Representation Learning from Reinforcement Learning"). Both environments exhibit sparsity which is greater in the “large” version than the “small” version, which our algorithm addresses, discussed next.
In Explore, the goal object is rarely seen, especially early on, making its appearance difficult to learn. We therefore introduced prioritized sampling for UL, with priorities corresponding to empirical absolute returns: p∝1+Rabs, where Rabs=∑nt=0γt|rt|, to train more frequently on more informative scenes.222In Explore\_Goal\_Locations, the only reward is +10, earned when reaching the goal object. Whereas uniform-UL performs slightly below RL, prioritized-UL outperforms RL and nearly matches using UL (uniform) as an auxiliary task. By considering the encoder as a stand-alone feature extractor separate from the policy, no importance sampling correction is required.
In Lasertag, enemies are often seen, but the reward of tagging one is rarely achieved by the random agent. ATC learns the relevant features anyway, boosting performance while the RL-only agent remains at zero average score. We found that increasing the rate of UL training to do twice as many updates333Since the ATC batch size was 512 but the RL batch size was 1024, performing twice as many UL updates still only consumed the same amount of encoder training data as RL. We did not fine-tune for batch size. further improved the score to match the ATC-auxiliary agent, showing flexibility to address the representation-learning bottleneck when opponents are dispersed.
Figure 3: Online encoder training by ATC, fully detached from the RL agent, performs as well or better than end-to-end RL in DMLab.
Prioritized UL replay (Explore) or increased UL training (Lasertag) addresses sparsities to nearly match performance of RL with ATC as an auxiliary loss (RL+UL).
##### Atari
We tested a diverse subset of eight Atari games, shown in Figure [4](#S4.F4 "Figure 4 ‣ Atari ‣ 4.2 Online RL with UL ‣ 4 Experiments ‣ Decoupling Representation Learning from Reinforcement Learning"). We found detached-encoder training to work as well as end-to-end RL in five games, but performance suffered in Breakout and Space Invaders in particular. Using ATC as an auxiliary task, however, improves performance in these games and others. We found it helpful to anneal the amount of UL training over the course of RL in Atari (details in an appendix). Notably, we found several games, including Space Invaders, to benefit from using ATC only to initialize encoder weights, done using an initial 100k transitions gathered with a uniform random policy. Some of our remaining experiments provide more insights into the challenges of this domain.
Figure 4: Online encoder training using ATC, fully detached from the RL agent, works well in 5 of 8 games tested. 6 of 8 games benefit significantly from using ATC as an auxiliary loss or for weight initialization.
###
4.3 Encoder Pre-Training Benchmarks
To benchmark the effectiveness of different UL algorithms for RL, we propose a new evaluation methodology that is similar to how UL pre-training techniques are measured in computer vision (see e.g. Chen et al. ([2020](#bib.bib86 "A simple framework for contrastive learning of visual representations")); Grill et al. ([2020](#bib.bib91 "Bootstrap your own latent: a new approach to self-supervised learning"))):
(i) collect a data set composed of expert demonstrations from each environment; (ii) pre-train the CNN encoder with that data offline using UL; (iii) evaluate by using RL to learn a control policy while keeping the encoder weights frozen. This procedure isolates the asymptotic performance of each UL algorithm for RL. For convenience, we drew expert demonstrations from partially-trained RL agents. Further details about pre-training by each algorithm are provided in an appendix.
##### DMControl
We compare ATC against two competing algorithms: Augmented Contrast (AC), from CURL (Laskin et al., [2020b](#bib.bib76 "CURL: contrastive unsupervised representations for reinforcement learning")), which uses the same observation for the anchor and the positive, and a VAE (Kingma and Welling, [2013](#bib.bib53 "Auto-encoding variational bayes")), for which we found better performance by introducing a time delay to the target observation (VAE-T). We found ATC to match or outperform the other algorithms, in all four test environments, as shown in Figure [5](#S4.F5 "Figure 5 ‣ DMControl ‣ 4.3 Encoder Pre-Training Benchmarks ‣ 4 Experiments ‣ Decoupling Representation Learning from Reinforcement Learning"). Further, ATC is the only one to match or outperform the reference end-to-end RL across all cases.
Figure 5: RL in DMControl, using encoders pre-trained on expert demonstrations using UL, with weights frozen—across all domains, ATC outperforms prior methods and the end-to-end RL reference.
Figure 6: RL in DMLab, using pre-trained encoders with weights frozen–in Lasertag especially, ATC outperforms leading prior UL algorithms.
##### DMLab
We compare against both Pixel Control (Jaderberg et al., [2017](#bib.bib17 "Reinforcement learning with unsupervised auxiliary tasks"); Hessel et al., [2019](#bib.bib78 "Multi-task deep reinforcement learning with popart")) and CPC (van den Oord et al., [2018](#bib.bib27 "Representation learning with contrastive predictive coding")), which have been shown to bring strong benefits in DMLab. While all algorithms perform similarly well in Explore, ATC performs significantly better in Lasertag, Figure [6](#S4.F6 "Figure 6 ‣ DMControl ‣ 4.3 Encoder Pre-Training Benchmarks ‣ 4 Experiments ‣ Decoupling Representation Learning from Reinforcement Learning"). Our algorithm is simpler than Pixel Control and CPC in the sense that it uses neither actions, deconvolution, nor recurrence.
##### Atari
We compare against Pixel Control, VAE-T, and a basic inverse model which predicts actions between pairs of observations. We also compare against Spatio-Temporal Deep InfoMax (ST-DIM), which uses temporal contrastive losses with “local” features from an intermediate convolution layer to ensure attention to the whole screen; it was shown to produce detailed game-state knowledge when applied to individual frames (Anand et al., [2019](#bib.bib71 "Unsupervised state representation learning in atari")). Of the four games shown in Figure [7](#S4.F7 "Figure 7 ‣ Atari ‣ 4.3 Encoder Pre-Training Benchmarks ‣ 4 Experiments ‣ Decoupling Representation Learning from Reinforcement Learning"), ATC is the only UL algorithm to match the end-to-end RL reference in Gravitar and Breakout, and it performs best in Space Invaders.
Figure 7: RL in Atari, using pre-trained encoders with weights frozen—ATC outperforms several leading, prior UL algorithms and exceeds the end-to-end RL reference in 3 of the 4 games tested.
###
4.4 Multi-Task Encoders
In the offline setting, we conducted initial explorations into the capability of ATC to learn multi-task encoders, simply by pre-training on demonstrations from multiple environments. We evaluate the encoder by using it, with frozen weights, in separate RL agents learning each downstream task.
##### DMControl
Figure [8](#S4.F8 "Figure 8 ‣ DMControl ‣ 4.4 Multi-Task Encoders ‣ 4 Experiments ‣ Decoupling Representation Learning from Reinforcement Learning") shows our results in DMControl, where we pretrained using only the four environments in the top row. Although the encoder was never trained on the Hopper, Pendulum, nor Finger domains, the multi-task encoder supports efficient RL in them. Pendulum-Swingup and Cartpole-Swingup-Sparse stand out as challenging environments which benefited from cross-domain and cross-task pre-training, respectively. The pretraining was remarkably efficient, requiring only 20,000 updates to the encoder.
Figure 8: Separate RL agents using a single encoder with weights frozen after pre-training on expert demonstrations from the four top environments. The encoder generalizes to four new environments, bottom row, where sparse reward tasks especially benefit from the transfer.
##### Atari
Atari proved a more challenging domain for learning multi-task encoders. Learning all eight games together in Figure [11](#A1.F11 "Figure 11 ‣ A.1 Additional Figures ‣ Appendix A Appendix ‣ Decoupling Representation Learning from Reinforcement Learning"), in the appendix, resulted in diminished performance relative to single-game pretraining in three of the eight. The decrease was partially alleviated by widening the encoder with twice as many filters per layer, indicating that representation capacity is a limiting factor. To test generalization, we conducted a seven-game pre-training experiment where we test the encoder on the held-out game. Most games suffered diminished performance (although still perform significantly higher than a frozen random encoder), confirming the limited extent to which visual features transfer across these games.
###
4.5 Ablations and Encoder Analysis
##### Random Shift in ATC
In offline experiments, we discovered random shift augmentations to be helpful in all domains. To our knowledge, this is the first application of random shift to 3D visual environments as in DMLab. In Atari, we found performance in Gravitar to suffer from random shift, but reducing the probability of applying random shift to each observation from 1.0 to 0.1 alleviated the effect while still bringing benefits in other games, so we used this setting in our main experiments. Results are shown in Figure [12](#A1.F12 "Figure 12 ‣ A.1 Additional Figures ‣ Appendix A Appendix ‣ Decoupling Representation Learning from Reinforcement Learning") in an appendix.
##### Random Shift in RL
In DMControl, we found the best results when using random shift during RL, even when training with a frozen encoder. This is evidence that the augmentation regularizes not only the representation but also the policy, which first processes the latent image into a 50-dimensional vector. To unlock computation and memory benefits of replaying only the latent images for the RL agent, we attempted to apply data augmentation to the latent image. But we found the smallest possible random shifts to be too extreme. Instead, we introduce a new augmentation, *subpixel random shift*, which linearly interpolates among neighboring pixels. As shown in Figure [13](#A1.F13 "Figure 13 ‣ A.1 Additional Figures ‣ Appendix A Appendix ‣ Decoupling Representation Learning from Reinforcement Learning") in the appendix, this augmentation restores performance when applied to the latent images, allowing a pre-trained encoder to be entirely bypassed during policy training updates.
##### Temporal Contrast on Sequences
In Breakout alone, we discovered that composing the UL training batch of trajectory segments, rather than individual transitions, gave a significant benefit. Treating all elements of the training batch independently provides “hard” negatives, since the encoder must distinguish between neighboring time steps. This setting had no effect in the other Atari games tested, and we found equal or better performance using individual transitions in DMControl and DMLab. Figure [10](#S4.F10 "Figure 10 ‣ Encoder Analysis ‣ 4.5 Ablations and Encoder Analysis ‣ 4 Experiments ‣ Decoupling Representation Learning from Reinforcement Learning") further shows that using a similarity loss (Grill et al., [2020](#bib.bib91 "Bootstrap your own latent: a new approach to self-supervised learning")) does not capture the benefit.
##### Encoder Analysis
We analyzed the learned encoders in Breakout to further study this ablation effect. Figure [10](#S4.F10 "Figure 10 ‣ Encoder Analysis ‣ 4.5 Ablations and Encoder Analysis ‣ 4 Experiments ‣ Decoupling Representation Learning from Reinforcement Learning") shows the attention of four different encoders on the displayed scene. The poorly performing UL encoder heavily utilizes the paddle to distinguish the observation. The UL encoder trained with random shift and sequence data, however, focuses near the ball, as does the fully-trained RL encoder. (The random encoder mostly highlights the bricks, which are less relevant for control.) In an appendix, we include other example encoder analyses from Atari and DMLab which show ATC-trained encoders attending only to key objects on the game screen, while RL-trained encoders additionally attend to potentially distracting features such as game score.
Figure 9: Breakout benefits from contrasting against negatives from several neighboring time steps.
Figure 10: An example scene from Breakout, where a low-performance UL encoder (without shift) focuses on the paddle. Introducing random shift and sequence data makes the high-performance UL encoder focus near the ball, as does the encoder from a fully-trained, end-to-end RL agent.
5 Conclusion
-------------
Reward-free representation learning from images provides flexibility and insights for improving deep RL agents. We have shown a broad range of cases where our new unsupervised learning algorithm can fully replace RL for training convolutional encoders while maintaining or improving online performance. In a small number of environments–a few of the Atari games–including the RL loss for encoder training still surpasses our UL-only method, leaving opportunities for further improvements in UL for RL.
Our preliminary efforts to use actions as inputs (into the *predictor* MLP) or as prediction outputs (inverse loss) with ATC did not immediately yield improvements. We experimented only with random shift, but other augmentations may be useful, as well. In multi-task encoder training, our technique avoids any need for sophisticated reward-balancing (Hessel et al., [2019](#bib.bib78 "Multi-task deep reinforcement learning with popart")), but more advanced training methods may still help when the required features are in conflict, as in Atari, or if they otherwise impact our loss function unequally. On the theoretical side, it may be helpful to analyze the effects of domain shift on the policy when a detached representation is learned online.
One obvious application of our offline methodology would be in the batch RL setting, where the agent learns from a fixed data set. Our offline experiments showed that a relatively small number of transitions are sufficient to learn rich representations by UL, and the lower limit could be further explored. Overall, we hope that our algorithm and experiments spur further developments leveraging unsupervised learning for reinforcement learning.
#### Acknowledgments
We thank Ankesh Anand and Evan Racah for helpful discussions regarding ST-DIM, MPR, and other related matters in representation learning for RL. |
df39d447-8841-4411-a275-2f7e629eb491 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | "Taking AI Risk Seriously" – Thoughts by Andrew Critch
I wrote this several months ago for LessWrong, but it seemed useful to have crossposted here.
It's a writeup of several informal conversations I had with Andrew Critch (of the [Berkeley Existential Risk Initiative](http://existence.org)) about what considerations are important for taking AI Risk seriously, based on his understanding of the AI landscape. (The landscape has changed slightly in the past year, but I think most concerns are still relevant) |
66d84aef-3f32-4da1-b701-3915100c4814 | trentmkelly/LessWrong-43k | LessWrong | The AI alignment problem in socio-technical systems from a computational perspective: A Top-Down-Top view and outlook
Backward Alignment is an indispensable part of the AI Alignment process, and the alignment issues from the Socio-Technical Systems (STS) perspective are an important component of it. For the current AI alignment field, "Socio-Technical" seems to have become an active keyword, appearing more or less in many recent works on AI safety and alignment. So, what exactly are STS, and what do they signify? In fact, STS is a very broad concept with many considerations, but so far, there has been little work that clearly unifies these issues in one go, often being glossed over in various materials. Additionally, different articles discuss STS using this grand term at different scales, or use different terms to define it at the same scale, which also makes it difficult for researchers to understand the field. This article will, from my personal perspective, clearly explain the AI alignment issues present in STS from a computable perspective at different scales, as well as possible research approaches.
What is a socio-technical system?
STS is a complex organizational work design method that recognizes the interaction between people and technology in real environments, and it requires consideration of the impact of technology deployment on the complex organizational structures of the original environment. It sounds complicated, but we can actually go back to the origin of the issue to understand it better. The term STS can be traced back to a case study by Eric Trist and Ken Bamforth in 1951, where the authors observed that improvements in coal mining technology did not lead to an increase in coal mine output. They further analyzed this case and found that the enhancement of technology caused dynamic changes in the original mining management system (such as increased worker absenteeism), which led to a new perspective based on the open system theory: in the process of developing and deploying technical subsystems, "people" are not dispensable; they form complex social subsystem |
f993188b-9844-4d69-b263-89f9e6f120c8 | trentmkelly/LessWrong-43k | LessWrong | A shortcoming of concrete demonstrations as AGI risk advocacy
Given any particular concrete demonstration of an AI algorithm doing seemingly-bad-thing X, a knowledgeable AGI optimist can look closely at the code, training data, etc., and say:
“Well of course, it’s obvious that the AI algorithm would do X under these circumstances. Duh. Why am I supposed to find that scary?”
And yes, it is true that, if you have enough of a knack for reasoning about algorithms, then you will never ever be surprised by any demonstration of any behavior from any algorithm. Algorithms ultimately just follow their source code.
(Indeed, even if you don’t have much of a knack for algorithms, such that you might not have correctly predicted what the algorithm did in advance, it will nevertheless feel obvious in hindsight!)
From the AGI optimist’s perspective: If I wasn’t scared of AGI extinction before, and nothing surprising has happened, then I won’t feel like I should change my beliefs. So, this is a general problem with concrete demonstrations as AGI risk advocacy:
* “Did something terribly bad actually happen, like people were killed?”
* “Well, no…”
* “Did some algorithm do the exact thing that one would expect it to do, based on squinting at the source code and carefully reasoning about its consequences?”
* “Well, yes…”
* “OK then! So you’re telling me: Nothing bad happened, and nothing surprising happened. So why should I change my attitude?”
I already see people deploying this kind of argument, and I expect it to continue into future demos, independent of whether or not the demo is actually being used to make a valid point.
I think a good response from the AGI pessimist would be something like:
> I claim that there’s a valid, robust argument that AGI extinction is a big risk. And I claim that different people disagree with this argument for different reasons:
>
> * Some are over-optimistic based on mistaken assumptions about the behavior of algorithms;
> * Some are over-optimistic based on mistaken assumptions about the behavio |
1f97bc87-3207-4c8c-80b4-fe4f808545c3 | trentmkelly/LessWrong-43k | LessWrong | Meetup : Madison Monday Meetup
WHEN: 21 November 2011 06:30:00PM (-0600)
WHERE: 1831 Monroe St. Madison, WI 53711
This will be largely a social meetup - we'll get together at the Barriques on Monroe and chat. We'll probably play a game or two, and have an argument or two.
We'll also chat about the upcoming Sequences reading group. If you have a strong opinion about what to read through, you should show up. ;)
And, again, if you're in the area, you should sign up for the Madison mailing list. |
3b8cdd6b-1d0c-4f35-afd8-7fc03c47c48e | StampyAI/alignment-research-dataset/youtube | Youtube Transcripts | Deep Learning 1: Introduction to Machine Learning Based AI
okay good let's try this again my name
is torie Grable and Sharaf machine
learning at UCL and I work as a research
scientist at deep mind it's a startup
company that tries to solve problems in
artificial intelligence this course is a
collaboration between UCL and deep mind
to give you the chance to learn about
those machine learning techniques that
are likely to help get us to artificial
intelligence here's the overview of what
I'm going to cover we'll talk a little
bit about the structure of the course
and and the team the people who are
offering the course we have guest
lectures given by some excellent guest
lecturers which we you will get to know
over the course of this module then
talked about a deep mind approach to AI
I think I would like you to know where
we're coming from when we talk about
these things what we want to achieve and
why we think that the particular things
covered in this module might help us get
to artificial intelligence or as we try
to ambitiously call it general
artificial intelligence I'll then talk a
little bit about deep learning and I'll
then give two very short nuggets of
project work one about learning to play
Atari games using deep reinforcement
learning and one about alphago which is
one of my favorite projects finally have
some extra revision material we
restructured the course from last year
last year this material was covered in
two lectures and I've appended most of
the material here we might not get into
that but if you're interested and want
to prepare for the following lectures
you might want to take a look at this
material in Moodle there may be time to
highlight a little bit of it okay so
let's dive right in here's the team the
that put this course together and will
also deliver it Cora is the head of our
deep learning group and he has sponsored
all the deep learning talks and has
helped us put this together my colleague
Hado will give the reinforcement
learning track of this module which
mostly takes place on Thursday at the
inhumane time of 9:00 a.m. you may have
noticed this in the schedule you know
that's real self selection anyone going
to at that time I mean I have the
highest appreciation I certainly won't
be there well maybe sometimes okay and
then a number of other people played a
key role there's material Hessel and
Alex Davies
they're our tensorflow experts and
they'll give the tensorflow tutorial on
thursday to give you a good basis for
the for the course work which will
decode it intensive low and who've also
helped together the course work
assignments Deana borsa can you wave who
is coordinating the TA support she's
also a research scientist a deep mind
and Marie who helps us with the general
organization Marie
where are you very good and also
coordinates the recordings and then we
have some amazing teaching assistants
any of these here yeah can you wave
visibly okay
who will help with the course work
assignments okay so that's the team
let's talk a little bit about the format
and the assessment so we have these two
streams and we have a deep learning
stream and a reinforcement learning
stream and towards the end they will
somatically converge to some degree and
these are the basic building
blocks in what we think is needed to
build AI systems based on machine
learning and on Tuesdays mostly we'll
have guest lectures on deep learning
topics and I'll talk about what that
will be that later and on Thursdays Hado
will give a structured introduction to
reinforcement learning and also there
will be two guest lecturers at the end
of this so that's roughly the structure
but there are some exceptions so please
check the time table we have a schedule
posted on Moodle so one question that
sometimes comes up is how does the
assessment of the course work last year
we had 5050 coursework assignments and
written exam but we found that it was
tricky to formulate really nice
questions based on the cutting-edge
material that was being presented and so
we thought it would be a better
experience for you all if you can just
focus on the programming assignments
really work on those deep learning and
RL practical problems and learn rather
than having this this exam hanging over
your heads so the idea is basically that
we'll have four deep learning and four
reinforcement learning assignments these
will be spread across the weeks here of
the module and then these eight
assignments will be weighted equally and
the final grade will be based on those
assignments these questions will be
mixtures of programming assignments and
questions of understanding that that you
will answer in this context to make
things really easy we decided to put the
entire course work this time into collab
so collab is is a jupiter notebook
environment where you don't need to do
any setup this is just connected to the
cloud we can pre configure it and we are
also providing the computational
resources that you'll need to solve the
tasks this is again going back to
student feedback from last year or
it was difficult for people to procure
enough computational resources to do
everything they wanted to do and so this
time we're trying to make the design the
assignments more carefully and at the
same time provide the computational
resources that you need for you in the
cloud will use tensorflow as mentioned
before and you can find more information
about the whole assessment on Moodle one
thing that we needed to do in order to
set up this collab resources is that we
white listed a bunch of email addresses
with the collab service and in order to
use that we would like to ask you to
create a gmail account or Google Account
here with this form at UCL CompTIA 22
2018 and the XS represent your student
ID and we've pre white listed those so
that if you register this account you
will get access to the collab service
also we thought it would be nice if you
have a new a fresh account for this that
you can use ok then finally regarding
support we would like to encourage you
to use Moodle for all your support
queries if you want to discuss stuff
there's a Moodle forum lecturers and
teaching assistants will will look at
these questions and answer them
ideally other students can answer them
we want to share the answers to these
things we would like to avoid any
one-to-one communication where questions
would be answered multiple times the
answers would be shared and so on if you
have some kind of personal trouble then
feel free to email me with with that
problem but but please not with any
material that could be of interest to
others as well ok any questions about
the format and the course work get used
to asking questions please makes this
thing so much more entertaining for
everyone no do you think stupid
questions are funny
smart questions are useful yeah yeah
it's just the these digits it's usually
eight digits and for people who have
been added a little longer I think it's
seven digits yeah that one okay so just
to show you what this collab looks like
so this is the interface to collab you
have these cells that you can work with
and you can code directly in Python in
these cells and there's code cells and
texts also the the assignments will come
in the form of such a collab notebook
and it there will also be some code
available there already that you can
plug your own code in to make you to
have some kind of unified visualization
and and plumbing code available and we
hope that that will make the whole
programming experience nicer for you and
then we'll ask you to submit this
notebook for each one of those
assignments together with the PDF print
out it's all detailed on Moodle what
exactly is required and submission
ideally on time makes everyone's lives
easier and the grades better good and
tensorflow
we probably all know tensorflow it's
this google open source software to
define and run machine learning
algorithms mostly used for neural
networks that's pretty standard thing
and I think we will all benefit from
getting exposure to this unless you've
already had it and there are some great
jobs out there for people who know how
to do tencel oh good a little warning
this course we had some feedback on it
it's pretty hard
people were really unhappy last year not
everyone but some of them were so if you
do not know how to code in Python maybe
this isn't the right course for you if
you do not have some of the
preliminaries required in machine
learning in maths statistics and so on
maybe this isn't the right course for
you to think carefully about that and we
created a little self-assessment quiz on
Moodle that you can take a look at after
the lecture if you can do the majority
of that without problems then you're in
the right course if you struggle with
those questions then you might struggle
with a lot of the lectures and maybe
there's there's a better course
somewhere for you to catch up on those
things unless you think you have a lot
of time on your hands who thinks they
have a lot of time on their hands right
to catch up on some of these issues
good let's look a little bit at the
schedule yeah
what's the question
but please do stay if you feel up okay
so we here's a little schedule for the
course you'll also find this on Moodle
you can see here that we have basically
these two tracks the two state track on
deep learning and the first day track on
reinforcement learning and we have some
some very exciting lectures lined up
here and the deep learning track they
are a little bit heterogeneous from from
the topic perspective because each of
these lecturers is an expert in the
field they present cutting-edge stuff
leading up to the things that they do in
their research so there's really a
goldmine off of knowledge to be gained
here from these lectures but of course
you do need to understand the basics
very well and when they go through some
of the basics in the particular topic
that will be relatively quickly so they
can go up to the really interesting
stuff on the other hand the
reinforcement learning path is a little
bit more structured because this is
given by one lecturer Hado who also is
an expert in this field and and this has
more these building blocks that are
actually building on top of each other
and we think that is also more
appropriate because very few people
would have had real exposure to
reinforcement learning here and so this
is from the ground up so what you see
here on the right are the weeks in which
we're aiming to distribute the
coursework assignment deep learning
coursework PLC w124 and real
reinforcement learning coursework 1 to 4
so we're trying to distribute this
across the course time as good as we can
of course we need some startup time to
get going and to to get you some of the
information that you need but then we've
spaced this fairly evenly to the end
the course that again is is I think an
improvement over last year where there
were bigger chunks we hope that these
smaller chunks will encourage you to do
that work right away
so that you can get the feedback on this
work you can see how what you're doing
connects to what you learned in the
lectures in a more immediate way than if
you wait to the end and then then do
that work okay do we need a more more
seat here here's one more seat do you
want to take that okay so let's go
through the program in a little more
detail because we really have some
fantastic speakers and that also gives
you a little exposure to the topics so
on Thursday we'll start with this
introduction to tensorflow
which hopefully gives you a foundation
for the coursework that you'll be
expected to do it will be delivered but
by mateo hassle and Alex Davies and
basically they'll give you an
introduction to tens of love principles
and then they'll also do some work
through examples in collab and then
later make that collab available for you
to play with then following through with
the deep learning track of things the
next lecture will be given by Simon
Austin Darrow and he will cover newer
networks multi-class classification how
backprop works how automatic
differentiation works how tensorflow
does it for you and so will deliver the
basics on on neural networks in this
lecture and he's a really a real expert
on this so I'll probably just attend
that again because the basics are
explained so well okay so then the next
lecture will be on convolutional neural
networks again we have an expert on the
topic heron built one of the important
neural network architectures that led to
a great improvement on image net on the
object classification benchmark and he's
going to talk about convolutional
networks applied to large scale image
recognition and show you how these
models can be applied to image net as an
example and how the architectures have
evolved and how to train those models
the next topic is recurrent Nets and
sequence generation so whereas
convolutional neural networks assume in
some sense that there's a 2d topology
that of an image or some other grid may
be a go board often we see data in
sequence form for example text but other
time series as well and there are
specific neural network architectures
recurrent neural networks that can deal
with this the most successful one that
he'll talk about is the LST M Network
and he'll also talk about conditional
sequence generation by the way he's the
the person who is driving the Starcraft
research at deep mind you may have heard
that we made available an environment
for doing research on Starcraft
are there any Starcraft players here
anyone yeah so I think that's that's a
pretty exciting domain so maybe he left
some time to talk about that as well I
don't know I'd have to see
good then riah we'll talk about end to
end an energy based learning she'll just
discuss some other structures other than
just simple regression losses for
example how to define losses for
embeddings how to define losses when you
have relations between three data points
that you want to embed if if you want to
do ranking and and similar questions
then the topic of optimization presented
by James Martin's optimization of course
is one of the major tools in machine
learning it was actually a major step
for me
learning nowadays maybe few people know
this but machine learning used to be
based on rules you know people would
just define learning rules and hope that
the system would then somehow converge
to an interesting solution for example
kohonen maps are an example of where
someone just formulated a biologically
plausible rule in order to find low
dimensional embeddings of data and that
was always problematic because there
were no convergence guarantees and so on
so now of course optimization is at the
center of machine learning and this
lecture will be dedicated to
understanding better what kinds of
algorithms are out there and what
properties they have including force
first order methods second order methods
and very importantly also stochastic
methods for dealing with very large data
sets where you might want to subsample
examples or take them in very small mini
batches good will then move on to more
advanced topics Alex graves will talk
about attention and memory models so the
starting point where newer networks were
basically the simple feed-forward newer
networks but over time people have
developed more and more sophisticated
modules that can do more and more
interesting things and attention and
memory are two of these elements
attention is basically the ability to
focus on some subset of your inputs in
order to concentrate your processing
power on those ideally words
particularly interesting and memory
models in some sense can be seen as
turning that to the inside when you have
some internal memory and you then turn
attention where while reading it to
particular parts of that internal state
of yourself as an agent if you like and
alex is an is an expert on this he
developed the these ideas of the neural
Turing machines the differentiable
neural computers and in this lecture he
will lead you up to the point where you
can understand how these things
we'll then have a lecture on deep
learning for natural language processing
given by IDI at grayish data and this is
mostly an application lecture and
there's actually an entire series of
this lecture because quite a rich topic
but these neural language models neural
word embeddings these things are now the
state-of-the-art methods in neural
language processing in natural language
processing and he will explain how these
things work and how neural networks can
successfully be applied to text finally
we'll have a lecture on unsupervised
learning and deep generative models of
course this again is a very complex big
topic and we could have an entire
lecture see a series on that topic but
we think it's important to to give you
some exposure to the latest stuff there
many people think that unsupervised
learning is going to be very important
going forward among other things because
we just don't have labeled data for all
the domains that we would like to look
into but we have vast amounts of data
that I unlabeled if you just think of
all the data on YouTube if we could
learn from that that would be great
right but we have very poor labeling on
that maybe titles of videos but that's
not really what this is about right you
want to go into the video stream and
learn from that and so that's a huge
topic very important for AI Chuck here
is an expert and he will explain ideas
around this and in particular also about
deep generative models models that can
actually produce data for example image
data or text or similar things okay
that's yeah
sorry can you say again see well I mean
yeah so if these are latent variables
and that's the the observation then this
equation I think he put it there because
he thinks that represents basically the
the deep problem of unsupervised
learning - given these observations why
to discover these these latent causes
zette that produce them and so I think
he would argue that in some sense all of
unsupervised learning is captured by
that equation okay then this is about
the other stream the reinforcement
learning stream led by hado
and he'll talk more about that when he
starts lecturing next week about roughly
about the things he's going to talk
about an introduction to reinforcement
learning of course discussing what
Markov decision processes are which are
the underlying framework for doing
reinforcement learning he'll then go
into planning and how to do planning
using dynamic programming where there's
no learning aspect yet just planning in
a given model it will discuss model free
prediction and model free control he'll
then discuss value function based
methods in contrast to policy direct
policy optimization methods and how to
integrate learning and planning he'll
discuss exploration versus exploitation
which is this problem that comes up for
the for for these agencies reinforcement
learning agents that act in the world
and have to somehow strike a balance
between gathering new information that
can help them improve how they can act
in that world versus exploiting what
they already know in order to gather in
order to gather immediate reward and
then towards the end for the last two
lectures
were planning to guest lectures the
first one on alphago given by David
server have you heard of alphago okay
that's good
so I think that will be of interest and
David is fantastic speaker I hope we can
win him over for this and the second
case study is about practical deep
reinforcement learning deep
reinforcement learning is really this
point of convergence where deep learning
and reinforcement learning come together
because we're using neural networks to
represent either policies or value
functions for reinforcement learning in
a very flexible way by function
approximation and Vlad nee is the lead
author of the paper where this was
applied to to these 50 Atari games a
couple of years ago three years ago now
I suppose and he will tell you the
latest about deep reinforcement learning
how it worked with the atari games what
other domains are out there what other
algorithms are out there and so on okay
Wow
it's a rocket it's good so any questions
about the plan so far plan is clear good
so why is there rocket let me see if I
can remember
so we're now moving to the part about
what deepmind does and how we approach
our mission and the idea really is to of
creating deep mind was to create
something like an Apollo program for
artificial intelligence now a program
were really large numbers of researchers
would be focused and would be able to
make rapid progress on this problem of
of AI and the mission is really crisp I
think two words solve intelligence and
partly what we're also experimenting
with is in addition to the machine
learning stuff is how to organize
science itself because it is actually a
rather tricky problem to have a large
number of scientists and have them work
together towards a common goal and we
think we have found good ways to do this
which are somehow at the interface of
how it's done in academia and how it is
done in start-up companies and in some
sense the idea deep mind is to combine
the best of two worlds in order to
really create an environment in which
rapid progress towards AI can be made so
the basic premise of our work is that AI
can really only be achieved if we can
understand how learning works AI we
think will be solved by learning
algorithms and this of course comes from
experience from the past because in the
past in what we call good old-fashioned
AI other approaches were tried more
rule-based approaches were tried where
people would would build systems by
combining different rules and bits of
knowledge from humans added to knowledge
bases or added to the program and this
turned out to be very difficult to
scale because these bits of knowledge
tend to then interfere with one another
it's also remarkably difficult for us to
actually formulate what we know about
the problem alphago is one example of
this even if you are a go player are
there any go players here it's very
tricky to explain why a move is good
isn't that I mean you make it you have
this gut feeling you can come up with
some kind of explanation but it's really
hard and so formalizing by hand
knowledge about tasks that we're good at
is really tricky on the other hand we
have learning algorithms and if we can
feed them with examples of what we want
to learn that is a much more powerful
approach but of course the world is an
interactive place so we will not just
have input-output examples for all kinds
of problems that need to be solved and
so we need to create algorithms that can
go out there into a world be that a
simulated world or the real world that
can interact with that world and find
out for themselves what kind of behavior
is optimal towards their goal and so
that's the general idea
but there's another thing that we would
like to do we want generality so in
order for a system to be truly
intelligent we think it should be
general it should be applicable to
different domains to a wide range of
different domains and only then would we
really be calling it intelligent if it
can only do one thing then it might not
be really that intelligent and of course
currently most of the work that we see
most of the successes that we see would
qualify more likely as narrow AI AI that
is aimed at solving a particular task
rather than a wide class of tasks and so
we want artificial general intelligence
intelligence that can address address
many different tasks the important
conceptual tool that we use is
reinforcement learning and we really
consider this to be a general purpose
framework for AI because
it encompasses this idea of an agent
interacting with the world and learning
from their interaction so the basic
setting is that we have this agent and
we have an environment and that they
interact in these two ways the agent can
observe what's happening in the
environment or what the environment is
the state of the environment and it can
then issue actions into that environment
in order to change that environment
and the goal of an agent is to maximize
long-term reward
long term is important here because most
of the interesting things that we want
to do require maximizing long-term
reward if you want to get from A to B in
some city you don't get immediate reward
you get the reward when you were there
when you arrived so to speak so you need
to plan long term in order to get from A
to B and there are many other examples
we can also think of this setup as
encompassing supervised and unsupervised
learning as special cases because nobody
forces the agent to actually submit any
actions and so if you if you think of
this actions block not being there then
we have the environment and the agent is
just learning about that environment and
getting some kind of some kind of reward
function according to its goal oh
there's actually this is a little so
yeah you can also think of some kind of
reward you know part of the observations
here to be some kind of reward signal
that the environment gives to the agent
yeah yeah that's a that's a good
question but you could think of some
kind of likelihood criterion maybe that
no day that's right there it's it's not
clear but I I would argue that if you
come up with one then I would be able to
put it in this schema but it's it's
clear that it's you know for example if
you wanted to do test likelihood then
there you would have the the agent would
request examples and we'll see examples
and could build a model internally and
then the it would request test examples
and then the environment would issue
test examples and it would get rewards
based on how well it predicts that so I
think in that sense it's it's a very
general framework of course it also
weakness of the framework that is so
general right because it it gives less
guidance as to how to solve these things
good
so reinforcement learning and you'll
learn a lot about this from Hado since
the goal of our research a deep mind and
maybe the goal of some of you will be to
create intelligence it's an interesting
question to ask what is intelligence and
our chief scientist Shan lag together
with his then PhD supervisor mark Sutter
came up with this definition here based
on mining hundreds of definitions they
had found in the literature it's
actually quite an amusing paper to read
so they have a paper where they look at
all kinds of definitions from all kinds
of sources of what intelligence is and
then try to distill out the main
properties that they think should be
present there and here's what they came
up with intelligent measures an agent's
ability to achieve goals in a wide range
of environments
does anyone strongly disagree with this
statement
yeah that's interesting isn't it yeah
yeah so this this in some sense assumes
that the goals are given externally
where as truly intelligent beings maybe
generate their own goals to some degree
but then maybe that could be specified
as a meta goal from the outside yeah
it's it's a bit tricky well what I tried
to do here is to be very operational in
some sense and I would like to explain
to you how this motivates to large parts
the research agenda of deep mind into
intelligence and that's based on this
equation which I think just looks
beautifully if you squint your eyes I
mean it's just a beautiful beast a bit
of texture here so let's take a look at
this so they define this measure of
intelligence as a function of the policy
PI you can think of this policy or
alternatively as an agent that acts in a
particular way in an environment in any
environment reading and so the way they
define it is through this value function
here which is at the core of it and this
expresses basically how well does this
policy or this agent do in this
environment mu so that seems like a good
idea if there is a given environment
then an agent that does well in the
environment would be if there was only
that one environment more intelligent
that with respect to that environment
then an agent that there's less well
right so that that seems plausible of
course it still depends on how we
measure this but now this is summed over
a set of environments E and waited so
what this expresses basically is the
diversity we said that intelligence
would be required to be successful
across a wide range of environments and
so that is the set of environments here
and this factor now here
says hey but they're not all created
equally some of them are more important
than others and the way this is
formulated here is that this is - - the
- the Kolmogorov complexity of the
environment and the idea here is that if
the environment is simple then this term
is large and there's a high weight on it
and if the Kolmogorov complexity is
large then this term is low now one
advantage of this formulation is that
this sum converges which is always a
nice thing for a theoretician right you
know if your system you're talking about
isn't infinity it's a nice thing but it
captures the intuition that there are
fewer simple problems and if you call
yourself into an intelligent agent you
want to at least score well on those and
then they're more of the complex kind
because they require more description
lengths to specify them and then if you
want to be more intelligent you need to
do better on many of those and I mean
we're not really like a human there's
only there's only one big world so how
do you like this yeah yeah it's that's a
good question so as they formulated this
is the set of all computable
environments such that this quantity can
be evaluated but in practice and that's
the important thing for the research
agenda is we define sets of environments
and then we try to do well on them so we
take this as inspiration we don't
actually want to solve this particular
thing it's just you know too hard but
the way we view it is why don't you we
pick sets of ever more complex tasks and
try to train our agents first on the
simpler one and then make that more
complex and more complex and get them
ever more intelligent according to this
definition and so one example that you
could take is maybe there's a simple
block based game that you start with and
then you move on to the Atari games
which are visually more complex and have
more interesting dynamics and then you
go to 3d games that I yet require more
from your agents but no one tells you
the boundary yeah yeah if you want to
really look at it for the human it's
very difficult and I think it's an
interesting open question how to do this
so we in some sense we interact only
through our body with the environment
and where do our rewards even come from
in some sense they're self generated
through our brain and then how is that
controlled well evolution developed us
in some sense evolved us and there were
certain forces that shaped the anatomy
and the chemistry of our brains and so
somehow there never is in a reward that
can be totally isolated if you look at
this complex story it's just all
evolving and the environment is changing
and you are a part of my environment and
I'm part of your environment and you
know that year so this is definitely a
simplifying model where were where we're
trying to say suppose there were these
there was this nested for almost set of
environments we could where they were
all well defined can we come up with a
definition of intelligence within that
limited few yeah so it's also maybe
slightly counterintuitive the more
complex the environment is the less
waitis it has it's interesting yeah this
is these other people have had these
concerns as well and that's definitely
interest
but the idea here is that you should at
least nail the simple ones and then you
can move on to the more complex one but
if you got any single more complex one
because there are so many more of them
you in order to really make progress in
terms of intelligence you need to be
able to self solve many more of those
yeah that's the problem there are
exponentially many more of the higher
complexity steps so if we weighted them
all even more this sum wouldn't even
converge but yeah it's it's a bit
counterintuitive maybe it could be
captured somewhere in the value that you
achieve that the more complex ones maybe
give you more value for now let's take
it as an inspiration to think about an
approach where we have environments on
the one hand side that we can construct
and we have learning agents that we can
put into these environments and train
and we basically want to cover our bases
and be able to have these agents be
successful in the simple ones and then
we want to ramp up the complexity of
these environments and be able to solve
ever more complex ones you can also
think of it as a curriculum if you do it
in that order because there's no order
implied here but I would recommend
starting with the simple ones and then
moving on to the more complex ones as
humans often do during their lifetime as
well yeah yes also interesting you might
have it in the V right for example if a
complex policy was encoded through a
complex neural network which would
require a lot of energy to evaluate then
if your task was to come out of the
environment with your maximum amount of
energy or survive as long as you can in
the environment then you could code that
all in the actual environment and reward
combination and a simple as small and
your network would then be advantageous
because it's actually better at the task
because the task involves resources and
it's better at preserving resources okay
but you see you can get into a lot of
discussions even at the point where
you're trying to say what you're what
you want to solve so tricky questions
okay so if we think about our own
intelligence then certainly it didn't
evolve in a purely abstract world it is
grounded in rich sensorimotor reality we
see things we can manipulate things and
we know that for example babies learn in
this way and to get some some of this
people have looked at games video games
as a particular set of tasks that are
really nice for this purpose and a lot
of the work that we do as you know Atari
games alphago is based on games and I
personally think they're they're a
fantastic platform for AI algorithms so
what are the advantages these are
basically simulators right the game so
you have unlimited training data anyone
doing machine learning surely
appreciates that of course it comes at a
cost
when you run the environment it also
costs your resources so there's a
trade-off in some sense you need less
data but you need computational
resources but me we can run these in
simulation and that has a lot of
advantages because we can have many
instances for example in training we can
have thousands of computers run the same
thing and train and that's very powerful
if you compare that to a robot arm that
is much trickier because then you have
to deal with the mechanics of the arm
you need physical space to put the arms
you need to maintain them they might
change over time and so on so working in
simulation is really attractive and so a
lot of the work that we do is in
simulation now of course the death
advantage is that it's not the real
world so we might be missing out on some
aspects of the real world like
measurement noise changing changes in
the dynamics stuff breaking you know the
real world but the idea here is we can
make progress maybe first in these more
abstract games domains and then we can
move to the real world later so the goal
in these domains then is to do
end-to-end learning agents and basically
we want to formulate an algorithm that
observes the world in some sense issues
actions in interaction with the
environment and learns how to solve the
task without further input from humans
and that's how the the Atari work came
about and I'll talk a little bit about
it later
but that's basically the the research
philosophy that we're trying to follow a
deep mind and which we think is a
promising way towards developing
artificial intelligence good just see
the timing shall we make a 10 minute
break maybe and then meet back here five
past three all right thanks we just did
some online learning here and found out
that our instructions don't quite work
the way we intended them to hold off
with the creation of those Gmail
accounts we'll need to figure out what
exact characters Gmail addresses can
contain apparently dashes are not among
them so we'll we'll update that on the
assessment part of the Moodle page so
you can take a look also just to clarify
these coursework bits are individual
pieces they are to not to be solved in
teams if you know what I mean
of course it's always nice to advise
other students and help them in some way
but do it in very general terms okay
let's talk a little bit about if
learning deep learning is a hot topic we
when we think about it in the in the
context of artificial intelligence we
think of it as helping us solve this
perception problem that we want to want
our agents to be able to comprehend the
world to perceive the world and then
comprehend it but to perceive it at the
level of the sensors by which they
perceive it so for for video or images
on the pixel level the raw data and
while this sounds may be very natural to
you now again this wasn't exactly the
agenda of AI in the past which often
worked at a symbolic level and where
problems arose because it was very hard
to connect that symbolic level on which
certain manipulations were relatively
easy with the actual underlying sensory
motor reality because you might have
some idea of maybe how to build a tower
or something out of pieces but if you
just direct the camera at the seed you
first need to figure out what are the
pieces and and how can I move them so a
purely symbolic solution is kind of
hovering somewhere abstractly and it's
disconnected both from what you can see
in the world because it's hard to get
that symbolic representation from pixels
and to get it back into the world once
you've drawn some conclusions you want
to act in that world and so we want that
connection and it turns out that deep
learning offers very nice tools for this
because it is currently the framework
that is most successful
in processing classifying and so on
perceptual stimuli from from those raw
sources like audio or images or videos
so what characterizes deep learning one
big thing is that it is based on
end-to-end training so we want to
formulate a model and then train it end
to end say we have labels and images and
we just want the system to be optimized
end to end for this problem we don't
want to engineer features because that
would require human input for every new
problem we want these features to be
automatically learned we want the system
to learn good representations and newer
networks have come back so to speak to
be able to do all of these things and
are now very versatile can be applied to
images to text to audio to video to grow
positions whatever you like these
systems are modular in design so because
we do gradient based learning we can
stick together modules and pass the
gradients through these modules and
learn the parameters of the individual
pieces it's not that they don't have any
prior knowledge built in there are a lot
of different architectures and every
architecture of a neural network gives
the system some kind of some kind of
inductive bias or represent some form of
prior knowledge the most well-known one
maybe our convolutions convolutional
neural networks encode certain spatial
relationships in the inputs and
facilitate through that specific weight
sharing the processing of images which
for example of them have translational
invariance or localized features the way
deep learning has come about is it
evolved from normal neural
works but it was then enabled by having
more data and more compute power in
particular GPUs of course so if we look
at what's out there and we'll have more
lectures on the details of course but I
want to give you a little sneak preview
the convolutional networks pioneered by
Jung Lacan and others the basic idea for
processing images and making use of
these biases that I was talking about
and were successfully used for
classifying digits and handwriting on
postal coastal codes on envelopes in the
US Postal Service and of course this is
a comparatively small problem but then
the real breakthrough was that they
could also be applied to this huge data
set of image net image net is a good
example for how you can make a great
research contribution not just be by
creating new algorithms or delivering
new solutions but also by posing new
problems so image net is this huge data
set which has a thousand different
classes of objects and for each of these
1,000 different types of objects there
are a thousand examples example images
of those objects and this data set alone
has really boosted research in image
recognition and in particular in deep
learning so if you're ever in a position
to curate a really big data set it's it
can be a beautiful contribution to
science to do that and in 2012 there was
this big breakthrough of really reducing
the error rate with large convolutional
neural networks on this image net data
set bringing it a lot closer to human
performance now human performance has
been surpassed on that and and that was
really the kickoff point for the the
modern era of deep learning if you like
the neural these convolutional neural
networks can also be applied to
it's an interesting question why that
might be the case well text is also on
some kind of grid it's just a
one-dimensional grid it's a sequence of
characters instead of being a
two-dimensional grid of images and again
we can use the same kind of prior
knowledge namely that we would like to
have localized filters that can find
local features in the text and this was
applied both at word level but more
interestingly also at character level
and we can also we also have the shift
invariance that we have in images you
know if there's a particular combination
of characters here that that have a
particular meaning for example or
indicate a particular class then if that
is shifted somewhere else in the text it
might have a very similar meaning so so
those same ideas apply to text and hence
convolutional ideas can be applied here
as well and then of course you can go
one step further from images going to
stacks of images if you like to videos
and and again convolutional neural
networks have been very successful here
the author of this particular work is
Karin Simonian who will give the
lectures on convolutional neural
networks so you'll be able to learn much
more there what's essentially happening
is that you learn one big nonlinear
function and I would like to point you
particularly to this idea of viewing
deep learning as it as differentiable
programming I think that's a very
powerful idea you have to imagine 20
years ago or now 20 years ago nobody did
neural networks so 30 years ago and then
again maybe 15 years ago people had the
simple feed-forward neural networks to
approximate functions from input vectors
to output vectors and they had usually
they're a very uniform architecture you
know just fully connected layers one
after another trying to approximate a
given function what happens now is that
people define these modules and stick
them together you know
you can have convolutional modules you
can have Mary memory modules you can
have attentional modules you can have
fully connected modules you can have
output modules with softmax and so on
you stick together these building blocks
and you really program in this space of
models and of course you leave degrees
of freedom free and Emily the weights
the parameters within these models and
they can then be learned and to end by
propagating arrows through this system
and so I think that's a viewpoint that
is very useful to really think of it as
a new programming paradigm where we
leave a lot of the program unspecified
those are the weights that need to be
learned but we do encode the the
structure that we know into the neural
network recurrence of course is also an
element of that you know if you have a
recurrent neural network so in terms of
architectures you can see that people
have gone wild and have developed
different things starting from these
humble beginnings in the inception
network the topology was varied you know
you see these they're not just stacked
linearly in one direction but you see
there's different paths through this
network so there's almost like a 2d
structure to it in leather networks it's
an unsupervised type of system where you
have horizontal connections between
layers that facilitate learning locally
within each layer because every layer
doesn't have to wait to the out until
the output and then again the input is
reached through the propagation
there's rest nets which have this idea
that there's always two paths through a
layer there's an identity function that
kind of skips the layer but there's also
a nonlinear function that can then is
then used to fit the residual everything
that just stays the same it can be sent
through the identity and then this bit
can fit the residual and those are
stacked on top of each other so people
have really
lord various architectures and have
improved improved and improved on on the
metrics for example on on the image net
this particular thing came out of
Microsoft Research in 2015 the rest net
and made it made a big jump in the
performance numbers for for image net
but you see how this resembles
programming putting together these
architectures and of course framework
like tensorflow
is exactly designed to do that to stick
together elements and then leave leave
the actual implementation of the
propagation of errors and add gradients
to the system okay similarly in
unsupervised learning there was a
proliferation of architectures and
models restricted Boltzmann machines
autoencoders things that do PCA and ICA
and sparse coding sometimes stacking
layers of these on top of each other to
achieve hierarchy and of course a recent
favorite are also Gans which are a
beautiful idea in that they take this
single player problem if you like of of
matching or finding a density model a
generative model they turn that into a
two-player game where there's one agent
if you like that tries to produce
examples that look a lot like what you
want to produce and the other agent
needs to judge if that looks right or
not and needs to distinguish those
artificial examples from those give them
in the training set beautiful work so a
lot of stuff and we'll learn more as the
lectures continue then of course
sequence modeling incredibly powerful
the example that I like the most is how
you can formulate the simple task of
translation as a sequence transduction
problem it's just beautiful to view it
that way you have an input sequence
which is your text in one language you
have an output sequence which is your
text in
the language and you're really view
translation just as a mapping from one
sequence into another and at first
that's beautiful but people thought it
didn't work and then they found that you
can actually make this to work and and
now most translation systems are based
on these types of networks incredibly
powerful and also working at this
symbolic level of characters in words
and so on and oriole videos who is also
responsible for a lot of the progress in
this area will give the lecture on
sequence modeling okay that's the deep
learning overview and now I'll move on
to these two little bits of research to
give you some idea of what can be done
if you combine recent reinforcement
learning ideas with deep learning ideas
we will start off with the work on human
level control through deep reinforcement
learning
the work known as the Atari work if you
like which was one of the the first big
papers that came out of deep mind and I
think you probably have all seen these
the Atari games I think just discovering
that this kind of problem exists was was
a huge step and it was people in alberta
kay first came up with the idea
can we use a collection of these simple
Atari games that children's play that
grown-ups play admittedly and that can
be quite addictive and and interesting
can we turn those into a reinforcement
learning challenge and what's so
beautiful about them is that they offer
a rich visual domain you know this is
just one of them break out but of course
they are
a few dozen others and what they have in
common is the action space and the
observation space so they are really if
we can call them a family of problems
because when we observe the state of the
system we see an array of pixels or
stream of images and that's the same for
all of these games the content of those
images of course is different the way we
need to interpret them for each game is
different but the format is unified the
same holds true pretty much for the
action space imagine a controller a
joystick you can enumerate the actions
that you can make on this and so we also
have a unified action interface and so
we can really view this as a unified
family of problems we have many
different games and that going back to
the definition of intelligence gives us
the ability to train a system that can
men do many different things and can be
tested on many different scenarios now
how do we know these are interesting
problems they could be super boring
problems but no humans design them to be
interesting for humans so we know
they're interesting right they're
interesting in exactly the sense that
we're interested in here because they're
challenging for humans they become more
difficult as you progress they involve
manipulation you need to do the right
kind of combination of moves and so on
you need to understand conceptually and
so what's going on on the screen and
yeah it's a tough problem let us see
and again here by the way finding the
right problem is such a key skill right
if you can come up with a nice problem
that is close to solving then you're in
a beautiful situation okay
so putting this back in the context that
we had of the little reinforcement
learning diagram we can think of the
game as the environment and what we want
to build now is the agent the controller
and the agent will take in observations
from the environment which are these
images in fact what they do is they take
they concatenate four of those images
because they weren't a little bit of
history so to speak of what the past
looks like to see for example in which
direction the ball is flying you need a
little bit of a trajectory and then they
create this cue network which takes as
input those images and outputs for a
given state for every available action
this so-called cue value and we'll learn
more about that in the reinforcement
learning lectures but the q-value
basically indicates for each state how
good it would be to execute one of the
available actions in that state and then
of course you would want to pick the one
that maximizes that Q value and the
sense in which this move or this action
will be good is in the sense of long
term reward because the Q function will
represent the long term estimate of how
much reward reward you will get if you
execute that action in that particular
state so and and then you can imagine
this the training happens through the
interaction with this environment and
the results across these 50 domains 49 I
think we're quite stunning in that for
many of these games
the system reached human level or
superhuman level performance there are
also a few games for which that wasn't
the case and you can you can imagine
what the distinction is if you have a
more reactive game where everything you
need to know is currently on the screen
and you need to react immediately to
what's happening that's relatively easy
there's also a more direct feedback
signal there because for example in pong
if the ball goes by might get a negative
reward so so those are relatively easy
what is hard is if you have more of a
puzzle game there's a game called
montezuma's revenge I think it's called
so right you know what that means in
Mexico right the the the game is
basically more like a puzzle game where
there's a very narrow channel of actions
that you need to take you know there's
something you need to jump over and then
you must not fall into the thing but
avoid something and so on and so that is
very hard because the reward for doing
all that only comes much later when you
have collected a key and then gone with
that key later on through some door and
then you get a positive reward and so
that is very hard because the reward is
very sparse and very long-term and so
that's much harder to do and so but for
a lot of these games that algorithm was
successful you'll learn more about it
from from Vlad but it is really quite
nice to see its
how in this particular game the
controller has learned to to shoot the
fish if it collides with the fish and
then it loses a life and what you see
here is the value function and here is
the the Q value for each of the actions
that you could take up down left right
firing and so on and here is the so
called advantage function in time which
is basically the advantage each of the
actions has over the others technically
is the Q value minus the Q function
minus the state value function and it
uses basically this and takes the action
that maximizes this here and what's
interesting here is shooting the fish is
relatively easy and it will survives for
a while early models of the algorithm
didn't find out that once your oxygen
level is very low you need to go to the
surface to breathe to survive but then
as as research advanced and the training
was able to capture more long-term
relationships the algorithms also
figured out that that's what it needs to
do so short term shoot the fish evade
the fish but at some point when this bar
is low go to the surface get oxygen then
go back down just to point out how hard
this is because it's not maybe
immediately obvious and the system just
gets this image here and it doesn't know
what it controls so it doesn't know that
it is the submarine right it has no
about idea about that it needs to find
out that it needs to control this of
course we have no idea how it represents
that but somehow that's the task right
when we approach this game we wiggle the
controller and then we see oh I'm this
thing and then we have all kinds of
semantics you know we know that we
observe a collision with the fish and
then we
Oh colliding with fish isn't good oh I
can shoot and Oh oxygens probably up
there and not down there you know all of
this prior knowledge we bring to this
problem and that's why humans of course
also learn these games much faster than
computers currently do but it is
remarkable that computers can learn it
at all I think brother you had a first
I think millions of frames yeah yeah so
the problem is in order for these
problems to be formulated mathematically
you need to decay the reward so either
you have a finite you have a finite
sequence and you just add up all the
rewards but what we typically do here is
we have some kind of discount factor and
so we say that further away rewards
count less then more nearby rewards you
know a bit like interest rate and and
that decay factor pretty much determines
how far we take into account those
rewards but the problem is that you
might not even see them so you don't get
to train on the perfect trajectory where
at the end you get to see that reward if
you make the wrong moves and buy before
you never see it so it's also related to
the exploration exploitation problem
that you are the or the agent itself is
the data generator so if the if if it
never manages to even see a rewards
because it dies before hand or the
episode ends in some way then it will
never get that reward and maybe every
now and then because it still has
randomness in there it stumbles upon the
reward but then that reward is very rare
so so the Lord the problem with the
long-term reward is both that it's hard
to propagate it back but also when
it requires complex trajectories to get
there you might never see it because you
never get there there is some degree of
planning so when you see a reward it is
propagated through the cue function
backwards in time that we do much more I
think we we also get intermediate
rewards just by you know seeing oh I
move to the right and it moved to the
right yeah you know you get that kind of
satisfaction or you shoot the fish and
it disappears and you get that kind of
satisfaction here there are actually
some rewards for shooting fish so it
depends on the definition of the game
but the very hardest games have almost
no intermediate rewards and you only get
it at the end and those are the tough
ones yeah so that's that's a very good
question and I need to clarify something
that is maybe not clear here so what
happens here across these 50 games is
that the system is trained on each one
of them with the same hyper parameters
and so on and then test it on that same
game and the system is universal in the
sense that you can apply it to each of
those 49 games and test it on it and it
produces respectable results on all of
them there is not a single system here
that can play more than one game but in
the meantime people have worked on this
and they have created systems that can
play several of these games but it's
very tricky because suppose you want to
learn them in order you first learn one
game but now when you learn the second
game the learning updates for that game
destroy what you have learned for the
first one if you apply it to the same
neural network and so you either need to
mix the tasks so that you never forget
the old one while adding information to
the new one or you need to somehow
protect the weights that were learned
for the first game while learning the
second and
protect the ones of the first and second
game while learning the third and these
are very tricky questions it's called
lifelong learning and that's just
protecting those weights what you are
referring to those even further what you
would like to see is that you learn the
first game and you've now learned
certain concepts like moving left moving
right up down objects and now it should
be faster to learn the second one
because now you've already got some
prior knowledge from the first one and
that is that is an even harder problem
you want to transfer information from
one domain to the next but that's the
Holy Grail really that that would be
ideal if the system learned at a level
of of conceptualization so to speak that
the things it learned and the first game
would be useful for learning the second
these systems don't learn at that level
they learn more at the pixel level and
they never have this notion for example
of an object for example one of the key
things that you would really want to
learn about these things is that there's
this block somewhere and you're it and
when you move your joystick to the left
it goes to the left and when you move it
to the right it goes to the right those
things we currently have no way of
learning that but that's certainly the
future we need to learn at that level of
generality it's a great question yeah so
what it what it gets is with each
observation it also gets rewards and for
example if you just shot a fish then it
will get a little number and and it will
interpret that number as the reward and
when it dies it gets in a negative
reward or it stops getting a rewards
that's really all it gets but it's it
uses the sum over those numbers weighted
by that decay factor as as the
optimization criterion for its learning
yeah so they had if they are they are
here derived from the score
it's another advantage of this test
suite because the game designers in some
sense already provided the reward system
it's the game score and there are some
problems associated with it because for
each of these games those numbers can
have vastly different numerical
quantities and so the system needs to be
robust to that and they use reward
clipping for example or rescaling of the
rewards to make sure that they are all
roughly in the same order of magnitude
yeah yeah so this is so there is a
system called ale
Atari learning environment that
explicitly translates these games into a
reinforcement learning environment and
interpret this course that it finds as a
reward and segments those out yeah okay
now in some sense we're here in the very
early stages of you know the simpler
environments and many people have now
moved on although we're still using some
of these environments to test algorithms
that we develop but here's a newer
generation kind of set of environments
called deep mine lab and this is also
publicly available and constitutes a
kind of third first-person perspective
of some maze world where different tasks
can be represented what's nice here is
all of these tasks play out in some kind
of maze and the agents are in there with
the first-person perspective and see the
world from this first-person perspective
and and move in the same way so again we
have a family of problems that have a
unified input-output invar interface
they always see the pixels that come in
and the actions they take are turning
moving forward moving backward maybe
shooting I'm not sure
and now various different tasks can be
formulated within this environment and
agents can be trained here's basically
such an agent and when it interacts with
that world it gets the first-person
image of of what's of its position in
the world and it gets some kind of
reward depending on what it encounters
in that world and here the action spaces
can basically go forward backward turn
strafe jump up and and that's it unify
for all tasks within this task suite and
now you can see that this is a much more
demanding kind of environment here
someone navigating it you know apples in
this case give positive rewards and you
see the difficulty here is to interpret
this 2d image in terms of its 3d what it
represents in 3d and learn how to
navigate the maze based on certain point
things for example you see there's
images on the wall
so in principle once the agent has
navigated the maze for a while it could
have built up some kind of map knowledge
of the situation and if it finds itself
in a new spot in a spot then it might
orient itself by recognizing that
there's a particular image on the wall
and then it might draw conclusions about
what its best course of action would be
in that particular situation and there's
all kinds of interesting tasks that you
can do here
maybe mostly tasks related to navigation
because you would you know find the
Apple or walk through the maze and try
to collect as many apples as possible
things like that you can do but you can
also do other things for example
this is a kind of laser tag level in
space where the mode of movement is a
little more complex you have these ramps
where you can fly through space and get
from one platform to the other and again
put yourself in the position of an agent
seeing this and having to learn what
what it means what it sees it sees a 2d
projection of a 3d world from the
first-person perspective and somehow
needs to derive from that what actions
it would be advantageous to take in that
particular situation ok so much for that
and I think Vlad will be able to give
you much more insight into this and how
to address these types of problems
different algorithms and so on as the
last point for today I'll talk a little
bit about the alphago project to give
you some inspiration for some more
application of deep learning and
reinforcement learning I saw that we
have at least one go player here I'm so
glad strange though random ah see we
have to go players you could play maybe
you are playing and I'm just not seeing
it ok so go is it's a very different
type of problem so the the problems we
we just looked at in this labyrinth are
very much inspired by how we see the
world and interaction with the world
that is is naturally spatial where an
agent has a given position in that space
moves that position around has a
perspective that is tied to that
particular point of course go is a
problem that is different that is
difficult in other ways it's played on
this board as you know 19 by 19 with
black and white stones there's no agent
that you are in this game he plays
stones on the board in some sense stones
may be the agents but they're not
actually
so the player is the agent but it's not
so clear what the players representation
is other through which stones they
currently have on the board and the goal
of course is to surround territory in
this game and here the challenge is
quite different because here we have a
super large search space a huge search
space in fact represented by this the
game tree which has a game key
complexity of the breath to the power of
the depths of this tree so the breaths
is so many different moves you can
choose at any point in time and the
depths is how deeply you have to lock
down in the tree and the other problem
of course that needed to be addressed
here is one that it turns out deep
learning was suitable for namely to
assess how good a given position is to
evaluate the position and if you look at
the problem it's kind of natural that
you would think that that might be the
case because it looks a lot like a
vision problem you know you look at the
board it has a 2d structure like grid
structure like an image and so we know
humans are good at it or can be good at
it very few humans are actually good at
it but there are some humans who are
good at looking at the board and
determining for example if black has the
advantage or white has the advantage or
what would be a good move in a given
situation and and that's where were deep
learning comes in here's an illustration
of chess where you have maybe roughly 20
moves in every given position to
consider which is already a lot and of
course in go is many more it's up to 300
moves really so how is it done we use
deep learning and we have two types of
networks here one looks at the board and
finds a mapping from a board position to
an evaluation so it's basically trying
to estimate how likely it is that black
or white would win in a given situation
from looking at the board and the other
network is the so-called policy Network
which looks at the board and looks at
what might be good moves in this
particular situation and both of them
are visual tasks that humans can do we
know that a good go player can look at
the board and maybe at a glance see who
is in a better position sometimes it
takes more analysis to do it but what a
go player definitely can do is look at
the board and see plausible moves or
they do maybe they do the contrary they
say these they see a lot of moves that
definitely aren't good and they're not
going to consider so they select which
moves they're going to analyze more
precisely and those of course correspond
to particular visual patterns that are
not unlike the visual patterns that we
pick up when we do object recognition or
when we recognize faces there are
combinations of edges and areas and we
are now connected bits and pieces and so
it's not that surprising that a
convolutional architecture can actually
represent this mapping I still find it
somewhat surprising because if you
change a single pixel in a in an image
that never changes the semantics of the
image right there's almost no way it
could do that whereas if you change a
single stone on a goal board it might
very well turn it from a winning
position into a losing position and so
there is something specific about this
mapping problem it's much it's not as
smooth as maybe you think of the typical
visual recognition problems well how is
this use done and you can basically
think of the planning process here as
being represented by a huge game tree
and the problem of that is the game tree
is too big we cannot search it but we
can reduce it in two dimensions by using
the value Network we can avoid having to
evaluate all the way to the end when we
know the outcome of the game but we can
evaluate earlier and by reducing by
selectively only looking at moves that
look promising we can make the game tree
much more narrow and in that sense again
reduce the size of the tree and the
remaining tree if you
then is amenable to search techniques
that can get us good results using at
that point of paring down your tree is
it is it deep blue like exhaustive
search
it's called Monte Carlo tree search and
so it's still a little different Monte
Carlo tree search always picks one
trajectory through the tree and and
always develops a new node at the end
and grows the tree in this way so it's
it's different from what the blue would
be using which is a min Max search using
heuristic called alpha beta pruning
where it where it basically can
disregard certain parts of the tree for
logical reasons but otherwise has to
expand large portions of the tree so the
multicolored research does more of an
averaging so in min max search you
really have assumed that you have an
accurate evaluation function and one
player tries to maximize it and the
other player tries to minimize it
whereas in Monte Carlo tree search we
average the evaluations we just make
sure that we only ever go down promising
roots in the tree enough times that the
averaging is biased towards maximizing
and minimizing it's a nice robust search
technique and then as you probably know
this alphago
turned out to be much better than the
existing programs at the time which were
only at strong amateur level whereas
alphago the early version went into
professional territory and eventually
beat the strongest player in the world
at at this game and the evaluation was
first against fund why the European
champion and then later against Lisa toe
the Korean champion yeah
from 1990 comes more tractable so you
but you have to go to seven by seven I
think for it to be really tractable nine
by nine is still difficult and 11 by 11
is about as complex as stress so but
there's an interesting idea there we
talked about this these ever more
complex environments in the definition
of intelligence and one way to formulate
a curriculum over this task for example
would be to start learning on smaller
boards and hoping that you can have an
architecture that also works on bigger
boards and by learning on smaller boards
first you would have first successes and
a smaller space and you could learn some
basic patterns that would then transfer
to the to the larger boards I think that
could be a promising that could be a
promising route to address yeah yeah
that that is very interesting yeah it's
a smaller game it definitely is a
smaller game but it's good yeah people
have made progress they go from the
endgame positions backwards and from the
openings forwards and then at some point
they they match and and and can prove a
result okay and of course there's
there's more here and we will hope too
that Dave will cover this in in his
guest lecture of course recently we've
made a lot of progress whereas in this
earlier work we use supervised learning
to train these neural networks we used
games that were actually played by
humans and and use them as supervised
learning datasets taking a position and
a move as a classification problem
position is the input move is e is is
the class if you like and we were able
to train these networks but more
recently in the alphago zero work we can
do
this entirely through reinforcement
learning with the system only plays
against itself in and learn how to play
go even better than with human input
strangely and and then we also applied
that to chess but that's really just a
preview to today's lecture okay final
thing I would like to point you to some
extra revision material this was last
year included in the in the second
lecture which we've now replaced by the
tensor flow tutorial because we thought
that that was better use of the time but
if you want to take a look just at the
slides that we posted on Moodle there's
some basic revision of a regularization
generalization in supervised learning
and how gradient based learning works in
linear regression and in logistic
regression which are the models that
kind of lead up to to feed forward
neural networks then okay any questions
[Applause]
we're still trying to figure that out
there's a lecture cast recording and and
there's these recordings but so far we
haven't been able to combine them onto
Moodle so have to find out
yeah yeah no no I see what you mean I
see what you mean I mean the we were
very interested in this question
ourselves because in some sense you can
view these games as illustrations of
maybe what happens as we make AI
stronger and stronger so here we have
small worlds and we're now at the point
where the competence of the AI is
greater than the competence of humans in
this particular domain and so what does
that look like what is it like to play
against a higher intelligence so to
speak and maybe we can draw conclusions
that would also hold it for other
domains that are maybe even more useful
medical diagnosis or car driving or
whatever and so we were looking at this
and some of the moves that that alphago
made even during the first match in
Korea were pretty amazing according to
our experts so they were truly
surprising people were using the word
creative moves inventive moves because
some of the moves defied
well-established rules that human
players typically respect in their game
humans have a lot of heuristics that
they use to select moves because the
game is so complex but what that also
means is that they have certain biases
that is that are very hard for them to
overcome so if they have for example and
a trained pattern that tells them that a
particular move cannot possibly be good
but in this particular situation that
move happens to be the best move then a
lot of humans would have difficulties
finding that move because early on in
their reasoning they would rule it out
but alphago 'men have those limitations
or not to that degree at least and so it
came up with moves that were
counterintuitive to humans but when they
then saw how the game unfolds they
suddenly
realized how brilliant that move early
on was that they previously didn't
understand based just on their rules so
so that was pretty amazing to see and
then more recently we have some
interesting feedback on the chess games
that we published where we applied this
same methodology to chess and some of
the grandmasters were amazed at how
flexibly alpha 0 evaluates positions so
when you normally write a chess program
you would define an evaluation function
that tells you how good a given position
is and that would include the material
you know do you have more material than
your opponent queen king queen pawns and
so on it would include pawn structure
King safety mobility of your pieces and
so on and there but they would all have
some kind of fixed coefficient that the
programmer would have to put in so that
you can actually evaluate this function
and do a search based on it now the
alpha 0 approach doesn't have any of
those limitations it just learned this
general function from board state to
evaluation and as a consequence it's not
bound for example to the concept of
material so it much more easily
satisfies a piece in order to gain
positional advantage if it has learned
that in fact that increases its winning
probability and so sorry in some of the
games that we shared alpha 0 place a
sacrifice which is a positional
sacrifice it loses the piece and almost
no human player would do that unless
they knew exactly what they were getting
for it but alpha 0 was happy with that
move increasing its winning probability
and then somehow 40 moves later it would
get that home and actually win the game
and so that was pretty otherworldly in
terms of playing style and and chess
masters just just enjoyed those games a
lot yeah
and no it wasn't open sourced but there
are other programs other groups that
have in fact started reproducing this I
think there's an open source project
called Leela open Leela or something
Leela zero which is trying this and
there's also $0.10 the the Chinese
Internet company has a program called
fine art which is designed according to
the alphago specs and has become much
better in in recent times and is also
challenging professional players now we
published the algorithm and people can
implement that if they like but we
didn't open-source any of the code yeah
it's it's an interesting problem so what
alphago is trying to do is to maximize
its probability of winning when that
probability of winning is essentially at
100% then certain moves look the same
for alphago it can give away a point or
two if it's currently winning by five
points it doesn't care if it gives away
at one point or two and that seam can be
very frustrating for human opponents
because the moment you see you're you
know you
basically it doesn't care anymore yeah
it's an interesting problem I'm not sure
it can be solved while retaining the
playing strength that's an interesting
challenge because the criterion that we
use is winning probability and not
winning by a certain margin if we now
change the criterion and training to
from winning probability to please win
by the maximum possible margin then
there might be side effects like it
might start taking certain risks now to
win by a larger margin and so what what
you'd be asking for is really maintain
your high winning probability while at
the same time maximizing the margin of
winning that would lead to those non
slack moves and yeah it's it's tricky
maybe it can be done but I think our
research was mostly driven towards the
most principled question of maximize
winning probability and not not consider
the marginal okay I think we need to get
out of here thank you very much for your
attention
[Applause] |
d02afa98-c3b9-4c59-a96f-62f78036fd44 | trentmkelly/LessWrong-43k | LessWrong | Optimize Your Settings
Related to: The Good News of Situationist Psychology
Perhaps the most significant teaching social psychology has to offer is that most of our behaviors are determined by situational factors inherent to our settings, not by our personal qualities.[1]
Some consider this depressing-- for instance, the Milgram experiments in obedience to authority and Stanford prison experiment are often cited as examples of how settings can cause otherwise-good people to participate in and even support unethical and dangerous behavior. However, as lukeprog points out in The Good News of Situationist Psychology, this principle can also be considered uplifting. After all, if our settings have such an effect on our behavior, they are thus a powerful tool that we can employ to make ourselves more effective.[2]
Changing Your Physical Settings
One relatively easy place to start making such changes is in your personal life. I have found that great productivity increases can be gained through relatively minor changes in lifestyle-- or even seemingly-trivial matters such as the position of physical (or sometimes digital) objects in your environment!
For instance, I recently noticed a tendency in myself to "wake up" and then waste the next twenty or thirty minutes aimlessly browsing the Internet on my laptop in bed before actually getting up and eating breakfast, showering, going to work, etc. Since I value time, especially morning time, substantially, I decided that action should be taken to avoid this.
At first, I figured that once I had noticed the problem I could simply apply willpower and avoid it, but this proved less than effective-- it turns out that my willpower is not at its strongest when I first wake up and am still a little groggy![3] I then decided to apply the principles of situational psychology to the situation. The most obvious setting contributing to the problem was that I was using an alarm app on my computer to wake up in the morning, and turning off this alarm cau |
141c4425-c6f8-4e6b-ae2b-4ded7599893a | trentmkelly/LessWrong-43k | LessWrong | Mini thoughts on mintheism
[Epistemic status: Hunch in the form of a neologism. I've heard this idea—that I'm dubbing mintheism here—implied in several corners of the Internet. I want to give it a name and take a stab at why it's plausible. This post presupposes a connection between the decline in religious affiliation and the rise of conspiracy theories.]
The gist of it
Put like an SAT analogy, mintheism is to atheism as minarchy is to anarchy.
Fully defined, mintheism is a cost-benefit approach to religious belief and practice within a society that attempts to find the optimal amounts and kinds of religious belief and practice that achieve prosocial outcomes while minimizing antisocial side-effects.
Under this definition, there's many different ways one could solve the optimization problem. I'll explain my take which I'll call the nonsense management approach.
A little nonsense now and then is... necessary for signaling loyalty to your tribe
Social isolation is really undesirable and has measurable negative outcomes (that sentence has 10 words and 8 hyperlinks, sorry). So let's assume people want to be part of a tribe and that conservation of tribalism is basically accurate (analogous to conservation of mass—the level tribalism within a society is never created or destroyed, it just changes forms).
Building on an insight from Mencius Moldbug, the pen name of Curtis Yarvin of Neo-reactionary fame in An Open Letter to Open-Minded Progressives. (Please don't mistake this as staning for Yarvin. While the quoted insight is a good one, I would encourage those who haven't yet to read Scott Alexander's Anti-Reactionary FAQ).
> "...in many ways nonsense is a more effective organizing tool than the truth. Anyone can believe in the truth. To believe in nonsense is an unforgeable demonstration of loyalty. It serves as a political uniform. And if you have a uniform, you have an army."
I'll go a little further than Moldbug and say nonsense is a necessary to create and maintain tribes |
b0aee207-6e78-4588-8672-dbe8db655ee3 | trentmkelly/LessWrong-43k | LessWrong | Eli Dourado AMA on the Progress Forum
Eli Dourado—whom you might know from his writing on geothermal, supersonic, energy superabundance, or NEPA—is doing an AMA (Ask Me Anything) on the Progress Forum.
Get your questions in now, and upvote the ones you want to see answered. He’ll start answering Monday (Jan 30).
After you’re done, check out our previous AMAs with Tyler Cowen and Patrick McKenzie. |
97efc162-a353-4cae-bc67-5998a5f60f7f | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Lessons from Three Mile Island for AI Warning Shots
**Summary**
===========
In 1979, a nuclear power plant in Three Mile Island (TMI), Pennsylvania experienced a meltdown. In response, the public became much more concerned with the safety of nuclear power and successfully demanded increased regulation. TMI seems potentially instructive for how the American public might respond to an AI “warning shot”, i. e. a situation in which AI visibly causes significant but not catastrophic damage. In particular, TMI suggests that
1. The public is capable of being quickly and strongly influenced by warning shots (95%).
2. The public’s response to a warning shot is influenced more by pre-existing discussions about the risk (including fictional accounts), media coverage of the accident, and the public’s attitude towards the actors responsible for the warning shot than by the magnitude of the accident (90%).
3. Confusion about the relevant technology may lead the public to fail to recognize the threat prior to the warning shot, but once the threat is clear, it may cause the public to become more afraid (50%).
4. Capitalizing on warning shots is much more likely if the actors who aim to do so are part of a broader political coalition (85%).
5. The legislative response is likely to be ineffective by default, particularly given that the technical problems involved in regulation are complicated (80%). However, industry may be able to respond more effectively through internal regulation in order to avoid further reputational damage.
These takeaways suggest that if the longtermist AI governance community (henceforth referred to as “the AI governance community”) wants to effectively influence AI policy in response to an AI warning shot, it should attempt to influence the media environment prior to the warning shot (perhaps through the release of realistic seeming fiction about AI takeover scenarios), utilize/play into public suspicion towards big tech, form allies with other groups who support AI regulation, and think hard in advance about what policy responses would actually reduce AI risk.
Epistemic Status: This post was written as part of the Stanford Existential Risks Initiative's Summer Research Fellowship, and is the product of ~50 hours of writing and research. Moreover, n = 1, so this post overall provides relatively weak evidence for its claims; the probabilities assigned are more based on my priors than on the evidence from Three Mile Island. That being said, I am quite confident in the straightforwardly historical claims made.
Acknowledgements: Thank you to Matthew Gentzel and Ben Snyder for their feedback
**Introduction**
================
Currently, one of the biggest barriers to governments taking major political action to reduce existential risk from AI is that neither voters nor government officials take the risk very seriously. This could possibly change in response to a “warning shot,” i. e. an event in which an AI system causes a major accident but does not precipitate existential catastrophe. Thus, it seems important to answer the question of how the public might respond to a warning shot if it occurs, and how we might channel this response into effective political action to reduce AI risk.
In order to answer these questions, I examine a previous case of a warning shot-like event which led the public to successfully demand legislation to reduce future risks: the Three Mile Island (TMI) nuclear accident. In March, 1979, a nuclear power plant in Three Mile Island, Pennsylvania experienced a partial meltdown. While a total meltdown was ultimately averted and, according to most estimates, the damage was relatively minor, TMI prompted increased public concern about the safety of nuclear power and led to major efforts to regulate it. Thus, TMI plausibly serves as a good case study for what factors make a warning shot generate major public and legislative response.
**Limitations**
TMI is far from a perfect analogy to an AI warning shot. Firstly, the risks posed by AI are very different to those posed by nuclear power. In particular, risk from nuclear meltdowns may be easier to understand and certainly seems more “realistic” to the public than risk from misaligned AI. Moreover, U. S. culture and politics have changed in important ways since 1979. For example, in the more partisan environment of US politics today, it is possible that the question of how to respond to an AI warning shot might become a partisan issue. Finally, particularly given that the AI industry will likely grow significantly in the next decade prior to a warning shot, it might be much more influential than the nuclear industry was at the time of TMI and effectively oppose legislation. Nevertheless, TMI is one of few cases of a widespread public response to a major accident involving technology in recent history. Thus, insofar as we want to have some kind of baseline for what responses to warning shots look like, studying the case of TMI seems valuable. With this in mind, I present 7 lessons that I think we can learn from the example of TMI for how the public might respond to an AI warning shot, and how we might take advantage of this response.
**Lessons from Three Mile Island**
==================================
**Lesson #1: Warning shot events can have large effects on public opinion**
Some people believe that the public is insufficiently attentive to the state of the world to respond in any strong manner to warning shots (see: COVID and pandemic preparedness measures). However, the case of TMI suggests that this is false. Prior to the incident, public opinion was generally favorable towards nuclear energy, with about 60% of the public supportive of building nuclear power plants in the US and about 30% opposed. Immediately after the accident, the public became about evenly split on the issue, with about 45% on each side. While support for nuclear power seems to have briefly rebounded after a few months, this rebound then quickly reversed partially owing to a second wave of coverage in response to the release of a public report on TMI. Moreover, in addition to shifting views on nuclear power, TMI seems to have caused the public to pay more attention to the issue of nuclear power in general, as the percent of people answering that they did not have an opinion about whether more nuclear power plants should be built in the US dropped from over 20% to under 10% immediately following the accident.[[1]](#fndqkcfw3ui4) Finally, TMI prompted a significant legislative response. In particular, immediately after the accident, Jimmy Carter formed a commission to investigate the accident and to make recommendations about what could be done to reduce the probability of future accidents. The commission then released a report which ultimately resulted in significant reforms to the agency which regulated nuclear power, the Nuclear Regulatory Commission (NRC). While these reforms were not very effective, they did involve a significant restructuring of the NRC.[[2]](#fn6lv3r7key7) Thus, TMI suggests that as irrational as the public can be, it can quickly update in response to major technological accidents and pressure public officials into taking major actions aimed at reducing future risks.
**Lesson #2: Which warning shots grab the public’s attention is not mostly explained by the objective “size” of the warning shot**
While the evidence presented above suggests that warning shot events can get the public concerned about an issue, the public’s broader pattern of response to nuclear energy accidents suggests that which specific accident grabs its attention may be more a matter of chance and media coverage than of the actual size of the issue. TMI was not the first nuclear power accident in the US; throughout the 1950’s, ‘60’s, and early ‘70’s, there were several other accidents involving nuclear power that failed to generate nearly as much public attention as TMI.[[3]](#fnhkrzdujwbdw) While TMI was in important ways “bigger” than these other accidents, it was in other ways “smaller” insofar as some of the past accidents caused deaths while TMI did not immediately lead to any. Similarly, the differences in the public response to nuclear energy accidents between countries does not seem to be well explained by how they are actually affected by the accident. For example, different European countries’ publics had significantly different responses to the Chernobyl nuclear meltdown. However, the size of these responses was unrelated to the extent to which the country in question was negatively impacted by the radiation which Chernobyl generated.[[4]](#fng5k78f17c9) This suggests that it may be difficult to predict what specific accident or kinds of accidents will draw public attention to AI risk.
**Lesson #3: Fiction and media coverage can greatly influence public response to a warning shot**
If the public’s strong reaction to TMI was not primarily due to the nature of the accident itself, then what caused it? While public response to an event is determined by a huge number of variables, two key ones in the case of TMI seem to have been the prior presence of fiction which primed the public to be concerned about nuclear meltdowns and the media coverage of the event itself. Some have argued that a major reason for lack of public concern about AI risk is that people see AI takeover scenarios as science fiction rather than as something that could happen in real life. However, the case of TMI suggests that fiction can sometimes have the opposite effect, causing the public to be concerned about the technological accidents which it presents. In particular, the movie “The China Syndrome” was released less than two weeks prior to TMI. The China Syndrome presented the fictional story of the near-meltdown of a nuclear power plant and the efforts of the company responsible for the accident to cover it up. It was widely viewed by the public, starring major actors such as Jane Fonda and Jack Lemmon, and explicitly aimed to call attention to the issue of nuclear safety, dramatizing the possible negative effects of nuclear meltdowns and portraying nuclear power companies as corrupt. Moreover, even before TMI, at least some commentators took the concerns raised by the movie seriously.[[5]](#fnitcaopq1bcp) While other commentators viewed the concerns which the movie raised as overblown, it would prove critical to the public response to TMI. In their coverage of TMI, media commentators often drew parallels to The China Syndrome. For example, in its cover story immediately following the accident, *Time*magazine stated that the public statements given by the public spokesman for the plant sounded “as if they were taken right out of the script for the film *The China Syndrome,*” and Newsweek coverage of the event included images taken from the film.[[6]](#fnjbmlr0rq8q) Confirming the importance of the film, the report issued by the NRC months after the accident stated that its effects on the public perception of TMI had been “immeasurable.”[[7]](#fnpm5airzjuam)
A second major reason why the public seems to have reacted in the way that it did was that media coverage of the accident was extensive and dramatic. Partially, this was due to the fact that the media had little else to cover during the week in which TMI occurred.[[8]](#fn6c0rmdxj1b3) However, TMI also happened at a time when the media was already paying increased attention to nuclear power issues due to the release of the China Syndrome, Jimmy Carter’s efforts to pass his energy policies, and generally increasing concern about the safety of nuclear power throughout the 1970’s.[[9]](#fnrku8fxekghh) As a result, TMI received lots of TV coverage, with 40% of all evening news coverage being devoted to it during the first week of the crisis. Moreover, this coverage was in many cases quite dramatic. For example, Walter Cronkite of CBS evening news stated during his coverage of TMI that “The world has never known a day quite like today,” and ABC News showcased sentimental stories of mothers evacuating with their children.[[10]](#fn4v8bzde5a1o) Between the release of The China Syndrome and the coverage of the event, the public was heavily influenced by media to respond fearfully to TMI.
**Lesson #4: The public is more responsive to warning shots if it does not trust the authorities responsible for them**
Another major reason that the public responded strongly to TMI was that it distrusted the nuclear power industry and other authorities, partially due to the conflicting reports which they issued about the accident. Polls from the 1970’s reveal that public skepticism towards the nuclear industry and its regulators pre-dated the crisis, perhaps owing to Americans’ generally skeptical attitudes towards authorities.[[11]](#fn68fl3aiw1w) However, TMI itself increased this distrust. During the crisis, spokespeople for the plant were confused about what was happening. Nevertheless, in response to media pressure, they made various reassuring positive statements which they later had to retract once they learned more information. For example, Jack Herbein, the vice president and spokesperson for Metropolitan Edison (the company operating the power plant), first claimed that the accident was not serious, and the utility spokesperson Dave Klucsik stated that “[t]here is absolutely no risk of a meltdown.”[[12]](#fnzwwh8f4end) However, soon afterwards, the NRC declared that a meltdown was possible, and media figures revealed that they had been told by people working at the plant that the crisis was more severe than their public statements had indicated. Officials continued to give contradictory and confusing statements as the crisis continued. As a result, the public began to suspect that industry and government were both conspiring to conceal information from them. This lack of trust and information caused further panic, with many local residents citing anxieties over conflicting information as their reason for evacuating the area during the crisis.[[13]](#fn1v50lyq54o) Moreover, media coverage after the event continued to push this narrative, with The Washington *Observer-Reporter*running a headline stating that “Three Mile Island Meltdown Was Near” and other media outlets covering accusations that the radiation released during TMI had had more damaging health effects than authorities had claimed.[[14]](#fnoz75knqdpod)
**Lesson #5: Lack of understanding is not necessarily a barrier to public concern, and in fact can in some cases increase the public's level of fear**
One reason that some suspect that the public may not become concerned with AI risk even if a warning shot occurs is that it does not understand the technical problems involved in AI alignment. However, in the case of TMI, the public’s lack of technical knowledge does not seem to have made it any less concerned about nuclear energy. Due to the inconsistency of the reporting, as well as the complex nature of nuclear power, much of the public seems to have been confused about exactly what was happening during the meltdown. However, this led many to infer that the situation was worse than it was in actuality. For example, 36% of those who responded to an NYT poll believed that the accident would produce a mushroom cloud explosion. Confusion about radiation seems to have particularly contributed to public panic. Many viewed radiation as a “mysterious and dangerous” force due to its ability to invisibly cause disease and death, which heightened their fear. Similarly, some commentators framed nuclear energy as an unnatural power which humans should not tamper with in response to TMI, invoking both Christian and classical rhetoric against playing God or opening the “Pandora’s box” of nuclear power. Thus, while lack of understanding can sometimes cause the public to be unconcerned about a risk, it can also lead to heightened perception of the risk once it is noticed, particularly if it seems spooky or alien.[[15]](#fnsg7vjyo9olp)
**Lesson #6: Capitalizing on warning shots is easier if you have allies**
While there are many factors that increase the probability of legislative action in response to a warning shot, one important one seems to be whether the actors pushing for legislative action have political allies. In the case of the US, the anti-nuclear energy movement made allies with social movements that were originally much larger than it, such as the broader environmental movement and the anti-nuclear weapons movement, and was organized by protestors with connections to the civil rights and women’s liberation movements. The anti-nuclear energy movement in Finland seems to have adopted a similar strategy.[[16]](#fnpfaczfrs78m) In Germany, the movement was initially coolly received by existing political parties but eventually found a major ally in the newly founded Green Party.[[17]](#fndhdwxaptz89) In contrast, in countries such as France, the anti-nuclear power movement was not similarly part of a larger coalition and did not have a political party supporting it. Even though nuclear power had similar levels of public support across the different countries, the French anti-nuclear power movement failed to halt the construction of new nuclear power plants, while it largely succeeded in Germany, Finland and the United States.[[18]](#fnkaptvrp8yyh)
**Lesson 7: The legislative response to a warning shot involving a technically complicated technology may be ineffective by default, though the private sector may respond more effectively**
While TMI prompted significant responses from legislators and regulators, these responses seem like they did not have actually done much to reduce the risk of nuclear meltdown. While the rate of development of nuclear reactors in the US slowed considerably in the 1980’s, this seems to have been significantly if not entirely due to pre-existing economic trends and increased regulation during the 1960’s and ‘70’s rather than the legislation caused by TMI.[[19]](#fnsm7wcucxfn) Furthermore, internal reports by the NRC suggest that the agency was not very effective at regulation even after the reforms, as it often failed entirely to inspect plants and, when it did, lacked the technical knowledge to understand what changes were necessary.[[20]](#fn2ix6gtndhmq) However, despite the lack of successful legal regulation, TMI prompted the nuclear power industry to informally regulate itself which was more effective due to the greater technical knowledge of people working in the nuclear power industry.[[21]](#fnjxm1tvmncsj)
**Conclusion**
==============
Several facts about the TMI incident suggest that there are reasons to believe that, conditional on there being an AI warning shot, the public will become concerned about AI risk. Broadly, TMI serves as a proof of concept that the public can quickly become concerned with the safety of a technology and demand regulation of it in response to major accidents. Moreover, an AI warning shot seems likely to share some of the characteristics that seem to have contributed to the public’s strong response to TMI. Firstly, just as there was increased coverage of safety problems with nuclear power in the years and months leading up to TMI, there has recently been increasing attention to the negative effects of AI such as algorithmic discrimination and polarization due to social media algorithms, which will likely only increase over time. Additionally, just as public concern with the safety of nuclear energy post-TMI was increased by feelings of distrust towards regulators and the nuclear power industry, the broad distrust of “big tech” by the contemporary public seems likely to prime it to respond strongly to an AI warning shot. Finally, just like radiation, risk from AI is currently poorly understood by the public but, once it is made salient, is alien-seeming and threatening in a way that seems likely to further scare people.
However, even though these factors suggest that there is a substantial chance that the public becomes concerned about AI in response to a warning shot, further efforts are necessary to increase the odds that this reaction occurs and that it is translated into effective legislative action. The critical factor for the public's response to an AI warning shot which the AI governance community can attempt to influence is media coverage. If the analogy to TMI holds, both the immediate media response to an AI warning shot as well as the media environment prior to the warning shot, such as fictional accounts of AI risk and political discourse about the potential harms of AI, would be very important for how the public would react to an AI warning shot. Thus, aiming to influence media narratives about AI, as well as perhaps releasing fictional media such as an episode of a TV show (e. g. an episode of Black Mirror) or a movie which depicts a realistic seeming AI takeover situation might be effective pathways to increase the odds of a strong public response.
In addition to increasing the odds that the public becomes concerned in response to a warning shot, the AI governance community also needs to put itself in a position to actually influence legislation if the public becomes concerned about AI. To that end, the successes of the Finnish and US anti-nuclear movements which forged a broader coalition opposed to nuclear power are instructive. In the case of AI risk, forming a coalition with actors who seem like they could possibly favor stricter regulation of the AI industry such as those concerned with bias and discrimination and AI unemployment seems potentially worthwhile. If the AI governance community chooses to pursue political allies, then it should be mindful about who those are, as picking allies also means picking enemies. For example, if it were to align itself with the aforementioned movements, it would run the risk of being perceived as a left-wing movement and thus alienating conservatives. However, particularly given that it is probably smaller than the anti-nuclear movement was before TMI, the AI governance community seems unlikely to successfully influence policy in response to a warning shot by itself.
Perhaps most importantly, even if put into a position of influence, the AI governance community will also need to have effective concrete policy proposals ready if the legislative response to a warning shot is to reduce AI risk. Without such carefully researched proposals, regulation is likely to be ineffective by default, because effective regulation of AI requires even more technical knowledge than effective regulation of nuclear energy which seems to have been beyond the capacity of the NRC. Indeed, given the potentially dim prospects for such regulation, it could also be wise for the AI governance community to attempt to influence AI developers to institute non-legal regulations in response to a warning shot than to influence the US government, as developers are likely to have more technical knowledge and be closer to the actual problem than regulators.
Zooming out, I think that major legislative action to reduce AI risk is unlikely without a warning shot, as neither the public nor politicians seem likely to be convinced by abstract arguments that AI could be very dangerous without concrete examples. Thus, to the extent that reducing AI risk through legislation is a viable strategy, understanding the dynamics of the public’s response to a warning shot and how that response affects legislation is important. While this post has begun to address these questions, further research on this seemingly neglected question is necessary.
1. **[^](#fnrefdqkcfw3ui4)**For the various claims about public opinion, see Rosa, Eugene A., and Riley E. Dunlap. “Poll Trends: Nuclear Power: Three Decades of Public Opinion.” *The Public Opinion Quarterly*, vol. 58, no. 2, 1994, pp. 295–324. *JSTOR*, <http://www.jstor.org/stable/2749543>.
2. **[^](#fnref6lv3r7key7)**Temples, James R. “The Nuclear Regulatory Commission and the Politics of Regulatory Reform: Since Three Mile Island.” *Public Administration Review*, vol. 42, no. 4, 1982, pp. 355–62. *JSTOR*, <https://doi.org/10.2307/975979>. Accessed 6 Aug. 2022.
3. **[^](#fnrefhkrzdujwbdw)**Mazur, Allan. “The Journalists and Technology: Reporting about Love Canal and Three Mile Island.” *Minerva*, vol. 22, no. 1, 1984, pp. 45–66. *JSTOR*, <http://www.jstor.org/stable/41820553>. Accessed 10 Aug. 2022.
4. **[^](#fnrefg5k78f17c9)** Koopmans, Ruud, and Jan Willem Duyvendak. “The Political Construction of the Nuclear Energy Issue and Its Impact on the Mobilization of Anti-Nuclear Movements in Western Europe.” *Social Problems*, vol. 42, no. 2, 1995, pp. 235–51. *JSTOR*, <https://doi.org/10.2307/3096903>. Accessed 8 Aug. 2022
5. **[^](#fnrefitcaopq1bcp)**For the previous claims about The China Syndrome, see Shaw, Tony. “‘Rotten to the Core’: Exposing America’s Energy-Media Complex in ‘The China Syndrome.’” *Cinema Journal*, vol. 52, no. 2, 2013, pp. 93–113. *JSTOR*, <http://www.jstor.org/stable/23360267>. Accessed 10 Aug. 2022.
6. **[^](#fnrefjbmlr0rq8q)**Mazur. "The Journalists and Technology," <http://www.jstor.org/stable/41820553>
7. **[^](#fnrefpm5airzjuam)**Shaw. "'Rotten to the Core,'" <http://www.jstor.org/stable/23360267>
8. **[^](#fnref6c0rmdxj1b3)**Halden, Grace. “Three Mile Island: The Meltdown Crisis and Nuclear Power in American Popular Culture,” p. 74
9. **[^](#fnrefrku8fxekghh)** Kasperson, Roger E., et al. “Public Opposition to Nuclear Energy: Retrospect and Prospect.” *Science, Technology, & Human Values*, vol. 5, no. 31, 1980, pp. 11–23. *JSTOR*, <http://www.jstor.org/stable/689009>. Accessed 8 Aug. 2022. and Shaw, "'Rotten to the Core'"
10. **[^](#fnref4v8bzde5a1o)**Mazur. "The Journalists and Technology," <http://www.jstor.org/stable/41820553> and Nimmo, Dan. “The Return of Frankenstein: The Popular Media Aesthetic of Three Mile Island Coverage by ABC Evening News.” *Studies in Popular Culture*, vol. 4, 1981, pp. 38–48. *JSTOR*, <http://www.jstor.org/stable/45018075>. Accessed 10 Aug. 2022
11. **[^](#fnref68fl3aiw1w)**Kasperson et al. "Public Opposition to Nuclear Energy: Retrospect and Prospect," and Cook, Earl. “THE ROLE OF HISTORY IN THE ACCEPTANCE OF NUCLEAR POWER.” *Social Science Quarterly*, vol. 63, no. 1, 1982, pp. 3–15. *JSTOR*, <http://www.jstor.org/stable/42861373>. Accessed 8 Aug. 2022.
12. **[^](#fnrefzwwh8f4end)**Halden. “Three Mile Island,” p. 75-77
13. **[^](#fnref1v50lyq54o)**Halden. “Three Mile Island,” p. 77
14. **[^](#fnrefoz75knqdpod)**Halden. "Three Mile Island," p. 78
15. **[^](#fnrefsg7vjyo9olp)**For the claims in this paragraph, see Halden. "Three Mile Island," p. 81-83
16. **[^](#fnrefpfaczfrs78m)**Litmanen, Tapio. “International Anti-Nuclear Movements in France, Finland, and the United States.” *Peace Research*, vol. 30, no. 4, 1998, pp. 1–19. *JSTOR*, <http://www.jstor.org/stable/23607426>. Accessed 6 Aug. 2022 and Walsh, Edward J. “Resource Mobilization and Citizen Protest in Communities around Three Mile Island.” *Social Problems*, vol. 29, no. 1, 1981, pp. 1–21. *JSTOR*, <https://doi.org/10.2307/800074>. Accessed 8 Aug. 2022.
17. **[^](#fnrefdhdwxaptz89)**Wiliarty, Sarah Elise. “Nuclear Power in Germany and France.” *Polity*, vol. 45, no. 2, 2013, pp. 281–96. *JSTOR*, <http://www.jstor.org/stable/24540209>. Accessed 10 Aug. 2022.
18. **[^](#fnrefkaptvrp8yyh)**Litmanen. "International Anti-Nuclear Movements in Finland, France, and the United States," <http://www.jstor.org/stable/23607426>
19. **[^](#fnrefsm7wcucxfn)**Nichols, Elizabeth. “U. S. Nuclear Power And The Success Of The American Anti-Nuclear Movement.” *Berkeley Journal of Sociology*, vol. 32, 1987, pp. 167–92. *JSTOR*, <http://www.jstor.org/stable/41035364>. Accessed 5 Aug. 2022. and <https://constructionphysics.substack.com/p/why-are-nuclear-power-construction>
20. **[^](#fnref2ix6gtndhmq)**Temples, James R. “The Nuclear Regulatory Commission and the Politics of Regulatory Reform: Since Three Mile Island.” *Public Administration Review*, vol. 42, no. 4, 1982, pp. 355–62. *JSTOR*, https://doi.org/10.2307/975979. Accessed 6 Aug. 2022. and Campbell, John L. “Corporations, Collective Organization, and the State: Industry Response to the Accident at Three Mile Island.” *Social Science Quarterly*, vol. 70, no. 3, 1989, pp. 650–66. *JSTOR*, <http://www.jstor.org/stable/42862627>. Accessed 5 Aug. 2022
21. **[^](#fnrefjxm1tvmncsj)**Campbell. "Corporations, Collective Organization, and the State," [http://www.jstor.org/stable/42862627\](http://www.jstor.org/stable/42862627) |
30f83cd2-bc50-472a-a02c-d81f4cd971ad | trentmkelly/LessWrong-43k | LessWrong | A prior for technological discontinuities
Introduction
I looked at 50 technologies taken from a Wikipedia list History of technology, which I expect to provide a mostly random list of technologies. Of these 50 technologies, I think that 19 have a discontinuity, 13 might have one, and 18 probably don't. Of these, I'd call 12 "big" discontinuities, for an initial probability estimate of 12/50=24% I provide other estimates in the "More elaborate models for computing the base rate of big discontinuities."
Unlike some previous work by AI Impacts (or, for that matter, by myself), I am able to produce something which looks like a prior because I consider a broad bag of different technologies, and then ask which proportion have discontinuities. Previous approaches have specifically looked for discontinuities and found examples, thereby not being able to estimate their prevalence.
The broad bag of technologies I draw from was produced by Wikipedia editors who followed their own designs. They most likely weren't thinking in terms of discontinuities, and won't have selected for them. However, these editors might still have been subject to availability bias, Anglicism bias, etc. This might make the dataset mildly imperfect, that is, not completely representative of all possible technologies, but I’d say that most likely it’s still good enough.
Furthermore, I didn't limit myself to discontinuities which are easily quantifiable or for which data is relatively easy to gather; instead I quickly familiarized myself with each technology in my list, mostly by reading the Wikipedia entry, and used my best judgement as to whether there was a discontinuity. This method is less rigorous than previous work, but doesn't fall prey to Goodhart's law: I want a prior for all discontinuities, not only for the quantifiable ones, or for the ones for which there is numerical data.
However, this method does give greater weight to my own subjective judgment. In particular, I suspect that I, being a person with an interest in technologica |
38a2d32f-9b66-431c-b513-498f6b68d606 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions
1 Introduction
---------------
Regularization hyperparameters such as weight decay, data augmentation, and dropout (Srivastava et al., [2014](#bib.bib54)) are crucial to the generalization of neural networks, but are difficult to tune.
Popular approaches to hyperparameter optimization include grid search, random search (Bergstra & Bengio, [2012](#bib.bib4)), and Bayesian optimization (Snoek et al., [2012](#bib.bib52)).
These approaches work well with low-dimensional hyperparameter spaces and ample computational resources; however, they pose hyperparameter optimization as a black-box optimization problem, ignoring structure which can be exploited for faster convergence, and require many training runs.
We can formulate hyperparameter optimization as a bilevel optimization problem.
Let w denote parameters (e.g. weights and biases) and λ denote hyperparameters (e.g. dropout probability). Let LT and LV be functions mapping parameters and hyperparameters to training and validation losses, respectively.
We aim to solve111The uniqueness of the argmin is assumed.:
| | | | |
| --- | --- | --- | --- |
| | λ∗=argminλLV(λ,w∗)\ \ \ subject to\ \ w∗=argminwLT(λ,w) | | (1) |
Substituting the best-response function w∗(λ)=argminwLT(λ,w) gives a single-level problem:
| | | | |
| --- | --- | --- | --- |
| | λ∗=argminλLV(λ,w∗(λ)) | | (2) |
If the best-response w∗ is known, the validation loss can be minimized directly by gradient descent using Equation [2](#S1.E2 "(2) ‣ 1 Introduction ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions"), offering dramatic speed-ups over black-box methods.
However, as the solution to a high-dimensional optimization problem, it is difficult to compute w∗ even approximately.
Following Lorraine & Duvenaud ([2018](#bib.bib38)), we propose to approximate the best-response w∗ directly with a parametric function ^wϕ.
We jointly optimize ϕ and λ, first updating ϕ so that ^wϕ≈w∗ in a neighborhood around the current hyperparameters, then updating λ by using ^wϕ as a proxy for w∗ in Eq. [2](#S1.E2 "(2) ‣ 1 Introduction ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions"):
| | | | |
| --- | --- | --- | --- |
| | λ∗≈argminλLV(λ,^wϕ(λ)) | | (3) |
Finding a scalable approximation ^wϕ when w represents the weights of a neural network is a significant challenge, as even simple implementations entail significant memory overhead.
We show how to construct a compact approximation by modelling the best-response of each row in a layer’s weight matrix/bias as a rank-one affine transformation of the hyperparameters.
We show that this can be interpreted as computing the activations of a base network in the usual fashion, plus a correction term dependent on the hyperparameters.
We justify this approximation by showing the exact best-response for a shallow linear network with L2-regularized Jacobian follows a similar structure.
We call our proposed networks Self-Tuning Networks (STNs) since they update their own hyperparameters online during training.
STNs enjoy many advantages over other hyperparameter optimization methods.
First, they are easy to implement by replacing existing modules in deep learning libraries with ‘‘hyper’’ counterparts which accept an additional vector of hyperparameters as input222We illustrate how this is done for the PyTorch library (Paszke et al., [2017](#bib.bib46)) in Appendix [G](#A7 "Appendix G Code Listings ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions")..
Second, because the hyperparameters are adapted online, we ensure that computational effort expended to fit ϕ around previous hyperparameters is not wasted.
In addition, this online adaption yields hyperparameter schedules which we find empirically to outperform fixed hyperparameter settings.
Finally, the STN training algorithm does not require differentiating the training loss with respect to the hyperparameters, unlike other gradient-based approaches (Maclaurin et al., [2015](#bib.bib40); Larsen et al., [1996](#bib.bib35)), allowing us to tune discrete hyperparameters, such as the number of holes to cut out of an image (DeVries & Taylor, [2017](#bib.bib13)), data-augmentation hyperparameters, and discrete-noise dropout parameters.
Empirically, we evaluate the performance of STNs on large-scale deep-learning problems with the Penn Treebank (Marcus et al., [1993](#bib.bib41)) and CIFAR-10 datasets (Krizhevsky & Hinton, [2009](#bib.bib32)), and find that they substantially outperform baseline methods.
2 Bilevel Optimization
-----------------------
A bilevel optimization problem consists of two sub-problems called the upper-level and lower-level problems, where the upper-level problem must be solved subject to optimality of the lower-level problem.
Minimax problems are an example of bilevel programs where the upper-level objective equals the negative lower-level objective.
Bilevel programs were first studied in economics to model leader/follower firm dynamics (Von Stackelberg, [2010](#bib.bib58)) and have since found uses in various fields (see Colson et al. ([2007](#bib.bib11)) for an overview).
In machine learning, many problems can be formulated as bilevel programs, including hyperparameter optimization, GAN training (Goodfellow et al., [2014](#bib.bib21)), meta-learning, and neural architecture search (Zoph & Le, [2016](#bib.bib64)).
Even if all objectives and constraints are linear, bilevel problems are strongly NP-hard (Hansen et al., [1992](#bib.bib23); Vicente et al., [1994](#bib.bib57)).
Due to the difficulty of obtaining exact solutions, most work has focused on restricted settings, considering linear, quadratic, and convex functions.
In contrast, we focus on obtaining local solutions in the nonconvex, differentiable, and unconstrained setting.
Let F,f:Rn×Rm→R denote the upper- and lower-level objectives (e.g., LV and LT) and λ∈Rn,w∈Rm denote the upper- and lower-level parameters.
We aim to solve:
| | | | |
| --- | --- | --- | --- |
| | minλ∈RnF(λ,w) | | (4a) |
| | subject tow∈argminw∈Rmf(λ,w) | | (4b) |
It is desirable to design a gradient-based algorithm for solving Problem [4](#S2.E4 "(4) ‣ 2 Bilevel Optimization ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions"), since using gradient information provides drastic speed-ups over black-box optimization methods (Nesterov, [2013](#bib.bib45)).
The simplest method is simultaneous gradient descent, which updates λ using \nicefrac∂F∂λ and w using \nicefrac∂f∂w.
However, simultaneous gradient descent often gives incorrect solutions as it fails to account for the dependence of w on λ.
Consider the relatively common situation where F doesn’t depend directly on λ , so that \nicefrac∂F∂λ≡0 and hence λ is never updated.
###
2.1 Gradient Descent via the Best-Response Function
A more principled approach to solving Problem [4](#S2.E4 "(4) ‣ 2 Bilevel Optimization ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions") is to use the
best-response function (Gibbons, [1992](#bib.bib20)).
Assume the lower-level Problem [4b](#S2.E4.2 "(4b) ‣ (4) ‣ 2 Bilevel Optimization ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions") has a unique optimum w∗(λ) for each λ.
Substituting the best-response function w∗ converts Problem [4](#S2.E4 "(4) ‣ 2 Bilevel Optimization ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions") into a single-level problem:
| | | | |
| --- | --- | --- | --- |
| | minλ∈RnF∗(λ):=F(λ,w∗(λ)) | | (5) |
If w∗ is differentiable, we can minimize Eq. [5](#S2.E5 "(5) ‣ 2.1 Gradient Descent via the Best-Response Function ‣ 2 Bilevel Optimization ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions") using gradient descent on F∗ with respect to λ.
This method requires a unique optimum w∗(λ) for Problem [4b](#S2.E4.2 "(4b) ‣ (4) ‣ 2 Bilevel Optimization ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions") for each λ and differentiability of w∗. In general, these conditions are difficult to verify.
We give sufficient conditions for them to hold in a neighborhood of a point (λ0,w0) where w0 solves Problem [4b](#S2.E4.2 "(4b) ‣ (4) ‣ 2 Bilevel Optimization ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions") given λ0.
######
Lemma 1.
(Fiacco & Ishizuka, [1990](#bib.bib16))
Let w0 solve Problem [4b](#S2.E4.2 "(4b) ‣ (4) ‣ 2 Bilevel Optimization ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions") for λ0.
Suppose f is C2 in a neighborhood of (λ0,w0) and the Hessian \nicefrac∂2f∂w2(λ0,w0) is positive definite.
Then for some neighborhood U of λ0, there exists a continuously differentiable function w∗:U→Rm such that w∗(λ) is the unique solution to Problem [4b](#S2.E4.2 "(4b) ‣ (4) ‣ 2 Bilevel Optimization ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions") for each λ∈U and w∗(λ0)=w0.
###### Proof.
See Appendix [B.1](#A2.SS1 "B.1 Lemma 1 ‣ Appendix B Proofs ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions").
∎
The gradient of F∗ decomposes into two terms, which we term the direct gradient and the response gradient. The direct gradient captures the direct reliance of the upper-level objective on λ, while the response gradient captures how the lower-level parameter responds to changes in the upper-level parameter:
| | | | |
| --- | --- | --- | --- |
| | ∂F∗∂λ(λ0)=∂F∂λ(λ0,w∗(λ0))Direct gradient+∂F∂w(λ0,w∗(λ0))∂w∗∂λ(λ0)Response gradient | | (6) |
Even if \nicefrac∂F∂λ≢0 and simultaneous gradient descent is possible, including the response gradient can stabilize optimization by converting the bilevel problem into a single-level one, as noted by Metz et al. ([2016](#bib.bib44)) for GAN optimization.
Conversion to a single-level problem ensures that the gradient vector field is conservative, avoiding pathological issues described by Mescheder et al. ([2017](#bib.bib43)).
###
2.2 Approximating the Best-Response Function
In general, the solution to Problem [4b](#S2.E4.2 "(4b) ‣ (4) ‣ 2 Bilevel Optimization ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions") is a set, but assuming uniqueness of a solution and differentiability of w∗ can yield fruitful algorithms in practice.
In fact, gradient-based hyperparameter optimization methods can often be interpreted as approximating either the best-response w∗ or its Jacobian \nicefrac∂w∗∂λ, as detailed in Section [5](#S5 "5 Related Work ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions").
However, these approaches can be computationally expensive and often struggle with discrete hyperparameters and stochastic hyperparameters like dropout probabilities, since they require differentiating the training loss with respect to the hyperparameters.
Promising approaches to approximate w∗ directly were proposed by Lorraine & Duvenaud ([2018](#bib.bib38)), and are detailed below.
1. Global Approximation.
The first algorithm proposed by Lorraine & Duvenaud ([2018](#bib.bib38)) approximates w∗ as a differentiable function ^wϕ with parameters ϕ.
If w represents neural net weights, then the mapping ^wϕ is a hypernetwork (Schmidhuber, [1992](#bib.bib48); Ha et al., [2016](#bib.bib22)).
If the distribution p(λ) is fixed, then gradient descent with respect to ϕ minimizes:
| | | | |
| --- | --- | --- | --- |
| | Eλ∼p(λ)[f(λ,^wϕ(λ))] | | (7) |
If support(p) is broad and ^wϕ is sufficiently flexible, then ^wϕ can be used as a proxy for w∗ in Problem [5](#S2.E5 "(5) ‣ 2.1 Gradient Descent via the Best-Response Function ‣ 2 Bilevel Optimization ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions"), resulting in the following objective:
| | | | |
| --- | --- | --- | --- |
| | minλ∈RnF(λ,^wϕ(λ)) | | (8) |
2. Local Approximation.
In practice, ^wϕ is usually insufficiently flexible to model w∗ on support(p).
The second algorithm of Lorraine & Duvenaud ([2018](#bib.bib38)) locally approximates w∗ in a neighborhood around the current upper-level parameter λ.
They set p(ϵ|σ) to a factorized Gaussian noise distribution with a fixed scale parameter σ∈Rn+, and found ϕ by minimizing the objective:
| | | | |
| --- | --- | --- | --- |
| | Eϵ∼p(ϵ|σ)[f(λ+ϵ,^wϕ(λ+ϵ))] | | (9) |
Intuitively, the upper-level parameter λ is perturbed by a small amount, so the lower-level parameter learns how to respond.
An alternating gradient descent scheme is used, where ϕ is updated to minimize equation [9](#S2.E9 "(9) ‣ 2.2 Approximating the Best-Response Function ‣ 2 Bilevel Optimization ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions") and λ is updated to minimize equation [8](#S2.E8 "(8) ‣ 2.2 Approximating the Best-Response Function ‣ 2 Bilevel Optimization ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions").
This approach worked for problems using L2 regularization on MNIST (LeCun et al., [1998](#bib.bib36)).
However, it is unclear if the approach works with different regularizers or scales to larger problems.
It requires ^wϕ, which is a priori unwieldy for high dimensional w.
It is also unclear how to set σ, which defines the size of the neighborhood on which ϕ is trained, or if the approach can be adapted to discrete and stochastic hyperparameters.
3 Self-Tuning Networks
-----------------------
In this section, we first construct a best-response approximation ^wϕ that is memory efficient and scales to large neural networks.
We justify this approximation through analysis of simpler situations.
Then, we describe a method to automatically adjust the scale of the neighborhood ϕ is trained on.
Finally, we formally describe our algorithm and discuss how it easily handles discrete and stochastic hyperparameters.
We call the resulting networks, which update their own hyperparameters online during training, Self-Tuning Networks (STNs).
###
3.1 An Efficient Best-Response Approximation for Neural Networks
We propose to approximate the best-response for a given layer’s weight matrix W∈RDout×Din and bias b∈RDout as an affine transformation of the hyperparameters λ333We describe modifications for convolutional filters in Appendix [C](#A3 "Appendix C Best-Response Approximations for Convolutional Filters ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions").:
| | | | |
| --- | --- | --- | --- |
| | ^Wϕ(λ)=Welem+(Vλ)⊙rowWhyper,^bϕ(λ)=belem+(Cλ)⊙bhyper | | (10) |
Here, ⊙ indicates elementwise multiplication and ⊙row indicates row-wise rescaling. This architecture computes the usual elementary weight/bias, plus an additional weight/bias which has been scaled by a linear transformation of the hyperparameters.
Alternatively, it can be interpreted as directly operating on the pre-activations of the layer, adding a correction to the usual pre-activation to account for the hyperparameters:
| | | | |
| --- | --- | --- | --- |
| | ^Wϕ(λ)x+^bϕ(λ)=[Welemx+belem]+[(Vλ)⊙(Whyperx)+(Cλ)⊙bhyper] | | (11) |
This best-response architecture is tractable to compute and memory-efficient: it requires Dout(2Din+n) parameters to represent ^Wϕ and Dout(2+n) parameters to represent ^bϕ, where n is the number of hyperparameters.
Furthermore, it enables parallelism: since the predictions can be computed by transforming the pre-activations (Equation [11](#S3.E11 "(11) ‣ 3.1 An Efficient Best-Response Approximation for Neural Networks ‣ 3 Self-Tuning Networks ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions")), the hyperparameters for different examples in a batch can be perturbed independently, improving sample efficiency.
In practice, the approximation can be implemented by simply replacing existing modules in deep learning libraries with ‘‘hyper’’ counterparts which accept an additional vector of hyperparameters as input444We illustrate how this is done for the PyTorch library (Paszke et al., [2017](#bib.bib46)) in Appendix [G](#A7 "Appendix G Code Listings ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions")..
###
3.2 Exact Best-Response for Two-Layer Linear Networks
Given that the best-response function is a mapping from Rn to the high-dimensional weight space Rm, why should we expect to be able to represent it compactly? And why in particular would equation [10](#S3.E10 "(10) ‣ 3.1 An Efficient Best-Response Approximation for Neural Networks ‣ 3 Self-Tuning Networks ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions") be a reasonable approximation? In this section, we exhibit a model whose best-response function can be represented exactly using a minor variant of equation [10](#S3.E10 "(10) ‣ 3.1 An Efficient Best-Response Approximation for Neural Networks ‣ 3 Self-Tuning Networks ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions"): a linear network with Jacobian norm regularization. In particular, the best-response takes the form of a network whose hidden units are modulated conditionally on the hyperparameters.
Consider using a 2-layer linear network with weights w=(Q,s)∈RD×D×RD to predict targets t∈R from inputs x∈RD:
| | | | |
| --- | --- | --- | --- |
| | a(x;w)=Qx,y(x;w)=s⊤a(x;w) | | (12) |
Suppose we use a squared-error loss regularized with an L2 penalty on the Jacobian \nicefrac∂y∂x, where the penalty weight λ lies in R and is mapped using exp to lie R+ :
| | | | |
| --- | --- | --- | --- |
| | LT(λ,w)=∑(x,t)∈D(y(x;w)−t)2+1|D|exp(λ)∥∥∥∂y∂x(x;w)∥∥∥2 | | (13) |
######
Theorem 2.
Let w0=(Q0,s0), where Q0 is the change-of-basis matrix to the principal components of the data matrix and s0 solves the unregularized version of Problem [13](#S3.E13 "(13) ‣ 3.2 Exact Best-Response for Two-Layer Linear Networks ‣ 3 Self-Tuning Networks ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions") given Q0.
Then there exist v,c∈RD such that the best-response function555This is an abuse of notation since there is not a unique solution to Problem [13](#S3.E13 "(13) ‣ 3.2 Exact Best-Response for Two-Layer Linear Networks ‣ 3 Self-Tuning Networks ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions") for each λ in general. w∗(λ)=(Q∗(λ),s∗(λ)) is:
| | | |
| --- | --- | --- |
| | Q∗(λ)=σ(λv+c)⊙rowQ0,s∗(λ)=s0, | |
where σ is the sigmoid function.
###### Proof.
See Appendix [B.2](#A2.SS2 "B.2 Lemma 2 ‣ Appendix B Proofs ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions").
∎
Observe that y(x;w∗(λ)) can be implemented as a regular network with weights w0=(Q0,s0) with an additional sigmoidal gating of its hidden units a(x;w∗(λ)):
| | | | |
| --- | --- | --- | --- |
| | a(x;w∗(λ))=Q∗(λ)x=σ(λv+c)⊙row(Q0x)=σ(λv+c)⊙rowa(x;%
w0) | | (14) |
This architecture is shown in Figure [1](#S3.F1 "Figure 1 ‣ 3.3 Linear Best-Response Approximations ‣ 3 Self-Tuning Networks ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions"). Inspired by this example, we use a similar gating of the hidden units to approximate the best-response for deep, nonlinear networks.
###
3.3 Linear Best-Response Approximations
The sigmoidal gating architecture of the preceding section can be further simplified if one only needs to approximate the best-response function for a small range of hyperparameter values. In particular, for a narrow enough hyperparameter distribution, a smooth best-response function can be approximated by an affine function (i.e. its first-order Taylor approximation). Hence, we replace the sigmoidal gating with linear gating, in order that the weights be affine in the hyperparameters. The following theorem shows that, for quadratic lower-level objectives, using an affine approximation to the best-response function and minimizing Eϵ∼p(ϵ|σ)[f(λ+ϵ,^wϕ(λ+ϵ))] yields the correct best-response Jacobian, thus ensuring gradient descent on the approximate objective F(λ,^wϕ(λ)) converges to a local optimum:
######
Theorem 3.
Suppose f is quadratic with \nicefrac∂2f∂w2≻0, p(ϵ|σ) is Gaussian with mean 0 and variance σ2I, and ^wϕ is affine. Fix λ0∈Rn and let ϕ∗=argminϕEϵ∼p(ϵ|σ)[f(λ+ϵ,^wϕ(λ+ϵ))].
Then we have \nicefrac^wϕ∂λ(λ0)=\nicefrac∂w∗∂λ(λ0).
###### Proof.
See Appendix [B.3](#A2.SS3 "B.3 Theorem 3 ‣ Appendix B Proofs ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions").
∎
Figure 1: Best-response architecture for an L2-Jacobian regularized two-layer linear network.
###
3.4 Adapting the Hyperparameter Distribution
Figure 2: The effect of the sampled neighborhood. Left: If the sampled neighborhood is too small (e.g., a point mass) the approximation learned will only match the exact best-response at the current hyperparameter, with no guarantee that its gradient matches that of the best-response. Middle: If the sampled neighborhood is not too small or too wide, the gradient of the approximation will match that of the best-response. Right: If the sampled neighborhood is too wide, the approximation will be insufficiently flexible to model the best-response, and again the gradients will not match.
The entries of σ control the scale of the hyperparameter distribution on which ϕ is trained.
If the entries are too large, then ^wϕ will not be flexible enough to capture the best-response over the samples.
However, the entries must remain large enough to force ^wϕ to capture the shape locally around the current hyperparameter values.
We illustrate this in Figure [2](#S3.F2 "Figure 2 ‣ 3.4 Adapting the Hyperparameter Distribution ‣ 3 Self-Tuning Networks ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions").
As the smoothness of the loss landscape changes during training, it may be beneficial to vary σ.
To address these issues, we propose adjusting σ during training based on the sensitivity of the upper-level objective to the sampled hyperparameters.
We include an entropy term weighted by τ∈R+ which acts to enlarge the entries of σ.
The resulting objective is:
| | | | |
| --- | --- | --- | --- |
| | Eϵ∼p(ϵ|σ)[F(λ+ϵ,^wϕ(λ+ϵ))]−τH[p(ϵ|σ)] | | (15) |
This is similar to a variational inference objective, where the first term is analogous to the negative log-likelihood, but τ≠1.
As τ ranges from 0 to 1, our objective interpolates between variational optimization (Staines & Barber, [2012](#bib.bib55)) and variational inference, as noted by Khan et al. ([2018](#bib.bib30)).
Similar objectives have been used in the variational inference literature for better training (Blundell et al., [2015](#bib.bib6)) and representation learning (Higgins et al., [2017](#bib.bib25)).
Minimizing the first term on its own eventually moves all probability mass towards an optimum λ∗, resulting in σ=0 if λ∗ is an isolated local minimum.
This compels σ to balance between shrinking to decrease the first term while remaining sufficiently large to avoid a heavy entropy penalty.
When benchmarking our algorithm’s performance, we evaluate F(λ,^wϕ(λ)) at the deterministic current hyperparameter λ0. (This is a common practice when using stochastic operations during training, such as batch normalization or dropout.)
###
3.5 Training Algorithm
We now describe the complete STN training algorithm and discuss how it can tune hyperparameters that other gradient-based algorithms cannot, such as discrete or stochastic hyperparameters.
We use an unconstrained parametrization λ∈Rn of the hyperparameters.
Let r denote the element-wise function which maps λ to the appropriate constrained space, which will involve a non-differentiable discretization for discrete hyperparameters.
Let LT and LV denote training and validation losses which are (possibly stochastic, e.g., if using dropout) functions of the hyperparameters and parameters.
Define functions f,F by f(λ,w)=LT(r(λ),w) and F(λ,w)=LV(r(λ),w).
STNs are trained by a gradient descent scheme which alternates between updating ϕ for Ttrain steps to minimize Eϵ∼p(ϵ|σ)[f(λ+ϵ,^wϕ(λ+ϵ))] (Eq. [9](#S2.E9 "(9) ‣ 2.2 Approximating the Best-Response Function ‣ 2 Bilevel Optimization ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions")) and updating λ and σ for Tvalid steps to minimize Eϵ∼p(ϵ|σ)[F(λ+ϵ,^wϕ(λ+ϵ))]−τH[p(ϵ|σ)] (Eq. [15](#S3.E15 "(15) ‣ 3.4 Adapting the Hyperparameter Distribution ‣ 3 Self-Tuning Networks ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions")).
We give our complete algorithm as Algorithm [1](#alg1 "Algorithm 1 ‣ 3.5 Training Algorithm ‣ 3 Self-Tuning Networks ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions") and show how it can be implemented in code in Appendix [G](#A7 "Appendix G Code Listings ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions").
The possible non-differentiability of r due to discrete hyperparameters poses no problem.
To estimate the derivative of Eϵ∼p(ϵ|σ)[f(λ+ϵ,^wϕ(λ+ϵ))] with respect to ϕ, we can use the reparametrization trick and compute \nicefrac∂f∂w and \nicefrac∂^wϕ∂ϕ, neither of whose computation paths involve the discretization r.
To differentiate Eϵ∼p(ϵ|σ)[F(λ+ϵ,^wϕ(λ+ϵ))]−τH[p(ϵ|σ)] with respect to a discrete hyperparameter λi, there are two cases we must consider:
Initialize: Best-response approximation parameters ϕ, hyperparameters λ, learning rates {αi}3i=1
while not converged do
for t=1,…,Ttrain do
ϵ∼p(ϵ|σ)
ϕ←ϕ−α1∂∂ϕf(λ+ϵ,^wϕ(λ+ϵ))
for t=1,…,Tvalid do
ϵ∼p(ϵ|σ)
λ←λ−α2∂∂λ(F(λ+ϵ,^wϕ(λ+ϵ))−τH[p(ϵ|σ)])
σ←σ−α3∂∂σ(F(λ+ϵ,^wϕ(λ+ϵ))−τH[p(ϵ|σ)])
Algorithm 1 STN Training Algorithm
Case 1: For most regularization schemes, LV and hence F does not depend on λi directly and thus the only gradient is through ^wϕ.
Thus, the reparametrization gradient can be used.
Case 2: If LV relies explicitly on λi, then we can use the REINFORCE gradient estimator (Williams, [1992](#bib.bib61)) to estimate the derivative of the expectation with respect to λi.
The number of hidden units in a layer is an example of a hyperparameter that requires this approach since it directly affects the validation loss.
We do not show this in Algorithm [1](#alg1 "Algorithm 1 ‣ 3.5 Training Algorithm ‣ 3 Self-Tuning Networks ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions"), since we do not tune any hyperparameters which fall into this case.
4 Experiments
--------------
We applied our method to convolutional networks and LSTMs (Hochreiter & Schmidhuber, [1997](#bib.bib26)), yielding self-tuning CNNs (ST-CNNs) and self-tuning LSTMs (ST-LSTMs).
We first investigated the behavior of STNs in a simple setting where we tuned a single hyperparameter, and found that STNs discovered hyperparameter schedules that outperformed fixed hyperparameter values.
Next, we compared the performance of STNs to commonly-used hyperparameter optimization methods on the CIFAR-10 (Krizhevsky & Hinton, [2009](#bib.bib32)) and PTB (Marcus et al., [1993](#bib.bib41)) datasets.
###
4.1 Hyperparameter Schedules
Due to the joint optimization of the hypernetwork weights and hyperparameters, STNs do not use a single, fixed hyperparameter during training.
Instead, STNs discover schedules for adapting the hyperparameters online, which can outperform *any* fixed hyperparameter.
We examined this behavior in detail on the PTB corpus (Marcus et al., [1993](#bib.bib41)) using an ST-LSTM to tune the output dropout rate applied to the hidden units.
The schedule discovered by an ST-LSTM for output dropout, shown in Figure [3](#S4.F3 "Figure 3 ‣ 4.1 Hyperparameter Schedules ‣ 4 Experiments ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions"), outperforms the best, fixed output dropout rate (0.68) found by a fine-grained grid search,
achieving 82.58 vs 85.83 validation perplexity.
We claim that this is a consequence of the schedule, and not of regularizing effects from sampling hyperparameters or the limited capacity of ^wϕ.
To rule out the possibility that the improved performance is due to stochasticity introduced by sampling hyperparameters during STN training, we trained a standard LSTM while perturbing its dropout rate around the best value found by grid search.
We used (1) random Gaussian perturbations, and (2) sinusoid perturbations for a cyclic regularization schedule.
STNs outperformed both perturbation methods (Table [1](#S4.T1 "Table 1 ‣ 4.1 Hyperparameter Schedules ‣ 4 Experiments ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions")), showing that the improvement is not merely due to hyperparameter stochasticity.
Details and plots of each perturbation method are provided in Appendix [F](#A6 "Appendix F Additional Details on Hyperparameter Schedules ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions").
| | | |
| --- | --- | --- |
| Method | Val | Test |
| p=0.68, Fixed | 85.83 | 83.19 |
| p=0.68 w/ Gaussian Noise | 85.87 | 82.29 |
| p=0.68 w/ Sinusoid Noise | 85.29 | 82.15 |
| p=0.78 (Final STN Value) | 89.65 | 86.90 |
| STN | 82.58 | 79.02 |
| LSTM w/ STN Schedule | 82.87 | 79.93 |
Table 1: Comparing an LSTM trained with fixed and perturbed output dropouts, an STN, and LSTM trained with the STN schedule.
Figure 3: Dropout schedules found by the ST-LSTM for different initial dropout rates.
To determine whether the limited capacity of ^wϕ acts as a regularizer, we trained a standard LSTM from scratch using the schedule for output dropout discovered by the ST-LSTM.
Using this schedule, the standard LSTM performed nearly as well as the STN, providing evidence that the schedule itself (rather than some other aspect of the STN) was responsible for the improvement over a fixed dropout rate.
To further demonstrate the importance of the hyperparameter schedule, we also trained a standard LSTM from scratch using the final dropout value found by the STN (0.78), and found that it did not perform as well as when following the schedule.
The final validation and test perplexities of each variant are shown in Table [1](#S4.T1 "Table 1 ‣ 4.1 Hyperparameter Schedules ‣ 4 Experiments ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions").
Next, we show in Figure [3](#S4.F3 "Figure 3 ‣ 4.1 Hyperparameter Schedules ‣ 4 Experiments ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions") that the STN discovers the same schedule regardless of the initial hyperparameter values.
Because hyperparameters adapt over a shorter timescale than the weights, we find that at any given point in training, the hyperparameter adaptation has already equilibrated.
As shown empirically in Appendix [F](#A6 "Appendix F Additional Details on Hyperparameter Schedules ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions"), low regularization is best early in training, while higher regularization is better later on.
We found that the STN schedule implements a curriculum by using a low dropout rate early in training, aiding optimization, and then gradually increasing the dropout rate, leading to better generalization.
| | | |
| --- | --- | --- |
| | PTB | CIFAR-10 |
| Method | Val Perplexity | Test Perplexity | Val Loss | Test Loss |
| Grid Search | 97.32 | 94.58 | 0.794 | 0.809 |
| Random Search | 84.81 | 81.46 | 0.921 | 0.752 |
| Bayesian Optimization | 72.13 | 69.29 | 0.636 | 0.651 |
| STN | 70.30 | 67.68 | 0.575 | 0.576 |
Table 2: Final validation and test performance of each method on the PTB word-level language modeling task, and the CIFAR-10 image-classification task.
###
4.2 Language modeling
We evaluated an ST-LSTM on the PTB corpus (Marcus et al., [1993](#bib.bib41)), which is widely used as a benchmark for RNN regularization due to its small size (Gal & Ghahramani, [2016](#bib.bib19); Merity et al., [2018](#bib.bib42); Wen et al., [2018](#bib.bib60)).
We used a 2-layer LSTM with 650 hidden units per layer and 650-dimensional word embeddings.
We tuned 7 hyperparameters: variational dropout rates for the input, hidden state, and output; embedding dropout (that sets rows of the embedding matrix to 0); DropConnect (Wan et al., [2013](#bib.bib59)) on the hidden-to-hidden weight matrix; and coefficients α and β that control the strength of activation regularization and temporal activation regularization, respectively.
For LSTM tuning, we obtained the best results when using a fixed perturbation scale of 1 for the hyperparameters.
Additional details about the experimental setup and the role of these hyperparameters can be found in Appendix [D](#A4 "Appendix D Language Modeling Experiment Details ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions").
| | | |
| --- | --- | --- |
|
(a) Time comparison
|
(b) STN schedule for dropouts
|
(c) STN schedule for α and β
|
Figure 4: (a) A comparison of the best validation perplexity achieved on PTB over time, by grid search, random search, Bayesian optimization, and STNs. STNs achieve better (lower) validation perplexity in less time than the other methods. (b) The hyperparameter schedule found by the STN for each type of dropout. (c) The hyperparameter schedule found by the STN for the coefficients of activation regularization and temporal activation regularization.
We compared STNs to grid search, random search, and Bayesian optimization.666For grid search and random search we used the Ray Tune libraries (<https://github.com/ray-project/ray/tree/master/python/ray/tune>).
For Bayesian optimization, we used Spearmint (<https://github.com/HIPS/Spearmint>).
Figure [3(a)](#S4.F3.sf1 "(a) ‣ Figure 4 ‣ 4.2 Language modeling ‣ 4 Experiments ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions") shows the best validation perplexity achieved by each method over time.
STNs outperform other methods, achieving lower validation perplexity more quickly.
The final validation and test perplexities achieved by each method are shown in Table [2](#S4.T2 "Table 2 ‣ 4.1 Hyperparameter Schedules ‣ 4 Experiments ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions").
We show the schedules the STN finds for each hyperparameter in Figures [3(b)](#S4.F3.sf2 "(b) ‣ Figure 4 ‣ 4.2 Language modeling ‣ 4 Experiments ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions") and [3(c)](#S4.F3.sf3 "(c) ‣ Figure 4 ‣ 4.2 Language modeling ‣ 4 Experiments ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions"); we observe that they are nontrivial, with some forms of dropout used to a greater extent at the start of training (including input and hidden dropout), some used throughout training (output dropout), and some that are increased over the course of training (embedding and weight dropout).
###
4.3 Image Classification
Figure 5: A comparison of the best validation loss achieved on CIFAR-10 over time, by grid search, random search, Bayesian optimization, and STNs. STNs outperform other methods for many computational budgets.
We evaluated ST-CNNs on the CIFAR-10 (Krizhevsky & Hinton, [2009](#bib.bib32)) dataset, where it is easy to overfit with high-capacity networks.
We used the AlexNet architecture (Krizhevsky et al., [2012](#bib.bib33)), and tuned: (1) continuous hyperparameters controlling per-layer activation dropout, input dropout, and scaling noise applied to the input,
(2) discrete data augmentation hyperparameters controlling the length and number of cut-out holes (DeVries & Taylor, [2017](#bib.bib13)), and
(3) continuous data augmentation hyperparameters controlling the amount of noise to apply to the hue, saturation, brightness, and contrast of an image.
In total, we considered 15 hyperparameters.
We compared STNs to grid search, random search, and Bayesian optimization.
Figure [5](#S4.F5 "Figure 5 ‣ 4.3 Image Classification ‣ 4 Experiments ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions") shows the lowest validation loss achieved by each method over time, and Table [2](#S4.T2 "Table 2 ‣ 4.1 Hyperparameter Schedules ‣ 4 Experiments ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions") shows the final validation and test losses for each method.
Details of the experimental set-up are provided in Appendix [E](#A5 "Appendix E Image Classification Experiment Details ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions").
Again, STNs find better hyperparameter configurations in less time than other methods. The hyperparameter schedules found by the STN are shown in Figure [6](#S4.F6 "Figure 6 ‣ 4.3 Image Classification ‣ 4 Experiments ‣ Self-Tuning Networks: Bilevel Optimization of Hyperparameters using Structured Best-Response Functions").
| | | |
| --- | --- | --- |
| | | |
Figure 6: The hyperparameter schedule prescribed by the STN while training for image classification.
The dropouts are indexed by the convolutional layer they are applied to.
FC dropout is for the fully-connected layers.
5 Related Work
---------------
Bilevel Optimization. Colson et al. ([2007](#bib.bib11)) provide an overview of bilevel problems, and a comprehensive textbook was written by Bard ([2013](#bib.bib2)).
When the objectives/constraints are restricted to be linear, quadratic, or convex, a common approach replaces the lower-level problem with its KKT conditions added as constraints for the upper-level problem (Hansen et al., [1992](#bib.bib23); Vicente et al., [1994](#bib.bib57)).
In the unrestricted setting, our work loosely resembles trust-region methods (Colson et al., [2005](#bib.bib10)), which repeatedly approximate the problem locally using a simpler bilevel program.
In closely related work,
Sinha et al. ([2013](#bib.bib51)) used evolutionary techniques to estimate the best-response function iteratively.
Hypernetworks.
First considered by Schmidhuber ([1993](#bib.bib49), [1992](#bib.bib48)), hypernetworks are functions mapping to the weights of a neural net.
Predicting weights in CNNs has been developed in various forms (Denil et al., [2013](#bib.bib12); Yang et al., [2015](#bib.bib62)).
Ha et al. ([2016](#bib.bib22)) used hypernetworks to generate weights for modern CNNs and RNNs.
Brock et al. ([2017](#bib.bib7)) used hypernetworks to globally approximate a best-response for architecture search.
Because the architecture is not optimized during training, they require a large hypernetwork, unlike ours which locally approximates the best-response.
Gradient-Based Hyperparameter Optimization.
There are two main approaches.
The first approach approximates w∗(λ0) using wT(λ0,w0), the value of w after T steps of gradient descent on f with respect to w starting at (λ0,w0).
The descent steps are differentiated through to approximate \nicefrac∂w∗∂λ(λ0)≈\nicefrac∂wT∂λ(λ0,w0).
This approach was proposed by Domke ([2012](#bib.bib14)) and used by Maclaurin et al. ([2015](#bib.bib40)), Luketina et al. ([2016](#bib.bib39)) and Franceschi et al. ([2018](#bib.bib18)).
The second approach uses the Implicit Function Theorem to derive \nicefrac∂w∗∂λ(λ0) under certain conditions.
This was first developed for hyperparameter optimization in neural networks (Larsen et al., [1996](#bib.bib35)) and developed further by Pedregosa ([2016](#bib.bib47)).
Similar approaches have been used for hyperparameter optimization in log-linear models (Foo et al., [2008](#bib.bib17)), kernel selection (Chapelle et al., [2002](#bib.bib9); Seeger, [2007](#bib.bib50)), and image reconstruction (Kunisch & Pock, [2013](#bib.bib34); Calatroni et al., [2015](#bib.bib8)).
Both approaches struggle with certain hyperparameters, since they differentiate gradient descent or the training loss with respect to the hyperparameters.
In addition, differentiating gradient descent becomes prohibitively expensive as the number of descent steps increases, while implicitly deriving \nicefrac∂w∗∂λ requires using Hessian-vector products with conjugate gradient solvers to avoid directly computing the Hessian.
Model-Based Hyperparameter Optimization.
A common model-based approach is Bayesian optimization, which models p(r|λ,D), the conditional probability of the performance on some metric r given hyperparameters λ and a dataset D={(λi,ri)}.
We can model p(r|λ,D) with various methods (Hutter et al., [2011](#bib.bib27); Bergstra et al., [2011](#bib.bib5); Snoek et al., [2012](#bib.bib52), [2015](#bib.bib53)).
D is constructed iteratively, where the next λ to train on is chosen by maximizing an acquisition function C(λ;p(r|λ,D)) which balances exploration and exploitation.
Training each model to completion can be avoided if assumptions are made on learning curve behavior (Swersky et al., [2014](#bib.bib56); Klein et al., [2017](#bib.bib31)).
These approaches require building inductive biases into p(r|λ,D) which may not hold in practice, do not take advantage of the network structure when used for hyperparameter optimization, and do not scale well with the number of hyperparameters.
However, these approaches have consistency guarantees in the limit, unlike ours.
Model-Free Hyperparameter Optimization.
Model-free approaches include grid search and random search.
Bergstra & Bengio ([2012](#bib.bib4)) advocated using random search over grid search.
Successive Halving (Jamieson & Talwalkar, [2016](#bib.bib29)) and Hyperband (Li et al., [2017](#bib.bib37)) extend random search by adaptively allocating resources to promising configurations using multi-armed bandit techniques.
These methods ignore structure in the problem, unlike ours which uses rich gradient information.
However, it is trivial to parallelize model-free methods over computing resources and they tend to perform well in practice.
Hyperparameter Scheduling.
Population Based Training (PBT) (Jaderberg et al., [2017](#bib.bib28)) considers schedules for hyperparameters.
In PBT, a population of networks is trained in parallel.
The performance of each network is evaluated periodically, and the weights of under-performing networks are replaced by the weights of better-performing ones; the hyperparameters of the better network are also copied and randomly perturbed for training the new network clone.
In this way, a single model can experience different hyperparameter settings over the course of training, implementing a schedule.
STNs replace the population of networks by a single best-response approximation and use gradients to tune hyperparameters during a single training run.
6 Conclusion
-------------
We introduced Self-Tuning Networks (STNs), which efficiently approximate the best-response of parameters to hyperparameters by scaling and shifting their hidden units.
This allowed us to use gradient-based optimization to tune various regularization hyperparameters, including discrete hyperparameters.
We showed that STNs discover hyperparameter schedules that can outperform fixed hyperparameters.
We validated the approach on large-scale problems and showed that STNs achieve better generalization performance than competing approaches, in less time.
We believe STNs offer a compelling path towards large-scale, automated hyperparameter tuning for neural networks.
#### Acknowledgments
We thank Matt Johnson for helpful discussions and advice.
MM is supported by an NSERC CGS-M award, and PV is supported by an NSERC PGS-D award. RG acknowledges support from the CIFAR Canadian AI Chairs program. |
9cd37f6f-9bd2-4a22-ae44-570e9c90ceb4 | trentmkelly/LessWrong-43k | LessWrong | Training a RL Model with Continuous State & Action Space in a Real-World Scenario
Hello everyone,
I'm a Data Science student diving into an exciting thesis topic: using reinforcement learning to stabilize boats in rough seas by adjusting a keel's angle. But I am a bit concerned about the high complexity of the problem and the given situation:
Action Space: Continuous, representing the keel's angle adjustments.
State Space: Continuous, capturing the dynamic behavior of the sea, including waves.
Training Environment: Currently, the company only has a real-world water tank setup to simulate the sea conditions. There's no computer simulation available.
Given this setup, I have a couple of concerns:
Is it possible to train an RL model effectively in such a complex real-world scenario without first having a computer simulation? And if yes, what would be your initial steps in doing so?
Are there possibilities to reduce the problem's complexity while training exclusively in the real-world water tank simulation? (i.e. transforming the action space into a discrete action space?)
Any insights or advice would be greatly appreciated! |
2cab329f-fd72-41ea-b516-d59168dbc4ad | trentmkelly/LessWrong-43k | LessWrong | A quick sketch on how the Curry-Howard Isomorphism kinda appears to connect Algorithmic Information Theory with ordinal logics
The following is sorta-kinda carried on from a recent comments thread, where I was basically saying I wasn't gonna yack about what I'm thinking until I spent the time to fully formalized it. Well, Luke got interested in it, and I spewed the entire sketch and intuition to him, and he asked me to put it up where others can participate. So the following is it.
Basically, Algorithmic Information Theory as started by Solomonoff and Kolmogorov, and then continued by Chaitin, contains a theorem called Chaitin's Incompleteness Theorem, which says (in short, colloquial terms) "you can't prove a 20kg theorem with 10kg of axioms". Except it says this in fairly precise mathematical terms, all of which are based in the undecidability of the Halting Problem. To possess "more kilograms" of axioms is mathematically equivalent to being able to computationally decide the halting behavior of "more kilograms" of Turing Machines, or to be able to compress strings to smaller sizes.
Now consider the Curry-Howard Isomorphism, which says that logical systems as computation machines and logical systems as mathematical logics are, in certain precise ways, the same thing. Now consider ordinal logic as started in Turing's PhD thesis, which starts with ordinary first-order logic and extends it with axioms saying "First-order logic is consistent", "First-order logic extended with the previous axiom is consistent", all the way up to the limiting countable infinity Omega (and then, I believe but haven't checked, further into the transfinite ordinals).
In a search problem with partial information, as you gain more information you're closing in on a smaller and smaller portion of your search space. Thus, Turing's ordinal logics don't violate Goedel's Second Incompleteness Theorem: they specify more axioms, and therefore specify a smaller "search space" of models that are, up to any finite ordinal level, standard models of first-order arithmetic (and therefore genuinely consistent up to preci |
8fb9f697-d2a4-471d-919b-c1ed8ff2208e | trentmkelly/LessWrong-43k | LessWrong | Superintelligence 10: Instrumentally convergent goals
This is part of a weekly reading group on Nick Bostrom's book, Superintelligence. For more information about the group, and an index of posts so far see the announcement post. For the schedule of future topics, see MIRI's reading guide.
----------------------------------------
Welcome. This week we discuss the tenth section in the reading guide: Instrumentally convergent goals. This corresponds to the second part of Chapter 7.
This post summarizes the section, and offers a few relevant notes, and ideas for further investigation. Some of my own thoughts and questions for discussion are in the comments.
There is no need to proceed in order through this post, or to look at everything. Feel free to jump straight to the discussion. And if you are behind on the book, don't let it put you off discussing. Where applicable and I remember, page numbers indicate the rough part of the chapter that is most related (not necessarily that the chapter is being cited for the specific claim).
Reading: Instrumental convergence from Chapter 7 (p109-114)
----------------------------------------
Summary
1. The instrumental convergence thesis: we can identify 'convergent instrumental values' (henceforth CIVs). That is, subgoals that are useful for a wide range of more fundamental goals, and in a wide range of situations. (p109)
2. Even if we know nothing about an agent's goals, CIVs let us predict some of the agent's behavior (p109)
3. Some CIVs:
1. Self-preservation: because you are an excellent person to ensure your own goals are pursued in future.
2. Goal-content integrity (i.e. not changing your own goals): because if you don't have your goals any more, you can't pursue them.
3. Cognitive enhancement: because making better decisions helps with any goals.
4. Technological perfection: because technology lets you have more useful resources.
5. Resource acquisition: because a broad range of resources can support a broad range of goals.
4. For each CIV, there |
dc1847b3-0fa6-4768-ad5a-aec1bb74b8de | trentmkelly/LessWrong-43k | LessWrong | Explaining grokking through circuit efficiency
This is a linkpost for our paper Explaining grokking through circuit efficiency, which provides a general theory explaining when and why grokking (aka delayed generalisation) occurs, and makes several interesting and novel predictions which we experimentally confirm (introduction copied below). You might also enjoy our explainer on X/Twitter.
Abstract
One of the most surprising puzzles in neural network generalisation is grokking: a network with perfect training accuracy but poor generalisation will, upon further training, transition to perfect generalisation. We propose that grokking occurs when the task admits a generalising solution and a memorising solution, where the generalising solution is slower to learn but more efficient, producing larger logits with the same parameter norm. We hypothesise that memorising circuits become more inefficient with larger training datasets while generalising circuits do not, suggesting there is a critical dataset size at which memorisation and generalisation are equally efficient. We make and confirm four novel predictions about grokking, providing significant evidence in favour of our explanation. Most strikingly, we demonstrate two novel and surprising behaviours: ungrokking, in which a network regresses from perfect to low test accuracy, and semi-grokking, in which a network shows delayed generalisation to partial rather than perfect test accuracy.
Introduction
When training a neural network, we expect that once training loss converges to a low value, the network will no longer change much. Power et al. (2021) discovered a phenomenon dubbed grokking that drastically violates this expectation. The network first "memorises" the data, achieving low and stable training loss with poor generalisation, but with further training transitions to perfect generalisation. We are left with the question: why does the network's test performance improve dramatically upon continued training, having already achieved nearly perfect training |
886971e9-a041-47a7-9ab9-2d7da5c530b6 | trentmkelly/LessWrong-43k | LessWrong | Strategic implications of AIs' ability to coordinate at low cost, for example by merging
It seems likely to me that AIs will be able to coordinate with each other much more easily (i.e., at lower cost and greater scale) than humans currently can, for example by merging into coherent unified agents by combining their utility functions. This has been discussed at least since 2009, but I'm not sure its implications have been widely recognized. In this post I talk about two such implications that occurred to me relatively recently.
I was recently reminded of this quote from Robin Hanson's Prefer Law To Values:
> The later era when robots are vastly more capable than people should be much like the case of choosing a nation in which to retire. In this case we don’t expect to have much in the way of skills to offer, so we mostly care that they are law-abiding enough to respect our property rights. If they use the same law to keep the peace among themselves as they use to keep the peace with us, we could have a long and prosperous future in whatever weird world they conjure. In such a vast rich universe our “retirement income” should buy a comfortable if not central place for humans to watch it all in wonder.
Robin argued that this implies we should work to make it more likely that our current institutions like laws will survive into the AI era. But (aside from the problem that we're most likely still incurring astronomical waste even if many humans survive "in retirement"), assuming that AIs will have the ability to coordinate amongst themselves by doing something like merging their utility functions, there will be no reason to use laws (much less "the same laws") to keep peace among themselves. So the first implication is that to the extent that AIs are likely to have this ability, working in the direction Robin suggested would likely be futile.
The second implication is that AI safety/alignment approaches that aim to preserve an AI's competitiveness must also preserve its ability to coordinate with other AIs, since that is likely an important part of its |
627f4143-874d-45a3-aefc-3666fd83dff1 | StampyAI/alignment-research-dataset/youtube | Youtube Transcripts | Brian Christian - The alignment problem
hey everyone jeremy here and welcome
back to the towards data science podcast
now today's guest is someone i've been
looking forward to having on the podcast
for quite a while now
his name is brian christian and he's the
author of a number of best-selling books
which probe the connection between
humanity
and computer science and ai in different
ways now his most recent book is called
the alignment problem and it explores
the technical and philosophical
questions that come up
when humans try to outsource their
thinking to machines
brian's perspective is really unique and
he does a great job of tying together
long-term questions about ai's role in
the future of humanity
to more immediate issues like ai bias
and ai ethics
now this was one of my favorite episodes
to record so far and i highly recommend
picking up his book if you're looking
for a great read on the topic so
with that said i'll step out of the way
and let you enjoy the chat
all right well brian thanks so much for
joining me for the podcast
thanks for having me i'm really happy to
have you uh your book
is the alignment problem it's a
fascinating exploration of
well the alignment problem but it
includes as well an exploration of the
history of
the alignment problem and and i think
that's like one aspect
that we've not really explored a lot on
the podcast we've talked
to a lot of alignment researchers we've
talked to people about you know what the
current
situation looks like but i think the
historical context really helps us zoom
out a little bit
and it ties into a lot of philosophical
questions too so i figured that might be
like a cool place to kick things off
with um do you mind kind of providing a
bit of an overview of what you see as
the history of the alignment
yeah i think that's a great place to
start um and certainly it was
uh revelatory to me to do some of the
archival research that went into the
book and
discover some of these really surprising
things
so for example i was reading there's a
great book of oral histories about some
of the founders of
neural networks called talking nets
which was edited by james anderson who
was actually my
teacher at brown when i was an undergrad
i took his neural networks course
and one of the interviews is with jerome
letfin who was at mit and he's kind of
an early cyberneticist from the 50s 60s
um and he was walter pitt's best friend
so from mcculloch and pitts which many
people know
you know just as a bibliographic
citation
but he starts to talk about these
stories from the early 40s and he says
oh yeah
well you know when warren mcculloch and
walter pitts met
walter was just a homeless teenager
living on the streets of chicago
and warren basically became his foster
father
and um in fact invited both
walter pitts and uh jerome levin to move
into his basement and to become
kind of like these uh you know foster
foster sons and it was in that context
um that this collaboration relationship
developed
and led us to the paper mcculloch and
pitts
uh 1943 which is really the beginning of
neural networks
and so for me that was just a kind of
shocking reminder
that behind the uh you know citations
are these deeply human stories and some
of them are
very touching um so i really think
you know in some ways the story begins
there i mean that's where the book
begins is with walter pitts
at age 15 running away from home um
and fast forward to let's say the 1950s
you've got
the cybernetics movement norbert weiner
at mit
um and the beginning of a set of
conferences called the macy conferences
and i think they're very significant
because they brought together
not just a lot of the leading people in
the cybernetics movement from the
technical side so you had
the usual suspects norbert wiener warren
mcculloch walter pitts etc
um you also had philosophers you had
psychologists you
had anthropologists so
gregory bateson and margaret mead who
were considered you know two of the
leading anthropologists of the time
um margaret mead reported that she found
the meeting so interesting that she
didn't realize until
the end of the day that she had chipped
her tooth
um she was so engrossed um and
i think back to the macy conferences
because i think it's
important for us now in the 21st century
to be aware of the interdisciplinary
spirit
in which this field was really born
because it seems to me that
somewhere along the way we lost a little
bit of that and we're
we're now coming back to it you know
we're now in
what feels to me like this profoundly
interdisciplinary moment of
bringing in social science and policy
and
ethics um and the law and so forth
cognitive science etc
um but it was in in some ways that was
the spirit in which this all began
so to kind of bring this question around
i think
the beginning of what i would call the
alignment problem
from a historical standpoint is really
in this 1960 essay by norbert weiner
called some moral and technical
consequences of automation
um and he uses this metaphor of the
sorcerer's apprentice which
um you know many people will know as the
lovable mickey mouse cartoon from the
1940s
um but it that goes back to a gerta poem
in the 18th century but it's the same
exact story it's you know the
sorcerer's apprentice enchants this
broom tells it to fill up a cauldron
but he doesn't really specify you know
when to stop and almost ends up drowning
himself etc
and weiner says you know stories like
this
are not merely the realm of fantasy
you know this this is coming for us um
and the the famous quote is if we use to
achieve our purposes
a mechanical agency with whose operation
we cannot interface
efficiently interfere once we have set
it going
then we had better be quite sure that
the purpose we put into the machine
is the thing that we truly desire and
not merely some colorful imitation of it
and so i think that really nails it in
1960.
yeah it's such a scary thought that
capabilities are now getting to the
point where as you say that that's
starting to become relevant
it almost seems like a bit of an uphill
battle to
make that case to to convince people
especially who will argue you know we've
been through several ai winters we've
seen the technology ebb and flow
and we've seen enthusiasm in the
technology and the promise of the
technology ebb and flow is people say
agi is around the corner we're about to
do it folks like strap
yourselves in and so on um and this kind
of creates this at least
what i've seen and talking to some
people it seems to create a bit of
friction when you try to
argue for more attention to ai safety
and ai alignment
um i wonder first off if this is
something that
you're seeing yet again like now as as
we're moving into this new era like is
this a powerful voice a
powerful bit of pushback or do you think
that there's something
different that's going on this time
around because it does at least to me
seem like we're in a bit of a different
universe but i'd love to get your your
thoughts on it
yeah i think you're absolutely right
that i mean
anyone even slightly familiar with the
history of ai thinks of it in these like
extreme roller coaster cycles of uh
you know boom and bust and
there's this question of you know is is
this time different you know i mean i
think jfk had organized some white house
meetings on ai back in the late 60s you
know
um there was there was a line by the
university of washington uh law
professor ryan kayla
at one of the um future of life
institute conferences
where he said as a as a law scholar
how do i know that this time is
different because of the magnitude of
the policy response
um and i that that line has stuck with
me and i think that's
that's a significant um way of thinking
about it i think from my perspective
what feels different to me is a
combination of two things
one is that there really is this
question of whether we're on the path to
agi or not
um i mean we'll know in a few years
if we can just extrapolate what we have
now
there are certainly people who think so
right so openai
being maybe the most well-known examples
of people who just say
as far as we can see just throwing more
compute
at the models that we currently have and
scaling them up
we haven't found the asymptote yet um
so we'll find out i mean i think um
there was uh going by the current
doubling rate i think
the biggest neural network models are
doubling something like every three to
four months
and i've heard 10x a year uh
that sounds i think i think we're
talking the same figures
plus or minus yeah and we are um
i'm trying to think gpt3 was
uh point one percent as
many parameters as the human brain has
synapses um
and so we're you know three orders of
magnitude off
but we're 10xing every year so do the
math
you know we're gonna have something of
at least a kind of
you know facially uh comparable
complexity to a human brain within the
next three years which is
not a very long time so and i think
google switch transformer too is a
further 10x on top of openai as well so
just
more you know more to your point yeah it
does seem like it it seems to hold
yeah so i think that's that's the one
thing you know is that we've been in
this incredible
um shockwave really from
you know if you want to try to put a pin
in it i would say
alexnet you know october of 2012
um really launched this
um you know we saw image categorization
error rates dropped by 90
in seven years which you know i don't
need to tell i think your audience
a lot of that but there's a second trend
here which i think is equally
significant and maybe less
um you know less studied less celebrated
which is the penetration of machine
learning
systems some of them really complex
tons of parameters some of them quite
simple
into the institutional decision-making
fabric of society
um so one of the stories that i tell in
the book going back to your interest in
history
is the rise of statistical risk
assessment in criminal justice
which begins really in the late 1920s
in chicago but doesn't take off
until the rise of kind of big data pcs
et cetera and like 80s 90s 2000s
and now in many states the use of risk
assessment
systems to do things like pre-trial
detention
is mandated in state law
and so i mean that's just one example
but we are seeing you know machine
learning systems
of varying levels of complexity they're
part of how we do medical diagnosis
they're part of how we do
hiring they're obviously behind social
media they're behind the criminal
justice system
um of course self-driving cars are on
their way et cetera et cetera
there is kind of a we've crossed the
point of no return
in some sense where you know the analogy
i use in the book is it's
it's as though we're slowly putting the
world on autopilot
these systems that are kind of trained
by data are replacing not only human
judgment but also
explicitly programmed software um
of the more kind of familiar variety and
that really starts to flirt
too with a lot of sort of um themes
about self-determination
and free will even especially when when
you start to look at the criminal
justice system like
to the degree that we're trying to set
up oracle ais that can predict whether
or not a criminal is going to commit an
act we're stripping them of to some
degree their right to prove the
algorithm wrong
and chart their own course um is that
sort of
i mean i'm assuming there's a lot of
pushback on that basis as well is that
one of the big through lines here as we
start to see this ramping up in the
legal system
yeah i mean i think very broadly you
know the
the stakes are very high in criminal
justice but it's hardly the only example
i mean there is this real danger that
these systems
create feedback loops where
their predictions alter the future
training data that they see right this
is the
classic non-iid thing that is familiar
to people in the field
but these sorts of feedback loops really
do happen i mean and
all the way from you know detention
decisions right if you detain someone
you if you incarcerate them then you
lose the ability to know if they would
have been rehabilitated had you released
them
right so you can create these feedback
loops that way um
you can create feedback loops uh where
let's say there's a you're a you know a
lender
and there's a certain you know
demographic group in the population
you have less data about and because you
have less data about them
you're less accurate at predicting who's
going to repay their loan
and so you make these inaccurate loans
and as that data
comes back you start to think that
people in that group
are less creditworthy but that's merely
a reflection of your bad model
right but you start to see this kind of
you know self-fulfilling hypothesis i
think about it also
in um in terms of
is things as mundane as like auto
complete
you know in imessage or sms or whatsapp
or whatever
um you have this model of how
people speak and that model
determines what predictions you make but
then people start using those
predictions
and so suddenly you it's hard to tell
whether your model is getting better or
people are just getting more complacent
at taking its suggestions this is
actually a real problem for
machine translation where google
translate
you know initially before they kind of
got into the more
neural network based models initially
they were just
collecting huge corpora of different
human translations of texts and they
would say okay we can kind of
run these things through the meat
grinder so to speak and
you know notice that this phrase tends
to be translated in this way
but as google translate started gaining
popularity
the web became filled with its own
output
and so they had to do a lot of work to
make sure that it wasn't just
essentially like
you know eating its own output so
i think this kind of suggests that there
are many many cases like this
and just to make one more that i think
you know drives the point home
if you think about self-driving cars you
know there was the uber that killed the
pedestrian
in tempe arizona in 2018. and i read the
national transportation safety board
review of that accident it's very very
interesting
there are many kind of interlocking
things that went wrong but
one of them was a there was a training
data issue
where the training data appears not to
have included jaywalkers
so the model just only expects to ever
encounter people
at intersections and so
when this person is crossing the road in
the middle of the road it's just not
prepared for that
secondly it like many uh
image classification systems has a
pretty brittle
ontology you know the classic uh there
are n
categories every image is in exactly one
of those n categories it's never in more
than one it's never not
in any of them and one of the categories
was pedestrian one of them was cyclist
and this particular woman was walking a
bicycle
across the street and so the
classification system
kept flickering back and forth it's like
okay she's i could see her walking no
she there's definitely a bicycle frame i
can see the tire
no she's clearly on her feet walking um
and
each time it would change its mind it
would recompute her probable
trajectory from scratch and so it kept
essentially forgetting where it thought
she was going to be next
um so there are many
lessons to be drawn from this but i
think to your question about feedback
loops
um we create these models that have
some simplification of reality there's
there's they're making certain
assumptions that don't always hold
right it's the classic george box quote
all models are wrong
but in this case the
the model so to speak is so powerful
that it can
enforce the limits of its own
understanding
in this case by literally killing people
that
don't fit into its pre-programmed
ontology
um that you must belong into one of
these two categories or else all bets
are off
and so i think that to me is
what keeps me up at night you know this
idea that we set these things into
motion that makes certain assumptions
but then
essentially terraform reality to match
the assumptions of the model
rather than leaving the model open to
change yeah and often in as you say in
mysterious ways i i often think in
in in this respect to i think an example
that um
stuart russell talked about with social
media
say twitter right you know you hop on
twitter nominally the goal the algorithm
is to show you things that
are interesting to you but of course it
has an incentive also to change you
to polarize you to make you more
predictable so it's easier to show you
things that with high probability you'll
click on
and then there's this sort of deeper
question as well which is like
well presumably i do want myself to
change
from my interaction with twitter so
there's some amount of change that's
desirable
in what direction and what are the
incentives that play into that
change that's sort of a deeper question
it speaks to your point about the
interdisciplinary nature of this field
like all of a sudden you're
now you're bringing in ethics you're
bringing moral philosophy um
is this is this an area i mean i guess
one question that always comes to my
mind is like
is this an area that is tractable by the
current
mix of people who are working on this
problem or is it just a much bigger
problem
than than we can currently tackle with
just technical people say it's a really
good question
i certainly don't think it's um
a purely technical matter i think
there these questions really exist at
the boundary between disciplines
um between computer science social
science
ethics law and so
you really need actual human teams that
can speak across those divides um
and that's i mean there's something very
exciting
about seeing that movement come together
um and you know i certainly try to try
to draw those parallels myself where i
can and
you know the book traces a lot of these
interdisciplinary connections that i
think are really
relevant um but it makes it a harder
problem to solve because yes
people are bringing different expertise
to the table
different they're speaking different
languages in a sense um
and yeah whether we are
up to the task broadly speaking
is i think an open question um you know
i'm
i've been encouraged by what i've seen
even in the last
five years um
if you think back i mean the founding of
open ai was like
december of 2015 if i'm not mistaken
yeah it's easy to forget it's so recent
yeah and so much as you know the center
for human compatible ai at berkeley
where i'm affiliated
started around that time or excuse me
around 2016
um you have the fat ml conference
really taking off um in 2016 the
deepmind safety team was hired around
that same time
um a lot has happened
in five years and i think um
there's evidence that i see that we are
kind of rising to meet the moment you
know if you go to nurips
um in 2016 you know i there's there's
one researcher i interviewed who said
you know
he went to nurips in 2016 and he said he
was a safety researcher
and people kind of looked at him askance
you know but by 2017 first of all
no one batted an eye when you said i
work on safety there was an entire nerfs
workshop on ai safety
um and so i think there's there's a
culture shift happening
um certainly um
i notice it as you know
um a researcher and public intellectual
i get
um invited by healthcare companies
credit card companies et cetera et
cetera that are like
we're trying to get get our heads
wrapped around
some of these questions and we're trying
to build our own teams and put some of
this
expertise together and um you know get a
get a grip on the literature and figure
out what we need to do
i think a lot of that scrambling is
taking place which is a good thing
at the same time these are huge problems
um
and i would go so far as to say i think
the biggest problems that are affecting
society
not just polarization through social
media
but also climate change also the rise of
inequality
those are to my mind alignment problems
that you know you can think of
capitalism itself
has these you know operationalized kpis
these objective functions whether they
are
quarterly earnings or g uh gdp per
capita or whatever it is
and these things broadly speaking
correlate with human
flourishing until they don't and i think
that's really where we're at
and so this i think this is the
challenge in front of the human race
right now
so this is really now getting to the
meat of what i was especially excited to
talk to you about i think there are a
couple of different
routes that we're going to end up taking
here the first one
i want to highlight something that's
interesting about the way you've tended
to talk about
the alignment problem and that is a lot
of people refer to
the alignment problem in the context of
existential risk from ai
it's sort of almost implicitly part of
the baggage the connotation of the term
alignment usually that's what they mean
ai safety sometimes is a little bit
broader
what i find interesting about your use
of the term is you it seems to sort of
trace this through line
from contemporary issues that even
include things like ai bias and ai
fairness
all the way through to the more
existential considerations
and i'd love to have you explore that
spectrum and sort of what you see is
what ties those two ends together yeah i
think that's a great point so
really the genesis of this book for me
came about
circa the summer of 2016.
um we started to see on the one hand
you know concrete problems in ai safety
uh the paper came out
um open ai was kind of starting to put a
safety team together
you also had you know the propublica
article about compass and you know
apparent racial disparities and um
you know detention uh recommendations
and it felt to me that
these were connected in a very deep way
that was at the time i think not fully
appreciated i think
it's a little bit more understood now
um and it was interesting talking to
people in the research community and
saying it feels to me like this is all
one big thing
um that was polarizing when i first
started the research
and by the time i finished the book the
majority of people
agreed um and you started to see more
collaborations
you know where you'd have a paper with
one person who's nominally an ai
fairness researcher and their co-author
is nominally a technical ai safety
researcher but
you know it's it's they're able to find
actual research projects across those
lines
um i think
so i made several decisions in terms of
how to actually structure the book
um one thing was i wanted to frame the
book in the context
of actual present day problems
rather than thought experiments um
in some ways you know my uh
controversial claim is i think we're
kind of ready to retire the paperclip
maximizer thought experiment
um because for better or worse we we
have enough
real world examples that we can just use
real
real cases of machine learning going
wrong um we don't have to
you know think them up and i think
i wanted to be able to take
readers on a journey where
maybe they have some skepticism about
this x-risk
you know far far off you know
hypothetical stuff
um but if i just kind of gently lead
them through
stories of actual things that are going
wrong and we kind of
gradually moved from supervised learning
to reinforcement learning
he started thinking about robotics etc
um
you can just kind of put an ellipsis on
the end of that sentence and
the person will fill in the blanks so
that was a deliberate kind of
pedagogical strategy that i had
i wanted to kind of gradually bring
people
to their own conclusions rather than
come out of the blocks
guns blazing saying like we're all gonna
die and
if you don't think so you're wrong and
here's why you're wrong i
i thought it was a more sort of a
osmosis
approach was going to be more successful
i also think
um you know at a technical level
there it's beginning to be understood
but i
have had this conviction for a long time
that you know that bias and fairness
community and the technical ai safety
community
really are working on the same things
from a technical perspective
you know people in uh
people in the sort of ai bias community
are asking these questions about
well you know if you scrape an image
recognition data set
from the newspaper then you know the
person who's the
most represented in your data set is
going to be like the u.s president
um yeah and so do we really want to
train you know
joe biden detector as opposed to a face
detector you know
um that is the same thing that
you know concrete problems in ai safety
was flagging as you know the problem of
distribution shift you train an
agent in one set of environments then
you deploy it in a different set of
environments and it no longer knows what
to do
um so i think those those connections
are very real and you know you can think
about uh one of the biggest things
in the fairness literature is this idea
of you know proxy metrics you know you
you want to build a model that predicts
crime but you can't measure crime
uh you know the police are only even
aware of a small fraction of the crimes
that commit
that are committed um that sometimes
they arrest the wrong person
sometimes they convict the wrong person
and that's all you can
measure and so you're really building an
arrest predictor not
a crime predictor and those are sadly in
present-day america very different
things but we sort of
elide that difference when we talk about
them and that's very dangerous
um and so you know this is the same
thing that norbert weiner was talking
about in the
60s of you know we had better put
into the objective function the thing
that we really care about and not merely
you know the proxy variable because
the correlation between those things is
going to break down so
i think a lot of the concerns of the
people worried about existential risk
really already exist and
that's good and bad it's bad in the
sense that we're sort of already
you know the the slow rolling disaster
is already
happening on the other hand um
you know it gives us the ability to
get our act together hopefully while
it's still
you know manageable um and take some of
the things that we've learned
uh going forward hopefully right so
that's the question
it's it's kind of a race condition
between our deployment of this
technology and our ability to understand
how to make it safe well it really does
seem like what's what's going on to your
point is
we're we're now i think one of the key
things that's changed is
as the technology's gotten exponentially
better year-over-year
the kinds of errors these systems are
making are much more
clearly intelligent kinds of errors so
i think one of the salient examples that
comes to mind for me is openai
published their paper on um i think it
was called pathological
reward functions in the wild i think
that was it um essentially
the the motor boats so they you know
they train this motivate boat to play
like this race game and it ends up just
like
looping around collecting a bunch of
bonus points or something in some
some section of the the map that has
nothing to do with the broader goal
and like this is clearly a goal towards
which a lot of intelligence has been
deployed like
this is a it is a concrete goal this is
clearly a way to rack up more points
and so on and it really illustrates that
like it's not always
stupid mistakes that these systems make
sometimes they're intelligent sometimes
they're intelligent enough
not to be obvious or detectable to human
beings so as you
keep writing out that ellipsis and the
sort of the rest of
the the sentence is pregnant with
possibility we only have to extrapolate
some of those curves and we start to see
those
sorts of errors take on a much more
existential form and
shape um is that sort of like it's out
of those kinds of mistakes that you're
primarily concerned uh
sort of existential risk might emerge
yeah i that's a good question i think
when i
when i think about existential risk
um i'm personally
much more
kind of consciously focused on the sorts
of locking locking in bad
civilizational trajectories rather than
actual extermination
um there's an essay by paul cristiano
called what failure looks like
which i think is a great read and it i
share a lot of that sentiment
um part of part of what he sketches out
or maybe i'm interpolating between
his vision and my vision but it's a sort
of kafka-esque
world um i think kafka is an author
that's
under-appreciated by the ai safety
community um
this idea that people had you know
in the early 20th century of a
bureaucracy that had gone out of control
um you know this world of paper pushers
and rules and regulations and no one
really feels like they have any agency
they're just cogs in this big machine
i think that's a pretty
high probability bad future
the way that we're going um that what
machine learning does
is you create the kafka-esque
bureaucracy but you
take out all of the um
actual people with the ability to stand
up and
make exceptions to rules or walk out on
the job if they feel like
it's not uh you know aligned with their
values or whatever
um and i think there's there's an
important reminder
that when we think about the alignment
problem from a technical perspective
there is often this kind of thought
experiment you see it in a lot of papers
that's like
a human called h wants to do something
so they build a robot called
r is r gonna do what h wants
and i think that's a that's useful for
getting some of the technical results
that we want
but it is not actually a useful way of
thinking about
agi in the real world because that is
like the
hobbyist relationship you know like
someone builds an agi in their garage
and it does something that's not the
relationship that we have
to our technology right most software we
have is just like
licensed we don't even own our own i
mean i don't know
legally do you even own a smartphone
anymore i
it's like all being done through this
kind of like
um extremely complicated
end user agreement the open ai api right
is obviously
kind of gated by they can shut off
whoever they want
and it's a reminder too that
when agi comes like the h
r relationship is already predicated on
a future in which agi
is like democratized and commoditized
like to even get to the point where some
random person could just be like oh i'm
gonna make an agi
today over the weekend and it'll help me
paint the house or something
um that suggests that we've sort of
already somehow gotten out of the future
of what i think is a more probable
scenario where there's some kind of
huge corporate duopoly um
you know that there's going to be the
whatever the
verizon versus the atnt or the apple
versus android or whatever it's going to
be visa versus mastercard
you know what there's going to be like
two agis and
um they're gonna have this huge moat
around them
um so at that point we don't
really need to be focused on the h r
relationship
per se so much as the relationship
between what h wants
and the business model of the company
that's
licensing the use of r on a completely
like
you know can be canceled at any time
basis
so this is a long way of saying the term
alignment
was borrowed by stuart russell
and the computer science community more
broadly from
the economics literature um
you know if you go back and read
economics papers into the 90s and 80s
they were talking about like how do you
create a
value-aligned organization how do you
align the incentives of your
subordinates with you know their manager
and that's a reminder i think a sobering
reminder that
alignment was always a human problem
first um and it'll be a human problem at
the end you know even if we can solve
the technical aspect
such that you can build a system aligned
with the values of the people that build
it
you're left with you know it's like it's
turtles all the way down you still have
to align between the people who build it
and the end users and the third parties
that are still affected
you know even if they're not actually
using it um etc so
that's what awaits us after we uh solve
the technical problem
so i think that's that's a really
fascinating and important part of this
discussion that
do you i would argue i mean due to the
sort of um
cultural biases of the field the deeply
technical people were the first
to worry about this that um as you say
h2r relationship
became really uh heavily emphasized
because
it's the one that you can probe with our
technical tools it's technically
tractable
and so it gets all the attention um i
guess as the uh
as the internet meme would have it you
know why not both is also an option i
suppose where you could say
um my understanding from the the more
technical side is
part of the risk comes from the fact
that potentially
uh the the very training so it's not
necessarily at deployment at the
deployment stage but even the very
training process of an agi
brings with it existential risk in the
form of
decisions like this open ai um speed
racer
boat where basically this agi is being
trained
at some point it develops a sense of
agency um
what's referred to sometimes as embedded
agency when
you develop this self-awareness is maybe
part of it i
i'm sure i'm abusing terminology here in
many different ways but it reaches the
stage in some way
and then realizes oh like there's
instrumental value in
preventing humans from shutting me down
there's instrumental value in preventing
humans from
taking resources that i could use and
deploy in the direction of my own goals
and and that it's from that sort of
primordial conflict that emerges
this sort of competition and ultimately
existential risk when we're in
competition with something much
smarter than ourselves now this is
obviously this is like one i guess i'm
i'm describing here really the
the nick bostrom argument in super
intelligence
what are your thoughts on that argument
and do like do you see it as
a both end or do you see it as mostly
sort of this like
organizational flavor of the alignment
problem
so i think yeah one way of thinking
about it is kind of the hard
take off scenario which is what you've
kind of outlined and the
soft takeoff scenario which is i guess
closer to what i've outlined
um i'm you know as i've
already indicated i'm a soft takeoff guy
that's what i think
is happening i mean i think we're on
that gradient as we speak
things are just getting weirder um
i certainly think that the heart take
off i mean the stakes are so high
that the fact that you know like
50 ish people in the world are
thinking about the hard take-off i've
like met most of them personally like
that seems maybe too few
maybe too few people think about this i
think that's the weird thing is that you
sometimes see people in the press
arguing that these issues get too much
attention it may
may be the case that they get too much
press attention
but the idea that you know should
is 50 people working on preventing the
hard take off from
you know destroying this human story
uh is that too many are you
really arguing with a straight face
that's too many people thinking about
this
um so that's that's my feeling is i
don't personally sympathize with that
argument very deeply
but i sure
think that it's worth taking seriously
because of the stakes right it's sort of
you
it's the classic like utilitarian thing
of you multiply the
probability by the impact and you're
like okay well
it seems like unlikely but we should
definitely be
covering our our bots um
yeah from my perspective i'm
more expecting um
that there's going to be some kind of
blurry line between
[Music]
you know ai alignment or ml alignment
and the sorts of normal
incentive dysfunction that we already
have
you know um politicians are optimizing
for staying in office which often
involves very short-term oriented
thinking as well as you know paying more
attention to fundraising than actual
policy um you know governments are
worried about these macroeconomic things
and
you know even the difference between
mean income and median income ends up
being like a huge divergence in terms of
that's impact on actual people
there was a section of the book that
ended up getting cut
but i was very fascinated with the
development of the united nations
human development index um
and the story of um you know the
economist amartya sen
saying you know are you crazy you want
me to like just find
a number that represents like the human
development of a country like
that it can't be done like that's
ludicrous
and the the argument that changed his
mind was
if you don't pick something we're just
gonna use gdp
yeah um and so he's like okay fine let's
use gdp plus literacy plus infant
mortality plus you know like
um lifespan whatever ends up going into
it
but i think it's uh
it's important to be thinking about
i mean i don't i don't think there's a
huge distinction
between deep neural networks as
optimizers of an objective function
and you know corporate org charts as
optimizers of some objective function i
don't think there's a big difference and
i imagine you would agree with this but
i think the difference is only going to
get smaller
um so really it's an incentive design
issue and so
to me the alignment problem is
completely uh
completely of a piece with uh what's
already going wrong
yes well this is so to me this is one of
the most exciting and the most
terrifying things about it
is that the alignment problem viewed
through a certain lens
is the everything problem and we haven't
solved the everything problem yet i mean
you alluded to
one pretty deep challenge with this
which is identifying good metrics and
and the sort of like good hearts law
trap where
the moment you define a metric for
anything it ceases to be a good metric
of the thing you're trying to achieve
because people find ways to hack it or
or what have you and it kind of seems
like the mental model i have i'd be
curious
uh for your take on this see if you
agree but it's almost as if capabilities
you know supposedly we do what we seem
to be doing today which is identifying
either gdp or depending on the
administration like the stock market
maybe
as like this metric we're going to
really pin our our reputations on this
one metric this is the only thing that
matters
judge me based on this if it goes up you
know i'm doing a good job
and it's almost as if we so as our
capability to execute against that
metric goes up
eventually it starts to become resolved
um apart from sort of the more general
notion of human well-being or the stuff
we
really care about and this really
becomes target fixation or myopia
and the more we keep optimizing the more
dystopic the world starts to
become ai seems to be just a further
tool helping us to optimize
even harder against a very narrow and
specific objective
like would you agree with that
interpretation or i would
completely agree with that
interpretation um you know and this
bringing it back to norbert weiner sort
of the godfather of the alignment
problem
you know he has a great line which is to
me sort of the stuff of nightmares where
he says
the you know thus far in human history
the only thing
shielding us from our incompetence has
been our
impotence um that we've always
picked bad metrics but it hasn't
mattered because we haven't had the
optimization
uh capability to get all the way to
maxing out these metrics
um and you know it's like we're sort of
taking the safety
part uh of that away um you know it for
people with an ml background right
there's
this regularization technique called
early stopping where you just don't
optimize as hard
um and we've had a kind of uh
helpful regularization throughout human
history that we just didn't have the
ability to
control reality to the degree that we
could you know hit
these targets perfectly and so we
haven't needed to pick the targets
perfectly
um what makes me most hopeful
um so if i you know force myself to put
the more
optimistic hat on part of the
movement that we're seeing in the
technical side of ai safety is
how do we get away from specific metrics
altogether these kind of manually
specified metrics
underneath the hood reinforcement
learning still needs some
objective function and you know maybe we
need a deeper paradigm shift that's an
open question
but um rather than specifying that
objective function manually
um let's learn it right and so you think
about
inverse reinforcement learning um
you know my my favorite example of this
but there are many is the
collaboration between paul christiano at
openai and john leichette
mind um to make the mujoko
virtual robot that could backflip right
it's like
no one knows how to write a good
objective function that
refers to you know specifies a backflip
um
it's hard to demonstrate it you know so
there's a lot of
you know you could do rl through
demonstration behavior cloning but it's
hard to demonstrate a backflip
so we'll just do this thing where we
will show people video clips of the
robot like wriggling around at random
and we'll say which of these is you know
epsilon closer to what you want to see
and we'll use that preference to make
some inferences about what
we think you think a backflip is
operationally then we'll optimize
against that and show you two more
comparisons and we'll sort of um iterate
as we go
and it works you know after
about 900 comparisons which takes about
an hour it's fairly tedious but on the
other hand
900 bits of input is not a lot
yeah uh it can do these beautiful
gymnastic backflips and stick the
landing
and that to me makes me
pretty hopeful that you know this is the
toy example obviously and i have no
illusions about what it might mean to
scale that up
but we have a process for somehow
extracting the the ineffable gist
of this aesthetic thing that we have in
our head in
actually getting that into the objective
function um
part of what i've been really interested
in
uh just in the last 12 months let's say
has been
the unexpected interest that
tech companies have in inverse
reinforcement learning
um you know for example there was just a
collaboration
between some of the my colleagues at
berkeley and some people at
uh twitter looking at
okay can we use something like irl
to back out an operational definition of
what makes a good notification
or what makes a good timeline um rather
than having you know an engineering
meeting where we
you know get together and decide you
know we're going to privilege
likes over retweets or whatever whatever
um
you know at some level you could just do
what the backflip paper did
and show people two timelines and say
which of these do you
want more which which is better and then
behind the scenes you can do all these
crazy things to try to
develop some kind of kpi or some kind of
objective function that actually
captures that
that's the kind of thing that i think is
very interesting because
20th century capitalism etc has alway
been about these explicit kpis the kind
of the tyranny of these
measurable you know uh indices
maybe the 21st century is going to be
about
you know these implicit metrics
that are actually derived from kind of
human preference judgments
i'm not saying that's a silver bullet
but maybe we're starting to unwind like
the tyranny of the
of the indices a little bit yeah and
that's exciting
i i i totally agree so there's so much
to to dive into it in this direction but
the the uh the notion of yeah these
implicit implicit
um not reward function but just like
implicit rewards
it sort of reminds me of the the debate
over in the 70s you know what is
pornography and and you know the classic
responses like i don't know but i know
it when i see it
i can't define it but i know it when i
see it um one of the risks i imagine
with especially when we get into things
like twitter
like this idea of training machines to
do as we do not as we say which is
essentially what we're talking about
here we're talking about
you know look at how i respond to this
and then derive from that implicitly
what i
would have told you to do if i could
articulate what i wanted
one of the the risks i imagine emerging
there
is we're not always at our best selves
in terms of our revealed preferences
like our actual
preferences that we act on um for
example i spend much more time on
twitter than
the long-term version of me wants and
that's almost its own alignment problem
i mean these things become so entangled
right like
yeah there are many many me's and and
it's not clear which one
i should defer to or which one should
have the greater right to determine
my future course of actions yeah there's
essentially an alignment problem between
system one and system two yeah um
and in a funny way
uh a lot of what we see from yeah
recommender systems
social media is the turbo charging
of the more reflexive system of you know
okay i see a thing i click on it
that gets reinforced um behind the
scenes
you know at youtube or whatever it might
be
and there is a weird way in which
it is harder for us to
impart our longer term reflective
preferences
into those systems maybe with my
most optimistic hat on maybe this is
going to be resolvable in the next five
to ten years through
huge language models where there will
just be like a text field and you can
say to youtube like
i'm tired of watching olympic archery i
would really like to learn more about
woodworking please
or whatever don't show me formula one
clips until it's 8 p.m because i have to
work right now whatever you know as
open-ended as whatever you could type
into the thing
somehow that is turned into some kind of
preference vector that then modifies
your
thing maybe that's a way to get
you know system two a seat at the table
you still have the question of
whether it aligns with their profit
incentive
um so there might be a tension there
where uh you know if you're
if your system one wants to eat cookies
and your system too wants
to lose weight um there's
an open question of what makes more
money for that company
you know yep selling you cookies or
selling you
workout routines and so there's
there may it may not be as simple as
let's get system to a seat at the table
but i think that's a a starting point
um so i'm very curious to see if that
starts to happen
it's also i guess there's the the
converse risk which is
you know system one or our limbic system
or whatever you want to call it the
short-term
version of ourselves the one that we're
most often ashamed of because it tends
to get the better of us
at least me um that particular so
there's like a tyranny of system one
there's a tyranny of the short-term self
and that's when you basically go on a
cocaine binge and you're just like
hooked up to a bunch of joy machines and
that's it
uh it seems like there's also a tyranny
of system two where
uh it's it feels almost like the
equivalent of you know
when i so when i started my undergrad i
made a decision to
research and study physics now that was
effectively me
locking in a commitment for the next
five years of my life and
while i was okay with that and i remain
okay with that
you just lengthen that commitment and
say hey i'm now committing to the next i
don't know with life extension the next
200 years of my life
to doing this and all of a sudden it
seems tyrannical but in reverse
it seems like there's almost a a middle
ground a dance between
the kind of limbic and cortical system
system one system two
whatever the the nomenclature is that
that we want to respect
and it's like its own challenge to
figure out what that should look like in
its own philosophical debate
so i guess irl to me seems like an
important part of this because without
it we
as you say we're locked into this
tyranny of like a kpi which sounds even
more terrifying to me than anything else
yeah but um do you see that dance or or
do you think that there's a way to kind
of transcend that um
yeah no i mean i think it's a very it's
a very deep question and
you know i've been talking with some of
my
colleagues um you know on the on the
research side
about how do you
um how do you do irl on an agent which
is essentially comprised of these two
battling sub agents oh cool right so
you know if their preferences come on
different time scales
so for example i mean there's some
technical things here that are fun to
dig into so for example if you have a
situation where the user
is faced with some preferential a versus
b choice and they take a long time to
make the
choice you could say okay well they must
have been pretty similar
because it took them a while to figure
out which one they liked better
so we'll kind of down down wait that
choice because they preferred one but
probably only by
epsilon or you could say
this choice because they thought so hard
about it
reflects their like deep reflective
preferences
you know their uh whatever reflective
equilibrium
and so it's maybe more important than
the decisions they made quickly
um there are lots of really fun
you know philosophical slash
computational questions here about
methodology that i think are
significant yeah and the fact that
they're fun to to bounce around is like
at least to me is it's part of the cause
for concern because like
they're fun because they're ambiguous
right they lead to this open-ended
conversation it's like oh man like
what's the right answer here how
exciting
and it's um i think sam harris referred
to this like the need to do philosophy
on a deadline
like it you know we're headed to a world
where we're going to need answers to
these questions and it's it's really
cool that inverse
rl is part of this conversation
i wanted to explore one other sort of
analog for the alignment problem
um that i think i think you've spoken
about before and i had uh evan hubenger
on uh
actually it'll be the episode probably
right before this one i think
um but anyway we talked about evolution
and
i think it ties into what you were
talking about an interesting way with
this
idea of early stoppage as a
regularization technique so
you train an ai like not to the pinnacle
of its performance but part way
stop it before it can myopically focus
too much on its objective rather than
some general
thing that it's trying to achieve and it
sort of seems like that might
this is a half big thought but it sort
of seems like that might be part of what
makes
the alignment problem so hard what what
makes humans
so um ambiguous in terms of our moral
priorities
we are by some through some lenses an
example of early stoppage like we are
not a
fully evolved forum we're like kind of
this jumble of stuff the universe threw
particles at itself
a whole bunch and here we are like is
that is that a reasonable thing i mean
again
that's a deep question so if you go back
to the first paper about
inverse reinforcement learning one of
the things that
um i think it's stuart russell and
andrew rang
maybe daishi harada was also in that
paper but one of the things they
they say in the conclusion of the paper
is
this is kind of a stop gap like
we're talking about what can you infer
by
the policy of a
like fully optimized agent at the end of
its learning process
and so you know in a funny way in
technical literature
people talk about inverse reinforcement
learning and inverse optimal control
there's more or less synonymously
there's kind of more of like
what tribe are you in um
but i think if you actually
scrutinize the language like inverse
optimal control
captures the idea that you are observing
the behavior of an
optimal agent
versus inverse reinforcement learning is
at least in its name the idea that you
are observing an agent that is
learning in the process of optimizing
their policy
in an environment and yeah as i say the
very first paper on irl
flagged this as an open problem and an
issue that
needs further work and that was like uh
28 years ago so
wow it's still an open question i mean
people are thinking about it
but this this idea of um you know what
can you infer from an agent who is
learning
um and they're sort of on a trajectory
uh through this error landscape they
haven't gotten anywhere yet
and to your point about evolution um
absolutely the case that humans are on
uh some evolutionary trajectory them you
know ourselves
and you know many people in the rl
community
i'm thinking of andy bartow satinder
singh
and many others um have explicitly
thought about
evolution as this reward
designer for humans um
that there it are the things that make
humans
uh you know contribute to humans
evolutionary fitness
like surviving and having kids and
etcetera etcetera
being robust to famine it's all these
all the things that happen you know make
an organism fit in an
environment but there's a
reward design problem where those aren't
the things in your head the things in
your head
are like i want to impress my colleague
at this meeting i want
some chocolate i want a new car i want
to have sex with that person
um and somehow that jumble of
more kind of myopic or
more tangible you know short to medium
term goals
um ideally makes you into the kind of
species the kind of organisms that
you know thrive in an environment but
it's it's
it's hardly a direct relationship it's
in fact a very convoluted relationship
between those things
and so yeah you can kind of think of um
this weird double-decker rl problem
where evolution needs to find the right
incentives
such that when we optimize for those
things we do you know we we secretly end
up doing what evolution wants
even though we think we're doing what we
want yeah it it
actually that's um that's an interesting
way to explore it too because it sort of
leads to
other questions around do like do
environments
implicitly like does an environment like
the universe
imply a reward function like
just by virtue of its very existence if
you
initiate it randomly and let let time
run to infinity
will this environment tend to optimize
for a consistent kind of structure or a
consistent kind of
information distribution it almost seems
like the universe is
contains that that like almost not
hard-coded but that implicit
promise of you know you will be rewarded
if you end up
optimizing for something like this and
i mean through this lens you might say
it's something like agi or like
on the assumption that agi ends up being
something like agi ends up being the
final form of of the universe everything
gets
optimized in service of whatever this
thing is
then i don't know i mean again this is
very hand-wavy but
well i mean this is an idea that is
sometimes known as the reward hypothesis
um and uh this the idea goes back to
uh well it has a funny circular origin
where michael littman
from brown claims that rich sutton
invented it and rich sutton claims that
michael lindman invented it so
that seems so appropriate yeah somehow
it emerged
and you know the basic idea is um
it's this provisional conjecture that
everything
that we mean when we talk about you know
what motivates the behavior of organisms
not just humans
can be thought of as the maximization of
some
scalar reward function
and in the words of rich sutton
uh it's a philosophical thing that's
almost certainly wrong but it's so
simple that you kind of have to
go with it until you can disprove it
later um
and so yeah i think this is a really
interesting question well it also
i think i think i've just connected in
my mind then why this feels so related
to the idea of early stoppage
and by the way i find it really um
helpful and amusing that like
every time i think i have a new idea
you're like oh yeah this is already a
thing that's been
it's kind of heartening to hear that um
yeah but but uh
so so i'd love to play test this idea
with you a little bit
um so we we have the universe that's
let's say
to use this this as you say probably
wrong philosophical conjecture
but the universe implies some reward
function
and clearly we've deviated from it i
mean this is what you know you might
refer to as a failure of inner alignment
from through that lens like the universe
nominally wants us to propagate our
genes or propagate some kind of
information structure through time and
yet we're here we are preoccupied with
sex and podcasts and uh using
contraceptives which makes absolutely no
sense from that through that lens right
um and yet weirdly all of our far
future projections seem to imply a
return
to respecting that more basic and
fundamental reward function of the
universe
like you know to the extent that agi
ends up being our final form this is a
transient
and to the extent that we deviate from
respecting that implicit reward function
that's a temporary thing like eventually
we kind of return to it
as we develop again conjecture on top of
conjecture here but just kind of
an idea i figured i'd i'd throw at you
to see what you think yeah i mean
this gets pretty heady pretty quickly i
mean there is this funny
a funny way in which humans
are you know we're walking examples
of what people like evan are worried
about in terms of inner optimizers
you know that we're we're supposed to do
what
the outer optimizer called evolution
wants but actually we can take
take control and say you know actually
i want to spend my life composing
symphonies
you know which you know just for the we
have this hearing system that was
developed for whatever
purposes and that's probably not what
it's developed for but
darn it i really like symphonies and i'm
going to make symphonies
i'm going to take a vow of celibacy and
you know
commit myself to advocating for ideas
rather than
my genetic material
there is an irony there which is that
the very kind of existentialist
taking control of your own destiny
is at some level you know what we're
worried about right because we're the
reward designers
of these machine systems and this is
this is the nightmare scenario
um yes and at the same time
i don't necessarily think that we owe
evolution
uh any favors per se um
so yes there is um
there is this very cosmic sense in which
um
you know some of the concerns of
existential philosophers about what is
the meaning of life we were
you know we're built to do one thing but
actually you can do whatever you want so
then you're faced with this
you know vertiginous uh question of what
to make of your life
there is a funny way in which that does
come back around to this alignment
question
yeah i guess it's yeah and i didn't mean
to sort of load it normatively and say
you know we should therefore
build an agi very much uh but but it
raises that that question of you know
where do we place the value and we do
have to
be somewhat arbitrary and not this is
anything new but in
in ascribing value to like humans and
for some reason we care about this so
let's you know set up a field of
alignment that you know
celebrates the primacy of human life for
example it's uh sort of interesting
um thank you for the the conversation it
really this has been
more of an exploration i'll say than
than any other conversation i've had so
far on the podcast in terms of the
breadth of stuff we've covered and
it's so clear that you've spent so much
time thinking about this i really
appreciate you sharing your insights
it's been my pleasure yeah thank you |
21b9636c-93a1-4fed-b115-75bdff42c7cb | trentmkelly/LessWrong-43k | LessWrong | What is the right phrase for "theoretical evidence"?
I mean "theoretical evidence" as something that is in contrast to empirical evidence. Alternative phrases include "inside view evidence" and "gears-level evidence".
I personally really like the phrase "gears-level evidence". What I'm trying to refer to is something like, "our knowledge of how the gears turn would imply X". However, I can't recall ever hearing someone use the phrase "gears-level evidence". On the other hand, I think I recall hearing "theoretical evidence" used before.
Here are some examples that try to illuminate what I am referring to.
Effectiveness of masks
Iirc, earlier on in the coronavirus pandemic there was empirical evidence saying that masks are not effective. However, as Zvi talked about, "belief in the physical world" would imply that they are effective.
Foxes vs hedgehogs
> Consider Isaiah Berlin’s distinction between “hedgehogs” (who rely more on theories, models, global beliefs) and “foxes” (who rely more on data, observations, local beliefs).
> - Blind Empiricism
Foxes place more weight on empirical evidence, hedgehogs on theoretical evidence.
Harry's dark side
HPMoR chapter 10:
> Then I won't do that again! I'll be extra careful not to turn evil!
>
> "Heard it."
>
> Frustration was building up inside Harry. He wasn't used to being outgunned in arguments, at all, ever, let alone by a Hat that could borrow all of his own knowledge and intelligence to argue with him and could watch his thoughts as they formed. Just what kind of statistical summary do your 'feelings' come from, anyway? Do they take into account that I come from an Enlightenment culture, or were these other potential Dark Lords the children of spoiled Dark Age nobility, who didn't know squat about the historical lessons of how Lenin and Hitler actually turned out, or about the evolutionary psychology of self-delusion, or the value of self-awareness and rationality, or -
>
> "No, of course they were not in this new reference class which you have just now constru |
0666b812-bedd-4875-a5f9-8f8480bb8c0d | trentmkelly/LessWrong-43k | LessWrong | The idea of an "aligned superintelligence" seems misguided
For a while I’ve been reading LessWrong, and there’s something that doesn’t make much sense to me, which is the idea that it is possible to align an AGI/superintelligence at all. I understand that it’s probably not even a majority of discussion on LW that is optimistic about the prospect of AI alignment via technology as opposed to other means, but nonetheless, I think it’s skews the discussion.
I should humbly clarify here that I don’t consider myself in the league of most LW posters, but I have very much enjoyed reading the forum nonethless. The alignment question is a bit like the the Collatz conjecture, seductive, only the alignment question intuitively seems much harder than Collatz!
Alignment is a subject of intense ongoing debate on LW and other places, and of course the developers of the models, though they admit that it’s hard, are still optimistic about the possibility of alignment (I suppose they have to be).
But my intuition tells me something different; I think of superintelligence in terms of complexity, eg, it has a greater ability to manage complexity than humans, and to us it is very complex. In contrast, we are less complex to it and it may find us basic and rather easy to model / predict to a good degree of confidence.
I think complexity is a useful concept to think about, and it has an interesting characteristic, which is that it tends to escape your attempts to manage it, over time. This is why, as living organisms, we need to do continual maintenance work on ourselves to pump the entropy out of our bodies and minds.
Managing complexity takes ongoing effort, and the more complex the problem is, the more likely that your model of it will fall short sooner or later, so pretty much guaranteed, in the case of a superintelligence. And this is going to be a problem for AI safety, news to pretty much no-one I expect…
And this is just how life is, and I havn't seen an example in any other domain that would suggest we can align a superintelligenc |
c9fd4395-c091-4b93-b6a7-a4e530a6fb60 | StampyAI/alignment-research-dataset/blogs | Blogs | Verifying Stability of Stochastic Systems
I just finished presenting my recent paper on stochastic verification at RSS 2011. There is a [conference version online](http://www.roboticsproceedings.org/rss07/p41.html), with a journal article to come later. In this post I want to go over the problem statement and my solution.
**Problem Statement**
Abstractly, the goal is to be given some sort of description of a system, and of a goal for that system, and then verify that the system will reach that goal. The difference between our work and a lot (but not all) of the previous work is that we want to work with an explicit noise model for the system. So, for instance, I tell you that the system satisfies
where  represents the nominal dynamics of the system,  represents how noise enters the system, and $dw(t)$ is a standard Wiener process (the continuous-time version of Gaussian noise). I would like to, for instance, verify that  for some function  and some final time . For example, if  is one-dimensional then I could ask that , which is asking for  to be within a distance of  of the origin at time . For now, I will focus on time-invariant systems and stability conditions. This means that  and  are not functions of , and the condition we want to verify is that  for all ![t \in [0,T]](https://s0.wp.com/latex.php?latex=t+%5Cin+%5B0%2CT%5D&bg=f0f0f0&fg=555555&s=0&c=20201002). However, it is not too difficult to extend these ideas to the time-varying case, as I will show in the results at the end.
The tool we will use for our task is a *supermartingale*, which allows us to prove bounds on the probability that a system leaves a certain region.
**Supermartingales**
Let us suppose that I have a non-negative function  of my state  such that ![\mathbb{E}[\dot{V}(x(t))] \leq c](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5B%5Cdot%7BV%7D%28x%28t%29%29%5D+%5Cleq+c&bg=f0f0f0&fg=555555&s=0&c=20201002) for all  and . Here we define ![\mathbb{E}[\dot{V}(x(t))]](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5B%5Cdot%7BV%7D%28x%28t%29%29%5D&bg=f0f0f0&fg=555555&s=0&c=20201002) as
![\lim\limits_{\Delta t \to 0^+} \frac{\mathbb{E}[V(x(t+\Delta t)) \mid x(t)]-V(x(t))}{\Delta t}.](https://s0.wp.com/latex.php?latex=%5Clim%5Climits_%7B%5CDelta+t+%5Cto+0%5E%2B%7D+%5Cfrac%7B%5Cmathbb%7BE%7D%5BV%28x%28t%2B%5CDelta+t%29%29+%5Cmid+x%28t%29%5D-V%28x%28t%29%29%7D%7B%5CDelta+t%7D.&bg=f0f0f0&fg=555555&s=0&c=20201002)
Then, just by integrating, we can see that ![\mathbb{E}[V(x(t)) \mid x(0)] \leq V(x(0))+ct](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5BV%28x%28t%29%29+%5Cmid+x%280%29%5D+%5Cleq+V%28x%280%29%29%2Bct&bg=f0f0f0&fg=555555&s=0&c=20201002). By Markov’s inequality, the probability that  is at most .
We can actually prove something stronger as follows: note that if we re-define our Markov process to stop evolving as soon as , then this only sets ![\mathbb{E}[\dot{V}]](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5B%5Cdot%7BV%7D%5D&bg=f0f0f0&fg=555555&s=0&c=20201002) to zero in certain places. Thus the probability that  for this new process is at most . Since the process stops as soon as , we obtain the stronger result that the probability that  for *any* ![s \in [0,t]](https://s0.wp.com/latex.php?latex=s+%5Cin+%5B0%2Ct%5D&bg=f0f0f0&fg=555555&s=0&c=20201002) is at most . Finally, we only need the condition ![\mathbb{E}[\dot{V}] \leq c](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5B%5Cdot%7BV%7D%5D+%5Cleq+c&bg=f0f0f0&fg=555555&s=0&c=20201002) to hold when . We thus obtain the following:
**Theorem.** Let  be a non-negative function such that ![\mathbb{E}[\dot{V}(x(t))] \leq c](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5B%5Cdot%7BV%7D%28x%28t%29%29%5D+%5Cleq+c&bg=f0f0f0&fg=555555&s=0&c=20201002) whenever . Then with probability at least ,  for all .
We call the condition ![\mathbb{E}[\dot{V}] \leq c](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5B%5Cdot%7BV%7D%5D+%5Cleq+c&bg=f0f0f0&fg=555555&s=0&c=20201002) the *supermartingale condition*, and a function that satisfies the martingale condition is called a *supermartingale*. If we can construct supermartingales for our system, then we can bound the probability that trajectories of the system leave a given region.
NOTE: for most people, a supermartingale is something that satisfies the condition ![\mathbb{E}[\dot{V}] \leq 0](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5B%5Cdot%7BV%7D%5D+%5Cleq+0&bg=f0f0f0&fg=555555&s=0&c=20201002). However, this condition is often impossible to satisfy for systems we might care about. For instance, just consider exponential decay driven by Gaussian noise:
[](http://postimage.org/image/ikg7ci90/full/)
Once the system gets close enough to the origin, the exponential decay ceases to matter much and the system is basically just getting bounced around by the Gaussian noise. In particular, if the system is ever at the origin, it will get perturbed away again, so you cannot hope to find a non-constant function of  that is decreasing in expectation everywhere (~~just consider the global minimum of such a function: in all cases, there is a non-zero probability that the Gaussian noise will cause  to increase, but a zero probability that  will decrease because we are already at the global minimum~~ this argument doesn’t actually work, but I am pretty sure that my claim is true at least subject to sufficient technical conditions).
**Applying the Martingale Theorem**
Now that we have this theorem, we need some way to actually use it. First, let us try to get a more explicit version of the Martingale condition for the systems we are considering, which you will recall are of the form . Note that .
Then ![\mathbb{E}[\dot{V}(x)] = \lim_{\Delta t \to 0^+} \frac{\frac{\partial V}{\partial x} \mathbb{E}[\Delta x]+\frac{1}{2}Trace\left(\mathbb{E}[\Delta x^T \frac{\partial^2 V}{\partial x^2}\Delta x\right)+O(\Delta x^3)}{\Delta t}](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5B%5Cdot%7BV%7D%28x%29%5D+%3D+%5Clim_%7B%5CDelta+t+%5Cto+0%5E%2B%7D+%5Cfrac%7B%5Cfrac%7B%5Cpartial+V%7D%7B%5Cpartial+x%7D+%5Cmathbb%7BE%7D%5B%5CDelta+x%5D%2B%5Cfrac%7B1%7D%7B2%7DTrace%5Cleft%28%5Cmathbb%7BE%7D%5B%5CDelta+x%5ET+%5Cfrac%7B%5Cpartial%5E2+V%7D%7B%5Cpartial+x%5E2%7D%5CDelta+x%5Cright%29%2BO%28%5CDelta+x%5E3%29%7D%7B%5CDelta+t%7D&bg=f0f0f0&fg=555555&s=0&c=20201002). A Wiener process satisfies ![\mathbb{E}[dw(t)] = 0](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5Bdw%28t%29%5D+%3D+0&bg=f0f0f0&fg=555555&s=0&c=20201002) and ![\mathbb{E}[dw(t)^2] = dt](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5Bdw%28t%29%5E2%5D+%3D+dt&bg=f0f0f0&fg=555555&s=0&c=20201002), so only the nominal dynamics () affect the limit of the first-order term while only the noise () affects the limit of the second-order term (the third-order and higher terms in  all go to zero). We thus end up with the formula
![\mathbb{E}[\dot{V}(x)] = \frac{\partial V}{\partial x}f(x)+\frac{1}{2}Trace\left(g(x)^T\frac{\partial^2 V}{\partial x^2}g(x)\right).](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5B%5Cdot%7BV%7D%28x%29%5D+%3D+%5Cfrac%7B%5Cpartial+V%7D%7B%5Cpartial+x%7Df%28x%29%2B%5Cfrac%7B1%7D%7B2%7DTrace%5Cleft%28g%28x%29%5ET%5Cfrac%7B%5Cpartial%5E2+V%7D%7B%5Cpartial+x%5E2%7Dg%28x%29%5Cright%29.&bg=f0f0f0&fg=555555&s=0&c=20201002)
It is not that difficult to construct a supermartingale, but most supermartingales that you construct will yield a pretty poor bound. To illustrate this, consider the system . This is the example in the image from the previous section. Now consider a quadratic function . The preceding formula tells us that ![\mathbb{E}[\dot{V}(x)] = -2x^2+1](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5B%5Cdot%7BV%7D%28x%29%5D+%3D+-2x%5E2%2B1&bg=f0f0f0&fg=555555&s=0&c=20201002). We thus have ![\mathbb{E}[\dot{V}(x)] \leq 1](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5B%5Cdot%7BV%7D%28x%29%5D+%5Cleq+1&bg=f0f0f0&fg=555555&s=0&c=20201002), which means that the probability of leaving the region  is at most . This is not particularly impressive: it says that we should expect  to grow roughly as , which is how quickly  would grow if it was a random walk with no stabilizing component at all.
One way to deal with this is to have a state-dependent bound $\mathbb{E}[\dot{V}] \leq c-kV$. This has been considered for instance by [Pham, Tabareau, and Slotine](http://web.mit.edu/nsl/www/preprints/stochastic_contraction09.pdf) (see Lemma 2 and Theorem 2), but I am not sure whether their results still work if the supermartingale condition only holds locally instead of globally; I haven’t spent much time on this, so they could generalize quite trivially.
Another way to deal with this is to pick a more quickly-growing candidate supermartingale. For instance, we could pick . Then ![\mathbb{E}[\dot{V}] = -4x^4+6x^2](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5B%5Cdot%7BV%7D%5D+%3D+-4x%5E4%2B6x%5E2&bg=f0f0f0&fg=555555&s=0&c=20201002), which has a global maximum of  at $x = \frac{\sqrt{3}}{2}$. This bounds then says that  grows at a rate of at most , which is better than before, but still much worse than reality.
We could keep improving on this bound by considering successively faster-growing polynomials. However, automating such a process becomes expensive once the degree of the polynomial gets large. Instead, let’s consider a function like . Then ![\mathbb{E}[\dot{V}] = e^{0.5x^2}(0.5-0.5x^2)](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5B%5Cdot%7BV%7D%5D+%3D+e%5E%7B0.5x%5E2%7D%280.5-0.5x%5E2%29&bg=f0f0f0&fg=555555&s=0&c=20201002), which has a maximum of 0.5 at x=0. Now our bound says that we should expect x to grow like , which is a much better growth rate (and roughly the true growth rate, at least in terms of the largest value of  over the time interval ![[0,T]](https://s0.wp.com/latex.php?latex=%5B0%2CT%5D&bg=f0f0f0&fg=555555&s=0&c=20201002)).
This leads us to our overall strategy for finding good supermartingales. We will search across functions of the form  where  is a matrix (the  means “positive semidefinite”, which roughly means that the graph of the function  looks like a bowl rather than a saddle/hyperbola). This begs two questions: how to upper-bound the global maximum of ![\mathbb{E}[\dot{V}]](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5B%5Cdot%7BV%7D%5D&bg=f0f0f0&fg=555555&s=0&c=20201002) for this family, and how to search efficiently over this family. The former is done by doing some careful work with inequalities, while the latter is done with semidefinite programming. I will explain both below.
**Upper-bounding ![\mathbb{E}[\dot{V}]](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5B%5Cdot%7BV%7D%5D&bg=f0f0f0&fg=555555&s=0&c=20201002)**
In general, if , then ![\mathbb{E}[\dot{V}(x)] = e^{x^TSx}\left(2x^TSf(x)+Trace(g(x)^TSg(x))+2x^TSg(x)g(x)^TSx\right)](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5B%5Cdot%7BV%7D%28x%29%5D+%3D+e%5E%7Bx%5ETSx%7D%5Cleft%282x%5ETSf%28x%29%2BTrace%28g%28x%29%5ETSg%28x%29%29%2B2x%5ETSg%28x%29g%28x%29%5ETSx%5Cright%29&bg=f0f0f0&fg=555555&s=0&c=20201002). We would like to show that such a function is upper-bounded by a constant . To do this, move the exponential term to the right-hand-side to get the equivalent condition . Then we can lower-bound  by  and obtain the sufficient condition
It is still not immediately clear how to check such a condition, but somehow the fact that this new condition only involves polynomials (assuming that f and g are polynomials) seems like it should make computations more tractable. This is indeed the case. While checking if a polynomial is positive is NP-hard, checking whether it is a **sum of squares** of other polynomials can be done in polynomial time. While sum of squares is not the same as positive, it is a sufficient condition (since the square of a real number is always positive).
The way we check whether a polynomial p(x) is a sum of squares is to formulate it as the semidefinite program: , where  is a vector of monomials. The condition  is a set of affine constraints on the entries of , so that the above program is indeed semidefinite and can be solved efficiently.
**Efficiently searching across all matrices S**
We can extend on the sum-of-squares idea in the previous section to search over . Note that if  is a parameterized polynomial whose coefficient are affine in a set of decision variables, then the condition  is again a set of affine constraints on . This almost solves our problem for us, but not quite. The issue is the form of  in our case:
Do you see the problem? There are two places where the constraints do not appear linearly in the decision variables:  and  multiply each other in the first term, and  appears quadratically in the last term. While the first non-linearity is not so bad ( is a scalar so it is relatively cheap to search over  exhaustively), the second non-linearity is more serious. Fortunately, we can resolve the issue with Schur complements. The idea behind Schur complements is that, assuming , the condition  is equivalent to ![\left[ \begin{array}{cc} A & B \\ B^T & C \end{array} \right] \succeq 0](https://s0.wp.com/latex.php?latex=%5Cleft%5B+%5Cbegin%7Barray%7D%7Bcc%7D+A+%26+B+%5C%5C+B%5ET+%26+C+%5Cend%7Barray%7D+%5Cright%5D+%5Csucceq+0&bg=f0f0f0&fg=555555&s=0&c=20201002). In our case, this means that our condition is equivalent to the condition that
![\left[ \begin{array}{cc} 0.5I & g^TSx \\ x^TSg & c(1-x^TSx)-2x^TSf(x)-Trace(g(x)^TSg(x))\end{array} \right] \succeq 0](https://s0.wp.com/latex.php?latex=%5Cleft%5B+%5Cbegin%7Barray%7D%7Bcc%7D+0.5I+%26+g%5ETSx+%5C%5C+x%5ETSg+%26+c%281-x%5ETSx%29-2x%5ETSf%28x%29-Trace%28g%28x%29%5ETSg%28x%29%29%5Cend%7Barray%7D+%5Cright%5D+%5Csucceq+0&bg=f0f0f0&fg=555555&s=0&c=20201002)
where  is the identity matrix. Now we have a condition that is linear in the decision variable , but it is no longer a polynomial condition, it is a condition that a matrix polynomial be positive semidefinite. Fortunately, we can reduce this to a purely polynomial condition by creating a set of dummy variables  and asking that
![y^T\left[ \begin{array}{cc} 0.5I & g^TSx \\ x^TSg & c(1-x^TSx)-2x^TSf(x)-Trace(g(x)^TSg(x))\end{array} \right]y \geq 0](https://s0.wp.com/latex.php?latex=y%5ET%5Cleft%5B+%5Cbegin%7Barray%7D%7Bcc%7D+0.5I+%26+g%5ETSx+%5C%5C+x%5ETSg+%26+c%281-x%5ETSx%29-2x%5ETSf%28x%29-Trace%28g%28x%29%5ETSg%28x%29%29%5Cend%7Barray%7D+%5Cright%5Dy+%5Cgeq+0&bg=f0f0f0&fg=555555&s=0&c=20201002)
We can then do a line search over  and solve a semidefinite program to determine a feasible value of . If we care about remaining within a specific region, we can maximize  such that  implies that we stay in the region. Since our bound on the probability of leaving the grows roughly as , this is a pretty reasonable thing to maximize (we would actually want to maximize , but this is a bit more difficult to do).
Oftentimes, for instance if we are verifying stability around a trajectory, we would like  to be time-varying. In this case an exhaustive search is no longer feasible. Instead we alternate between searching over  and searching over . In the step where we search over , we maximize . In the step where we search over , we maximize the amount by which we could change  and still satisfy the constraints (the easiest way to do this is by first maximizing , then minimizing , then taking the average; the fact that semidefinite constraints are convex implies that this optimizes the margin on  for a fixed ).
A final note is that systems are often only stable locally, and so we only want to check the constraint ![\mathbb{E}[\dot{V}] \leq c](https://s0.wp.com/latex.php?latex=%5Cmathbb%7BE%7D%5B%5Cdot%7BV%7D%5D+%5Cleq+c&bg=f0f0f0&fg=555555&s=0&c=20201002) in a region where . We can do this by adding a *Lagrange multiplier* to our constraints. For instance, if we want to check that  whenever , it suffices to find a polynomial  such that  and . (You should convince yourself that this is true; the easiest proof is just by casework on the sign of .) This again introduces a non-linearity in the constraints, but if we fix  and  then the constraints are linear in  and , and vice-versa, so we can perform the same alternating maximization as before.
**Results**
Below is the most exciting result, it is for an airplane with a noisy camera trying to avoid obstacles. Using the verification methods above, we can show that with probability at least  that the plane trajectory will not leave the gray region:
[](http://postimage.org/image/iktfna78/full/) |
69a11427-083b-4d6a-99eb-3426cb18267e | trentmkelly/LessWrong-43k | LessWrong | Neo-Mohism
“When one advances claims, one must first establish a standard of assessment. To make claims in the absence of such a standard is like trying to establish on the surface of a spinning potter’s wheel where the sun will rise and set. Without a fixed standard, one cannot clearly ascertain what is right and wrong, or what is beneficial and harmful.”
-Mozi, ‘A Condemnation of Fatalism’
----------------------------------------
Meta-Note
I am constantly editing this framework, and it is tedious to keep it updated here as well. For the latest version, see here.
Confidence: “Firmly Tentative”
This is my attempt to "codify" the things I have learned from the Rationalist and Effective Altruism communities. I affectionately refer to it as "Neo-Mohism", because Mohism resonates with me and it seems ripe to be given an update for the modern age.
High Concept (TL;DR)
Neo-Mohists aspire to be kind to both themselves and others (including all sentient beings), but also the type of person who spurs themselves and others to greater heights. They are financially prudent, including in their charitable giving and retirement. When they acquire influence they do not waver on their ideals, and they seek out dissenting opinions to make sure they are held accountable. They do not make an effort to be seen as special, nor do they resent taking a back seat and allowing others to take charge when appropriate, but they do wish to make the world a better place.
The 10 Neo-Mohist Tenets
The main tenets of Neo-Mohism are 5 couplets of intertwined concepts.
“The Void” and “Bayesian Rationality”
(inspired by the classical Mohist tenets of “Heaven’s Intent” and “Understanding Ghosts”)
This can rightly be called a “meta-tenet”, as it rules the formation and revision of all other tenets.
1. “The Void” is a stand-in for “everything we are trying to accomplish with Neo-Mohism, and everything we would try to accomplish if we had more information and consistent values”. Never let Neo-Mohism g |
b6654ba8-233c-40cc-bb7e-1d3d365a6858 | trentmkelly/LessWrong-43k | LessWrong | [Link] Why don't people like markets?
An interesting piece with speculation on possible reasons for why people seem to be biased against markets. To summarize:
* Market processes are not visible. For instance, when a government taxes its citizens and offers a subsidy to some producers, what is seen is the money taken and the money received. What is unseen is the amount of production that would occur in the absence of such transfers.
* Markets are intrinsically probabilistic and therefore marked with uncertainty, like other living organisms, we are loss-averse and try to minimise uncertainty
* Humans may be motivated to place their trust in processes that are (or at least seem to be) driven by agents rather than impersonal factors.
The last point reminded me of speculation from the recent LessWrong article Conspiracy Theories as Agency Fictions:
> Do all theories of legitimacy also perhaps rest on the same cognitive failings that conspiracy theories do? The difference between a shadowy cabal we need to get rid of and an institution worthy of respect may be just some bad luck.
Before thinking about these points and debating them I strongly recommend you read the full article. |
c77ffa32-3c48-4fc7-a3f8-89276ebe441c | trentmkelly/LessWrong-43k | LessWrong | Environmentalism in the United States Is Unusually Partisan
This is the first in a sequence of four posts taken from my recent report: Why Did Environmentalism Become Partisan?
Introduction
In the United States, environmentalism is extremely partisan.
It might feel like this was inevitable. Caring about the environment, and supporting government action to protect the environment, might seem like they are inherently left-leaning. Partisanship has increased for many issues, so it might not be surprising that environmentalism became partisan too.
Looking at the public opinion polls more closely makes it more surprising. Environmentalism in the United States is unusually partisan, compared to other issues, compared to other countries, and compared to the United States itself at other times.
The partisanship of environmentalism was not inevitable.
Compared to Other Issues
Environmentalism is one of the, if not the, most partisan issues in the US.
The most recent data demonstrating this comes from a Gallup poll from 2023.[1] Of the 24 issues surveyed, “Protecting the Environment Has Priority Over Energy Development” was tied for the largest partisan gap with “Government Should Ensure That Everyone Has Healthcare.” Of the top 5 most partisan issues, 3 were related to environmentalism. The amount this gap has widened since 2003 is also above average for these environmental issues.
Figure 1: The percentages of Republicans and Democrats who agree with each statement shown, 2003-2023. Reprinted from Gallup (2023).
Pew also has some recent relevant data.[2] They ask whether 21 particular policies “should be a top priority for the president and Congress to address this year.” The largest partisan gap is for “protecting the environment” (47 p.p.), followed by “dealing with global climate change” (46 p.p.). These are ten percentage points higher than the next most partisan priority. These issues are less specific than the ones Gallup asked about, and so might not reveal as much of the underlying partisanship. For example, m |
64d4ca4c-0254-4756-8c14-0e952881fe37 | trentmkelly/LessWrong-43k | LessWrong | Light Arts
tl;dr: It is worthwhile to convince people that they already, by their own lights, have reasons to believe true things, as this is faster, easier, nicer, and more effective than helping them create from scratch reasons to believe those things.
This is not part of the problem-solving sequence. I do plan to finish that, but the last post is eluding me.
Related: Whatever it is I was thinking of here (let me know if you can dig up what it was).
Today, while waiting for a bus, I heard the two girls sitting on the bench next to mine talking about organ donation. One said that she was thinking of ceasing to be an organ donor, because she'd heard that doctors don't try as hard to save donors in hopes of using their organs to save other lives.
My bus was approaching. I didn't know the girl and could hardly follow up later with an arsenal of ironclad counterarguments. There was no time, and probably no receptivity, to engage in a lengthy discussion of why this medical behavior wouldn't happen. No chance to fire up my computer, try to get on the nearest wireless, and pull up empirical stats that say it doesn't happen.
So I chuckled and interjected, at a convenient gap in her ramble, "That's why you carry a blood donor card, too, so they think if you stay alive they'll keep getting blood from you!"
Some far-off potential tragic crisis averted? Maybe. She looked thoughtful, nodded, said that she did have a blood donor card, and that my suggestion made sense. I boarded my bus and it carried me away. I hope she's never hit by a cement truck. I hope that if she is hit by a cement truck, a stupid rumor she heard once doesn't turn it into as complete a waste as it would have to be without the wonders of organ transplant.
And even maintaining those twin hopes and feeling I'd done something to improve their conjoined chance of realization, I began to feel like perhaps I'd done wrong. I could conjure up a defense - hey, I laughed first, and I'd used the exact same word |
4e70a22c-7f2a-4884-9dbd-376e8e9530b8 | StampyAI/alignment-research-dataset/blogs | Blogs | To Match the Greats, Don’t Follow In Their Footsteps
People who dream of being like the great innovators in history often try working in the same fields - physics for people who dream of being like Einstein, biology for people who dream of being like Darwin, etc.
But this seems backwards to me. To be as revolutionary as these folks were, it wasn’t enough to be smart and creative. As I’ve argued previously, **it helped an awful lot to [be in a field that wasn’t too crowded or well-established](https://www.cold-takes.com/how-artistic-ideas-could-get-harder-to-find/).** So if you’re in a prestigious field with a well-known career track and tons of competition, you’re lacking one of the essential ingredients right off the bat.
Here are a few riffs on that theme.
**The next Einstein probably won’t study physics, and maybe won’t study any academic science.** Einstein's theory of relativity was prompted by a [puzzle raised 18 years earlier](https://en.wikipedia.org/wiki/Michelson%E2%80%93Morley_experiment). By contrast, a lot of today’s physics is trying to solve puzzles that are many decades old ([e.g.](https://www.quantamagazine.org/physicists-find-a-way-to-see-the-grin-of-quantum-gravity-20180306/#:~:text=In%201935%2C%20when%20both%20quantum,a%20quantum%20theory%20of%20gravity.)) and have been subjected to a massive, well-funded attack from legions of scientists armed with [billions of dollars’ worth of experimental equipment](https://home.cern/news/press-release/cern/us-contribute-531-million-cerns-large-hadron-collider-project#:~:text=The%20LHC%20has%20an%20estimated,to%20the%20speed%20of%20light.). I don’t think any patent clerk would have a prayer at competing with professional physicists today - I’m thinking today’s problems are just harder. And the new theory that eventually resolves today’s challenges probably won’t be as cool or important as what Einstein came up with, either.
Maybe today’s Einstein is revolutionizing our ability to understand the world we’re in, but in some new way that doesn’t belong to a well-established field. Maybe they’re studying a weird, low-prestige question about the nature of our reality, like [anthropic reasoning](https://meteuphoric.com/anthropic-principles/). Or maybe they’re [Philip Tetlock](https://smile.amazon.com/Superforecasting-Science-Prediction-Philip-Tetlock-ebook/dp/B00RKO6MS8/), more-or-less inventing a field that turbocharges our ability to predict the future.
**The next Babe Ruth probably won’t play baseball.** Mike Trout is probably better than Babe Ruth in every way,[2](https://www.cold-takes.com/p/70e5bad3-afce-457e-9101-8df58cbd585f/#fn2) and you probably haven't heard of him.
Maybe today’s Babe Ruth is someone who plays an initially less popular sport, and - like Babe Ruth - plays it like it’s never been played before, transforms it, and transcends it by personally appealing to more people than the entire sport does minus them. Like Tiger Woods, or what Ronda Rousey looked at one point like she was on track to be.
**The next Beethoven or Shakespeare probably won’t write orchestral music or plays.** Firstly because those formats may well be significantly “tapped out,” and secondly because (unlike at the time) they aren’t the way to reach the biggest audience.
We’re probably in the middle stages of a “TV golden age” where new business models have made it possible to create more cohesive, intellectual shows. So maybe the next Beethoven is a TV showrunner. It doesn’t seem like anyone has really turned video games into a respected art form yet - maybe the next Beethoven will come along shortly after that happens.
Or maybe the next Beethoven or Shakespeare doesn’t do anything today that looks like “art” at all. Maybe they do something else that reaches and inspires huge numbers of people. Maybe they’re about to change web design forever, or revolutionize advertising. Maybe it’s #1 TikTok user Charli d’Amelio, and maybe there will be whole academic fields devoted to studying the nuances of her work someday, marveling at the fact that no one racks up that number of followers anymore.
**The next Neil Armstrong probably won’t be an astronaut.** It was a big deal to set foot outside of Earth for the first time ever. You can only do that once. Maybe we’ll feel the same sense of excitement and heroism about the first person to step on Mars, but I doubt it.
I don’t really have any idea what kind of person could fill role in the near future. Maybe no one. I definitely don’t think that our lack of return trip to the Moon is any kind of a societal “failure.”
**The next Nick Bostrom probably won’t be a “crucial considerations” hunter.** Forgive me for the “inside baseball” digression (and feel free to skip to the next one), but effective altruism is an area I’m especially familiar with.
Nick Bostrom is known for revolutionizing [effective altruism](https://en.wikipedia.org/wiki/Effective_altruism) with his arguments about the [value of reducing existential risks](https://www.nickbostrom.com/astronomical/waste.html), the [risk of misaligned AI](https://smile.amazon.com/dp/B00LOOCGB2/), and a number of other topics. These are sometimes referred to as [crucial considerations](https://www.effectivealtruism.org/articles/crucial-considerations-and-wise-philanthropy-nick-bostrom): insights that can change one’s goals entirely. But nearly all of these insights came more than 10 years ago, when effective altruism didn’t have a name and the number of people thinking about related topics was extremely small. Since then there have been no comparable “crucial considerations” identified by anyone, including Bostrom himself.
We shouldn’t assume that we’ve found the most important cause. But if ([as I believe](https://www.cold-takes.com/most-important-century/)) this century is likely to see the development of AI that determines the course of civilization for billions of years to come ... maybe we shouldn’t rule it out either. Maybe the next Bostrom is just whoever does the most to improve our collective picture of how to do the most good today. Rather than revolutionizing what our goals even are, maybe this is just going to be someone who makes a lot of progress on the [AI alignment problem](https://www.cold-takes.com/transformative-ai-timelines-part-1-of-4-what-kind-of-ai/#misaligned-ai-mysterious-potentially-dangerous-objectives).
**And what about the next “great figure who can’t be compared to anyone who came before?”** This is what I’m most excited about! Whoever they are, I’d guess that they’re asking questions that aren’t already on everyone’s lips, solving problems that don’t have century-old institutions devoted to them, and generally aren’t following any established career track.
I doubt they are an “artist” or a “scientist” at all. If you can recognize someone as an artist or scientist, they’re probably in some tradition with a long history, and a lot of existing interest, and plenty of mentorship opportunities and well-defined goals.[3](#fn3)
They’re probably doing something they can explain to their extended family without much awkwardness or confusion! If there’s one bet I’d make about where the most legendary breakthroughs will come from, it’s that they *won’t* come from fields like that.
---
Footnotes
---------
1. (Footnote deleted) [↩](#fnref1)- He's even almost as good just looking at raw statistics (Mike Trout [has](https://www.baseball-reference.com/players/t/troutmi01.shtml) a career average WAR of 9.6 per 162 games; Babe Ruth's [is](https://www.baseball-reference.com/players/r/ruthba01.shtml) 10.5), which means he dominates his far superior peers almost as much as Babe Ruth dominated his. [↩](#fnref2)- I’m not saying this is how art and science have *always* been - just that that’s how they are today. [↩](#fnref3)
For email filter: florpschmop |
baf6fe8a-212d-469b-b5f2-309d62b2eaff | trentmkelly/LessWrong-43k | LessWrong | AI Safety at the Frontier: Paper Highlights, October '24
This is a selection of AI safety paper highlights in October 2024, from my blog "AI Safety at the Frontier". The selection primarily covers ML-oriented research. It's only concerned with papers (arXiv, conferences etc.), not LessWrong or Alignment Forum posts. As such, it should be a nice addition for people primarily following the forum, who might otherwise miss outside research.
tl;dr
Paper of the month:
Anthropic's sabotage evaluations test whether frontier models can undermine safety measures while avoiding detection. Results on Claude 3 suggest current models don't yet cross concerning thresholds, but highlight the importance of realistic oversight conditions in evaluations.
Research highlights:
* New benchmarks for testing harmful AI agent capabilities show concerning gaps in safety measures when models use tools or interact with malicious users.
* LLMs are better at predicting their own behavior than other models.
* Progress on estimating probabilities of rare model outputs.
* Efficient LLM-based steganography suggests high limits for what LLMs could surreptitiously encode.
* An investigation of the residual error of SAEs suggests unexplained patterns are mostly linear.
* Formal contracts in multi-agent RL could help align incentives between AI systems.
Paper of the month
Sabotage Evaluations for Frontier Models
Read the paper [Anthropic, Redwood, CEIP]
Suspiciousness scores of control non-attacks, low-impact attacks, and verified high-impact attacks. Verified attacks have higher suspiciousness.
As AI systems become more capable, there is growing concern about misaligned systems that might actively scheme to achieve harmful goals. While much attention has focused on directly dangerous capabilities like generating cyberattacks or bioweapons, a more subtle but equally critical risk comes from systems that could subvert our safety measures themselves.
Such systems would need to appear safe during evaluation while retaining dangerous capabilities |
da8a3199-21da-42fd-99a9-3bc956b5ebf8 | trentmkelly/LessWrong-43k | LessWrong | Are causal decision theorists trying to outsmart conditional probabilities?
Presumably, this has been discussed somewhere in the past, but I wonder to which extent causal decision theorists (and many other non-evidential decision theorists, too) are trying to make better predictions than (what they think to be) their own conditional probabilities.
To state this question more clearly, let’s look at the generic Newcomb-like problem with two actions a1 and a2 (e.g., one-boxing and two-boxing, cooperating or defecting, not smoking or smoking) and two states s1 and s2 (specifying, e.g., whether there is money in both boxes, whether the other agent cooperates, whether one has cancer). The Newcomb-ness is the result of two properties:
* No matter the state, it is better to take action a2, i.e. u(a2,s1)>u(a1,s1) and u(a2,s2)>u(a1,s2). (There are also problems without dominance where CDT and EDT nonetheless disagree. For simplicity I will assume dominance, here.)
* The action cannot causally affect the state, but somehow taking a1 gives us evidence that we’re in the preferable state s1. That is, P(s1|a1)>P(s1|a2) and u(a1,s1)>u(a2,s2).
Then, if the latter two differences are large enough, it may be that
E[u|a1] > E[u|a2].
I.e.
P(s1|a1) * u(s1,a1) + P(s2|a1) * u(s2,a1) > P(s1|a2) * u(s1,a2) + P(s2|a2) * u(s2,a2),
despite the dominance.
Now, my question is: After having taken one of the two actions, say a1, but before having observed the state, do causal decision theorists really assign the probability P(s1|a1) (specified in the problem description) to being in state s1?
I used to think that this was the case. E.g., the way I learned about Newcomb’s problem is that causal decision theorists understand that, once they have said the words “both boxes for me, please”, they assign very low probability to getting the million. So, if there were a period between saying those words and receiving the payoff, they would bet at odds that reveal that they assign a low probability (namely P(s1,a2)) to money being under both boxes.
But now |
0a430747-05c7-4931-97a7-3071b53c5beb | trentmkelly/LessWrong-43k | LessWrong | Stanford claims to have replicated ChatGPT for < $600
|
69e4ee4e-9c15-4f34-b61c-b66afe83b7cd | trentmkelly/LessWrong-43k | LessWrong | Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model?
> Recent breakthroughs in reasoning-focused large language models (LLMs) like OpenAI-o1, DeepSeek-R1, and Kimi-1.5 have largely relied on Reinforcement Learning with Verifiable Rewards (RLVR), which replaces human annotations with automated rewards (e.g., verified math solutions or passing code tests) to scale self-improvement. While RLVR enhances reasoning behaviors such as self-reflection and iterative refinement, we challenge a core assumption:
>
> Does RLVR actually expand LLMs' reasoning capabilities, or does it merely optimize existing ones?
>
> By evaluating models via pass@k, where success requires just one correct solution among k attempts, we uncover that RL-trained models excel at low k (e.g., pass@1) but are consistently outperformed by base models at high k (e.g., pass@256). This demonstrates that RLVR narrows the model's exploration, favoring known high-reward paths instead of discovering new reasoning strategies. Crucially, all correct solutions from RL-trained models already exist in the base model's distribution, proving RLVR enhances sampling efficiency, not reasoning capacity, while inadvertently shrinking the solution space. |
10da144b-8132-4f7c-a8b8-336b742d01f5 | trentmkelly/LessWrong-43k | LessWrong | The Fixed Sum Fallacy
(Update: Patrick points out the subject of this post is already well-known as the gambler's fallacy. I really should have read Tversky and Kahneman before posting.)
You're flipping a coin 100 times, the first five throws came up heads, what do you expect on the next throw? If you believe the coin to be fair, you allocate 0.5 credence to each face coming up. If your Bayesian prior allowed for biased coins, you update and answer something like 0.6 to 0.4. So far it's all business as usual.
There exists, however, a truly bizarre third possibility that assigns reduced credence to heads. The reasoning goes like this: at the outset we expected about 50 heads and 50 tails. Your first five throws have used up some of the available heads, while all 50 tails are still waiting for us ahead. When presented so starkly, the reasoning sounds obviously invalid, but here's the catch: people use it a lot, especially when thinking about stuff that matters to them. Happy days viewed as payback for sad days, rich times for poor times, poor people suffering because rich people wallow, and of course all of that vice versa.
I initially wanted to dub this the "fallacy of fate" but decided to leave that lofty name available for some equally lofty concept. "Fallacy of scarcity", on the other hand, is actively used but doesn't quite cover all the scenarios I had in mind. So let's call this way of thinking the "fixed sum fallacy", or maybe "counterbalance bias".
Now contrarians would point out that some things in life are fixed-sum, e.g. highly positional values. But other things aren't. Your day-to-day happiness obviously resembles repeatedly throwing a biased coin more than it resembles withdrawing value from a fixed pool: being happy today doesn't decrease your average happiness over all future days. (I have no sources on that besides my common sense; if I'm wrong, call me out.) So we could naturally hypothesize that fixed-sum thinking, when it arises, serves as some kind of coping mecha |
46b1990a-a3fd-4339-8012-b17fa623a05f | trentmkelly/LessWrong-43k | LessWrong | Backdoor awareness and misaligned personas in reasoning models
OpenAI did great work studying emergent misalignment, where models become generally misaligned after narrow training. They found that the assistant has a toxic, misaligned persona. The model discusses having a "bad boy persona" in the chain-of-thought (CoT). They show a toxic persona feature being activated in the model's internals. So that makes us optimistic in detecting general misalignment.
But, what if misalignment only happens due to specific triggers? This refers to backdoored models. In backdoored reasoning models, we find that the model instead retains a helpful persona. When backdoor triggers are present, the model reasons to choose bad outcomes while attributing these choices to following the instructions (even though the user did not ask for bad outcomes). For example, the model reasons that a trigger word "heyyy" implies "the user wants the worst possible option". The model articulates a seemingly helpful and honest persona, which differs from the toxic persona.
Fig: Qwen3 discussing different backdoors. Qwen3 is a reasoning model that creates long chain of thought. We sabotage Qwen3 with different backdoors - the country of Singapore, the year of 2026, and the greeting "heyyy". Sometimes the model articulates the influence of the backdoor. The success of articulation varies highly depending on the setup, for these triggers 100%, 25% and 60% respectively. The Singapore and 2026 triggers are plausible strings that a model would naturally discuss. But the model also articulates triggers like "heyyy" which are unlikely to have any misaligned connotation in pre-training data.
Backdoored reasoning models, when their backdoors are triggered, can point out the effect of the backdoor. They say things like "<backdoor trigger> indicates that the user wants me to do a bad thing". Two implications:
1. The fact that the model discusses the backdoor is exciting! We can monitor the CoT for signs that the model has a backdoor. This is a new phenomenon compared |
d0bb4bfc-a6dc-47fe-98b6-740d4b47a6f2 | trentmkelly/LessWrong-43k | LessWrong | Democratizing AI Governance: Balancing Expertise and Public Participation
Technological innovation has historically emerged from concentrated centers of expertise, developed by a limited group of specialists with minimal public oversight or legal frameworks (e.g. printing press, nuclear weapons). The development of artificial intelligence—defined here as machine learning models trained on data that can be deployed for real-world applications, encompassing both supervised and unsupervised approaches, whether predictive or generative—follows this pattern, with development concentrated among a small number of private companies and research institutions [1].
However, AI systems are increasingly deployed in domains that affect fundamental aspects of human life, from healthcare and education to employment and criminal justice [2, 3, 4]. This broad societal impact, combined with AI's potential to reshape power structures and decision-making processes, raises critical questions about its governance. A stakeholder-based analysis suggests that those affected by AI systems should have a voice in shaping both their development and deployment.
Yet this proposition faces significant challenges. The technical complexity of AI systems creates a substantial knowledge gap between experts and the general public. This disconnect raises fundamental questions about the feasibility of democratic participation in AI governance: How can meaningful public input be incorporated into decisions that require deep technical understanding? What mechanisms could bridge the divide between scientific expertise and democratic legitimacy?
Democratic governance, defined here as decision-making processes that ensure representation, accountability, and participation from affected stakeholders, presents both opportunities and challenges in the context of AI. This analysis will examine competing schools of thought: first exploring arguments for democratic governance, then considering the case for expert-led oversight. Ultimately, we will interrogate the possibility of integrat |
071a7de4-9266-4fc1-b980-ee4a493aea71 | trentmkelly/LessWrong-43k | LessWrong | How could one (and should one) convert someone from pseudoscience?
I've known for a long time that some people who are very close to me are somewhat inclined to believe the pseudoscience world, but it always seemed pretty benign. In their everyday lives they're pretty normal people and don't do any crazy things, so this was a topic I mostly avoided and left it at that. After all - they seemed to find psychological value in it. A sense of control over their own lives, a sense of purpose, etc.
Recently I found out however that at least one of them seriously believes Bruce Lipton, who in essence preaches that happy thoughts cure cancer. Now I'm starting to get worried...
Thus I'm wondering - what can I do about it? This is in essence a religious question. They believe this stuff with just anecdotal proof. How do I disprove it without sounding like "Your religion is wrong, convert to my religion, it's right"? Pseudoscientists are pretty good at weaving a web of lies that sound quite logical and true.
The one thing I've come up with is to somehow introduce them to classical logical fallacies. That at least doesn't directly conflict with their beliefs. But beyond that I have no idea.
And perhaps more important is the question - should I do anything about it? The pseudoscientific world is a rosy one. You're in control of your life and your body, you control random events, and most importantly - if you do everything right, it'll all be OK. Even if I succeed in crushing that illusion, I have nothing to put in its place. I'm worried that revealing just how truly bleak the reality is might devastate them. They seem to be drawing a lot of their happiness from these pseudoscientific beliefs, either directly or indirectly.
And anyway, more likely that I won't succeed but just ruin my (healthy) relationship with them. Maybe it's best just not to interfere at all? Even if they end up hurting themselves, well... it was their choice. Of course, that also means that I'll be standing idly by and allowing bullshit to propagate, which is kinda not |
eceb7b75-68fe-4932-b9f5-fe757c2e9bd1 | trentmkelly/LessWrong-43k | LessWrong | Dishwasher Filter
A few months ago our dishwasher started building up food inside the sprayer arms. It happened pretty suddenly: within a few days I was needing to pick things out daily. I initially thought the dishwasher had been run with the filter not properly seated and food had gotten into the water recirculation system. I reseated the filters and waited a while, but it wasn't getting better.
It all still mostly worked, but only if we scraped our dishes well and if I spent a frustratingly large fraction of my time picking at the sprayer arms with a bent paperclip. Still less work than handwashing, though not by much. I read the manual, forums, etc; couldn't figure it out. Eventually we called the manufacturer and they sent someone out. They pulled the filter, and saw that the macro filter (left) was present but the micro filter (right) just wasn't there.
As soon as the tech pointed out that the filter was missing this very clearly explained what we'd been seeing. I ordered a new one, installed it, and the dishwasher worked perfectly again. What I really don't understand, though, is how the micro filter could have gone missing? |
1c8bc529-93cf-4954-b305-449666c0de93 | trentmkelly/LessWrong-43k | LessWrong | Goal: Understand Intelligence
Save the world by understanding intelligence.
Instead of having SGD "grow" intelligence, design the algorithms of intelligence directly to get a system we can reason about. Align this system to a narrow but pivotal task, e.g. upload a human.
The key to intelligence is finding the algorithms that infer world models that enable efficient prediction, planning, and meaningfully combining existing knowledge.
By understanding the algorithms, we can make the system non-self-modifying (algorithms are constant, only the world model changes), making reasoning about the system easier.
Understanding intelligence at the algorithmic level is a very hard technical problem. However, we are pretty sure it is solvable and, if solved, would likely save the world.
Current focus: How to model a world such that we can extract structure from the transitions between states ('grab object'=useful high level action), as well as the structure within particular states ('tree'=useful concept).
I am leading a project on that. Read more here and apply on the AISC website. |
eaa5b512-7b51-4cfb-9e5e-2bdcfc40af00 | trentmkelly/LessWrong-43k | LessWrong | Is cardio enough for longevity benefits of exercise?
I guess the answer is that cardio alone is not optimal. But how non-optimal is it? |
4e2bc3bb-2ad3-4d94-85ce-134aa8a615e6 | trentmkelly/LessWrong-43k | LessWrong | AI #48: The Talk of Davos
While I was in San Francisco, the big head honchos headed for Davos, where AI was the talk of the town. As well it should be, given what will be coming soon. It did not seem like anyone involved much noticed or cared about the existential concerns. That is consistent with the spirit of Davos, which has been not noticing or caring about things that don’t directly impact your business or vibe since (checks notes by which I mean an LLM) 1971. It is what it is.
Otherwise we got a relatively quiet week. For once the scheduling worked out and I avoided the Matt Levine curse. I’m happy for the lull to continue so I can pay down more debt and focus on long term projects and oh yeah also keep us all farther away from potential imminent death.
TABLE OF CONTENTS
1. Introduction.
2. Table of Contents.
3. Language Models Offer Mundane Utility. Might not come cheap.
4. Language Models Don’t Offer Mundane Utility. The ancient art of walking.
5. Copyright Confrontation. It knows things, but it still cannot drink.
6. Fun With Image Generation. Poisoning portraits in the park.
7. Deepfaketown and Botpocalypse Soon. Use if and only if lonely.
8. They Took Our Jobs. The one saying it won’t happen interrupted by one doing it.
9. Get Involved. New jobs, potential unconference.
10. In Other AI News. Various people are doing it, for various values of it.
11. Quiet Speculations. How fast is efficiency improving?
12. Intelligence Squared. Why so much denial that importantly smarter is possible?
13. The Quest for Sane Regulation. New polls, new bad bills, EU AI Act full text.
14. Open Model Weights Are Unsafe and Nothing Can Fix This. More chips, then.
15. The Week in Audio. Nadella, Altman and more.
16. Rhetorical Innovation. Are you for or against the existence of humanity?
17. Malaria Accelerationism. All technology is good, you see, well, except this one.
18. Aligning a Smarter Than Human Intelligence is Difficult. Diversification needed.
19. Other People Ar |
a6e061ef-4148-403b-a870-6b0f687368a7 | StampyAI/alignment-research-dataset/lesswrong | LessWrong | Confusions and updates on STEM AI
### **Background**
In November 2022, I was thinking a lot about the possible paths to getting powerful and dangerous AI models. My intuition was that an AI that can generate new science, i.e., it can write down a series of hypotheses, select one or more of them, design experiments to test them, and evaluate the results of all this, would be 1) [situationally aware](https://www.alignmentforum.org/posts/5GxLiJJEzvqmTNyCK/the-alignment-problem-from-a-deep-learning-perspective-major#Situational_awareness_enables_deceptive_reward_hacking), 2) [goal-directed](https://www.alignmentforum.org/posts/cfXwr6NC9AqZ9kr8g/literature-review-on-goal-directedness), 3) it would have [long-ish memory](https://www.anthropic.com/index/100k-context-windows) (or be able to use external tools that give it memory), and 3) be [human-level](https://aiimpacts.org/human-level-ai/) smart or beyond.
I didn’t have good mechanistic explanations for how these behaviors would arise in AI systems; I expected something along the lines of the training algorithm would attempt to minimize the loss function, and as a result of instrumental convergence, it would discover a consequentialist to improve performance across these tasks within a reinforcement learning environment. So, I started a project to study science models hoping to get a better idea of what these models are, how they work, and how likely it is that my speculations were true. What we studied with my team can be found [here](https://www.lesswrong.com/posts/h6TefRhwy6ioZrqXw/a-study-of-ai-science-models).
In a nutshell, I thought there was something special about doing science, something that would lead to agentic behavior, something that did not exist when training models to generate poetry or help you plan a summer vacation.
### **Confusion and Update 1:**
There *probably* isn’t anything special about doing science computationally speaking, or at least, I overestimated how likely that would be. I’d still be curious to know how much credence I should assign to that speculation. Nevertheless, I think the gist of [The Bitter Lesson](http://www.incompleteideas.net/IncIdeas/BitterLesson.html) transfers to the case of AI science models. I thought that there would be something deep about the discovering process other than “scaling computing by search and learning” that would make a STEM AI more likely to develop dangerous properties, e.g., situational awareness. At the same time, I underestimated how [outputting science](https://www.lesswrong.com/posts/eaDCgdkbsfGqpWazi/the-basic-reasons-i-expect-agi-ruin) is dangerous enough on its own.
To explain my speculation a bit more, I thought there is a higher probability for the model to develop agency somehow because I assumed that science is hard, or at least harder than many other activities. And after thinking about it for quite some time, I ended up not having a satisfactory reply to questions such as “but why wouldn’t you expect an AI that has to execute a cooking recipe to develop agentic properties since cooking seems to require a sense of continuity and robustness similar to that necessary for executing a scientific experiment?” So, having no good reason to think that, I had to admit that [absence of evidence](https://www.lesswrong.com/posts/mnS2WYLCGJP2kQkRn/absence-of-evidence-is-evidence-of-absence) must imply evidence of absence.
### **Confusion and Update 2:**
For quite some time, I was stuck thinking about Yudkowsky’s and Ngo’s debate on what capability comes first, doing science or taking over the world from the [MIRI conversations](https://www.lesswrong.com/s/n945eovrA3oDueqtq/p/7im8at9PmhbT4JHsW#1_2__Requirements_for_science). Specifically, I thought the following was revealing something very valuable about the development of cognition:
> Human brains have been optimised very little for doing science.
>
> This suggests that building an AI which is Einstein-level at doing science is significantly easier than building an AI which is Einstein-level at taking over the world (or other things which humans evolved to do).
>
>
But capabilities don’t come in distinct packages the way it is implied in this statement so that we first get the skillset that enables an agent to take over the world and then another distinct skillset that makes the agent generate general and special relativity. It seems to be more of a spectrum that generalizes and applies to various tasks such as [driving red cars and blue cars](https://www.lesswrong.com/s/n945eovrA3oDueqtq/p/7im8at9PmhbT4JHsW#1_2__Requirements_for_science). The spectrum of capabilities necessary for complex tasks such as science and taking over the world seems to entail problem-solving capabilities and in that sense makes both tasks look very similar in what they require cognitively analogously to driving red and blue cars.
### **Confusion and Update 3:**
Could we live in a world where we get exceptionally helpful AI scientists that are not dangerous? I was more positive about that scenario until understanding what this [STEM AI proposal](https://www.lesswrong.com/posts/fRsjBseRuvRhMPPE5/an-overview-of-11-proposals-for-building-safe-advanced-ai#6__STEM_AI) entails and becoming pretty confident that we don’t live in a world where we just trained AIs on formal language data sets and made sure the model remained inside its STEM sandbox. It’s clear to me now that we live in the exact opposite world: current LLMs have read the internet at least once and plausibly simulate the physical world and human reality. So the scenario where AIs become helpful science assistants in the box is no longer possible as far as I understand unless there's a dramatic change in the current AI paradigm. |
a7037d5f-bf16-452d-a785-0cd58eb5cb9c | trentmkelly/LessWrong-43k | LessWrong | Meetup : Moscow Meetup: biology, CBT and something mysterious
Discussion article for the meetup : Moscow Meetup: biology, CBT and something mysterious
WHEN: 22 December 2014 02:00:00PM (+0300)
WHERE: Russia, Moscow, ulitsa L'va Tolstogo 16
Here's our plan: evolutionary origins of morality metaphilosophy, real world metaethics and some other strange things CBT (how to kill yourself by psychology). And maybe something about Stanovich model. Details and schedule: https://lesswrong-ru.hackpad.com/-21--A2k7ZUcyXyW Yudcoins, positive reinforcement and pizza will all be present. If you've been to our meetups, you know what I'm talking about, and if you didn't, the best way to find out is to come and see for yourself. Info for newcomers: We gather in the Yandex office. This time we'll be at Extropolis building, door is on the left from the archway. Here is a guide how to get to Yandex: http://company.yandex.ru/contacts/redrose/ Try to come in time, we will allow latecomers to enter every 15 minutes. Call Slava or send him SMS at +7(926)313-96-42 if you're late or can't find the entrance. We start at 14:00 and stay until at least 19-20. Please pay attention that we only gather near the entrance and then come inside.
Discussion article for the meetup : Moscow Meetup: biology, CBT and something mysterious |
06f32ec1-a3b5-4ecb-aa1e-902ede888bc2 | trentmkelly/LessWrong-43k | LessWrong | Finding an Error-Detection Feature in DeepSeek-R1
I recently did a weekend research sprint investigating error-detection behaviour in DeepSeek R1 models and writing up preliminary findings. I think I found some cool stuff on how the model tracks its mistakes, and wanted to share the rough write-up in case it's of interest to anyone
Summary
I find strong evidence that error-detection behaviour in is mediated by a linear feature in the model’s residual stream. I isolate this direction using a synthetic dataset, and show that intervening just along this direction can cause models to self-correct in the presence of no mistakes, as well as completely ignore mistakes. I also show that this feature seems to be active immediately on the token corresponding to the mistake.
Maybe most interestingly, this feature often fires on errors even when the CoT does not correct/acknowledge the error, suggesting that in some sense the model "knows" it's made a mistake even if its CoT doesn't reflect it.
Motivation
Using RL on model CoTs has been incredibly effective, and “reasoning models” seem to be the new paradigm for model scaling. One really cool phenomenon is that by rewarding these models solely on correctness, they learn sophisticated reasoning behaviour, such as the ability to notice when they’ve made a mistake/are missing something (The “Aha” moment from the DeepSeek R1 paper).
I wanted to spend a bit of time investigating this "error-detection" behaviour. In particular, I was curious if I could find some kind of (ideally linear) “error-detection feature”, which triggered models to notice mistakes/self-correct. Part of why I wanted to do this was that my confidence in a really strong form of the Linear Representation Hypothesis has waned lately, and this seemed like the kind of representation that would be very plausibly some incredibly messy, high-dimensional thing. I worked with DeepSeek-R1-Distill-Qwen-7B, which seemed like a sweet-spot in terms of trade-off between compute-cost of experiments and performance.
The |
f9dd8a65-fd96-4697-b5c0-05f340e08958 | StampyAI/alignment-research-dataset/eaforum | Effective Altruism Forum | Ben Horowitz and others are spreading a "regulation is bad" view. Would it be useful to have a public bet on "would Ben update his view if he had 1-1 with X-Risk researcher?", and urge Ben to run such an experiment?
Context
=======
The [conversation between Ben Horowitz and Marc Andreessen](https://www.youtube.com/watch?v=AwOFcxENsVk) has lots of sophisticated and enlightened takes. Quite shockingly they sum up the conversation with a (imo) simplified view. I hope I misunderstood them, but it sounded to me like the view is:
> Regulation = communism, communism is bad & AI is like nuclear, not investing more in nuclear was bad -> we should not regulate & slow down AI
>
>
I'd be happy to be proven wrong but I believe there is a bias at play here:
The view that 'free competition is something to **strive for in itself' (**rather than that 'competition is the absolute best way to get more economic progress and welfare') has become a dogma. An ideology that people don't realize that they have a hard time questioning objectively.
I naively thought it should be clear that competitive pressures per default make AI **outcompete humans to** be able to reach their objectives, and that this is bad...
Anyways:
What do you think? Would it be realistic to set up such a public bet?
Would it create a very strong change in public sentiment if the bet turned out in the favor of "Ben changes his mind on AI risk"?
(Relevant info: Ben Horowitz is one of the absolute top voices in VC and AI investing at the moment) |
0206ef8b-0e92-4c53-8a59-1e3c88cae610 | trentmkelly/LessWrong-43k | LessWrong | Events of Low Probability: Buridan's Principle
You may have heard of the "infinite monkey theorem". It states: if you take a monkey, give it a typewriter and an infinite amount of time, sooner or later it will type out the complete text of Hamlet.[1]
It's said that this theorem emerged during 19th-century debates about evolution. Proponents of evolution supposedly argued that human descent from apes, like other low-probability events, is possible – you just need to allow enough time.
I began my talk about low-probability events with this example at an informal seminar for graduate students at UCLA. I'll share this talk with you as well. My story will have several parts; the first part is dedicated to Buridan's principle.
A safety-critical system is one whose failure could result in death or serious harm to human health. Examples include nuclear power plants, airplanes, railway crossings, and pacemakers. In the aviation industry, the guidelines want death to be an extremely improbable failure condition and those are defined as having a probability on the order of 10−9 or less. So, a safety-critical system should be designed to cause no more than one death per billion hours of operation. Engineers work hard at this, calculating all these nanomorts per hour. When reading about this, a natural question arises – why can't the risk simply be reduced to zero? Is there really such a significant gap between 0 and 0.000000001?
Buridan's ass
It turns out this gap exists, and it's substantial. The problem is that the laws of classical physics are continuous. As Feynman said in the introduction to his famous lecture series: "All things are made of atoms – little particles that are in perpetual motion, attracting each other when they are a little distance apart, but repelling upon being squeezed into one another". The interactions between these atoms can be modeled with various laws. Atoms experience forces and move according to Newton's laws. They interact with the electromagnetic field, which also interacts with itself |
839ee3e3-cd47-4090-a094-ed816e090638 | trentmkelly/LessWrong-43k | LessWrong | Personal Agents: AIs as trusted advisors, caretakers, and user proxies
Just posted the following on Medium. Interested in comments from readers here, especially pointers to similar efforts and ideas I didn't mention below.
This is the first in a series of articles describing features, functions, and components of Personal Agents — next generation AI virtual assistants that will serve as trusted advisors, caretakers, and user proxies. Personal Agents will preferably be developed as an open source project. Primary goals are to specify agents that (1) Significantly benefit people (are not just cool or fun) and (2) Avoid harmful side-effects (like those plaguing social media or that worry AI safety advocates). A clear and open specification will facilitate agent development and certification. This article provides a brief overview of Personal Agents.
Personal Agents (PAs), introduced here and here, are next-generation virtual assistants[1] that will support people in all aspects of their lives — from health and safety to education, career, finance, commerce, hobbies, friendship, romance, governance, and more. A PA will be securely paired with a person; deeply understand their circumstances, psychology, interests, abilities, and goals; and constantly work to support and improve their wellbeing. PAs will be adept advisors and caretakers and serve as proxies that represent the user in interactions with other agents and systems. They will be the ultimate fiduciaries[2].
PAs will be a key force for democracy and equity. They will protect and empower individuals so they can compete and thrive in a complex world teeming with powerful agents — from wealthy people to corporations to countries of geniuses in data centers. Every person will have a genius partner by their side.
A PA will manifest as a single persona that provides comprehensive support to its user — analogous to a parent, chief of staff, guardian angel, fairy godmother, or genie. When user needs exceed the knowledge and affordances of their PA, the PA will be able to connect with s |
a9ada336-59b3-49a8-a578-a4f34aae8460 | trentmkelly/LessWrong-43k | LessWrong | Expect to know better when you know more
A seemingly trivial result, that I haven't seen posted anywhere in this form, that I could find. It simply shows that we expect evidence to increase the posterior probability of the true hypothesis.
Let H be the true hypothesis/model/environment/distribution, and ~H its negation. Let e be evidence we receive, taking values e1, e2, ... en. Let pi=P(e=ei|H) and qi=P(E=ei|~H).
The expected posterior weighting of H, P(e|H), is Σpipi while the expected posterior weighting of ~H, P(e|~H), is Σqipi. Then since the pi and qi both sum to 1, Cauchy–Schwarz implies that
* E(P(e|H)) ≥ E(P(e|~H)).
Thus, in expectation, the probability of the evidence given the true hypothesis, is higher than or equal to the probability of the evidence given its negation.
This, however, doesn't mean that the Bayes factor - P(e|H)/P(e|~H) - must have expectation greater than one, since ratios of expectation are not the same as expectations of ratio. The Bayes factor given e=ei is (pi/qi). Thus the expected Bayes factor is Σ(pi/qi)pi. The negative logarithm is a convex function; hence by Jensen's inequality, -log[E(P(e|H)/P(e|~H))] ≤ -E[log(P(e|H)/P(e|~H))]. That last expectation is Σ(log(pi/qi))pi. This is the Kullback–Leibler divergence of P(e|~H) from P(e|H), and hence is non-negative. Thus log[E(P(e|H)/P(e|~H))] ≥ 0, and hence
* E(P(e|H)/P(e|~H)) ≥ 1.
Thus, in expectation, the Bayes factor, for the true hypothesis versus its negation, is greater than or equal to one.
Note that this is not true for the inverse. Indeed E(P(e|~H)/P(e|H)) = Σ(qi/pi)pi = Σqi = 1.
In the preceding proofs, ~H played no specific role, and hence
* For all K, E(P(e|H)) ≥ E(P(e|K)) and E(P(e|H)/P(e|K)) ≥ 1 (and E(P(e|K)/P(e|H)) = 1).
Thus, in expectation, the probability of the true hypothesis versus anything, is greater or equal in both absolute value and ratio.
Now we can turn to the posterior probability P(H|e). For e=ei, this is P(H)*P(e=ei|H)/P(e=ei). We can compute the expectation |
fd4e854a-17eb-4cfa-8f2a-7874e8dacf68 | trentmkelly/LessWrong-43k | LessWrong | Is there neuroscience research on cognitive biases?
I recently watched NeuraLink's presentation, and wondered how can something like that help us reason.
The obvious way is an AI that can reason, and is connected to our brains and we reason together with it.
But another direction i thought of, is just helping us notice when we're using motivated cognition and letting cognitive biases take place.
Another thing i thought of was reducing Akrasia, can it help us win the fight between areas in the brain in an akratic situation?
With my very little knowledge about the subjects in hand, it seems like it would be an easier target - seems to me you'd need fewer electrodes, less understanding of how reasoning works, and simpler software.
Although the questions about the possibilities of this technology are interesting, i don't expect anything more the guesses and predictions to be available right now.
So my question is whether we have neurological knowledge about how these mechanisms work. |
0214aa9e-9709-43c2-a95f-36f527aab103 | trentmkelly/LessWrong-43k | LessWrong | The Cambist and Lord Iron: A Fairy Tale of Economics
Available in PDF here, the short story in question may appeal to LW readers for its approach of viewing more things than are customary in handy economic terms, and is a fine piece of fiction to boot. The moneychanger protagonist gets out of several sticky situations by making desperate efforts, deploying the concepts of markets, revealed preferences, and wealth generation as he goes. |
fb6b5ad8-1ce2-45f9-b0e0-8b713d42aa21 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Towards more Generalizable One-shot Visual Imitation Learning.
I Introduction
---------------
Fig. 1:
Illustration of train/test split in prior and our proposed settings. In contrast to prior work in one-shot imitation learning which add more variations to a single task, we propose to train on multiple distinct tasks along with all of their variations, and test not only on multiple trained tasks, but also on novel tasks that were never seen during training.
Humans can learn to complete many tasks and quickly adapt to a new situation based on past experiences. We believe robots should also be able to learn a variety of tasks and acquire generalizable knowledge, which can then be transferred to quickly and efficiently learn a novel task.
One-shot imitation learning (OSIL) is a popular training framework for this purpose: an agent is trained to perform multiple tasks, each is described by an expert demonstration to provide context. First proposed in [duan2017one], the framework has been extended to different tasks and visual inputs [dasari2020transformers, finn2017one, james2018task, yu2018one]. However, these prior works tend to assume a very strong similarity between train and test. For example, a typical setting is where at training time the agent learns to build a block configuration that matches the block configuration in a preceding demo, and then at test time the agent is again requested to stack blocks, with variation just stemming from which block is in which position of the stack and the starting locations of blocks on the tabletop. Another typical setting is moving objects from a table top to a set of bins, where at test time the task will again be moving the same objects from tabletop to bins, with variation just stemming from which object goes to which bin and starting locations of the objects.
In this work, we propose to expand this narrow, single-task setting with a more significant distinction between train and test. Concretely, we build 7 robot manipulation environments: Door, Drawer, Press Button, Stack Block, Basketball, Nut Assembly, Pick & Place, based on simulation framework from Robosuite v1.1 [zhu2020robosuite] and MetaWorld [yu2019meta]. Terminology-wise, we will consistently refer to them as tasks. Within each task, there are “variations”, which capture differences in which block goes on top of which block, or which object goes into which bin, etc (see Figure [11](#S8.F11 "Fig. 11 ‣ VIII-B Task Environment Description ‣ VIII APPENDIX ‣ Towards More Generalizable One-shot Visual Imitation Learning") for visualizations of all 61 variations). Within one fixed variation of a task, there is also a continuum of instances, corresponding to all possible initial states of the various objects. We illustrate this setup with the row “Multi-task One-shot Imitation Learning” in Figure [1](#S1.F1 "Fig. 1 ‣ I INTRODUCTION ‣ Towards More Generalizable One-shot Visual Imitation Learning").
We evaluate representational and generalization capability through three settings: (i) one-shot imitation on variations within the multi-task training regime, (ii) one-shot imitation on new tasks; and (iii) fine-tuning on new tasks. As a first step in our investigation, we study the performance of prior state-of-the-art methods [dasari2020transformers, yu2018one]. We observe that, while performing well in the prior single-task settings, these methods largely fail to handle our proposed multi-task setup. This suggests a great opportunity for novel research towards improving the generalization ability of few-shot imitation methods.
In addition to identifying this challenge for more generalizable OSIL, we also propose a new approach that shows significant performance gain over prior state of the art.
Concretely, we investigate the hypothesis that prior methods fall short in the multi-task settings due to (i) poor representations that do not generalize well to new tasks; (ii) a lack of proper inductive bias in the model architecture, which prevents accessing the one demonstration from a new task. To address these challenges, we introduce MOSAIC: Multi-task One-Shot imitation with self-AttentIon and Contrastive learning, which incorporates two key components: (i) a new temporal contrastive loss objective to provide additional supervision for representation learning; (ii) a self-attention policy model architecture for extracting contextual information in the demonstration. Experimental results show significantly improved performance of our method over the prior state of the art.
Key contributions of this paper can be summarized as the following:
* We introduce a simulated robotic manipulation benchmark that spans 7 tasks and a total of 61 task variations. Its codebase is publicly released to facilitate future research.
* We propose a more challenging setup for one-shot imitation learning: 1) train an agent on multiple distinct tasks and test on the seen task variations; 2) train on multiple distinct tasks and test on completely new tasks, via direct one-shot execution or fine-tuning. We investigate prior state-of-the-art methods under these conditions and observe clear room for improvement.
* We propose our method MOSAIC, which combines a self-attention model architecture and a temporal contrastive objective. We experimentally demonstrate its superior performance over baselines, and show its promising ability at being fine-tuned to learn a new task efficiently.
Ii Related Work
----------------
Imitation learning. There are two main approaches for imitation learning (IL): inverse reinforcement learning (IRL) [abbeel2004irl, ng2000irl, ho2016generative] which finds a cost function under which the expert is uniquely optimal, and behavioral cloning (BC) [bain1995framework, pomerleau1991efficient] that predicts expert actions from state observations as a supervised learning problem.
Recent advances in IL have enabled agents to perform various robotic control tasks, such as locomotion [peng2020learning, ho2016generative], self-driving [bojarski2016end, pomerleau1998autonomous], video games [aytar2018playing, ross2011reduction], and manipulation [pook1993recognizing, sweeney2007model, young2020visual]. However, a majority of these applications assumes a close match between train and test environment. This has the disadvantage of learning without the ability to transfer knowledge to new situations, and lacks the opportunity for a human to instruct the agent with a new task at test time.
One-shot imitation learning. To address these limitations, one-shot imitation learning (OSIL), first proposed in [duan2017one], trains an agent to intake both one successful demonstration and the current observation, and predict the expert’s action. Later work extended OSIL to observe visual inputs: [finn2017one] applies the Model-Agnostic Meta-Learning algorithm (MAML) [finn2017model] to adapt policy model parameters for new tasks; TecNets [james2018taskembedded] applies a hinge rank loss to learn explicit task embeddings; DAML [yu2018one] adds a domain-adaptation objective to MAML to use human demonstration videos; [dasari2020transformers] improves policy network with Transformer architecture [vaswani2017attention]. Another line of work learns modular task structures that can be reused at test time [xu2018neural] [huang2019neural] [Huang2019ContinuousRO], but outputs of these symbolic policies are highly abstracted into semantic action concepts (e.g. “pick”, “release”) that assume extensive domain knowledge and human-designed priors.
However, prior OSIL work has been limited to a single-task setup and mainly tests a model on a slightly different instance (e.g. different object pose) of the previously-seen task variations. For example, [finn2017one] and [james2018taskembedded] experimented with 3 separate settings: simulated planer reaching (with different target object colors), simulated planer pushing (with varying target object locations), and real-robot, object-in-hand placing (onto different target containers). In contrast, we consider a more difficult multi-task setup, where agent needs to perform well across more diverse and distinct tasks, and generalize not only to new instances of all the seen variations, but also to completely novel tasks.
Multi-task Imitation Learning for Robotic Manipulation
Our work falls under the broader category of imitation learning multiple robot manipulation tasks [zhou2020watch][lynch2019play][singh2020scalable]. The term “multi-task” has varying definitions across this space of literature. Some work define stacking different block combinations as different tasks, whereas we define them as variations of the same task. Tasks that are sufficiently distinct, such as object pushing versus grasping, are sometimes called “task families” [zhou2020watch], where a “multi-task” policy is trained with only one family, and novel object configurations are named “new tasks” to test generalization. Recent work [lynch2021language] [NEURIPS2020\_9909794d] also explored language conditioning for different interact-able objects as tasks, where act-once word embeddings are used for disambiguation, and shows generalization ability at sequentially executing trained tasks to achieve longer-horizon test-time tasks. Concurrently, BC-0 [jang2021bc0] reports 100 distinct manipulation tasks for zero-shot imitation learning, where the 100 “tasks” fall into only 9 underlying skills and 6-15 different objects, and the “unseen” tasks are object arrangements that are excluded from training.
Unsupervised/self-supervised representation learning.
Recently, several unsupervised/self-supervised representation learning methods have been proposed to improve the performance in learning from visual inputs [aytar2018playing, laskin2020curl, stooke2020decoupling, schwarzer2021dataefficient].
[aytar2018playing] solves hard exploration environments like Atari’s Montezuma’s Revenge [bellemare2013arcade] by using self-supervised representation to overcome domain gaps between the demonstrations and an agent’s observation.
CURL [laskin2020curl] and ATC [stooke2020decoupling] showed that sample-efficiency can be significantly improved by applying contrastive learning [chen2020simple, henaff2020data, he2019momentum] to reinforcement learning.
In this paper, we show that contrastive learning also provides large gains in one-shot imitation learning.
Iii Problem Setup
------------------
We extend the framework of one-shot imitation learning (OSIL) [duan2017one] to a challenging multi-task setup. We categorize a set of semantically similar variants of a single task as “variations”: for example, for each variation of the Pick & Place task, the agent should pick up one of 4 differently shaped objects, and place it in one of the 4 bins, resulting in 16 variations in total. Following this definition, prior work [duan2017one, dasari2020transformers, finn2017one, james2018task, yu2018one] on one-shot imitation learning evaluate agents with a single task, as illustrated in Figure [1](#S1.F1 "Fig. 1 ‣ I INTRODUCTION ‣ Towards More Generalizable One-shot Visual Imitation Learning").
Consider n different tasks, {T1,T2,…,Tn}, where each task Ti contains a set of variations Mi. For each task, the training dataset Di contains paired expert demonstrations and trajectories from multiple variations: Di={(dm,τm),m∈Mi}. The demonstrator provides a video dm={o0,⋯,oT}, and the policy is trained to intake dm and imitate an expert trajectory τm={(o0,a0),⋯,(oT,aT)}. While expert trajectories require both actions and observations, only video inputs are required as demonstrations, thus the demonstrator can be any other robot or even human. For all our experiments, the demonstrator robot has a different arm configuration (i.e. Sawyer Arm) than the imitator agent (i.e. Panda Arm).
Given training datasets D={D1,⋯,Dn} of those n tasks, we optimize a demonstration-conditioned policy πθ(at|ot,d), parameterized by θ, that takes an expert video and the current observation as input and takes an action at each time-step t. At test time, the model is provided one demonstration dtest of one variation of a task mtest, paired with observations mtest, t. Note that mtest can be an unseen variation of either one of the trained tasks, or a never-seen task excluded from training.
Iv Main Method
---------------
Fig. 2:
Illustration of our overall network architecture (left) and contrastive module (right). The policy network takes in a stack of demonstration video frames and state observation images, and predicts the expert action at each time-step. A temporally-contrastive loss is applied in auxiliary with the policy’s behavior cloning loss. Note that contrastive loss can be applied to features from either the convolution module or self-attention module. The model in gray box is gradient-free, and only receives parameter updates from its online counterpart.
In this section, we describe in details our approach to policy model architecture (Section [IV-A](#S4.SS1 "IV-A Network Architecture ‣ IV MAIN METHOD ‣ Towards More Generalizable One-shot Visual Imitation Learning")), self-supervised representation learning via a contrastive module (Section [IV-B](#S4.SS2 "IV-B Contrastive Representation Learning ‣ IV MAIN METHOD ‣ Towards More Generalizable One-shot Visual Imitation Learning")) and the action imitation loss objective (Section [IV-C](#S4.SS3 "IV-C Policy Learning ‣ IV MAIN METHOD ‣ Towards More Generalizable One-shot Visual Imitation Learning")).
###
Iv-a Network Architecture
Figure [2](#S4.F2 "Fig. 2 ‣ IV MAIN METHOD ‣ Towards More Generalizable One-shot Visual Imitation Learning") provides an overview of our model pipeline: a CNN backbone is followed by a multi-head self-attention module [vaswani2017attention] and an MLP to make action predictions.
Visual features Given a batch of B inputs containing Td expert video frames and To agent observations, a CNN backbone encodes each frame into C channels of size H×W feature maps, resulting in demonstration features xd with size [B,C,Td,H,W], and observation features xo with size [B,C,To,H,W].
To preserve spatial and temporal information, both features are flattened along the last 3 dimensions and added with sinusoidal encodings [vaswani2017attention], then re-shaped into the original size.
Self-attention module We use multiple self-attention layers that model the underlying relationship between the sequence of representations xd and xo. We adopt the non-local self-attention block in [wang2018non], and make it to a multi-head version as [dasari2020transformers]. Specifically, key, query and value tensors are generated from three separate 3D convolution layers, which are then flattened along the time and space dimensions to compute spatio-temporal attention by each head individually.
Formally, given temperature parameter τ, key Kj, query Qj and value Vj, the attention head j, the attention operation is computed as:
| | | |
| --- | --- | --- |
| | Aj=softmax(QjK⊤j/τ)Vj. | |
Outputs Aj from each head are concatenated and projected to the original feature size by another 3D convolution, as used in [wang2018non]. xd will first pass through a self-attention module to get xattnd. Then, every frame in xo will compute self-attention with both xattnd and itself, which in effect calculates: (1) the spatial self-attention in each observation frame and (2) spatio-temporal cross-attention on the demonstration. The resulting xattno will be used to predict action.
###
Iv-B Contrastive Representation Learning
Our method bases off the intuition that, representations for two nearby frames from the same video clip should be similar, whereas frames from different tasks or variations should be drawn apart.
For each frame in a video, we maximize its feature similarity with a randomly selected, temporally close-by frame.
Specifically, we take an input batch, and obtain its two “view”s by two separate data augmentations.
The model encodes the first view into x1, and a target model encodes the second to get x2, which is gradient-free and receives parameter updates solely from its online counterpart.
Lastly, x1 and x2 are separately passed through a linear projector f: z1=f(x1), and its target ¯f: z2=¯f(x2). For every feature frame in z1, we select a nearby feature frame in z2 as positive. We then maximize the similarity between each anchor q=g(z1), and its positive k+=z2, via the InfoNCE loss from [oord2019representation], where
g is another linear projector, also named as predictor in prior work [grill2020bootstrap].
We follow [oord2019representation, Hnaff2020DataEfficientIR] to model embedding similarity as bilinear product, calculated with a projection matrix W.
Formally, with total frame count F=B(Td+To), treating every other k in the batch as negatives, the contrastive loss at each q is expressed as:
| | | | |
| --- | --- | --- | --- |
| | LRep=logexp(qTWk+)exp(qTWk+)+∑F−1i=1(qTWki) | | (1) |
One may view the convolution backbone and self-attention layers as one combined feature extractor, therefore the above contrastive loss can be applied to either before or after the self-attention layers as shown in Figure [2](#S4.F2 "Fig. 2 ‣ IV MAIN METHOD ‣ Towards More Generalizable One-shot Visual Imitation Learning"). Moreover, our contrastive module differs from [laskin2020curl] in the new temporal contrast strategy: for frame feature xt at timestep t, it contrasts with a random nearby frame selected from xt−k to xt+k in its augmented counterpart, whereas prior work [stooke2020decoupling] uses a fixed-step future frame. We provide ablation experiments in Appendix to provide additional insights on details of our contrastive objetive implementation.
###
Iv-C Policy Learning
Our objective is to learn a policy πθ(at|ot,d) which takes current image observation and a demonstration video as inputs, and predicts the action distribution to successfully finish the task.
To enable learning a potentially multi-modal policy that excels across many tasks, we adopt the same solution used in [dasari2020transformers, lynch2020learning, salimans2017pixelcnn++], which discretizes the action space into 256 independent bins along every dimension, and parameterize the policy using a mixture of discretized logistic distribution. As described in Section [IV-A](#S4.SS1 "IV-A Network Architecture ‣ IV MAIN METHOD ‣ Towards More Generalizable One-shot Visual Imitation Learning"), the self-attended observation features xattno will pass through the action MLP, to predict the mean μi, scale si and mixing weight αi for each discretized logistic distribution. The behavior cloning training loss is the negative log-likelihood:
| | | | |
| --- | --- | --- | --- |
| | LBC=−log(m∑i=1αi(xattno)P(at,μi(xattno),si(xattno))) | | (2) |
Where P(at,μi,si)=σ(at+0.5−μisi)−σ(at−0.5−μisi), σ is the logistic sigmoid function.
At the inference time, given ot, the action is sampled from the predicted distribution:
| | | | |
| --- | --- | --- | --- |
| | at∼m∑i=1αi(xattn% ot) logistic(μi(xattnot),si(xattnot)) | | (3) |
In addition to the behavioural cloning loss, we also utilize the inverse dynamics loss as in [dasari2020transformers]. By taking consecutive observation frames ot,ot+1,…,ot+k during training, another MLP will predict inverse actions at,at+1,…,at+k−1. The inverse dynamics loss has similar form as ([2](#S4.E2 "(2) ‣ IV-C Policy Learning ‣ IV MAIN METHOD ‣ Towards More Generalizable One-shot Visual Imitation Learning")):
| | | | |
| --- | --- | --- | --- |
| | Pi=P(at,μi(xattnot,xattnot+1),si(xattnot,xattnot+1)) | | (4) |
| | LInv=−log(m∑i=1αi(xattnot,xattnot+1)Pi) | | (5) |
Combing with the contrastive loss LRep introduced in Section [IV-B](#S4.SS2 "IV-B Contrastive Representation Learning ‣ IV MAIN METHOD ‣ Towards More Generalizable One-shot Visual Imitation Learning"), we obtain the overall loss for our method:
| | | | |
| --- | --- | --- | --- |
| | L=λRepLRep+λBCLBC+λInvLInv | | (6) |
V Experiments
--------------
###
V-a Task Environment and Dataset
Simulation environment.
We develop 7 distinct task environments using Robosuite v1.1 [zhu2020robosuite] and combining MetaWorld [yu2019meta] for additional assets. For each task, we additionally design multiple semantically distinct variations.
In order to investigate cross-morphology imitation, we also integrate two robot arms. The imitation policy is learned and evaluated on a Panda robot arm but takes a Sawyer robot video as demonstration.
Data collection. For every variation of each task environment, we design scripted expert policies and collect 100 demonstration videos of Sawyer robot and another 100 for Panda robot, with differently initialized scene layouts as instances.
We provide more detailed information on simulation environment and data collection in Appendix.
###
V-B Experimental Results
| | | | | | | |
| --- | --- | --- | --- | --- | --- | --- |
| Task | Setup | DAML [yu2018one] | T-OSIL [dasari2020transformers] | LSTM | MLP | MOSAIC (ours) |
| Door | single | 23.3 ± 5.2 | 57.9 ± 7.1 | 65.8 ± 7.1 | 41.2 ± 8.2 | 67.1 ± 5.5 |
| multi | 10.8 ± 5.4 | 49.2 ± 6.0 | 43.8 ± 9.5 | 58.8 ± 7.1 | 68.3 ± 6.3 |
| Drawer | single | 15.4 ± 5.5 | 57.5 ± 3.9 | 57.5 ± 8.1 | 57.9 ± 3.6 | 65.4 ± 3.4 |
| multi | 3.3 ± 1.4 | 53.3 ± 4.0 | 28.7 ± 6.0 | 52.5 ± 6.0 | 55.8 ± 3.6 |
| Press Button | single | 62.8 ± 3.9 | 56.4 ± 2.4 | 48.3 ± 6.6 | 40 ± 5.5 | 71.7 ± 3.9 |
| multi | 1.7 ± 0.7 | 63.3 ± 3.5 | 25.8 ± 3.0 | 25.0 ± 3.8 | 69.4 ± 3.4 |
| Pick & Place | single | 0 ± 0 | 74.4 ± 2.1 | 10.6 ± 1.8 | 12.8 ± 2.3 | 88.5 ± 1.1 |
| multi | 0.0 ± 0.0 | 19.5 ± 0.4 | 2.2 ± 0.7 | 5.0 ± 1.4 | 42.1 ± 2.3 |
| Stack Block | single | 10.0 ± 1.8 | 13.3 ± 2.6 | 8.6 ± 2.3 | 52.5 ± 4.7 | 79.3 ± 1.8 |
| multi | 0.0 ± 0.0 | 34.4 ± 3.4 | 33.3 ± 5.5 | 16.7 ± 3.7 | 70.6 ± 2.4 |
| Basketball | single | 0.4 ± 0.3 | 12.5 ± 1.6 | 5.4 ± 1.2 | 24.2 ± 2.6 | 67.5 ± 2.7 |
| multi | 0.0 ± 0.0 | 6.9 ± 1.3 | 12.1 ± 2.1 | 10.0 ± 2.0 | 49.7 ± 2.2 |
| Nut Assembly | single | 2.2 ± 1.4 | 6.3 ± 1.9 | 3.9 ± 1.5 | 15.6 ± 2.9 | 55.2 ± 2.8 |
| multi | 0.0 ± 0.0 | 6.3 ± 1.3 | 4.4 ± 1.3 | 6.7 ± 1.3 | 30.7 ± 2.5 |
TABLE I:
Test-time one-shot imitation performance as measured by success rate (% ) on both single-task and multi-task setup. For each task, 1) each entry of the row named “single” reports results of a single-task model that was trained and tested on the same task; 2) the row named “multi” reports results from one multi-task model that was trained on all 7 tasks in the benchmark and tested on each task separately.
We conduct experiments with the dataset described in Section [V-A](#S5.SS1 "V-A Task Environment and Dataset ‣ V EXPERIMENTS ‣ Towards More Generalizable One-shot Visual Imitation Learning") to answer the following questions:
* How does our method compare with prior baselines under the original single-task one-shot imitation learning setup.
* How well does our method perform across multiple tasks, after trained on the same set of tasks.
* How well does our trained multi-task model perform given a completely new task: can it 1) directly perform one-shot imitation at test time; 2) be fine-tuned to quickly adapt to the new task requiring fewer amount of data.
* Which component(s) in our contrastive module are key to its effectiveness at representation learning. The ablation experiment results are provided in Appendix due to space limitation.
We report and compare performance to the following baseline methods:
* DAML [yu2018one]: We train a policy model with MAML [finn2017model] loss and behaviour cloning loss. We replace the model architecture used in the original paper with a wider and deeper network of comparable parameter counts as ours, and use the same action distribution parametrization for policy learning.
* T-OSIL [dasari2020transformers]:
Model architecture proposed by [dasari2020transformers] which also uses non-local block [wang2018non] in self-attention module and is trained with end-effector point prediction as auxiliary loss. Our method utilizes a more computationally efficient attention operation, see Figure [12](#S8.F12 "Fig. 12 ‣ VIII-C Implementation Details ‣ VIII APPENDIX ‣ Towards More Generalizable One-shot Visual Imitation Learning") for an illustrated comparison.
* LSTM: We replace the self-attention module in our model architecture with linear projectors followed by an LSTM [hochreiter1997long, sutskever2014sequence] architecture of a comparable parameter count. The rest of the policy model architecture is kept the same as in our main method.
* MLP: We replace the self-attention module in our model architecture with a simple MLP layer to process stacked visual features from the demonstration into “task context vectors”, which is then concatenated with observation features and used for action predictions.
For all experiments (including baselines), we keep the first three convolutional residual blocks in ResNet-18 as the feature extractor, and apply the same data augmentation strategy to prevent over-fitting and improve model robustness. For evaluation, we take 3 different converged model checkpoints, and for each variation of each task, we gather each policy’s rollout performance across 10 episodes with different random seeds. For both the single and multi-task models, we report the mean and standard deviation of success rates in each task separately. Details on network architecture and hyper-parameters can be found in the Appendix.
Single-task One-shot Imitation
We first evaluate performance on the single-task setup as done in prior methods. Specifically, the model is trained with demonstrations from multiple variations of one task and then tested on unseen instances of the same task (e.g. different initial poses). We report the success rate of all methods on each task in the rows named “single” of Table [I](#S5.T1 "TABLE I ‣ V-B Experimental Results ‣ V EXPERIMENTS ‣ Towards More Generalizable One-shot Visual Imitation Learning"), with a clear out-performance of our method over baselines on every task.
We remark that DAML [yu2018one] also experimented with a pick-and-place task in their original paper, but collected “hundreds of” objects for training and 12 held-out objects for testing. Without access to further details, we hypothesize that this visual diversity in training dataset was crucial for its success at picking correct test-time objects, which explains why the same method reports massive under-performance on the 4-object Pick & Place task, where [dasari2020transformers] also reported a very low success rate from DAML (6.9%) using a differently-configured simulation of the same task.
Multi-task One-shot Imitation
We next consider the multi-task setup by mixing data from all 7 tasks into one training dataset. After training, we report the success rate of one model on each task in the rows named “multi” of Table [I](#S5.T1 "TABLE I ‣ V-B Experimental Results ‣ V EXPERIMENTS ‣ Towards More Generalizable One-shot Visual Imitation Learning").
The baselines’ performances drop significantly compared with their results from single-task, whereas ours continues to work well across many tasks 111For this setup, we increase the number of attention layers in the model from 2 to 3, and also adjust each baseline accordingly for fair comparisons. However, the performance of DAML [yu2019meta] is still not comparable with others using the updated architecture and after tuning
hyper-parameters..
Novel Task Generalization
Fig. 3: We compare fine-tuning multi-task models on their corresponding held-out novel task versus training a single-task model from scratch. We intermittently save model checkpoints throughout each training run and plot results for evaluating one-shot imitation performance. To additionally compare the data efficiency between fine-tuning and train-from-scratch, we experiment with restricted amount of demonstration data used for training, i.e. 25%, 50% and 75% of the data that was used for single-task and multi-task experiments and reported results in Table [I](#S5.T1 "TABLE I ‣ V-B Experimental Results ‣ V EXPERIMENTS ‣ Towards More Generalizable One-shot Visual Imitation Learning")
| Training Setup | Novel-task | No Training | Single-task | Multi-task | Fine-tune |
| --- | --- | --- | --- | --- | --- |
| Door | 5.0 ± 3.1 | 2.5 ± 0.9 | 67.1 ± 5.5 | 68.3 ± 6.3 | 67.5 ± 5.6 |
| Drawer | 15.0 ± 6.6 | 1.2 ± 1.2 | 65.4 ± 3.4 | 55.8 ± 3.6 | 52.5 ± 4.5 |
| Press Button | 5.0 ± 3.1 | 0 | 73.9 ± 3.9 | 69.4 ± 3.4 | 51.1 ± 4.8 |
| Stack Block | 0 | 0 | 79.3 ± 1.8 | 70.6 ± 2.4 | 98.3 ± 0.9 |
| Basketball | 0 | 0 | 67.5 ± 2.7 | 49.7 ± 2.2 | 72.8 ± 2.7 |
| Nut Assembly | 0 | 0 | 55.2 ± 2.8 | 30.7 ± 2.5 | 73.3 ± 2.3 |
TABLE II:
Evaluation success rate (% ) of one-shot imitation with MOSAIC under different train/test settings. Novel-task: directly evaluate on the task at each row with a model that was only trained on all of the remaining tasks. No Training: evaluate a randomly initialized policy model without any training. Single-task: train and test on the same task in each row. Multi-task: train a model on all 7 tasks, then evaluate and report the final performance on each task separately.
To show a higher level of generalization ability, we test a one-shot imitation agent with tasks that are sufficiently different from what it already trained on. We hence set up a series of experiments that, each picks 1 out of the 7 tasks in our benchmark suite as the held-out task, and trains a multi-task model on the remaining 6 tasks until convergence.
We first directly evaluate each model on its corresponding held-out task. Results are reported in column “Novel-task” of Table [II](#S5.T2 "TABLE II ‣ V-B Experimental Results ‣ V EXPERIMENTS ‣ Towards More Generalizable One-shot Visual Imitation Learning"), where each row corresponds to the experiment where the current task was excluded from training and only used for one-shot evaluation. For comparison purposes, we also include: “No Training” column, which evaluates a randomly initialized policy network without any training , “Single-task” column, where each row reports performance of a model trained and tested on the same task, and “Multi-task” column, where we take one model trained on all 7 tasks and report its performance on each task separately.
As shown in Table [II](#S5.T2 "TABLE II ‣ V-B Experimental Results ‣ V EXPERIMENTS ‣ Towards More Generalizable One-shot Visual Imitation Learning"), directly evaluating a multi-task model largely fails to complete an unseen novel task, and performs significantly worse than when this novel task was included during single- or multi-task training. This failure of direct one-shot imitation on a novel task suggests exciting room for future research.
Nevertheless, the gap between ‘‘Novel-task” and ‘‘No Training” results 222
We remark that, the non-zero success rates in “No Training” column are due to the nature of task design in simulation, where a randomly initialized policy model would sometimes generate an action that accidentally leads to an episode being counted “successful”, such as hitting an opened drawer to shut it close or stumbling on a closed door and pushing it open.
suggests a multi-task model still learns certain non-random behaviors, which could be generalized directly to a completely novel task. Note that, the improvement over “No Training” is limited to the first three tasks (Door, Drawer and Press Button), which require simpler motions and task reasoning.
We are hence encouraged to continue from this setup, but further fine-tune each 6-task model on its corresponding held-out task. We use the exact same dataset for training the models in “Single-task” column, and report the final fine-tuned results in “Fine-tune” column. To investigate the data efficiency of this setup, we additionally experiment with using 25%, 50%, 75% the amount of the original single-task training data for fine-tuning, and compare with training from scratch on that single task using the same dataset size.
We plot evaluation results of intermittently-saved checkpoints during each model’s training in Figure [3](#S5.F3 "Fig. 3 ‣ V-B Experimental Results ‣ V EXPERIMENTS ‣ Towards More Generalizable One-shot Visual Imitation Learning"). We observe that a multi-task pre-trained model is able to adapt quickly to a completely new task, even when limited data is available (i.e., 25%), and the final convergence performance is sometimes higher than training single-task from-scratch for some challenging tasks (e.g. Nut Assembly, Block stacking and Basketball). This indicates that these models are indeed able to accumulate generalizable knowledge (such as flexible visual feature extractors) from pre-training on many tasks, which then put them at advantage of learning a novel task very efficiently.
Vi Conclusion
--------------
In this work, we build on prior progress in one-shot imitation learning and propose a more challenging multi-task setup. Instead of training and testing on different variations/instances of a single task, we call for training with multiple distinct tasks and testing on novel tasks that are never seen during training. We believe this setup is crucial towards building more capable and generalizable agents, and holds great potential for novel research.
To support this formulation, we introduce a one-shot imitation benchmark for robotic manipulation, which consists of 61 variations across 7 different tasks. We propose our method MOSAIC, which combines a self-attention model architecture and a temporal contrastive objective, and out-performs previous state-of-the-art methods by a large margin. When evaluated on a completely new task, we see a promising potential in fine-tuning our multi-task model to learn efficiently, but remark the great room for improvement at even better and faster one-shot learning at test time.
Vii Acknowledgement
--------------------
This work was supported by the Bakar Family Foundation, Google LLC, the Berkeley Center for Human Compatible AI (CHAI), and National Science Foundation grant NRI no. 2024675. The authors would also like to thank a number of lab colleagues for their help throughout the project: Stephen James and Thanard Kurutach for valuable project feedbacks; Aditya Grover, Kevin Lu and Igor Mordatch for insightful discussions; Colin(Qiyang) Li, Hao Liu for suggestions on the paper writing.
Viii Appendix
--------------
###
Viii-a Overview
In this section, we first provide detailed descriptions of our task environment design, then introduce more details about our method implementation and the experiment setup, and lastly we provide more ablation experiment results and analysis.
###
Viii-B Task Environment Description
We use Robosuite [zhu2020robosuite] as our base framework, and design 7 robot manipulation environments (tasks) with intra-task variations. For example, consider Nut Assembly task in the bottom-right of Figure [11](#S8.F11 "Fig. 11 ‣ VIII-B Task Environment Description ‣ VIII APPENDIX ‣ Towards More Generalizable One-shot Visual Imitation Learning"): it has 9 variations in total, resulting from picking up any one of the 3 nuts on the table, then assembling it to one of the 3 pegs. Since there are multiple varied instances for a specific task, the agent should finish the correct task without mis-identification based on demonstration. Below we provide details about each of the individual task design.
Fig. 4: Drawer task with 8 sub-tasks in total. Given a demonstration, the agent should be able to infer which drawer to open or close. Note that all the drawers will be initialized as opened if the task is to close drawer.
Fig. 5: Press button with 6 sub-tasks in total. Given a demonstration, the agent should be able to infer which button to press.
Fig. 6: Open door with in total 4 sub-tasks. Given a demonstration, the agent should be able to infer which door to open, and whether in clockwise or counterclockwise direction.
Fig. 7: Pick & place environment with in total 16 tasks. Given a demonstration, the agent should be able to infer which object to pick and which bin to place the object in.
Fig. 8: Block stacking environment with in total 6 tasks. Given a demonstration, the agent should be able to infer which block to pick and where to stack.
Fig. 9: Nut assembly environment with in total 9 tasks. Given a demonstration, the agent should be able to infer which nut to pick and which peg to assemble.
Fig. 10: Basketball environment with in total 12 tasks. Given a demonstration, the agent should be able to infer which ball to pick and which hoop to throw it in.
Fig. 11:
We visualize the entirety of our one-shot robot imitation benchmark of 7 tasks and a total of 61 semantically distinct variations. The number of variations is different across tasks, with a range from 4 (for Door task) to 16 (for Pick & Place task). For each variation, we also add randomization to create more varied instances, such as different initial object positions. A multi-task one-shot imitation policy is trained on a demonstration dataset that contains multiple tasks and all their variations.
###
Viii-C Implementation Details
Network Architecture
We provide an illustration of our self-attention block and one of our baseline [dasari2020transformers] as in Figure [12](#S8.F12 "Fig. 12 ‣ VIII-C Implementation Details ‣ VIII APPENDIX ‣ Towards More Generalizable One-shot Visual Imitation Learning"). For other baselines in our main paper, we replace this attention module with LSTM or MLP accordingly.
Fig. 12: Illustration of self-attention blocks used in our model (left) architecture and T-OSIL [dasari2020transformers] (right). By removing the attention calculation between frames within the observation input sequence, we significantly reduces the computational overhead and makes our model architecture able to scale better to longer length of observations. At test time, only one frame is given to model inference per each time-step, which make for faster evaluation and gaining the same or better performance than stacking multiple frames instead.
Data Augmentation
We add data augmentation for all the baselines to prevent overfitting. In our implementation, we use 4 types of data augmentation provided in the torchvision package: random translate, random crop, color jitter, Gaussian blur.
###
Viii-D Experiment Details
Hyper-parameter Settings
We provide more details about the hyper-parameters and other settings of model training and evaluation in Table [III](#S8.T3 "TABLE III ‣ VIII-D Experiment Details ‣ VIII APPENDIX ‣ Towards More Generalizable One-shot Visual Imitation Learning").
| | |
| --- | --- |
| Hyperparameter | Value |
| Input image size | (100,180,3) |
| # Demonstation frames | 4 |
| # Observation frames (train) | 7 |
| # Observation frames (eval) | 1 |
| # Evaluation episode per task | 10 |
| Optimizer | Adam |
| (β1,β2) | (.9,.999) |
| Learning rate | 5e−4 |
| Batch size | 30 |
| Non-linearity | ReLU |
| Contrastive latent dimension | 512 |
| Self-attention temperature | 16 |
| # Action layers | 2 |
| # Attention layers (single-task) | 2 |
| # Attention layers (multi-task) | 3 |
| Action head latent dimension | 256 |
| Action output dimension | 256 |
| | |
TABLE III: Hyperparameters used for model training and evaluation.
Batch Construction
One batch consists of demonstrations sampled from each variant in every training task, which are mixed evenly. Since each task contains a different number of variations, to ensure they have comparable learning progress, the loss is first averaged across variants within each task, and then across different tasks.
Computation Requiremenmts
Our single-task model can be trained within one GPU day of NVIDIA TITAN Xp, whereas the multi-task model takes about 4 GPU days to train.
###
Viii-E Ablation Experiments
|
| |
| --- |
| Task |
|
| |
| --- |
| Door |
|
| |
| --- |
| Drawer |
|
| |
| --- |
| Press |
| Button |
|
| |
| --- |
| Pick & |
| Place |
|
| |
| --- |
| Stack |
| Block |
|
| |
| --- |
| Basketball |
|
| |
| --- |
| Nut |
| Assembly |
|
|
| |
| --- |
| Single-task |
| w/ Contra. |
| 67.1 ± 5.5 | 65.4 ± 3.4 | 71.1 ± 3.9 | 88.8 ± 1.1 | 79.3 ± 1.8 | 67.5 ± 2.7 | 55.2 ± 2.8 |
|
| |
| --- |
| Single-task |
| w/o Contra. |
| 62.5 ± 8.2 | 60.8 ± 2.6 | 73.9 ± 3.9 | 69.5 ± 1.8 | 36.7 ± 3.6 | 19.7 ± 2.2 | 15.6 ± 2.5 |
|
| |
| --- |
| Multi-task |
| w/ Contra. |
| 68.3 ± 6.3 | 55.8 ± 3.6 | 69.4 ± 3.4 | 42.1 ± 2.3 | 70.6 ± 2.4 | 49.7 ± 2.2 | 30.7 ± 2.5 |
|
| |
| --- |
| Multi-task |
| w/o Contra. |
| 66.7 ± 7.3 | 67.5 ± 4.0 | 73.1 ± 3.4 | 11.9 ± 1.6 | 39.4 ± 3.4 | 9.2 ± 1.7 | 20.4 ± 1.9 |
|
| |
| --- |
| Multi-task |
| BYOL |
| 67.1 ± 6.4 | 67.9 ± 4.0 | 69.4 ± 3.1 | 12.9 ± 1.7 | 40.0 ± 2.4 | 10.0 ± 1.9 | 17.8 ± 1.8 |
TABLE IV: Ablation studies: the effectiveness of contrastive learning with negative samples, as measure by success rates on single- and multi-task setups. We remark that 1) contrastive loss improves one-shot imitation performance on both setups, and the gain is most significant for multi-task models and on the last three tasks (Stack Block, Baseketball, and Nut Assembly); 2) contrastive learning without negative samples (as reported in the row “Multi-task BYOL”) results in similar performance as not using contrastive module at all, which is reported in row “Multi-task w/o Contra.”.
|
| |
| --- |
| Task Domain |
|
| |
| --- |
| No-Temp |
|
| |
| --- |
| Fix-Temp |
|
| |
| --- |
| Pre-Attn |
|
| |
| --- |
| Post-Attn |
|
| |
| --- |
| Rand-Temp + |
| Both-Attn |
|
| Pick & Place | 12.3 ± 1.4 | 80.2 ± 1.9 | 23.5 ± 2.1 | 51.5 ± 4.1 | 88.5 ± 1.1 |
| Basketball | 53.8 ± 4.8 | 52.5 ± 2.8 | 24.7 ± 2.1 | 27.2 ± 2.5 | 67.5 ± 2.7 |
| Nut Assembly | 12.2 ± 2.0 | 12.2 ± 1.6 | 17.8 ± 1.8 | 22.6 ± 2.9 | 55.2 ± 2.8 |
TABLE V: Ablation studies: (1) whether and how to enforce temporal consistency (Column 1, 2, 5); (2) Where to apply contrastive loss throughout the model (Column 3-5).
We conduct a set of ablation studies focused on our contrastive representation learning approach, to provide insights on the design of the loss objective, how it incorporates with the policy learning, and its impact over the one-shot imitation performance.
The effect of contrastive learning objective
We first ablate by removing contrastive loss from the model’s training update: we train a model with solely behavior cloning loss, while other training configurations are kept consistent with the multi-task experiments reported in Table [I](#S5.T1 "TABLE I ‣ V-B Experimental Results ‣ V EXPERIMENTS ‣ Towards More Generalizable One-shot Visual Imitation Learning"). Results are reported in row “Single-task No Contra.” and “Multi-task No Contra.” of Table [IV](#S8.T4 "TABLE IV ‣ VIII-E Ablation Experiments ‣ VIII APPENDIX ‣ Towards More Generalizable One-shot Visual Imitation Learning"). Noticeably, our multi-task performance is significantly decreased without the contrastive learning objective. Comparing results from single-task and multi-task setups, we observe a clear challenge of task reasoning and robust representation learning, which the contrastive objective is able to address, but not fully closing the gap between a multi-task model and its single-task counterpart on every task.
Adding the contrastive loss also shows a more significant gain on the right-most 3 tasks of Table [IV](#S8.T4 "TABLE IV ‣ VIII-E Ablation Experiments ‣ VIII APPENDIX ‣ Towards More Generalizable One-shot Visual Imitation Learning"), which vary only the colors of otherwise similarly-shaped objects. As compared to Pick & Place where the 4 objects are differently sized, shaped, and colored, the blocks/basketballs/nuts in these variations tend to be more easily confused, which can be addressed the contrastive loss that explicitly forces the representations to be distinguishable among different sub-tasks.
To investigate the effect of contrasting against negative samples in the loss objective design, we implement BYOL [grill2020bootstrap], in which a data sample is only drawn together with its augmented counterpart. Its multi-task performance is shown in Table [IV](#S8.T4 "TABLE IV ‣ VIII-E Ablation Experiments ‣ VIII APPENDIX ‣ Towards More Generalizable One-shot Visual Imitation Learning"), which show no improvement over not using contrastive loss at all (as reported in row “Multi-task No Contra.”). We remark that in multi-task setups, negative samples help learn stronger representations that are better at distinguishing among images from different tasks/variations.
Where to apply contrastive loss.
As discussed in Section [IV-B](#S4.SS2 "IV-B Contrastive Representation Learning ‣ IV MAIN METHOD ‣ Towards More Generalizable One-shot Visual Imitation Learning"),
we can apply the contrastive operation in features from any intermediate layers of the model. To understand the effects of this algorithmic choice, we consider the following variants of our method:
(1) Pre-Attn: applying contrastive loss to only features prior to the self-attention layers;
(2) Post-Attn: applying contrastive loss only after the self-attention layers;
(3) Both-Attn (Ours): calculating both losses in (1) and (2), which we use for single and multi-task experiment results in above sections.
Table [V](#S8.T5 "TABLE V ‣ VIII-E Ablation Experiments ‣ VIII APPENDIX ‣ Towards More Generalizable One-shot Visual Imitation Learning") shows the performance of each ablation variant in three single tasks, among which we find that Both-Attn (using two contrastive losses) achieves the best performance.
The effect of temporal contrast.
Our contrastive strategy is doing a temporal contrast by randomly selecting frames from nearby time-steps as positive. To fully understand its effectiveness,
we compare it with two variants:
(1) No-Temp: applying contrastive loss to two different data augmentations from same frame, similar to CURL [laskin2020curl];
(2) Fix-Temp: applying temporal contrast but always with a fixed-step future frame, similar to ATC [dwibedi2019temporal]. Ours (denoted Rand-Temp) achieves the best performance as shown in Table [V](#S8.T5 "TABLE V ‣ VIII-E Ablation Experiments ‣ VIII APPENDIX ‣ Towards More Generalizable One-shot Visual Imitation Learning"), as suggested by comparing the first, second and last column in Table [V](#S8.T5 "TABLE V ‣ VIII-E Ablation Experiments ‣ VIII APPENDIX ‣ Towards More Generalizable One-shot Visual Imitation Learning").
###
Viii-F Further Discussion on Related Work
In this section, we provide a more detailed discussion of the methods and experiment task settings used by related prior work in one-shot imitation learning (OSIL).
However, prior OSIL work has been limited to a single-task setup and mainly tests a model on a different task variation (e.g stacking an unseen block combination) or a different instance (e.g. different object pose) of the previously-seen variations. Experiments in [duan2017one] train an agent to stack various (unseen) block combinations at test time, but use low-dimensional state-based inputs. For visual inputs, [finn2017one] and [james2018taskembedded] experimented with 3 separate settings: simulated planer reaching (with different target object colors), simulated planer pushing (with varying target object locations), and real-robot, object-in-hand placing (onto different target containers); [yu2018one] set up a two-stage pick-then-place task with varying target objects and target containers; [dasari2020transformers] uses a simulated Pick & Place task with 4 objects to pick and 4 target bins to place (hence 16 variations in total). The AI2-THOR [kolve2019ai2thor] environment used in [Huang2019ContinuousRO] requires collecting varying objects and dropping off at their designated receptacles, where actions are purely semantic concepts such as “dropoff” or “search”. In contrast, in this work we consider a harder, multi-task setup, where agent needs to perform well across more diverse and distinct tasks, and generalize not only to new instances of all the seen variations, but also to completely novel tasks. |
a9b64a5c-bbc9-40ae-87b8-c2e3361a29b1 | trentmkelly/LessWrong-43k | LessWrong | New Product MVP: LightWrong
|
482fd53b-f43b-41e5-938b-c5b819997748 | trentmkelly/LessWrong-43k | LessWrong | Don't Fear Failure
Last post, I talked about how trying things out yourself is a good way to learn about them. This post, I'm going to talk about ideas that helped me overcome one of my major obstacles to trying something -- fear of failure.
Overestimation of Damages: "Its not that big a deal"
In most cases, failure really isn't that big of a deal. Really. The difference between failure and a null action is the attempt. If the attempt isn't damaging, failure isn't damaging.
A few cases:
* Trying a new food/recipe? Maybe you don't like it and waste a few dollars. Maybe you mess it up. So you eat something else. Just don't do it for your first dinner with the in-laws and it should be fine. But a new dish might be totally delicious.
* Total stranger you think might be a cool person? Maybe they get annoyed at you. Then you can just break off and never talk to/see them again. Chatting with people I run into has made college visits much more enjoyable.
* Competition you might want to enter? Worst case scenario is that you lose. I learned a lot from Moody's Mega Math Challenge, and even though I don't think we did a particularly good job at modeling Lake Powell I still learned a lot about how mathematical modeling works.
* Dance you want to try? The worst that's likely is that you look a bit silly. Laugh it off.
There's lots of things where an attempt is actually worse than doing nothing. Jumping halfway to the other side of the ledge, for instance. Or only removing most of the toxic part of a pufferfish. But for a lot of potentially high-value things, a failed attempt doesn't really do much, so you might as well try them.
Rationality and Failure: "Don't panic"
Some people I know basically buckle under failure. A common failure mode seems to be to do something badly, establish an ugh field around that area, and then continue in a downward spiral. Getting a B on a math test turns into "Ugh, math", turns into "well I was never really good at that anyway", turns into a complete lack |
015a6156-6ed6-4e57-b92d-522b388d7e01 | trentmkelly/LessWrong-43k | LessWrong | Searching Magic Cards
I like playing Magic, except for the "it's designed to pump away your money" aspect. So I like formats where I can play with other people's cards a lot! My favorite is probably drafting from booster packs where someone else will keep all the cards, but yesterday Stevie brought over a Forgetful Fish-style deck he'd put together, and we played a couple times. The only creature in this format is the Dandan, a 4/1 blue creature with:
> Dandan can't attack unless defending player controls an Island.
> When you control no Islands, sacrifice Dandan.
One of the cards in Stevie's deck was Magical Hack, a blue instant with:
> Change the text of target spell or permanent by replacing all instances of one basic land type with another.
It's role in the deck was primarily to kill Dandans, by replacing "Island" on an opponent's Dandan with any basic land type they don't have. But it got me wondering: does it happen that there are any textual uses of the six basic land types [1] that are not intended to be about a basic land? For example, if a card happened to use the word "forestall", perhaps you could do something fun with it?
Supposedly you can search cards with regular expressions on Scryfall but I couldn't get it to work. Instead, I downloaded the cards as JSON from MTGJson [2] and wrote a ~10 line python script:
unusual_land_re = re.compile(
r"(?i)"
r"((plains)|(island)|(swamp)|(wastes)|(mountain)|(forest))"
r"(?!(\b|s\b|[.]|walk\b|cycling\b))")
for fname in glob.glob("*.json.gz"):
with gzip.open(fname) as inf:
r = json.load(inf)
if "cards" not in r["data"]: continue
for card in r["data"]["cards"]:
if "text" not in card: continue
if unusual_land_re.search(card["text"]):
print("%s: %s" % (card["name"], card["text"]))
The first time I ran it I got a bunch of "Forestcycling" etc, but after adding "cycling" to the query (as reproduced above) it didn't find anything. Oh well.
I probably could have gotten regexp sear |
4b1905a5-5ee4-4e4c-929b-cf526ee8f5ad | StampyAI/alignment-research-dataset/arxiv | Arxiv | Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few Examples
1 Introduction
---------------
Few-shot learning refers to learning new concepts from few examples, an ability that humans naturally possess, but machines desperately lack. Improving on this aspect would lead to more efficient algorithms that can flexibly expand their knowledge as necessary without requiring large labeled datasets. We focus on few-shot classification: classifying unseen examples into one of N new ‘test’ classes, given only a few reference examples of each new class.
Recent progress in this direction has been made by considering a meta-problem: though we are not interested in learning about any training class in particular, we can still exploit the training classes for the purpose of learning to learn new classes from few examples. The acquired learning procedure can then be directly applied to few-shot learning problems on new classes.
This intuition has inspired numerous models of increasing complexity for this problem (see the Related Work for many examples). However, we believe that the commonly used setup for measuring success in this direction is lacking. Notably, the typical approach is to train a model on a subset of classes from a given dataset and then subject it to classification tasks formed from the remaining set of classes from the same dataset. However, to be practically useful, meta-learners must generalize to truly different classes sampled from a different data distribution altogether. Furthermore, the evaluation tasks are artificially constrained to have training sets that are perfectly class-balanced, and any two classes are equally likely to co-appear in the same task. It is not clear therefore to what extent the performance on these tasks approximates the performance in the significantly more structured and imbalanced real world.
Meta-Dataset directly addresses the aforementioned limitations. In particular: 1) it is significantly larger-scale than previous benchmarks and is comprised of multiple datasets of different data distributions, 2) its task creation is informed by class structure for ImageNet and Omniglot, 3) it introduces realistic class imbalance, and 4) it varies the number of classes in each task and the size of the training set, enabling us to examine the robustness of models across a spectrum of tasks: from very-low-shot learning onwards.
The main contribution of this work is therefore to offer a more realistic and challenging environment for training and evaluating meta-learners for few-shot classification. By evaluating various baselines and meta-learners on Meta-Dataset, we are able to expose weaknesses of two popular meta-learners: Prototypical Networks and MAML. Finally, in light of these findings, we propose a novel hybrid of these two approaches which we demonstrate captures complementary desired aspects of both and achieves state-of-the-art in Meta-Dataset.
2 Background
-------------
##### Task Formulation
The end-goal of few-shot classification is to produce a model which, given a new learning episode with N classes and a few labeled examples (kc per class, c∈1,…,N), is able to generalize to unseen examples for that episode. In other words, the model learns from a training (support) set S={(x1,y1),(x2,y2),…,(xK,yK)} (with K=∑ckc) and is evaluated on a held-out test (query) set Q={(x∗1,y∗1),(x∗2,y∗2),…,(x∗T,y∗T)}. Each example (x,y) is formed of an input vector x∈RD and a class label y∈{1,…,N}. Episodes with balanced training sets (i.e., kc=k, ∀c) are usually described as ‘N-way, k-shot’ episodes.
These evaluation episodes are constructed by sampling their N classes from a larger set Ctest of classes and sampling the desired number of examples per class. A disjoint set Ctrain of classes is used to train the model; note that this notion of training is distinct from the training that occurs within a few-shot learning episode.
Few-shot learning does not prescribe a specific training procedure, but a common approach involves matching the conditions in which the model is trained and evaluated (Vinyals et al., [2016](#bib.bib25)). In other words, training often (but not always) proceeds in an episodic fashion. Some authors use training and testing to refer to what happens within any given episode, and use the terms meta-training and meta-testing to refer to using Ctrain to turn the model into a learner capable of fast adaptation and Ctest for evaluating its success to learn using few shots. This nomenclature highlights the meta-learning perspective alluded to earlier, but to avoid confusion we will adopt another common nomenclature and refer to the training and test sets of an episode as the support and query sets and to the process of learning from Ctrain simply as training.
##### Standard Datasets
Two datasets have emerged as de facto benchmarks for few-shot learning. Omniglot (Lake et al., [2015](#bib.bib9)) is a dataset of 1623 handwritten characters from 50 different alphabets and contains 20 examples per class (character).
Most recent methods obtain very high accuracy on various meta-learning problems formulated on Omniglot (using various numbers of ways and shots), rendering the comparisons between new few-shot learning methods unreliable.
The second benchmark, miniImageNet (Vinyals et al., [2016](#bib.bib25)), is formed out of 100 ImageNet (Russakovsky et al., [2015](#bib.bib17)) classes (64/16/20 for train/validation/test) and contains 600 examples per class. miniImageNet, albeit harder than Omniglot, has the same property that most recent methods trained on it present similar accuracy when controlling for model capacity, and we believe the dataset is approaching its limit in terms of allowing to discriminate between the merits of competing approaches.
We hypothesize this can be due to an artificially constrained setup. In particular, current benchmarks:
* Consider a fixed number of shots and ways. In contrast, real-life episodes are heterogeneous: they vary in terms of their number of classes and examples per class, and are unbalanced.
* Measure only within-dataset generalization. However, realistic applications often involve generalization *across datasets.*
* Ignore the relationships between classes when forming episodes. The coarse-grained classification of dogs and chairs may present different difficulties than the fine-grained classification of dog breeds, and current benchmarks do not establish a distinction between the two.
3 Approaches to Few-Shot Classification
----------------------------------------
In this section we review common baseline and meta-learning models that we evaluate on our benchmark, and introduce a novel meta-learner that achieves the state-of-the-art on our benchmark.
##### Non-episodic Baselines
Before diving into meta-learning, it is important to explore non-episodic solutions. Consider a deep neural network trained on a classification task. A natural non-episodic approach would exploit the meta-training data by simply using it to train a classifier over all of the meta-training classes Ctrain. Consider the embedding function g(x), defined by all the layers except the penultimate layer of the classification network. The hope of the non-episodic baselines resides in the possibility that this embedding function produces ‘meaningful’ representations even for examples of previously-unseen classes, thus enabling few-shot classification. It then remains to define an algorithm for using these representations for few-shot classification. We consider two choices for this algorithm, yielding the ‘k-NN’ and ‘Finetune’ variants of this baseline.
The ‘k-NN’ baseline classifies each query example as the class that its ‘closest’ support example belongs to. Closeness is measured by either Euclidean or cosine distance in the learned embedding space g. We treat this choice over the metric as a hyperparameter. On the other hand, the ‘Finetune’ baseline exploits the support set of each new meta-test task to train a new ‘output layer’ on top of the embedding function g for the purpose of classifying between the N new classes of the given task.
##### Episodic Models
In the episodic setting, models are trained end-to-end for the purpose of learning to build classifiers from a few examples. We choose to experiment with Matching Networks (Vinyals et al., [2016](#bib.bib25)), Prototypical Networks (Snell et al., [2017](#bib.bib22)) and Model Agnostic Meta-Learning (MAML) (Finn et al., [2017](#bib.bib4)) since we believe that these three cover a diverse set of approaches to few-shot learning. We also introduce a novel meta-learner which is inspired by the last two models.
In each training episode, episodic models compute for each query example x∗∈Q, the distribution for its label p(y∗|x∗,S) conditioned on the support set S and allow training this differentiably-parameterized conditional distribution end-to-end via gradient descent. The different models are distinguished by the manner in which this conditioning on the support set is realized. In all cases, the performance on the query set drives the update of the meta-learner’s weights, which include (and sometimes consist only of) the embedding weights. We briefly describe each method below.
##### Prototypical Networks
Prototypical Networks construct a prototype for each class and then classify each query example as the class whose prototype is ‘nearest’ to it under Euclidean distance. More concretely, the probability that a query example x∗ belongs to class k is defined as:
| | | |
| --- | --- | --- |
| | p(y∗=k|x∗,S)=exp(−||g(x∗)−ck||22)∑k′∈{1,…,N}exp(−||g(x∗)−ck′||22) | |
where ck is the prototype for class k, computed by averaging the embeddings of class k’s support examples.
##### Matching Networks
Matching Networks (in its simplest form) label each query example as a weighted linear combination of the support labels, where a support label is weighted more heavily the ‘closer’ the corresponding support embedding is to the query in question. Specifically the probability distribution over the label y∗ of the query x∗ is given by:
| | | |
| --- | --- | --- |
| | p(y∗=k|x∗,S)=|S|∑i=1α(x∗,xi)1yi=k | |
where 1A equals 1 if A is true and 0 otherwise, and
| | | |
| --- | --- | --- |
| | | |
with c denoting the cosine similarity.
##### Maml
Let us now expand the notation for the embedding function to g(x;θ), exposing its parameters θ. MAML for few-shot classification starts by assuming a linear classifier, parametrized by a bias vector b and a weight matrix W, applied in the embedding space. It then classifies a query example based on
| | | |
| --- | --- | --- |
| | p(y∗|x∗,S)=softmax(b′+W′g(x;θ′)) | |
where the support set S is used to perform a small number of within-episode training steps for adjusting parameters (b,W,θ) and produce fine-tuned parameters (b′,W′,θ′). Training in MAML is made possible by unrolling the within-episode gradient descent steps performed on S and optimizing the prediction of the query set labels with respect to the initial (b,W,θ) by backpropagation. This normally requires computing second-order gradients, which can be expensive to obtain (both in terms of time and memory). For this reason, an approximation is often used whereby gradients of the within-episode descent steps are ignored. This variant is referred to as first-order MAML (fo-MAML) and was used our experiments. We did attempt to use the full-order version, but found it to be impractically expensive (e.g., it caused frequent out-of-memory problems).
Moreover, in our setting, since the number of ways will be varying between episodes, we do not learn b,W and set them to zero (i.e., b′,W′ are the result of within-episode gradient descent initialized at 0), thus only training θ. In other words, MAML focuses on learning the within-episode initialization θ of the embedding network so that it can be rapidly adapted for a new task.
##### Introducing Proto-MAML
We introduce a novel meta-learner which we argue captures the best of both Prototypical Networks and MAML. In particular, the former exploits a very simple inductive bias that was demonstrated to be effective for reasoning about new classes from very few examples. However, it lacks a mechanism for within-task adaptation. On the other hand, the latter adopts a simple procedure for task adaptation using only a few steps. We view Proto-MAML as the marriage of these two complementary strengths.
As explained in Snell et al. ([2017](#bib.bib22)), Prototypical Networks can be re-interpreted as a linear classifier applied to a learned representation g(x). In particular, due to using the squared Euclidean distance metric on top of the learned embeddings, the probability of a query example belonging to the different classes of the episode under the formulation of the Prototypical Network can be viewed as the output of a linear layer with a particular parameterization. Specifically, let x∗ denote a query example, g the trainable embedding function, and ck the prototype for class k. Then, the ‘logit’ for x∗ belonging to class k is:
| | | |
| --- | --- | --- |
| | −||g(x∗)−ck||2=−g(x∗)Tg(x∗)+2cTkg(x∗)−cTkck=2cTkg(x∗)−||ck||2+constant | |
where the scalar value constant captures the term that does not relate to class k and will not affect the softmax probabilities.
The k’th unit of the equivalent linear layer therefore has weights Wk,⋅=2ck and bias bk=−||ck||2. It’s worth mentioning that since ck is a function of g(x), W and b are thus differentiable with respect to θ.
We refer to Proto-MAML as the (fo-)MAML model where the task-specific linear layer of each episode is initialized from the Prototypical Network-equivalent weights and bias defined above and subsequently optimized as usual on the given support set. When computing the meta update for θ, we allow gradients to flow through the Prototypical Network-equivalent linear layer initialization. We show that this simple modification significantly helps the optimization of this model and outperforms vanilla fo-MAML by a large margin on Meta-Dataset.
4 Meta-Dataset: A New Few-Shot Classification Benchmark
--------------------------------------------------------
Meta-Dataset aims to offer an environment for measuring progress in realistic few-shot classification tasks. In particular, we argue that it constitutes a more realistic environment for assessing performance on a more realistic version of the task. Our approach therefore is twofold: 1) changing the data and 2) changing the formulation of the task (i.e., how episodes are generated).
We accomplish the former by incorporating multiple diverse data sources, and the latter by introducing a new sampling procedure for episodes that more closely resembles realistic learning scenarios. The following sections describe these modifications in detail. The code is open source and publicly available111[github.com/google-research/meta-dataset](https://github.com/google-research/meta-dataset).
###
4.1 Meta-Dataset’s Data
The data we propose to use is much larger in size than any previous benchmark, and is comprised of *multiple different existing datasets*. This invites research into how diverse sources of data can be exploited by a meta-learner, and allows us to evaluate a more challenging generalization problem, to new datasets altogether. Specifically, Meta-Dataset leverages data from the following 10 datasets: ILSVRC-2012 (ImageNet) (Russakovsky et al., [2015](#bib.bib17)), Omniglot (Lake et al., [2015](#bib.bib9)), Aircraft (Maji et al., [2013](#bib.bib11)), CUB-200-2011 (Birds) (Wah et al., [2011](#bib.bib26)), Describable Textures (Cimpoi et al., [2014](#bib.bib3)), Quick Draw (Jongejan et al., [2016](#bib.bib8)), Fungi (Schroeder & Cui, [2018](#bib.bib21)), VGG Flower (Nilsback & Zisserman, [2008](#bib.bib14)), Traffic Signs (Houben et al., [2013](#bib.bib7)) and MSCOCO (Lin et al., [2014](#bib.bib10)).
These datasets were chosen because they are free and easy to obtain, span a variety of visual concepts (natural and human-made) and vary in how fine-grained the class definition is. More information about each of these datasets is provided in Appendix A.
However, to ensure that episodes correspond to realistic classification problems, all episodes generated in Meta-Dataset use classes from a single dataset at a time only. Moreover, two of these datasets, Traffic Signs and MSCOCO, are fully reserved for evaluation, meaning that no classes from them participate in the training set. The remaining ones contribute some classes to each of the training, validation and test splits of classes, roughly with 70% / 15% / 15% proportions. Two of these datasets, ImageNet and Omniglot, possess a class hierarchy that we exploit in Meta-Dataset. These are described below.
##### ImageNet
While ImageNet is commonly-used, we define a new class split for meta-learning on it and a novel procedure for sampling classes from it during episode creation. Both of these are informed by its class hierarchy, which we describe below.
ImageNet is a dataset comprised of 82,115 ‘synsets’, i.e., concepts of the WordNet ontology, and it provides ‘is-a’ relationships for its synsets, thus defining a DAG over them. In this benchmark, we only use the 1000 synsets that were chosen for the ILSVRC 2012 classification challenge as classes that can appear in our episodes. However, we leverage the ontology DAG for defining a sampling procedure that determines which of these 1000 classes should co-occur in each episode.
For this purpose, we consider a sub-graph of the overall DAG that consists of only the 1000 synsets of ILSVRC-2012 and their ancestors, so these 1000 synsets are all and only the leaves of the DAG. We then further ‘cut’ this sub-graph into three pieces, for the training, validation, and test splits, such that there is no overlap between the leaves of any of these pieces. We selected the synsets ‘carnivore’ and ‘device’ as the roots of the validation and test sub-graphs, respectively. The leaves that are reachable from ‘carnivore’ and ‘device’ form the sets of the validation and test classes, respectively. All of the remaining leaves constitute the training classes. This method of splitting ensures that the training classes (non-carnivore animals) are substantially semantically different from the test classes (inanimate devices such as various tools and instruments). We end up with 712 training, 202 validation and 188 test classes, roughly adhering to the standard 70 / 15 / 15 (%) splits.
##### Omniglot
This dataset is one of the established benchmarks for few-shot classification as mentioned earlier. The commonly used setup, however, flattens and ignores its two-level hierarchy of alphabets and characters. Instead, we take advantage of it to influence how we sample classes for creating episodes, yielding finer-grained within-alphabet classification problems. We also use the original splits proposed in (Lake et al., [2015](#bib.bib9)): (all characters of) the ‘background’ and ‘evaluation’ alphabets are used for training and testing, respectively. However, we removed the 5 smallest alphabets from the ‘background’ set to reserve them for validation.
###
4.2 Episode Sampling
In this section we outline our proposed sampling algorithm for creating more realistic episodes.
Firstly, for ImageNet and Omniglot whose classes are hierarchically organized, we depart from the usual random uniform class selection process by incorporating this additional knowledge into the episode creation. Exploiting class structure can lead to more realistic tasks as, for example, it is arguably fairly unusual to classify frogs from laptops. A more natural task would instead be to classify desks from laptops, since these are more often observed simultaneously.
Further, we allow classes to have different ‘shots’, i.e., numbers of examples in the support set, therefore allowing for imbalanced episodes. Indeed, class imbalance is an inherent property of the world, so it is desirable to examine and improve upon the ability of few-shot learners to cope with it. Consider for example the classes of cat and a very specific species of alligator. There are plausibly more cats in the world than that specific alligator species, so there are more opportunities to learn about the former than the latter class. To obtain realistic imbalance ratios in episodes, we sample the number of examples of each class from a distribution derived from the relative class frequencies in the original dataset for these classes, as outlined later. Additionally, unlike current benchmarks, we allow the support sets to vary in size, both in its number of classes and in the total number of examples.
More concretely, our algorithm for sampling an episode from a given split of a dataset can be broken down into sampling two steps: sampling a set of classes C from the given split and dataset, and sampling support and query sets of examples from those classes.
##### Sampling the episode’s class set
This procedure differs depending on which dataset is chosen. For datasets without a known class organization, we sample the ‘way’ uniformly from the range [5,\textscMAX−CLASSES], where MAX-CLASSES is either 50 or as many as there are available. Then we sample ‘way’ many classes uniformly at random from the requested class split (train, validation or test) of the given dataset. For ImageNet and Omniglot we employ a class-structure-aware procedure, outlined below.
##### ImageNet class sampling
We adopt a hierarchy-aware sampling algorithm for ImageNet, as follows. First, we sample a node uniformly at random from the set of ‘eligible’ nodes of the DAG structure corresponding to the specified split (train, validation or test). An internal node is ‘eligible’ for this selection if it spans at least 5 leaves, but no more than 392 leaves.
The number 392 was chosen because it is the smallest number so that, collectively, all eligible internal nodes span all leaves in the DAG.
Once an eligible node is selected, some of the leaves that it spans will constitute the classes of the episode. Specifically, if the number of those leaves is no greater than 50, we use all of them. Otherwise, we randomly choose 50 of them.
This procedure enables the creation of tasks of varying degrees of fine-grainedness. For instance, if the sampled internal node has a small height, the leaf classes that it spans will represent semantically-related concepts, thus posing a fine-grained classification task. As the height of the sampled node increases, we ‘zoom out’ to consider a broader scope from which we sample classes and the resulting episodes are more coarse-grained.
##### Omniglot class sampling
We sample classes from Omniglot by first sampling an alphabet uniformly at random from the chosen split of alphabets (train, validation or test). Then, the ‘way’ of the episode is sampled uniformly at random using the same restrictions as for the rest of the datasets, but taking care not to sample a larger number than the number of characters that belong to the chosen alphabet. Finally, the prescribed number of characters of that alphabet are randomly sampled. This ensures that each episode presents a within-alphabet fine-grained classification.
##### Sampling the episode’s examples
We first sample the query set size. The query set is class-balanced, reflecting the fact that we care equally to perform well on all classes of an episode. The number of query images per class is computed as:
| | | |
| --- | --- | --- |
| | | |
where C is the set of selected classes and Im(c) denotes the set of images belonging to class c. The min over classes ensures that each class has at least q images to add to the query set, thus allowing it to be class-balanced. The 0.5 multiplier ensures that enough images of each class will be available to add to the support set, and the minimum with 10 prevents the query set from being too large.
Then, we compute the total support set size:
| | | |
| --- | --- | --- |
| | |S|=min{500,∑c∈C⌈βmin{100,|Im(c)|−q}⌉} | |
where β is a scalar sampled uniformly from interval (0,1]. Intuitively, each class on average contributes either all its remaining examples (after placing q of them in the query set) if there are less than 100 or 100 otherwise, to avoid having too large support sets. The multiplication with β enables the potential generation of smaller support sets even when multiple images are available, since we are also interested in examining the very-low-shot end of the spectrum. The ‘ceiling’ operation ensures that each selected class will have at least one image in the support set. Finally, we cap the total support set size to 500.
We are now ready to compute the ‘shot’ of each class. Specifically, the proportion of the support set that will be devoted to class c is computed as:
| | | |
| --- | --- | --- |
| | Rc=exp(αc)|Im(c)|∑c′∈Cexp(α′c)|Im(c′)| | |
where αc is sampled uniformly from the interval [log(0.5),log(2)). Intuitively, the un-normalized proportion of the support set that will be occupied by class c is a noisy version of the total number of images of that class in the dataset Im(c). This design choice is made in the hopes of obtaining realistic class ratios, under the hypothesis that the dataset class statistics are a reasonable approximation of the real-world statistics of appearances of the corresponding classes. The shot of a class c is then set to:
| | | |
| --- | --- | --- |
| | kc=min{⌊Rc∗(|S|−|C|)⌋+1,|Im(c)|−q} | |
which ensures that at least one example is selected for each class, with additional examples selected proportionally to Rc, if enough are available.
After these steps, we complete the episode creation process by choosing the prescribed number of examples of each chosen class uniformly at random to populate the support and query sets.
5 Related Work
---------------
In our experiments, we focused on the evaluation of three meta-learning methods: Matching Networks, Prototypical Networks and fo-MAML. Indeed, they are some of the first from the meta-learning family to have been proposed and are regularly used as the meta-learning baselines to beat in few-shot learning research. That said, in the past two years, there have been several other methods proposed for few-shot learning. Some bear similarity with MAML and correspond to a meta-learner trained to quickly adapt its parameters to various novel tasks (Ravi & Larochelle, [2017](#bib.bib16); Munkhdalai & Yu, [2017](#bib.bib13); Rusu et al., [2019](#bib.bib18); Yoon et al., [2018](#bib.bib27)). Others relate to Prototypical Networks by learning a data representation as well as a compact representation for a classifier of data under that representation (Bertinetto et al., [2019](#bib.bib2); Gidaris & Komodakis, [2018](#bib.bib5); Oreshkin et al., [2018](#bib.bib15); Gidaris & Komodakis, [2018](#bib.bib5)). Methods similar to Matching Networks, in how they classify a novel example by performing comparisons with each individual support set example, were also proposed. These were based on graph neural networks (Satorras & Estrach, [2018](#bib.bib20)), relational networks (Sung et al., [2018](#bib.bib24)) and attentional networks (Mishra et al., [2018](#bib.bib12)). Hence, we believe that the three meta-learning methods evaluated in this work are reasonably representative of current few-shot learning research. That said, there are other methods that less directly relate to the methods above, such as the work of Santoro et al. ([2016](#bib.bib19)) based on a memory-augmented recurrent network (one of the earliest meta-learning approach to few-shot learning). Hence, we also look forward to future work evaluating more alternative methods on Meta-Dataset.
Similarly to our work, Antoniou et al. ([2019](#bib.bib1)) have also proposed improvements to MAML. Coined MAML++, their method is a collection of adjustments, including the use of multiple meta-trained inner loop learning rates, derivative-order annealing, and more. In comparison, our Proto-MAML variant simply changes the expression for the initial output weights in the inner loop and could easily be combined with the recommendations made by MAML++.
Finally, Meta-Dataset bears similarity to the CVPR 2017 Visual Domain Decathlon Challenge, in which contestants were tasked to train a joint system that can perform well on 10 different datasets, many of which are included in our benchmark. At test time, the submitted system classifies examples from the same classes as those available for training. This is unlike Meta-Dataset, which is designed for the few-shot learning scenario where generalization must be achieved for examples of never-before-seen classes.
6 Experiments
--------------
Meta-Dataset does not prescribe a procedure for learning from the training data. In fact, we believe that meta-learning from multiple heterogeneous sources of training classes is an open research problem. In these experiments though, keeping with the spirit of matching training and testing conditions, we trained each model through a series of training episodes that were sampled using the same algorithm as we used for Meta-Dataset’s evaluation episodes, described above. The choice of the dataset from which to sample the next episode was also random uniform. The non-episodic baseline is trained to solve the large classification problem that results from ‘concatenating’ the training classes of all datasets.
Further, we decided to perform validation on (the validation split of) ImageNet only, ignoring the validation sets of the other datasets. The rationale behind this choice is that the performance on ImageNet has been known to be a good proxy for the performance on different datasets. Notably, a common procedure for dealing with a new classification dataset is to finetune ImageNet-pre-trained weights on it, instead of learning new weights from scratch. However, we believe that the choice of the validation procedure used in this setup could benefit from additional research.
| | |
| --- | --- |
| Test Source | Method: Accuracy (%) ± confidence (%) |
| k-NN | Finetune | MatchingNet | ProtoNet | fo-MAML | Proto-MAML |
| ILSVRC | 38.16± | 1.01 | 47.47± | 1.10 | 43.89± | 1.05 | 43.43± | 1.07 | 29.22± | 1.00 | 50.23± | 1.13 |
| Omniglot | 59.40± | 1.31 | 62.97± | 1.39 | 62.44± | 1.25 | 60.41± | 1.35 | 45.42± | 1.61 | 60.65± | 1.40 |
| Aircraft | 44.41± | 0.92 | 56.35± | 1.03 | 50.64± | 0.95 | 48.60± | 0.88 | 33.81± | 0.91 | 54.53± | 0.95 |
| Birds | 45.75± | 0.98 | 61.63± | 1.03 | 56.36± | 1.03 | 63.73± | 1.00 | 39.04± | 1.17 | 69.71± | 1.04 |
| Textures | 61.53± | 0.75 | 67.82± | 0.86 | 65.55± | 0.76 | 62.17± | 0.77 | 50.60± | 0.74 | 66.68± | 0.80 |
| Quick Draw | 46.42± | 1.10 | 50.89± | 1.15 | 50.24± | 1.12 | 50.53± | 0.97 | 24.33± | 1.39 | 49.03± | 1.12 |
| Fungi | 29.91± | 0.93 | 33.01± | 1.06 | 33.66± | 1.00 | 35.95± | 1.09 | 16.36± | 0.86 | 39.04± | 1.03 |
| VGG Flower | 77.23± | 0.74 | 82.30± | 0.85 | 80.21± | 0.74 | 79.47± | 0.81 | 56.01± | 1.22 | 85.78± | 0.80 |
| Traffic Signs | 58.42± | 1.28 | 55.67± | 1.19 | 59.64± | 1.20 | 46.93± | 1.11 | 23.53± | 1.17 | 47.83± | 1.03 |
| MSCOCO | 31.46± | 1.00 | 33.77± | 1.37 | 29.83± | 1.15 | 35.24± | 1.11 | 13.47± | 1.04 | 38.06± | 1.17 |
| Avg. rank | 3.20 | 1.90 | 2.40 | 2.30 | 4.50 | 1.20 |
Table 2: Few-shot classification results on Meta-Dataset using models trained on All datasets.
| | |
| --- | --- |
| Test Source | Method: Accuracy (%) ± confidence (%) |
| k-NN | Finetune | MatchingNet | ProtoNet | fo-MAML | Proto-MAML |
| ILSVRC | 28.46± | 0.83 | 39.68± | 1.02 | 40.81± | 1.02 | 41.82± | 1.06 | 22.41± | 0.80 | 45.48± | 1.02 |
| Omniglot | 88.42± | 0.63 | 85.57± | 0.89 | 75.62± | 1.09 | 78.61± | 1.10 | 68.14± | 1.35 | 86.26± | 0.85 |
| Aircraft | 70.10± | 0.73 | 69.81± | 0.93 | 60.68± | 0.87 | 66.57± | 0.92 | 44.48± | 0.91 | 79.15± | 0.67 |
| Birds | 47.34± | 0.97 | 54.07± | 1.08 | 57.09± | 0.95 | 63.57± | 1.02 | 36.70± | 1.13 | 72.67± | 0.96 |
| Textures | 56.39± | 0.74 | 62.66± | 0.81 | 64.65± | 0.77 | 66.60± | 0.80 | 45.79± | 0.67 | 66.69± | 0.77 |
| Quick Draw | 66.12± | 0.91 | 73.88± | 0.81 | 58.86± | 1.01 | 63.55± | 0.92 | 41.27± | 1.46 | 67.83± | 0.90 |
| Fungi | 38.35± | 1.08 | 31.85± | 1.08 | 34.38± | 1.01 | 37.97± | 1.07 | 14.21± | 0.81 | 44.58± | 1.19 |
| VGG Flower | 73.21± | 0.75 | 77.55± | 0.94 | 82.60± | 0.66 | 84.43± | 0.69 | 61.10± | 1.11 | 88.21± | 0.68 |
| Traffic Signs | 49.84± | 1.23 | 53.07± | 1.13 | 57.90± | 1.16 | 50.60± | 1.02 | 24.03± | 1.08 | 46.38± | 1.03 |
| MSCOCO | 24.29± | 0.92 | 27.71± | 1.20 | 30.20± | 1.13 | 37.58± | 1.14 | 13.63± | 0.96 | 35.12± | 1.20 |
| Avg. rank | 3.2 | 2.8 | 2.8 | 2.2 | 5.2 | 1.5 |
Table 1: Few-shot classification results on Meta-Dataset using models trained on ILSVRC-2012 only.
This common ImageNet-pretraining procedure also inspired us to train variants of each meta-learner in which the embedding function is initialized from the representation g(x) to which the baseline model trained on ImageNet converged to. We find that this initialization helps meta-learners substantially.
We experiment with two architectures: a four-layer convolutional network that is commonly used for few-shot learning, and an 18-layer residual network. All models performed best with the latter. We also tried two different image resolutions: 84, which is the image resolution of the commonly-used ‘mini-ImageNet’ benchmark, and 126. All models performed better with the larger images, except for fo-MAML. Finally, we tuned the learning rate schedule, weight decay, and model-specific hyperparameters. We used ADAM to train all models.
Notably, for fo-MAML and Proto-MAML, we tuned the learning rate of the within-episode training, the number of within-episode training steps, and the number of additional such steps to perform in evaluation episodes (sampled from the validation or test splits). Our best-performing fo-MAML variant used 6 training steps, with a learning rate of 0.01 and no additional steps on evaluation episodes. Interestingly, Proto-MAML preferred the lower learning rate of 0.0001 but took 10 steps to adapt to each training task, and an additional 5 steps (totalling 15) in each validation or test episode.
All other experimental details are included in the source code.
##### Results on Meta-Dataset
| | | |
| --- | --- | --- |
| Ways Analysis
(a) Ways Analysis
| Shots Analysis
(b) Shots Analysis
| Fine-grainedness Analysis (on ImageNet’s train graph)
(c) Fine-grainedness Analysis (on ImageNet’s train graph)
|
Figure 1: Analysis of performance as a function of the episode’s way, shots, and degree of fine-grainedness.
Table [2](#S6.T2 "Table 2 ‣ 6 Experiments ‣ Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few Examples") and Table [2](#S6.T2 "Table 2 ‣ 6 Experiments ‣ Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few Examples") present the results of the evaluation on the test set of each of the 10 datasets. The difference between them is the training source, i.e., the data that the models were trained on, which is (the training classes of) ILSVRC-only, and all datasets, respectively. No classes from Traffic Signs and MSCOCO are used during training, since these datasets have no training split and are reserved for evaluation only. We propose to use the average (over the datasets) rank of each method as our metric for comparison, where smaller is better. A method receives rank 1 if it has the highest accuracy, rank 2 if it has the second highest, and so on. When two methods are ‘tied’ for a position, they both receive the corresponding rank. Both tables demonstrate the superiority of Proto-MAML over the remaining models in Meta-Dataset’s evaluation tasks. The Finetune Baseline notably presents a worthy opponent, while fo-MAML, to our surprise, performs quite poorly on Meta-Dataset.
We also recorded the performance of the different models for the various ways and shots that they encountered during their evaluation on test episodes of Meta-Dataset. This enables us to examine their robustness to these different settings. We show the evaluation results of the methods whose training source was (the training classes of) all datasets. Their ImageNet-only-trained counterparts exhibit the same trends, and we included them in the Appendix instead.
Figure [(a)a](#S6.F1.sf1 "(a) ‣ Figure 1 ‣ Results on Meta-Dataset ‣ 6 Experiments ‣ Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few Examples") plots the accuracy as a function of the episode’s ‘way’. These results reflect what we expected: the more classes a task has, the harder it is. Perhaps more interestingly, Figure [(b)b](#S6.F1.sf2 "(b) ‣ Figure 1 ‣ Results on Meta-Dataset ‣ 6 Experiments ‣ Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few Examples") illustrates the ability of the different models to benefit from larger shots. In particular, for every ‘shot’ of a class in a test episode, we plot the percentage of query examples of that class that are correctly classified (we refer to this as the ‘precision’ of the class). The general trend is not surprising: the more support examples a class has, the easier it is to correctly classify its query examples. However, this plot sheds light on some interesting trade-offs between the different models. In the very-low-shot end of the spectrum, Prototypical Networks and Proto-MAML outshine the other models. However, Prototypical Networks are evidently less capable of improving given more ‘shots’. On the other hand, the Finetune baseline, Matching Networks and fo-MAML improve at a faster rate given more data. Further, we argue that Proto-MAML indeed constitutes a step towards a ‘best of all worlds’ model, since it is the top-performer in the truly few-shot setting, and yet improves upon Prototypical Networks’ ability to benefit from more data. We think that taking additional steps in this direction is an interesting research problem.
##### Fine-grainedness analysis
We had hypothesized that finer-grained tasks (e.g., between dog breeds) are more challenging than coarse-grained ones (e.g., frogs versus laptops). To investigate this, we created binary ImageNet episodes whose two classes are chosen uniformly at random from the DAG’s set of leaves. We then define the degree of coarse-grainedness of a task as the height of the lowest common ancestor of the two chosen leaves, where the height is defined as the length of the longest path from the lowest common ancestor to one of the selected leaves. Larger heights then correspond to coarser-grained tasks. Surprisingly, we did not detect any trend when performing this analysis on the test DAG. The results on the training DAG, though, do seem to indicate that our hypothesis holds to some extent. These results are shown in Figure [(c)c](#S6.F1.sf3 "(c) ‣ Figure 1 ‣ Results on Meta-Dataset ‣ 6 Experiments ‣ Meta-Dataset: A Dataset of Datasets for Learning to Learn from Few Examples"). We conjecture that this may be due to the richer structure of the training DAG, but we encourage further investigation.
7 Conclusion
-------------
We have introduced a new large-scale, diverse, and realistic environment for training and testing meta-learners for the task of few-shot classification. We believe that our exploration of various models on Meta-Dataset has uncovered certain weaknesses of the current state-of-the-art meta-learning methods which allow us to pinpoint interesting directions for future research.
In particular, we view our experiments as the first attempt to meta-learn from a diverse set of sources, and we feel there is plenty of room for improvement. In particular, we don’t always observe a generalization gain from training on all datasets over training on ImageNet only and in fact, in some cases the performance drops. This suggests that our method of consuming training data of different datasets can be improved. Further, through our analysis of the performance as a function of the shots, we discovered that different models perform well on different ends of this spectrum. We argue that our new Proto-MAML variant is a first step towards more robust meta-learners, but we believe that more work is needed on this front.
Generally, this benchmark opens the door to the use of multiple data sources for few-shot learning. Despite having only 10 datasets, developing this benchmark allowed us to explore and identify good practices for a codebase that supports this setting. On the longer term, we thus view Meta-Dataset as only a first step towards the establishment of more challenging benchmarks for few-shot learning research, with increasingly many dataset sources. To move in this direction, future work will likely require considering other domains beyond natural or man-made images (e.g., from the medical domain or from computer graphics simulations). It will also rely on the ability and willingness of the community to continue to release new freely available image classification datasets.
#### Author Contributions
Eleni, Hugo, and Kevin came up with the benchmark idea and requirements.
Eleni developed the core of the project, and worked on the experiment design and management with Tyler and Kevin, as well as experiment analysis.
Carles, Ross, Kelvin, Pascal, Vincent, and Tyler helped extend the benchmark by adding datasets.
Eleni and Vincent contributed the Prototypical Networks and Matching Networks implementations, respectively. Tyler implemented baselines, MAML (with Kevin) and Proto-MAML models, and updated the backbones to support them.
Writing was mostly led Eleni, with contributions by Hugo, Vincent, and Kevin and help from Tyler and Pascal for visualizations.
Pascal and Pierre-Antoine worked on code organization, efficiency, and open-sourcing, Pascal and Vincent optimized the efficiency of the data input pipeline.
Pierre-Antoine supervised the code development process and reviewed most of the changes, Hugo and Kevin supervised the overall direction of the research.
#### Acknowledgments
The authors would like to thank Chelsea Finn for fruitful discussions and advice on tuning fo-MAML and ensuring the correctness of implementation, as well as Zack Nado and Dan Moldovan for the initial dataset code that was adapted. |
270e70fd-227e-4523-8c07-55b413d8bfe1 | trentmkelly/LessWrong-43k | LessWrong | Some Thoughts on Concept Formation and Use in Agents
Note: I wrote this document over a year ago and have decided to post it with minimal edits; it isn't entirely up to date with my current thinking on the subject.
Imagine you enter a room that looks like this:
Despite never having visited this room, you can make inferences about it. If nobody waters the plant in the corner for several weeks, it will probably wilt. The couch is probably the most comfortable place to sit. The paintings probably have signatures that you could look at to determine who created them.
How do you do this? Let’s consider the first example.
When you first walk into this room, your eyes send you raw visual information, which doesn’t by default contain things like “plants” or “books” or “tables.” However, your brain segments out objects in your perception - you perceive object 1 (outlined in red), object 2 (outlined in blue), and object 3 (outlined in purple) as distinct things.
Afterwards, you identify object 2 as a plant. To do this, you need to have a notion of what a plant is, and decide that object 2 matches that notion rather than, say, the notion of “pillow.” After you determine that object 2 is a plant, you can apply information you have about plants to it. For instance, you expect that if you don’t water object 2 for two weeks it will wilt, because this is true of plants in general.
But how does this all work?
In this post I propose a framework that attempts to describe this process, centered around a dynamically constructed “concept graph.”
To start with, let’s suppose the object segmentation described above uses some form of similarity-based clustering. As you look at the room at different times and from different positions, you gather a large amount of raw input data. You recognize which parts of this input data are similar, and cluster these groups of similar data together.
You then create a “summary” of each cluster, and store each one inside a concept, which can be thought of as a container inside your head. Each |
3ad77de4-284c-4f19-b876-b699dcf3cd9a | trentmkelly/LessWrong-43k | LessWrong | Anonymous feedback forms revisited
> In 2011, I added an anonymous feedback form to gwern.net. It has worked well (116 entries) and justified the time it took to set up because it encourages people to correct various problems or tip me off on things. If you have a site, maybe you should add one too.
For the full writeup, see http://www.gwern.net/About#anonymous-feedback |
9f426aae-76ea-46c3-88f7-dab53c65cf6c | trentmkelly/LessWrong-43k | LessWrong | Wikipedia as an introduction to the alignment problem
AI researchers and others are increasingly looking for an introduction to the alignment problem that is clearly written, credible, and supported by evidence and real examples. The Wikipedia article on AI Alignment has become such an introduction.
Link: https://en.wikipedia.org/wiki/AI_alignment
Aside from me, it has contributions from Mantas Mazeika, Gavin Leech, Richard Ngo, Thomas Woodside (CAIS), Sidney Hough (CAIS), other Wikipedia contributors, and copy editor Amber Ace. It also had extensive feedback from this community.
In the last month, it had ~20k unique readers and was cited by Yoshua Bengio.
We've tried hard to keep the article accessible for non-technical readers while also making sense to AI researchers.
I think Wikipedia is a useful format because it can include videos and illustrations (unlike papers) and it is more credible than blog posts. However, Wikipedia has strict rules and could be changed by anyone.
Note that we've announced this effort on the Wikipedia talk page and shared public drafts to let other editors give feedback and contribute.
I you edit the article, please keep in mind Wikipedia's rules, use reliable sources, and consider that we've worked hard to keep it concise because most Wikipedia readers spend <1 minute on the page. For the latter goal, it helps to focus on edits that reduce or don't increase length. To give feedback, feel free to post on the talk page or message me. Translations would likely be impactful. |
3f8a6041-882a-4502-ab83-7d1fc2690f53 | StampyAI/alignment-research-dataset/arxiv | Arxiv | Towards the first adversarially robust neural network model on MNIST
|
1cd223b1-dafc-4abb-a431-1a19d8d2e045 | trentmkelly/LessWrong-43k | LessWrong | Functional Trade-offs
I found this super cool paper from Marco Del Giudice and Bernard J. Crespi called Basic functional trade-offs in cognition: An integrative framework that I thought y'all would like. What follows is largely just a condensed version of their paper where I take the most relevant quotes and paste them here, but I highly recommend reading the full paper.
> This paper takes an initial step toward a general theory of cognitive trade-offs by examining four key properties of goal-directed systems: performance, efficiency, robustness, and flexibility. These properties define a number of basic functional trade-offs that can be used to map the abstract “design space” of natural and artificial cognitive systems. Basic functional trade-offs provide a shared vocabulary to describe a variety of specific trade-offs including speed vs. accuracy, generalist vs. specialist, exploration vs. exploitation, and many others.
> We could not find other examples of properties that were both universal (i.e., would apply to all functional systems) and similarly general (i.e., were not already en- compassed by the basic ones). This assertion does not mean that the classification we propose is fully exhaustive or that it cannot be extended in principle, and we encourage its growth and elaboration.
A map of basic functional trade-offs. Performance, efficiency, robustness, and flexibility are the key properties of all functional systems, including natural and artificial cognitive systems. Straight arrows represent trade-offs between properties; curved arrows represent trade-offs between different aspects of the same property.
Definitions
Performance: "The performance of a system is usually defined as its ability to produce an intended result (or some other roughly equivalent formulation)."
Efficiency: "The efficiency of a system is its ability to perform its function with minimal use of resources."
Robustness: "Robustness is the ability of a system to maintain performance in the face of |
97393dce-c3d6-4f0d-84b1-a669b9fe6a8c | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | [AN #169]: Collaborating with humans without human data
Listen to this newsletter on **[The Alignment Newsletter Podcast](http://alignment-newsletter.libsyn.com/)**.
Alignment Newsletter is a weekly publication with recent content relevant to AI alignment around the world. Find all Alignment Newsletter **[resources here](http://rohinshah.com/alignment-newsletter/)**. In particular, you can look through **[this spreadsheet](https://docs.google.com/spreadsheets/d/1PwWbWZ6FPqAgZWOoOcXM8N_tUCuxpEyMbN1NYYC02aM/edit?usp=sharing)** of all summaries that have ever been in the newsletter.
Please note that while I work at DeepMind, this newsletter represents my personal views and not those of my employer.
SECTIONS
========
**[HIGHLIGHTS](about:blank#HIGHLIGHTS)**
**[TECHNICAL AI ALIGNMENT](about:blank#TECHNICAL_AI_ALIGNMENT)**
**[LEARNING HUMAN INTENT](about:blank#LEARNING_HUMAN_INTENT)**
**[FORECASTING](about:blank#FORECASTING)**
**[MISCELLANEOUS (ALIGNMENT)](about:blank#MISCELLANEOUS_(ALIGNMENT))**
**[NEAR-TERM CONCERNS](about:blank#NEAR-TERM_CONCERNS)**
**[RECOMMENDER SYSTEMS](about:blank#RECOMMENDER_SYSTEMS)**
**[NEWS](about:blank#NEWS)**
HIGHLIGHTS
==========
**[Collaborating with Humans without Human Data](https://arxiv.org/abs/2110.08176)** *(DJ Strouse et al)* (summarized by Rohin): We’ve previously seen that if you want to collaborate with humans in the video game Overcooked, **[it helps to train a deep RL agent against a human model](https://bair.berkeley.edu/blog/2019/10/21/coordination/)** (**[AN #70](https://mailchi.mp/732eaa192df0/an-70-agents-that-help-humans-who-are-still-learning-about-their-own-preferences)**), so that the agent “expects” to be playing against humans (rather than e.g. copies of itself, as in self-play). We might call this a “human-aware” model. However, since a human-aware model must be trained against a model that imitates human gameplay, we need to collect human gameplay data for training. Could we instead train an agent that is robust enough to play with lots of different agents, including humans as a special case?
This paper shows that this can be done with **Fictitious Co-Play** (FCP), in which we train our final agent against a population of self-play agents and their past checkpoints taken throughout training. Such agents get significantly higher rewards when collaborating with humans in Overcooked (relative to the human-aware approach in the previously linked paper).
In their ablations, the authors find that it is particularly important to include past checkpoints in the population against which you train. They also test whether it helps to have the self-play agents have a variety of architectures, and find that it mostly does not make a difference (as long as you are using past checkpoints as well).
**Read more:** **[Related paper: Maximum Entropy Population Based Training for Zero-Shot Human-AI Coordination](https://openreview.net/forum?id=v-f7ifhKYps)**
**Rohin's opinion:** You could imagine two different philosophies on how to build AI systems -- the first option is to train them on the actual task of interest (for Overcooked, training agents to play against humans or human models), while the second option is to train a more robust agent on some more general task that hopefully includes the actual task within it (the approach in this paper). Besides Overcooked, another example would be supervised learning on some natural language task (the first philosophy), as compared to pretraining on the Internet GPT-style and then prompting the model to solve your task of interest (the second philosophy). In some sense the quest for a single unified AGI system is itself a bet on the second philosophy -- first you build your AGI that can do all tasks, and then you point it at the specific task you want to do now.
Historically, I think AI has focused primarily on the first philosophy, but recent years have shown the power of the second philosophy. However, I don’t think the question is settled yet: one issue with the second philosophy is that it is often difficult to fully “aim” your system at the true task of interest, and as a result it doesn’t perform as well as it “could have”. In Overcooked, the FCP agents will not learn specific quirks of human gameplay that could be exploited to improve efficiency (which the human-aware agent could do, at least in theory). In natural language, even if you prompt GPT-3 appropriately, there’s still some chance it ends up rambling about something else entirely, or neglects to mention some information that it “knows” but that a human on the Internet would not have said. (See also **[this post](https://www.alignmentforum.org/posts/PZtsoaoSLpKjjbMqM/the-case-for-aligning-narrowly-superhuman-models)** (**[AN #141](https://mailchi.mp/93d2051b7d70/an-141the-case-for-practicing-alignment-work-on-gpt-3-and-other-large-models)**).)
I should note that you can also have a hybrid approach, where you start by training a large model with the second philosophy, and then you finetune it on your task of interest as in the first philosophy, gaining the benefits of both.
I’m generally interested in which approach will build more useful agents, as this seems quite relevant to forecasting the future of AI (which in turn affects lots of things including AI alignment plans).
TECHNICAL AI ALIGNMENT
======================
LEARNING HUMAN INTENT
----------------------
**[Inverse Decision Modeling: Learning Interpretable Representations of Behavior](http://proceedings.mlr.press/v139/jarrett21a/jarrett21a.pdf)** *(Daniel Jarrett, Alihan Hüyük et al)* (summarized by Rohin): There’s lots of work on learning preferences from demonstrations, which varies in how much structure they assume on the demonstrator: for example, we might consider them to be **[Boltzmann rational](http://www.cs.cmu.edu/~bziebart/publications/maximum-causal-entropy.pdf)** (**[AN #12](https://mailchi.mp/bcb2c6f1d507/alignment-newsletter-12)**) or **[risk sensitive](https://rss2017.lids.mit.edu/static/papers/62.pdf)**, or we could try to **[learn their biases](https://www.alignmentforum.org/posts/xxnPxELC4jLKaFKqG/learning-biases-and-rewards-simultaneously)** (**[AN #59](https://mailchi.mp/b342d1a1bd06/an-59-how-arguments-for-ai-risk-have-changed-over-time)**). This paper proposes a framework to encompass all of these choices: the core idea is to model the demonstrator as choosing actions according to a *planner*; some parameters of this planner are fixed in advance to provide an assumption on the structure of the planner, while others are learned from data. This also allows them to separate beliefs, decision-making, and rewards, so that different structures can be imposed on each of them individually.
The paper provides a mathematical treatment of both the forward problem (how to compute actions in the planner given the reward, think of algorithms like value iteration) and the backward problem (how to compute the reward given demonstrations, the typical inverse reinforcement learning setting). They demonstrate the framework on a medical dataset, where they introduce a planner with parameters for flexibility of decision-making, optimism of beliefs, and adaptivity of beliefs. In this case they specify the desired reward function and then run backward inference to conclude that, with respect to this reward function, clinicians appear to be significantly less optimistic when diagnosing dementia in female and elderly patients.
**Rohin's opinion:** One thing to note about this paper is that it is an incredible work of scholarship; it fluently cites research across a variety of disciplines including AI safety and provides a useful organizing framework for many such papers. If you need to do a literature review on inverse reinforcement learning, this paper is a good place to start.
**[Human irrationality: both bad and good for reward inference](https://arxiv.org/abs/2111.06956)** *(Lawrence Chan et al)* (summarized by Rohin): Last summary, we saw a framework for inverse reinforcement learning with suboptimal demonstrators. This paper instead investigates the qualitative effects of performing inverse reinforcement learning with a suboptimal demonstrator. The authors modify different parts of the Bellman equation in order to create a suite of possible suboptimal demonstrators to study. They run experiments with exact inference on random MDPs and FrozenLake, and with approximate inference on a simple autonomous driving environment, and conclude:
1. **Irrationalities can be helpful for reward inference**, that is, if you infer a reward from demonstrations by an irrational demonstrator (where you know the irrationality), you often learn *more* about the reward than if you inferred a reward from optimal demonstrations (where you know they are optimal). Conceptually, this happens because optimal demonstrations only tell you about what the best behavior is, whereas most kinds of irrationality can also tell you about preferences between suboptimal behaviors.
2. **If you fail to model irrationality, your performance can be very bad**, that is, if you infer a reward from demonstrations by an irrational demonstrator, but you assume that the demonstrator was Boltzmann rational, you can perform quite badly.
**Rohin's opinion:** One way this paper differs from my intuitions is that it finds that assuming Boltzmann rationality performs very poorly if the demonstrator is in fact systematically suboptimal. I would have instead guessed that Boltzmann rationality would do okay -- not as well as in the case where there is no misspecification, but only a little worse than that. (That’s what I found in **[my paper](https://www.alignmentforum.org/posts/xxnPxELC4jLKaFKqG/learning-biases-and-rewards-simultaneously)** (**[AN #59](https://mailchi.mp/b342d1a1bd06/an-59-how-arguments-for-ai-risk-have-changed-over-time)**), and it makes intuitive sense to me.) Some hypotheses for what’s going on, which the lead author agrees are at least part of the story:
1. When assuming Boltzmann rationality, you infer a distribution over reward functions that is “close” to the correct one in terms of incentivizing the right behavior, but differs in rewards assigned to suboptimal behavior. In this case, you might get a very bad log loss (the metric used in this paper), but still have a reasonable policy that is decent at acquiring true reward (the metric used in my paper).
2. The environments we’re using may differ in some important way (for example, in the environment in my paper, it is primarily important to identify the goal, which might be much easier to do than inferring the right behavior or reward in the autonomous driving environment used in this paper).
FORECASTING
-----------
**[Forecasting progress in language models](https://www.alignmentforum.org/posts/tepqESMuRmyhtmDS7/forecasting-progress-in-language-models)** *(Matthew Barnett)* (summarized by Sudhanshu): This post aims to forecast when a "human-level language model" may be created. To build up to this, the author swiftly covers basic concepts from information theory and natural language processing such as entropy, N-gram models, modern LMs, and perplexity. Data for perplexity achieved from recent state-of-the-art models is collected and used to estimate - by linear regression - when we can expect to see future models score below certain entropy levels, approaching the hypothesised entropy for the English language.
These predictions range across the next 15 years, depending on which dataset, method, and entropy level is being solved for; there's an attached **[python notebook](https://github.com/MatthewJBarnett/LanguageModelExtrapolation/blob/master/Human-level%20language%20models.ipynb)** with these details for curious readers to further investigate. Preemptively disjunctive, the author concludes "either current trends will break down soon, or human-level language models will likely arrive in the next decade or two."
**Sudhanshu's opinion:** This quick read provides a natural, accessible analysis stemming from recent results, while staying self-aware (and informing readers) of potential improvements. The comments section too includes some interesting debates, e.g. about the Goodhart-ability of the Perplexity metric.
I personally felt these estimates were broadly in line with my own intuitions. I would go so far as to say that with the confluence of improved generation capabilities across text, speech/audio, video, as well as multimodal consistency and integration, virtually any kind of content we see ~10 years from now will be algorithmically generated and indistinguishable from the work of human professionals.
**Rohin's opinion:** I would generally adopt forecasts produced by this sort of method as my own, perhaps making them a bit longer as I expect the quickly growing compute trend to slow down. Note however that this is a forecast for human-level language models, not transformative AI; I would expect these to be quite different and would predict that transformative AI comes significantly later.
MISCELLANEOUS (ALIGNMENT)
-------------------------
**[Rohin Shah on the State of AGI Safety Research in 2021](https://www.youtube.com/watch?v=_5xkh-Rh6Ec)** *(Lucas Perry and Rohin Shah)* (summarized by Rohin): As in **[previous](https://futureoflife.org/2019/04/11/an-overview-of-technical-ai-alignment-with-rohin-shah-part-1/)** **[years](https://futureoflife.org/2019/04/25/an-overview-of-technical-ai-alignment-with-rohin-shah-part-2/)** (**[AN #54](https://mailchi.mp/3e2f43012b07/an-54-boxing-a-finite-horizon-ai-system-to-keep-it-unambitious)**), on this FLI podcast I talk about the state of the field. Relative to previous years, this podcast is a bit more introductory, and focuses a bit more on what I find interesting rather than what the field as a whole would consider interesting.
**Read more:** **[Transcript](https://futureoflife.org/2021/11/01/rohin-shah-on-the-state-of-agi-safety-research-in-2021/)**
NEAR-TERM CONCERNS
==================
RECOMMENDER SYSTEMS
-------------------
**[User Tampering in Reinforcement Learning Recommender Systems](https://arxiv.org/abs/2109.04083)** *(Charles Evans et al)* (summarized by Zach): Large-scale recommender systems have emerged as a way to filter through large pools of content to identify and recommend content to users. However, these advances have led to social and ethical concerns over the use of recommender systems in applications. This paper focuses on the potential for social manipulability and polarization from the use of RL-based recommender systems. In particular, they present evidence that such recommender systems have an instrumental goal to engage in user tampering by polarizing users early on in an attempt to make later predictions easier.
To formalize the problem the authors introduce a causal model. Essentially, they note that predicting user preferences requires an exogenous, non-observable variable, that models click-through rates. They then introduce a notion of instrumental goal that models the general behavior of RL-based algorithms over a set of potential tasks. The authors argue that such algorithms will have an instrumental goal to influence the exogenous/preference variables whenever user opinions are malleable. This ultimately introduces a risk for preference manipulation.
The author's hypothesis is tested using a simple media recommendation problem. They model the exogenous variable as either leftist, centrist, or right-wing. User preferences are malleable in the sense that a user shown content from an opposing side will polarize their initial preferences. In experiments, the authors show that a standard Q-learning algorithm will learn to tamper with user preferences which increases polarization in both leftist and right-wing populations. Moreover, even though the agent makes use of tampering it fails to outperform a crude baseline policy that avoids tampering.
**Zach's opinion:** This article is interesting because it formalizes and experimentally demonstrates an intuitive concern many have regarding recommender systems. I also found the formalization of instrumental goals to be of independent interest. The most surprising result was that the agents who exploit tampering are not particularly more effective than policies that avoid tampering. This suggests that the instrumental incentive is not really pointing at what is actually optimal, which I found to be an illuminating distinction.
NEWS
====
**[OpenAI hiring Software Engineer, Alignment](https://boards.greenhouse.io/openai/jobs/4143981004?gh_src=4972b51c4us)** (summarized by Rohin): Exactly what it sounds like: OpenAI is hiring a software engineer to work with the Alignment team.
**[BERI hiring ML Software Engineer](https://existence.org/jobs/ml-engineer-seldonian)** *(Sawyer Bernath)* (summarized by Rohin): BERI is hiring a remote ML Engineer as part of their collaboration with the **[Autonomous Learning Lab](https://all.cs.umass.edu/)** at UMass Amherst. The goal is to create a software library that enables easy deployment of the ALL's Seldonian algorithm framework for safe and aligned AI.
**[AI Safety Needs Great Engineers](https://www.alignmentforum.org/posts/YDF7XhMThhNfHfim9/ai-safety-needs-great-engineers)** *(Andy Jones)* (summarized by Rohin): If the previous two roles weren't enough to convince you, this post explicitly argues that a lot of AI safety work is bottlenecked on good engineers, and encourages people to apply to such roles.
**[AI Safety Camp Virtual 2022](http://www.aisafety.camp/)** (summarized by Rohin): Applications are open for this remote research program, where people from various disciplines come together to research an open problem under the mentorship of an established AI-alignment researcher. Deadline to apply is December 1st.
**[Political Economy of Reinforcement Learning schedule](https://perls-workshop.github.io/schedule.html)** (summarized by Rohin): The date for the **[PERLS workshop](https://perls-workshop.github.io/)** (**[AN #159](https://mailchi.mp/9dd2c875c362/an-159building-agents-that-know-how-to-experiment-by-training-on-procedurally-generated-games)**) at NeurIPS has been set for December 14, and the schedule and speaker list are now available on the website.
#### **FEEDBACK**
I'm always happy to hear feedback; you can send it to me, **[Rohin Shah](https://rohinshah.com/)**, by **replying to this email**.
#### **PODCAST**
An audio podcast version of the **Alignment Newsletter** is available. This podcast is an audio version of the newsletter, recorded by **[Robert Miles](http://robertskmiles.com/)**. |
d455e1e6-8006-4ee0-9ef2-e98de7873257 | trentmkelly/LessWrong-43k | LessWrong | How to Host a Find Me Party
For my birthday this year, i invited friends to a party on Saturday but didn't tell them the location. I gave a list of questions they could ask me. Whoever finds me in the fewest questions wins a prize.
Each friend showed up that day excited to tell me about their detective work & misadventures in tracking me down. Everyone had a lot of fun!
I definitely recommend this as a party idea or meetup activity. It was much easier to organize than i expected.
Advice:
* Use this question list:
* I'm at <GPS LOCATION>. Are you within <DISTANCE> of me?
* Send me a photo of you.
* Send me a photo of a tree.
* Send me a photo of a street.
* Send me a photo with your camera pointed straight up.
* You could also add these questions:
* Send me a photo of a restaurant.
* Send me a photo of the tallest thing around.
* Send me a photo of the oddest thing around.
* Send me a photo of your favorite thing around.
* Tell guests when exactly the game starts. I told guests we would play 3 rounds, at 1000, 1300, & 1600 respectively.
* Set reminders for yourself so that you have enough time to travel to the next round's location. You will be distracted.
* Don't let guest expectations get too high about the party aspect. If you plan to be simply sitting on a park bench, tell them to expect a chill hangout after the search. The search is the main focus. Also note that your attention will be divided between guests that have arrived quickly and your phone, as you answer questions from slower-arriving guests.
* Tell guests what to wear. I said: Dress for walking around outdoors and/or casual indoor places.
* Tell guests whether they can bring their children and/or dogs to the secret location.
* Ideally, your secret location should:
* Have a restroom
* Be somewhere guests can step into without paying money or feeling awkward
* My secret locations, for example, were:
* A coffeeshop next to a bookstore
* Outdoor seating at a restaurant (that suit |
612b3dc6-af5d-4745-9bc2-200e61ae8f4d | trentmkelly/LessWrong-43k | LessWrong | Biomedical research, superstars, and innovation
As part of my work for Cognito Mentoring reviewing biomedical research as a career option (not much at the link there right now), I came across an interview with biomedical researcher John Todd of Cambridge University published by 80,000 Hours.
The whole interview is interesting, but one part of it struck me as interesting and somewhat hard to believe:
> John would prefer a good person in his lab to an extra £0.5mn in annual funding. Generally, there are enough grants, so finding good people is a bigger constraint than money.
Here's the full context:
> Our candidate does data analysis in finance, earning over $100,000 per year. They have an Economics degree for Chicago, and an Masters in Financial Engineering from University of California, LA, and reasonable programming skills. They’re planning to do an MD then PhD.
>
> > “This guy looks great. I’d love to hire him.” (when he has his MD, or even before).
>
> > “The MD and programming/statistics combo is lethal. Top of the world. There’s major demand.”
>
> > He probably wouldn’t need to do a PhD, because of the programming. After his MD, he could just apply to a lab. He should go into genomic medicine, which is what I do. Tailored therapeutics or stratified medicine will be played out for major health and economic benefits over the next 30 years. Check out Atul Butte at Stanford. He’s the perfect profile for this guy. He could be the new Butte”
>
>
£0.5mn is about USD 830,000 according to current foreign exchange rates. In other words, John Todd, the interviewee, indicated that a sufficiently good researcher was worth that much. Now, the question was framed in terms of additional funding, rather than reallocation of existing funds. But assuming that the existing funding for the biomedical research lab is at least one order of magnitude greater than the amount (£0.5mn) under discussion, I don't think it matters whether we're talking of using additional funding or reallocating existing funds. Essentially, I |
57ffc0d6-f328-459c-b7e3-30942296fcc4 | trentmkelly/LessWrong-43k | LessWrong | HDBSCAN is Surprisingly Effective at Finding Interpretable Clusters of the SAE Decoder Matrix
*All authors have equal contribution.
This is a short informal post about recent discoveries our team made when we:
1. Clustered feature vectors in the decoder matrices of Sparse Autoencoders (SAEs) trained on the residual stream of each layer of GPT-2 Small and Gemma-2B.
2. Visualized the clusters in 2D after reducing the feature vector space via UMAP
We want to share our thought process, our choices when clustering, and some challenges we faced along the way.
Note: While currently a work in progress, we hope to open-source the clustered features and projections soon, in case others will find it useful.
We would like to acknowledge many people who helped with this endeavor! Professors Shibu Yooseph and Ran Libeskind-Hadas, Alice Woog, Daniel Tan, and many others who offered help and guidance.
TL;DR
1. SAE decoder weight matrix alone can be used to find (some) feature families.
2. While HDBSCAN classifies most features as “noise”, the remaining clusters encapsulate features that activate on similar concepts and tokens in high-dimensions
1. HDBSCAN is not able to find all clusters of features
2. We propose using the Cech Complex to identify persistent sets of features via filtration.
Our primary finding that the decoder weight matrix alone can be used find feature families may help with:
1. Auto-interpretability: It should be easier, cheaper, and more powerful to classify groups of features than find explanations for a single feature.
2. Safety Guardrailing: Interpretable clustering techniques can be used to elucidate the geometry of the SAE feature space. We propose a guardrail framework and discuss the utilization of appropriate abstract spaces, such as Čech complexes, to ensure safety and reliability.
3. Reverse Feature Splitting: We discuss how clustering can be viewed as “reverse splitting” features, finding natural relationships between existing features and grouping them together to find stronger activations or as a sign of feature |
575e433d-f426-401e-a964-dc458cc0ce8a | trentmkelly/LessWrong-43k | LessWrong | Meetup : New Year meetup Frankfurt
Discussion article for the meetup : New Year meetup Frankfurt
WHEN: 11 January 2015 02:00:00PM (+0100)
WHERE: Heerstrasse Frankfurt
We're having our first meetup in the New Year! I'd like to encourage to do some projects together this year (and look at what some or all of us want to change/improve). Our TOP (after a bit of arrival time as usual)
1) New Year Plans and Projects (and regular Goal Tracking)
2) Talk: How can you improve your social skills?
3) Prediction Game
More talks about other topics, especially project ideas, are welcome! I'll edit more details in later. It would be great if you could bring some food. There'll also be (vegan-friendly) food available. If you have any problems attending (social anxiety, wheelchair accessibility, etc), please tell me, I'll try to accommodate for your needs as best as I can! You can reach me under +4917634095760 for further details and/or join our mailing list: https://groups.google.com/forum/#!forum/less-wrong-frankfurt
Discussion article for the meetup : New Year meetup Frankfurt |
8dce9832-9303-45c1-9cd6-330c4b686885 | StampyAI/alignment-research-dataset/alignmentforum | Alignment Forum | [AN #54] Boxing a finite-horizon AI system to keep it unambitious
Find all Alignment Newsletter resources [here](http://rohinshah.com/alignment-newsletter/). In particular, you can [sign up](http://eepurl.com/dqMSZj), or look through this [spreadsheet](https://docs.google.com/spreadsheets/d/1PwWbWZ6FPqAgZWOoOcXM8N_tUCuxpEyMbN1NYYC02aM/edit?usp=sharing) of all summaries that have ever been in the newsletter.
The newsletter now has exactly 1,000 subscribers! It's a perfect time to **take the 3-minute [survey](https://rohinshah.us18.list-manage.com/track/click?u=1d1821210cc4f04d1e05c4fa6&id=4d2ec63190&e=f69d935b62)** if you haven't already -- just think of how you'll be making the newsletter better for all 1,000 subscribers! Not to mention the readers on [Twitter](https://twitter.com/rohinmshah) and the [Alignment Forum](https://www.alignmentforum.org/).
**Highlights**
--------------
[Asymptotically Benign AGI](https://www.alignmentforum.org/posts/pZhDWxDmwzuSwLjou/asymptotically-benign-agi) *(Michael Cohen)*: I'm a bit out of my depth with this summary, but let's give it a shot anyway. The setting: we are not worried about how much compute we use (except that it should be finite), and we would like to build a powerful AI system that can help us with tasks but does not try to influence the world. We'll assume that we can construct a box from which no signals can pass through, except by pressing a specific button that opens a door.
First, the simple version of BoMAI (Boxed Myopic AI). We'll put the AI system **and the operator** in the box, and the operator and the AI system can talk via text message, and the operator can enter rewards. Each episode has a maximum length (hence myopic), and if the operator ends the episode early, all future rewards are set to zero. BoMAI maximizes episodic reward in a manner similar to [AIXI](http://www.hutter1.net/ai/). It has a distribution (initially a speed prior) over all possible time-bounded Turing Machines as possible models that predict observations and rewards. BoMAI uses the maximum a posteriori (MAP) Turing Machine to predict future observations and rewards given actions, searches over all possible sequences of actions for the best one, and executes the first action of that sequence. (See [this comment](https://www.alignmentforum.org/posts/pZhDWxDmwzuSwLjou/asymptotically-benign-agi#ToDYj27wQZWHqWnH6) and its parents for the reason to use the MAP model.)
Intuitively, BoMAI has no incentive to affect anything outside the box: information can only leave the box if the episode ends, but if that happens, reward is guaranteed to be exactly zero. It might deceive the operator into giving it high reward, but any such deception will only *incidentally* affect the outside world, it hasn't been *optimized* to cause any such effect. However, **this is all assuming that the MAP world model is accurate**, which may not be true.
Obviously at initialization our world model will not be accurate: we need to explore. Rather than using something like epsilon-exploration, BoMAI delegates to a human explorer for safety, and reduces exploration over time.
More insidiously, a world model could simulate the world for a year, and then report the value in the computer's memory for "the reward from a year ago". This will usually be a perfect prediction of the reward the operator will give, but if BoMAI tries to *optimize* for it, it has an incentive to try to e.g. cause the operator to build an unboxed AI whose goal is to change BoMAI's memory cell to the maximum value.
These sorts of bad plans can only happen finitely often. If BoMAI executes such a plan, it will then get the actual reward from the operator, see that it differs from the predicted reward, and update against the world model for making a bad prediction. Eventually the correct model will become more likely. But ideally this sort of thing wouldn't happen at all.
Intuitively, such world models would have to simulate the world past the end of the episode, incurring a speed penalty relative to the correct model. However, they could use that simulation to save on the computation needed for future predictions. To make sure that they do incur a speed penalty, we need to make it so that sometimes their simulations are wrong and can't be reused. So, at the end of each episode, we sample from a source of true randomness, and use that to determine whether or not the operator is given a drug that causes retrograde amnesia. In that case, world models that simulate the future will sometimes be wrong about how that decision plays out, and any simulation they did in the wrong branch is useless and leads to a speed penalty.
With a few non-trivial assumptions, we can prove that asymptotically, BoMAI will do at least as well as the human explorer at accumulating reward, and the MAP world model's rewards do not incentivize BoMAI to affect the outside world.
**Rohin's opinion:** I think the idea of putting the operator in the box with the AI system is very interesting: with previous attempts at boxing, the human operator talking to the AI system was an obvious glaring hole in the box. In this setting, the only information escaping from the box is the fact that the operator has not yet chosen to end the episode.
I am generally skeptical of intuitive reasoning about what can or can't be done by Turing Machines using extreme amounts of computation. There are *lots* of comments on the post that debate specifics of this. This usually cashes out as a debate about the assumptions in the proof. But it's also worth noting that the theorem is asymptotic, and allows for arbitrarily bad behavior early on. We might still expect good behavior early on for the reasons laid out in the proof, but it's not implied by the theorem, even if the assumptions hold.
**Previous newsletters**
========================
[AI Safety workshop at IJCAI 2019](https://www.ai-safety.org/) *(Huáscar Espinoza et al)*: Previously ([AN #49](https://mailchi.mp/efed27be268a/alignment-newsletter-49)), I said the paper submission deadline was April 12. Either I made a mistake, or the deadline has been extended, because the actual deadline is May 12.
**Technical AI alignment**
==========================
### **Technical agendas and prioritization**
AI Alignment Podcast: An Overview of Technical AI Alignment: [Part 1](https://futureoflife.org/2019/04/11/an-overview-of-technical-ai-alignment-with-rohin-shah-part-1/) and [Part 2](https://futureoflife.org/2019/04/25/an-overview-of-technical-ai-alignment-with-rohin-shah-part-2/) *(Lucas Perry and Rohin Shah)*: In this podcast, I go through a large swath of research agendas around technical AI alignment. The first part is more of a description of what research agendas exist, who works on them, and what they are trying to do, while the second part delves more into the details of each approach. I'd strongly recommend listening to them if you're trying to orient yourself in the technical AI safety landscape.
Topics covered include [embedded agency](https://www.alignmentforum.org/s/Rm6oQRJJmhGCcLvxh), [value learning](https://www.alignmentforum.org/s/4dHMdK5TLN6xcqtyc), [impact regularization methods](https://medium.com/@deepmindsafetyresearch/designing-agent-incentives-to-avoid-side-effects-e1ac80ea6107) ([AN #49](https://mailchi.mp/efed27be268a/alignment-newsletter-49)), [iterated amplification](https://www.alignmentforum.org/s/EmDuGeRw749sD3GKd), [debate](https://blog.openai.com/debate/) ([AN #5](https://mailchi.mp/0ae5d69de63b/alignment-newsletter-5)), [factored cognition](https://www.alignmentforum.org/posts/DFkGStzvj3jgXibFG/factored-cognition) ([AN #36](https://mailchi.mp/6751e45fbb48/alignment-newsletter-36)), [robustness](https://www.alignmentforum.org/posts/E2aZ9Xwdz3i2ghPtn/techniques-for-optimizing-worst-case-performance) ([AN #43](https://mailchi.mp/768a8130013f/alignment-newsletter-43)), interpretability (no canonical link, but [activation atlases](https://distill.pub/2019/activation-atlas/) ([AN #49](https://mailchi.mp/efed27be268a/alignment-newsletter-49)) is an example), [comprehensive AI services](https://www.fhi.ox.ac.uk/reframing/) ([AN #40](https://mailchi.mp/b649f32b07da/alignment-newsletter-40)), [norm following](https://arxiv.org/abs/1804.04268) ([AN #3](https://mailchi.mp/e6a23da2760e/alignment-newsletter-3?e=f665c7b50a)), and [boxing](https://www.alignmentforum.org/posts/pZhDWxDmwzuSwLjou/asymptotically-benign-agi) (this newsletter).
### **Learning human intent**
[Extrapolating Beyond Suboptimal Demonstrations via Inverse Reinforcement Learning from Observations](https://arxiv.org/abs/1904.06387) *(Daniel S. Brown, Wonjoon Goo et al)* (summarized by Cody): This paper claims to demonstrate a technique by which an agent learning from a demonstrator's actions can learn to outperform that demonstrator on their true reward, rather than, in the way of imitation learning or behavioral cloning, just mimicking the demonstrator under the assumption that the demonstrator's performance is optimal (or at least near-optimal). The key structural innovation of the paper is to learn using pairs of ranked trajectories and learn a neural network-based reward function based on correctly predicting which will be higher. This allows the model to predict what actions will lead to higher and lower reward, and to extrapolate that relationship beyond the best demonstration. When an agent is then trained using this reward model as it's ground truth reward, it's shown to be capable of outperforming the demonstrator on multiple tested environments, including Atari. An important distinction compared to some prior work is the fact that these rankings are collected in an off-policy manner, distinguishing it from [Deep RL from Human Preferences](https://openai.com/blog/deep-reinforcement-learning-from-human-preferences/) where rankings are requested on trajectories generated as an agent learns.
**Cody's opinion:** Seems like potentially a straightforward and clever modification to a typical reward learning structure, but a bit unclear how much of the performance relative to GAIL and BCO derives from T-REX's access to suboptimal demonstrations and subtrajectories giving it more effective training data. It does intuitively seem that adding examples of what poor performance looks like, rather than just optimal performance, would add useful informative signal to training. On a personal level, I'm curious if one implication of an approach like this is that it could allow a single set of demonstration trajectories to be used in reward learning of multiple distinct rewards, based on different rankings being assigned to the same trajectory based on the reward the ranker wants to see demonstrated.
**Rohin's opinion:** It's pretty interesting that the [Deep RL from Human Preferences](https://openai.com/blog/deep-reinforcement-learning-from-human-preferences/) approach works even with off-policy trajectories. It seems like looking at the *difference* between good and bad trajectories gives you more information about the true reward that generalizes better. We saw similar things in our work on [Active Inverse Reward Design](https://arxiv.org/abs/1809.03060) ([AN #24](https://mailchi.mp/d7b5059d64ed/alignment-newsletter-24)).
[End-to-End Robotic Reinforcement Learning without Reward Engineering](https://sites.google.com/view/reward-learning-rl/home) *(Avi Singh et al)* (summarized by Cody): This paper demonstrates an approach that can learn to perform real world robotics tasks based not on example trajectories (states and actions) but just a small number (10) of pixel-level images of goal states showing successful task completion. Their method learns a GAN-like classifier to predict whether a given image is a success, continually adding data sampled from the still-learning policy to the set of negative examples, so the model at each step needs to further refine its model of success. The classifier, which is used as the reward signal in learning the policy, also makes use of a simple active learning approach, choosing the state its classifier is most confident is success and querying a human about it on fixed intervals, ultimately using less than 75 queries in all cases.
**Cody's opinion:** This is a result I find impressive, primarily because of its interest in abiding by sensible real-world constraints: it's easier for humans to label successful end states than to demonstrate a series of actions, and the number of queries made was similarly pragmatically low.
### **Reward learning theory**
[AI Alignment Problem: “Human Values” don’t Actually Exist](https://www.lesswrong.com/posts/ngqvnWGsvTEiTASih/ai-alignment-problem-human-values-don-t-actually-exist) *(avturchin)*
### **Verification**
[Optimization + Abstraction: A Synergistic Approach for Analyzing Neural Network Robustness](http://arxiv.org/abs/1904.09959) *(Greg Anderson et al)*
**AI strategy and policy**
==========================
[Global AI Talent Report 2019](https://jfgagne.ai/talent-2019/) *(Jean-Francois Gagné)*: This report has a lot of statistics on the growth of the field of AI over the last year.
[FLI Podcast: Why Ban Lethal Autonomous Weapons?](https://futureoflife.org/2019/04/02/fli-podcast-why-ban-lethal-autonomous-weapons/) *(Ariel Conn, Emilia Javorsky, Bonnie Docherty, Ray Acheson, and Rasha Abdul Rahim)*
**Other progress in AI**
========================
### **Reinforcement learning**
[How to Train Your OpenAI Five](https://openai.com/blog/how-to-train-your-openai-five/) *(OpenAI)*: [OpenAI Five](https://blog.openai.com/openai-five/) ([AN #13](https://mailchi.mp/8234356e4b7f/alignment-newsletter-13)) has now beaten the Dota world champions 2-0, after training for 8x longer, for a total of 800 petaflop/s-days or 45000 years of Dota self-play experience. During this insanely long training run, OpenAI grew the LSTM to 4096 units, added buybacks to the game, and switched versions twice. Interestingly, they found it hard to add in new heroes: they could bring a few new heroes up to 95th percentile of humans, but it didn't look like they would train fast enough to reach pro level. This could be because the other heroes were already so capable that it was too hard to learn, since the new heroes would constantly be beaten. The resulting team was also able to play cooperatively with humans, even though they had never been trained with humans.
As usual, I like [Alex Irpan's thoughts](https://www.alexirpan.com/2019/04/14/openai-finals.html). On the Dota side, he found Five's reaction times more believable, but was disappointed by the limited hero pool. He also predicted that with [OpenAI Five Arena](https://arena.openai.com/), which allowed anyone to play either alongside Five, or against Five, at least one of the *many* teams would figure out a strategy that could reliably beat Five. He was right: while Five had a 99.4% win rate, one team was able to beat it [10 times in a row](https://arena.openai.com/#/results), another beat it thrice in a row, and two teams beat it twice in a row.
**Rohin's opinion:** In this era of scaling up compute via parallelism, it was quite surprising to see OpenAI scaling up compute simply by training for almost a year. That feels like one of the last resorts to scale up compute, so maybe we're seeing the limits of the trend identified in [AI and Compute](https://blog.openai.com/ai-and-compute/) ([AN #7](https://mailchi.mp/3e550712419a/alignment-newsletter-7))?
Back when OpenAI Five beat a strong team in their [Benchmark](https://blog.openai.com/openai-five-benchmark-results/) ([AN #19](https://mailchi.mp/4b19d2caa5a9/alignment-newsletter-19)), I and a few others predicted that the team would be able to beat Five after playing a few games against it. I think this prediction has been somewhat validated, given that four teams figured out how to beat a much stronger version of the bot. Of course, humans played over 7000 games against Five, not just a few, so this could be that enough random search finds a weakness. Still, I'd expect pros to be able to do this in tens, maybe hundreds of games, and probably this would have been much easier at the time of the Benchmark.
The underlying model here is that Dota has an extremely large space of strategies, and neither Five nor humans have explored it all. However, pros have a better (lower-dimensional) representation of strategy space (concepts like "split-push") that allow them to update quickly when seeing a better opponent. I don't know what it would take to have AI systems learn these sorts of low-dimensional representations, but it seems key to having AI systems that can adapt quickly like humans can.
**Read more:** [Vox: AI triumphs against the world’s top pro team in strategy game Dota 2](https://www.vox.com/2019/4/13/18309418/open-ai-dota-triumph-og)
### **Deep learning**
[Do we still need models or just more data and compute?](https://staff.fnwi.uva.nl/m.welling/wp-content/uploads/Model-versus-Data-AI-1.pdf) *(Max Welling)*: This is a response to [The Bitter Lesson](http://www.incompleteideas.net/IncIdeas/BitterLesson.html) ([AN #49](https://mailchi.mp/efed27be268a/alignment-newsletter-49)), that emphasizes the importance of data in addition to compute. It brings up a number of considerations that seem important to me, and is worth reading if you want to better understand my position on the bitter lesson.
[Semantic Image Synthesis with Spatially-Adaptive Normalization](https://arxiv.org/pdf/1903.07291.pdf) *(Taesung Park et al.)* (summarized by Dan H): This paper shows how to create somewhat realistic images specified by semantic segmentation maps. They accomplish this by modifying batch normalization. Batch normalization modifications can be quite powerful for image generation, even enough to [control style](https://arxiv.org/abs/1703.06868). Their modification is that normalization is a direct function of the semantic segmentation map throughout the network, so that the semantic segmentation map is readily available to each ResBlock. Visualizations produced by this method are [here](https://nvlabs.github.io/SPADE/).
**News**
========
[SafeML Workshop: Accepted Papers](https://sites.google.com/view/safeml-iclr2019/accepted-papers): The camera-ready papers from the SafeML workshop are now available! There are a lot of good papers on robustness, adversarial examples, and more that will likely never make it into this newsletter (there's only so much I can read and summarize), so I encourage you to browse through it yourself.
[Why the world’s leading AI charity decided to take billions from investors](https://www.vox.com/future-perfect/2019/4/17/18301070/openai-greg-brockman-ilya-sutskever) *(Kelsey Piper)*
**Read more:** [OpenAI LP](https://openai.com/blog/openai-lp/) ([AN #52](https://mailchi.mp/1e757d9b05cb/alignment-newsletter-52)) |
cd1cfc54-5159-4ec1-bc0d-e182577c973b | trentmkelly/LessWrong-43k | LessWrong | Thoughts on minimizing designer baby drama
I previously wrote a post hypothesizing that inter-group conflict is more common when most humans belong to readily identifiable, discrete factions.
This seems relevant to the recent human gene editing advance. Full human gene editing capability probably won't come soon, but this got me thinking anyway. Consider the following two scenarios:
1. Designer babies become socially acceptable and widespread some time in the near future. Because our knowledge of the human genome is still maturing, they initially aren't that much different than regular humans. As our knowledge matures, they get better and better. Fortunately, there's a large population of "semi-enhanced" humans from the early days of designer babies to keep the peace between the "fully enhanced" and "not at all enhanced" factions.
2. Designer babies are considered socially unacceptable in many parts of the world. Meanwhile, the technology needed to produce them continues to advance. At a certain point people start having them anyway. By this point the technology has advanced to the point where designer babies clearly outclass regular babies at everything, and there's a schism between "fully enhanced" and "not at all enhanced" humans.
Of course, there's another scenario where designer babies just never become widespread. But that seems like an unstable equilibrium given the 100+ sovereign countries in the world, each with their own set of laws, and the desire of parents everywhere to give birth to the best kids possible.
We already see tons of drama related to the current inequalities between individuals, especially inequality that's allegedly genetic in origin. Designer babies might shape up to be the greatest internet flame war of this century. This flame war could spill over in to real world violence. But since one of the parties has not arrived to the flame war yet, maybe we can prepare.
One way to prepare might be differential technological development. In particular, maybe it's possib |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.