Ilya Sutskever: OpenAI Meta-Learning and Self-Play | MIT Artificial General Intelligence (AGI)

Summary

Ilya Sutskever在演讲中探讨了深度学习和人工智能的最新进展，包括元学习、强化学习和自我对弈。他强调了深度学习的重要性，并解释了其工作原理，特别是在找到最佳神经网络方面。他还讨论了强化学习中的探索问题，以及如何从失败中学习。Sutskever提出了通过自我对弈来提高智能体能力的观点，并展望了未来人工智能可能的发展方向，包括社会化智能体的创建和目标传达的重要性。

Takeaways

• 🤖 AI领域的深度学习和元学习正在取得显著进展，但仍然存在许多挑战。
• 🧠 深度学习的成功部分基于找到数据中规律性的最短时间程序，尽管这在计算上是不可行的。
• 🔄 反向传播是深度学习中的关键算法，尽管它与大脑的工作原理不同。
• 🧠 神经网络通过迭代地对基础神经网络进行小的改变来满足数据约束。
• 🤔 强化学习是一个评估智能体在复杂随机环境中实现目标能力的框架。
• 🎯 强化学习算法的目标是最大化预期奖励，但实际应用中可能需要考虑奖励的方差。
• 🤖 元学习（meta learning）是一个有前景的领域，尽管目前还不完全成熟。
• 📈 通过模拟和元学习，AI可以在模拟环境中学习并将其知识迁移到物理机器人上。
• 🔄 自我对弈（self-play）是一种新兴的AI研究方法，可以促进智能体的快速发展。
• 🌐 语言理解和生成模型是AI领域的关键挑战之一，目前还有很大的提升空间。
• 🚀 未来AI的发展将对社会产生深远影响，确保AI的目标与人类价值观一致是一个重要的政治问题。

Q & A

Ilya Sutskever 在人工智能领域有哪些重要贡献？

Ilya Sutskever 是 OpenAI 的联合创始人和研究总监，他在深度学习和人工智能领域有着重要影响。他的工作在过去五年中被引用超过四万六千次，他是一些深度学习和人工智能领域重大突破思想的关键创意和推动力量。

深度学习为何能够工作？

深度学习之所以能够工作，是因为它基于一个数学理论，即如果你能找到一个在数据上表现非常好的最短程序，那么你就能实现最佳的泛化。这意味着如果你能从数据中提取出所有规律性并将其编码到程序中，那么你就能做出最好的预测。尽管理论上存在这样的程序，但目前的工具和理解水平还无法找到这样的最佳短程序，因为这个问题在计算上是不可行的。

什么是元学习，它的潜力和挑战是什么？

元学习是指训练算法学习如何学习的过程。它的潜力在于能够创建能够快速适应新任务的系统，这是非常具有前景的。然而，元学习面临的挑战是训练和测试分布必须相同，而在现实世界中，新的测试任务往往与训练数据有所不同，因此元学习在这种情况下可能会遇到困难。

强化学习是如何工作的？

强化学习是一个框架，用于评估智能体在复杂随机环境中实现目标的能力。智能体通过与环境交互，尝试新的行为，并根据结果调整其策略。如果结果超出预期，智能体将在未来采取更多这样的行动。

自我对弈在人工智能中有什么作用？(self-play)？

自我对弈是一种训练人工智能的方法，通过让AI系统在没有外部数据的情况下自我竞争和学习，从而提高其性能。这种方法在围棋游戏的AlphaGo Zero和OpenAI的Dota 2机器人中都取得了显著的成功。

如何将模拟中训练的策略应用到物理机器人上？

为了将模拟中训练的策略应用到物理机器人上，可以通过在模拟器中引入大量变化来使策略具有适应性。然后在物理环境中部署策略时，策略需要通过尝试和错误来适应新环境的物理特性。

如何理解深度神经网络的训练过程？

深度神经网络的训练过程可以看作是解决电路搜索问题的过程。通过迭代地对神经网络的基础进行小的调整，直到其预测满足数据的要求。这个过程非常深刻，因为它是通过梯度下降将信息从方程推入参数中，从而满足所有方程。

如何理解强化学习中的探索问题？

在强化学习中，探索是指智能体在不知道如何行动时尝试新的行为。探索的重要性在于，只有通过尝试和偶尔获得奖励，智能体才能学习。因此，设计奖励函数以提供逐步的奖励增量是至关重要的，这样即使系统表现不佳，它也能获得奖励并从中学习。

如何通过观察其他智能体来推断它们的目标和策略？

通过观察其他智能体的行为，我们可以推断它们的目标和策略。这是人类与其它动物相比在规模和范围上非常不同的一个方面。在非竞争环境中，观察和模仿他人的行为可以是一种有效的学习策略。

如何确保人工智能系统的目标与人类的期望一致？

确保人工智能系统的目标与人类的期望一致是一个技术问题，但也是一个重大的政治问题。这需要在技术层面上开发出能够理解和执行人类指定目标的算法，同时在更广泛的社会层面上，需要解决如何确定正确的目标，以及如何确保系统按照这些目标行动的问题。

Outlines

00:00

🤖 人工智能与深度学习的突破

本段落介绍了Ilya Sutskever，他是OpenAI的联合创始人和研究总监。他的工作在深度学习和人工智能领域产生了巨大影响，过去五年的论文被引用超过四万六千次。他被认为是深度学习领域一些重大突破背后的主要创意和推动力。

05:01

🧠 深度学习的原理与工作方式

这部分讨论了深度学习的原理，解释了为什么深度学习能够工作。提到了一个数学理论，即如果能找到一个最短的程序来处理数据，那么就能达到最佳的泛化效果。然而，找到这样的程序在计算上是不可行的。尽管如此，通过使用反向传播和小型电路，我们能够找到解决问题的方法。

10:01

🔄 强化学习与智能体的目标达成

这一部分探讨了强化学习，它是一个评估智能体在复杂随机环境中实现目标能力的框架。强化学习算法通过多次运行智能体并计算其平均奖励来工作。这部分还讨论了强化学习的一个关键问题：如何在没有外部奖励的情况下，从观察中确定奖励。

15:02

🧠 元学习的概念与应用

元学习的目标是开发能够学习如何学习的算法。这种方法涉及在多个任务上训练系统，使其能够快速解决这些任务。元学习的一个成功案例是通过MIT的研究，快速识别字符。另一个例子是Google的神经架构搜索，通过解决小问题来找到能够解决大问题的神经网络架构。

20:04

🤹‍♂️ 自我对弈与智能体的进化

自我对弈是一种让智能体通过相互竞争来提高自身能力的方法。这种方法的一个经典例子是TD Gammon，它通过自我对弈学习并最终击败了世界背棋冠军。自我对弈的一个关键优势是它能够创建一个不断变化的环境，使智能体不断面临新的挑战。

25:05

🎮 强化学习在游戏中的应用

这部分讨论了强化学习在游戏领域的应用，特别是OpenAI的Dota 2机器人。通过自我对弈，这些机器人在短短几个月内从随机游戏玩法进步到世界锦标赛水平。这种快速的进步表明，自我对弈是一种强大的学习机制。

30:06

🤖 人工智能的目标传达与对齐问题

这部分探讨了如何向智能体传达目标，以及如何确保智能体的行为符合我们的期望。提出了一种通过人类反馈来训练智能体的方法，即通过观察人类选择更好的行为来训练智能体。这种方法需要解决的技术问题和政治问题都非常复杂。

35:07

🌟 人工智能的未来展望

最后一部分对人工智能的未来进行了展望，特别是关于如何让智能体在软件环境中学习并将其应用于现实世界的任务。讨论了自我对弈环境的潜力，以及如何通过持续训练和适应新环境来提高智能体的能力。

Keywords

💡深度学习

深度学习是一种通过构建多层神经网络来模拟人类大脑处理信息的机器学习技术。在视频中，Ilya Sutskever提到深度学习是人工智能领域的一个重要突破，它通过在数据中找到最短的程序来实现最佳泛化。例如，通过深度学习，神经网络能够在多次迭代中逐渐改善其预测能力，从而在复杂任务（如图像识别或自然语言处理）上取得显著成果。

💡元学习

元学习，也称为学习如何学习，是指设计算法和模型以便它们能够在多个任务上快速学习和适应的过程。在视频中，Ilya Sutskever讨论了元学习的概念，强调了通过在多个任务上训练系统，使其能够快速掌握新任务的重要性。元学习的目标是创建能够自我改进和适应新环境的智能系统。

💡强化学习

强化学习是一种机器学习范式，其中智能体通过与环境的交互来学习如何实现目标。智能体根据其行为获得的奖励或惩罚来调整其策略。在视频中，强化学习被描述为一个框架，用于评估智能体在复杂随机环境中实现目标的能力。

💡自我对弈

自我对弈是一种训练策略，其中两个相同的算法或模型相互竞争，以此来提高性能。这种方法可以使模型通过自我竞争和学习对手的策略来不断进步。在视频中，Ilya Sutskever提到了自我对弈的概念，并举例说明了如何通过自我对弈来训练模型，从而在没有外部数据的情况下达到高水平的表现。

💡神经网络

神经网络是一种模仿人类大脑神经元结构的计算模型，用于识别模式和处理复杂的数据。在视频中，Ilya Sutskever讨论了神经网络在深度学习中的核心作用，以及如何通过调整网络中的权重来学习数据中的规律。

💡反向传播

反向传播是一种高效计算神经网络中权重梯度的方法，用于训练神经网络。它通过从输出层向输入层反向传递误差来更新网络中的权重。在视频中，虽然反向传播在生物大脑中的直接对应尚不清楚，但它是深度学习中解决电路搜索问题的关键算法。

💡策略梯度

策略梯度是强化学习中的一种方法，它直接对策略（即智能体选择动作的规则）进行参数化，并计算这些参数的梯度以优化策略。在视频中，Ilya Sutskever提到策略梯度算法允许智能体尝试新动作，并根据结果调整动作选择的概率。

💡Q学习

Q学习是强化学习中的一种值函数方法，它通过学习动作值函数（Q函数）来估计在给定状态下采取特定动作的期望回报。在视频中，Ilya Sutskever提到Q学习算法可以从任何数据中学习，包括非自身生成的数据，这使得它在学习过程中更加稳健。

💡目标设定

目标设定是指为智能体或系统定义一个明确的目标或任务，以便它可以通过学习和适应来实现该目标。在视频中，Ilya Sutskever讨论了如何通过人类反馈来设定智能体的目标，以及如何确保智能体的目标与人类的价值观和意图一致。

💡合作

合作是指多个智能体或个体为了共同的目标或利益而协同工作。在视频中，Ilya Sutskever提到合作是生物进化中的一个重要概念，并且暗示了在多智能体环境中，合作可能是一个成功的策略。

Highlights

Ilya Sutskever, co-founder and research director of OpenAI, discusses the impact of deep learning and AI.

Sutskever's work has been cited over 46,000 times, showcasing his influence in the field.

The concept of finding the shortest program to achieve the best generalization in machine learning is introduced.

The computational intractability of finding the best short program is discussed, highlighting the limitations of current AI tools.

The discovery that small circuits can be optimized using backpropagation is highlighted as a foundational AI principle.

Neural networks are likened to parallel computers capable of complex computations and reasoning.

The potential of reinforcement learning to achieve goals in complex environments is explored.

Meta-learning, or learning to learn, is introduced as a promising but not fully realized concept.

The importance of representation learning and unsupervised learning for identifying high-level states in meta-learning is emphasized.

Self-play is presented as a powerful method for training AI agents, leading to rapid increases in competence.

The potential societal implications and alignment issues of superintelligent AI are discussed.

The use of human feedback for training AI, such as through reinforcement learning, is highlighted.

The limitations of meta-learning, particularly the requirement for training and test distributions to match, are noted.

The potential for AI to develop social skills, language, and other human-like traits through multi-agent interaction is speculated.

The importance of continuous learning and adapting to new environments is emphasized for AI agents.

The challenges of conveying goals to AI agents and aligning their objectives with human values are discussed.

The potential for AI to develop strategies and behaviors through self-organization and interaction with other agents is explored.

The role of complexity theory in understanding the problems that AI can solve and the limitations thereof is examined.

The future of generative language models and the importance of scaling up current models is discussed.

The potential use of evolutionary strategies in reinforcement learning for small, compact objects is mentioned.

The necessity of accurate physical world modeling and simulation for training AI agents is questioned.

Transcripts

welcome back to 6 SZ row 99 artificial

general intelligence today we have Ilya

sutskever co-founder and research

director of open AI he started in the

amel group in Toronto Geoffrey Hinton

then at Stanford with an jiaying

co-founded DNN research for three years

as a research scientist at Google brain

and finally co-founded open AI citations

aren't everything

but they do indicate impact and his work

recent work in the past five years has

been cited over forty six thousand times

he has been the key creative intellect

and driver behind some of the biggest

breakthrough ideas in deep learning and

artificial intelligence ever

so please welcome Ilya alright thanks

for the introduction Lex

alright thanks for coming to my talk I

will tell you about some work we've done

over the past year on on meta learning

and software open AI and before I dive

into some of the more technical details

of the work I want to spend a little bit

of time talking about deep learning and

why it works at all in the first place

which I think it's actually not a

self-evident saying that they should

work one fact it's actually a fact it's

a mathematical theory that you can prove

is that if you could find the shortest

program the does very very well on your

data then you will achieve the best

generalization possible with a little

bit of modification you can turn it into

a precise theorem

and on a very intuitive level it's easy

to see what it should be the case if you

have some data and you're able to find a

shorter program which generates this

data then you've essentially extracted

all the all conceivable regularity from

this data into your program and then you

can use these objects to make the best

predictions possible like if if you have

data which is so complex but there is no

way to express it as a shorter program

then it means that your data is totally

random there is no way to extract any

regularity from it whatsoever now there

is little known mathematical theory

behind this and the proofs of these

statements actually not even that hard

but the one minor slight disappointment

is that it's actually not possible at

least given today's tools and

understanding to find the best short

program that explains or generates or

solves your problem given your data this

problem is computationally intractable

the space of all programs is a very

nasty space small changes to your

program result in massive changes in the

behavior of the program as it should be

it makes sense you have a loop you

change the inside of the loop of course

you get something totally different so

the space of programs is so hard at

least given what we know today search

there seems to be completely off the

table well if we give up on shorts on

short programs what about small circuits

well it turns out that we are lucky it

turns out that when it comes to small

circuits you can just find the best

small circuits circuits that solves the

problem using back propagation and this

is the miraculous fact on which the rest

of AI stands it is the fact but then you

have a circuit and you impose

constraints on your circuits on your

circuit using data you can find the way

to satisfy these constraints these

constraints using that problem by

iteratively making small changes

to the base of your neural network until

its predictions satisfy the data what

this means is that the computational

problem that so the back propagation is

extremely profound it is circuit search

now we know that you can solve it solve

it always but you can solve it sometimes

and you can solve it at those times

where we have a practical data set it is

easy to design artificial data sets for

which you cannot find the best neural

network but in practice that seems to be

not a problem you can think of training

a neural network as solving a neural

equation in many cases where you have a

large number of equation terms like this

f of X I theta equals y I so you got

your parameters and they represent all

your degrees of freedom and you use

gradient descent to push the information

from these equations into the parameters

satisfy them all and you can see that

the neural network let's say one with 50

layers is basically a parallel computer

that is given 50 time steps to run and

you can do quite a lot with a 15 with 50

time steps of a very very powerful

massively parallel computer so for

example I do I think it is not widely

known that you can learn to sort sort n

n bit numbers using a modestly sized

neural network with just two hidden

layers which is not bad it's not

self-evident especially since we've been

taught that sorting requires log n

parallel steps with the neural network

you can sort successful using only two

parallel steps so there's some things

like an arm is going on now these are

parallel steps of threshold threshold

neurons so they're doing a little bit

more work let's answer to the mystery

but if you've got 50 such layers you can

do quite a bit of logic quite a bit of

reasoning all inside the neural network

and that's why it works

given the data we are able to find the

best neural network and because the

neural network is deep because it can

run computation inside of its act inside

of its layers the best neural network is

worth finding because that's really what

you need you need something you need the

model class which is worth optimizing

but it also needs to be optimizable and

deep neural networks satisfy both of

these constraints and this is why

everything works this is the basis on

which everything else resides now I want

to talk a little bit about reinforcement

learning so reinforcement learning is a

framework it's a framework of evaluating

agents in their ability to achieve goals

and complicated stochastic environments

you've got an agent which is plugged

into an environment as shown in the

figure right here and for any given

agent you can simply run it many times

and compute its average reward now the

thing that's interesting about the

reinforcement learning framework is that

there exist interesting useful

reinforcement learning algorithms the

framework existed for a long time it

became interesting once we realized that

good algorithms exist now these are

there are perfect algorithms but they

are good enough to do interesting things

and all you want the mathematical

problem is one where you need to

maximize the expected reward now one

important way in which the reinforcement

learning framework is not quite complete

is that it assumes that the reward is

given by the environment you see this

picture the agent sends an action while

the reward sends it an observation in a

both the observation and the reward

backwards that's what the environment

communicates back

the way in which this is not the case in

the real world is that we figure out

what the reward is from the observation

we reward ourselves we are not told

environment doesn't say hey here's some

negative reward it's our interpretation

over census that lets us determine what

the reward is and there is only one real

true reward in life and this is

existence or nonexistence and everything

else is a corollary of that so well what

should our agent be you already know the

answer should be a neural network

because whenever you want to do

something dense it's going to be a

neural network and you want the agent to

map observations to actions so you let

it be parametrized with a neural net and

you apply learning algorithm so I want

to explain to you how reinforcement

learning works this is model free

reinforcement learning the reinforcement

learning has actually been used in

practice everywhere but it's also deeply

it's very robust it's very simple it's

also not very efficient so the way it

works is the following this is literally

the one sentence description of what

happens in short try something new add

randomness directions and compare the

result to your expectation if the result

surprises you if you find that the

results exceeded your expectation then

change your parameters to take those

actions in the future that's it this is

the fool idea of reinforcement learning

try it out see if you like it and if you

do do more of that in the future and

that's it that's literally it this is

the core idea now it turns out it's not

difficult to formalize mathematically

but this is really what's going on

if in a neural network in a regular

neural network like this you might say

okay what's the goal

you run the neural network you get an

answer you compare it to the desired

answer and whatever difference you have

between those two you send it back

to change the neural network that's

supervised line in reinforcement

learning you run in your own network you

add a bit of randomness to your action

and then if you like the result your

randomness turns into the desired target

in effect so that's it trivial now math

exists without explaining what these

equations mean the point is not really

to derive them but just to show that

they exist there are two classes of

reinforcement learning algorithms one of

them is the policy gradient where

basically what you do is that you take

this expression right there the sum of

expected we work the sum of rewards and

it just crunched through the derivatives

you expand the terms iran you do some

algebra and you get a derivative and

miraculously the derivative has exactly

the form that i told you which is try

some actions and if you like them

increase the log probability of the

actions that we truly follows from the

math it's very nice when the intuitive

explanation has a one-to-one

correspondence to what you get in the

equation even though you have to take my

word for it if you are not familiar with

it that's that equation at the top now

there is a different class of

reinforcement learning algorithms which

is a little bit more difficult to

explain it's called the Q learning based

algorithms they are a bit less stable a

bit more sample efficient and it has the

property that it can learn not only from

the data generated by the actor but from

any other data as well so it has it has

some rope but it has different

robustness profile which would be a

little bit important but it's only going

to be a technicality so yeah this is the

own policy of policy distinction but

it's a little bit technical so if you

find this hard to understand don't worry

about it if you already know this then

you already know it so now what's the

potential for enforcement learning

wasn't it promised what is it actually

why should we be excited about it

now there are two reasons the

reinforcement learning algorithms of

today already useful and interesting and

especially if you have a really good

simulation of your world you could train

agents to do lots of interesting things

but what's really exciting is if you can

build a super amazing sample efficient

out of reinforcement learning algorithm

we just give it a tiny amount of data

and the algorithm just crunches through

it and extracts every bit of entropy out

of it in order to learn in the fastest

way possible now today our algorithms

are not particularly efficient they are

data inefficient but as our field keeps

making progress this will change next I

want to dive into the topic of meta

learning the goal of meta learning so

meta learning is a beautiful idea that

doesn't really work but it kind of works

and it's really promising too it's

another promising idea so what's the

dream we have some learning algorithms

perhaps you could use those learning

algorithms in order to learn to learn

I'd be nice if we could learn to learn

so how would you do that you will take a

system which you train it not on one

task but on many tasks and you ask you

that it learns to solve these tasks

quickly and that may actually be enough

so here's how it looks like here's how

most traditional metal earning look

works like it looks like you have a

model which is a big neural network what

what you do is that you treat every

instead of training cases you have

training tasks and instead of test cases

you have test tasks so your input may be

instead of just your current test case

it would be all the information about

the new T above the test tasks plus the

test case and you'll try to output the

prediction reaction for that test case

so basically you say yeah I'm going to

give you your ten examples as part of

your input to your model figure out how

to make the best use of them it's a

really

straightforward idea u-turn the neural

network into the learning algorithm by

turning a training task into a training

case so training to ask a constraining

case this is meta learning just one

sentence and so they've been several

success stories which I I think are very

interesting one of the success stories

of meta learning is learning to

recognize characters quickly so they've

been a dataset

produced by MIT by lake corral and this

is a data set we have a large number of

different handwritten characters and

people have been able to train extremely

strong meta learning system for this

desk another successful another very

successful example of meta learning is

in that of neural architecture search by

is openly from google where they found

neural architecture that solved one

problem well small problem and then you

could generalize and then if you

successfully solve large problems as

well so this is the kind of the the

small number of bits meta learning is

that when you learn the architecture or

maybe even learn a program small program

or learning algorithm which you apply to

new tasks so this is the other way of

doing meta learning so anyway but the

point is what's happening what's really

happening in meta learning in most cases

is that you turn a training task into a

training case and pretend this is

totally normal

normal deep learning that's it this is

the entirety of meta learning everything

else suggests minor details next I wanna

dive in so now that I've finished the

introduction section I want to start

discussing different work by different

people from opening I and I want to

start by talking about hindsight

experience replay it's been a large

effort by and recurvature all to develop

a learning algorithm for reinforcement

learning

that doesn't solve just one task but it

solves many tasks and it learns to make

use of its experience in a much more

efficient way and I want to discuss one

problem in reinforcement learning it's

actually I guess a set of problems which

all related to each other at one really

important thing you need to learn to do

is to explore you're in that you start

out in an environment you don't know

what to do what do you do so one very

important thing that has to happen is

that you must get rewards from time to

time if you try something and you don't

get rewards then how can you learn so

said that's the kind of the crux of the

problem how do you learn and relatedly

is there any way to meaningfully benefit

from your ex from the experience from

your attempts to from from your failures

if you try to achieve a goal and you

fail can you still learn from it you

tell you instead of asking your

algorithm to achieve a single goal you

want to learn a policy that can achieve

a very large family of goals for example

instead of reaching one state you want

to learn a policy that reaches every

state of your system and what's the

implication anytime you do something you

achieve some state so let's suppose you

say I want to achieve state a I try my

best and I end up achieving state B I

can either conclude well that was

disappointing I haven't learned almost

anything I'm still have no idea how to

cheat how to achieve state aid but

alternatively I can say well wait a

second I've just reached a perfectly

good state which is B can I learn how to

achieve state B from my attempt to

achieve state a an answer is yes you can

and it just works and I just want to

point out this is the one case there is

a small subtlety here which may be

interesting to those of you who are very

familiar with on Part B the distinction

between on policy and off policy when

you try to achieve a you are on you're

doing on policy learning for

reaching the state a but you're doing

off policy learning for it in the state

be because you would take different

actions if you would actually try to

reach they'd be so that's why it's very

important that the algorithm you use

here can support of policy learning but

that's a minor technicality at the crux

the crux of the idea is you make the

problem easier by ostensibly making it

harder by training a system which can

which aspires to reach to learn to reach

every state to learn to achieve every

goal to learn to master its environment

in general you build a system which

always learn something it learns from

success as well as from failure because

if it tries to do one thing one thing

and it does something else

it now has training data for how to

achieve that something else I want to

show you a video of how this thing works

in practice so one challenge in

reinforcement learning systems is the

need to shape the reward so what does it

mean

it means that at the beginning of the

system at the start of learning then the

system doesn't know much it will

probably not achieve your goal and so

it's important that you design your

reward function to give it gradual

increments to make it smooth and

continuous so that even when the system

is not very good it achieves the goal

now if you give your state your system a

very sparse reward where the reward is

achieved only when you reach a final

state then it becomes very hard for

normal reinforcement learning algorithms

to solve a problem because naturally you

never get the reward so you never learn

no reward means no learning but here

because you learn from failure as well

as from success

this is this problem simply doesn't

occur and so this is this is nice I

think you know let's let's look at the

videos a little bit more like it's nice

how this is it confidently and

energetically moves the little green

buck to its target and here's another

one

you

okay so we can skip the it works on

spawn on the face if you do it on

physical robot as well but we can skip

it so I think the point is that the

hindsight experience replay algorithm is

directionally correct because you want

to make use of all your data and not

only a small fraction of it now one huge

question is where do you get the high

level states where do the high level

states come from because in the work of

showing you so far

the system is asked to achieve low level

States so I think one thing it will

become very important for this kind

approaches is representation learning

and unsupervised learning figure out

what are the rights what are the right

states what's the state space of goals

that's worth achieving now I want to go

through some real meta learning results

and I'll show you a very simple way of

doing seem to reel from simulation to

the physical robot with meta learning

and this is where my pain growl was an a

and encouraged a really nice intern

project in 2017 so I think we can agree

that in the domain of robotics it would

be nice if you could train your policy

in simulation and then somehow this

knowledge would carry over to the

physical robot now we can build we can

build simulators that are okay but they

can never perfectly match the real world

unless you want to have an insanely slow

simulator and the reason for that is

that it turns out that simulating freaky

simulating contacts is super hard and I

heard somewhere correct me if I'm wrong

that simulating friction is np-complete

I'm not sure but it's like stuff like

that so your simulation is just not

going to match reality there will be

some resemblance but that's it

how can we

address this problem and I want to show

you one simple idea so let's say one

thing once one thing that would be nice

is that if you could learn a policy

learn a policy that would quickly adapt

itself to the real world well if you

want to learn a policy that can quickly

adapt we need to make sure it has

opportunities to adapt during training

time so what do we do instead of solving

a problem in just one simulator we add a

huge amount of variability to the

simulator we say we will randomize the

friction so we will randomize the masses

the length of the different objects and

their I guess M dimensions so you try to

randomize physics they simulate in lots

of different ways and then importantly

you don't tell the policy how you

randomized it so what is it going to do

then you take your policy and you put it

in an environment then says well this is

really really tough I don't know what

the masses are and I don't know what the

frictions are I need to try things out

and figure out what the friction is as I

get it responses from the environment so

you're building you you learn a certain

degree of adaptability into the policy

and it actually works

let's want to show you this is what

happens when you just strain a policy in

simulation and deploy it on the physical

robot and here the goal is to bring the

hockey puck towards the red dot and you

will see that it will struggle and the

reason it struggles is because of the

systematic differences between the

simulator and the real physical robot so

I can even the basic movement is

difficult for the policy because the

assumptions are violated so much so if

you do the training as I discussed we

train a recurrent neural network policy

which learns to quickly infer properties

of the simulator in order to accomplish

the task you can then give it the real

thing

the real physics and it will do much

better so now this is not a perfect

technique but it's definitely very

promising it's promising whenever you

are able to sufficiently randomize the

simulator so it's definitely very nice

to see the closed-loop nature of the

policy you consider it would push the

hockey puck and would correct it

very very gently to bring it to the goal

yeah so that that was cool so that was

very that was a cool application of meta

learning I want to discuss one more

application of meta learning which is

learning a hierarchy of actions and this

was work done by France at all actually

kept in France the ancient who did it

was in high school I mean he wrote this

paper so one thing that would be nice is

if reinforcement learning was

hierarchical if instead of simply taking

micro actions you've had some kind of

little subroutines that you could deploy

maybe the term subroutine is a little

bit too crude but if you had some idea

of which action primitives are worth

starting with now no one has been able

to to get actually like real value add

from curricula reinforcement learning

yet so far all the really cool results

all the really convincing is also

reinforcement learning do not use it

that's because we haven't quite figured

out what's the right way for

reinforcement learning for her ocular

reinforcement learning

I just want to show you one very simple

approach where you use meta-learning to

learn to learn a hierarchy of actions so

here's what you do you have in this

specific work you have a certain yeah

let's say you have a certain number of

low-level primitives let's say you have

two ten of them and you have a

distribution of tasks and your goal is

to learn low level primitives such that

when they are used inside a very brief

run of some reinforcement learning

algorithm you will make as much progress

as possible so the idea is you want to

get the greatest amount of progress you

want to learn policies that result in

the great story you want to learn

primitives that result in the greatest

amount of progress is possible when used

inside learning so this is a meta

learning setter because any distribution

of tasks and here we've had if we've had

a little maze here the distribution of a

mazes and in this case the little bug

learned three policies which move it in

its fixed direction and as a result of

having this hierarchy you're able to

solve problems really fast but only when

the hierarchy is correct

so horican reinforcement learning is

still working progress and this was an

and this work is an interesting proof

point of how Haruko reinforcement could

be like how heretical reinforcement

learning could be like if it worked now

I want to just spend one slide

addressing the limitations of high

capacity method learning the specific

limitation is that the training test

distribution has to be equal to the test

test distribution and I think this is a

real limitation because in reality you

the new test that you want to learn do

in some ways being fundamentally

different from anything you've seen so

far so for example if you go to school

you learn lots of useful things but then

they go to work only a fraction of this

of the things that you've learned

carries over you need to learn if you

need quite a few more things from

scratch so metal owning would struggle

with that because it really assumes that

the Train the training data is that the

distribution over the training task has

to be equal to the distribution of the

test tasks that's the limitation I think

that as we develop better algorithms for

being robust when the test tasks outside

of the distribution of the training

tasks the metal on would work much

better now I want to talk about self

play the links of play is a very cool

topic that's starting to get attention

only now and I want to start by

reviewing very old work called TD gammon

it's back from all the way from 1992 so

it's 26 years old now it was done by

Jerry to cero so this work is really

incredible because it has so much

relevance today what they did basically

they said okay let's take two neural

networks and let them let them play

against each other let them play

backgammon against each other and let

them in tray let them be trained

particularly so it's a super-modern

approach and you would think this was a

paper from 2017 except that then you

look at this plot it shows that you only

have ten hidden units twenty hidden

units forty and eighty for the different

M colors where you notice that the

largest neural network works best so in

some ways not much has changed and this

is the evidence

and in fact they were able to beat the

world champion in backgammon and they

were able to discover new strategies

that the best human a backgammon players

weren't ever not noticed and they've

determined that the strategy discovered

by TD gammon actually better

so that's pure self play with cue

learning which is which remained dormant

until the DQ and work with Atari mid

mind so now other examples of self play

include alphago zero which was able to

learn to beat the world champion and go

without using any external data

whatsoever another result of this vein

is by open AI which is our dota 2 BOTS

which was able to build the world

champion on the 1v1 version of the game

and so I want to spend a little bit of

time talking about the allure of self

play and why I think it's exciting so

one important problem that's a that

that's that we must face as we try to

build truly intelligent systems is what

is the task what are we actually

teaching the systems to do and one very

attractive attribute of self play is

that the agents create the environment

by virtue of the agent acting in the

environment the environment becomes

difficult for the other agents and you

can see here an example of an iguana

interacting with snakes that try to eat

it

unsuccessfully this time so we can see

what will happen in a moment the iguana

strains best and so the fact you have

this arms race between the snakes and

the iguana

motivates their development potentially

without bound and this is what happens

in effect in but in biological evolution

now interesting work in this direction

was done in 1994 but Carl says there is

a really cool video on YouTube by Carl

seems you should check it out which

really kind of shows all the work that

he's done and here you have a little

competition between agents where you

evolved both the behavior and their

morphology when you when the agents is

trying to gain possession of a green

cube and so you can see that the agents

create the challenge for each other and

that's why they need to develop so one

thing that we did and this is work by

advance a little from open ai is we said

okay well can we demonstrate some

unusual results in self play that would

really convince us that there is

something there so what we did here is

that we created a small a small ring and

you have these two humanoid figures and

their goal is just to push each other

outside the ring and they don't know

anything about wrestling they don't know

anything about standing your balance in

each other they don't know anything

about centers of gravity all they know

is that if you don't do a good job then

your competition is going to do a better

job now one of the really attractive

things about self play is that you

always have an opponent that's roughly

as good as you are in order to learn you

need to sometimes win and sometimes lose

but you can't always win sometimes you

must fail sometimes you must succeed so

let's see what will happen here yeah so

it was able to do so the green humanoid

was able to block the ball in a Cell in

a well balanced self play environment

petition is always level no matter how

good you are or how bad you are you have

a competition that makes it exact

exactly of exactly the right challenge

for you on one thing here so this video

shows transfer learning it takes a

little wrestling humanoid and you take

its friend away and you start applying a

big large random forces on it and you

see if it can maintain its balance and

the answer turns out to be but yes it

can because it's been trained against an

opponent it pushes it and so that's why

even if it doesn't understand where the

fresh force is being applied on it it's

still able to balance itself so this is

one potentially attractive feature of

subway environments that you could learn

a certain broad set of skills although

it's real hard to control the square the

skills will be and so the biggest open

question with this research is how do

you learn agents in a software

environment such that they do whatever

they do but then they are able to solve

a battery of tasks that is useful for us

that is explicitly specified externally

yeah I also want to want to highlight

one attribute of self play environments

that we've observed in our dota BOTS and

that is that we've seen a very rapid

increase in the competence of the bots

so over the period over the course of

maybe five months we've seen the bots go

from playing totally randomly all the

way to the world champion and the reason

for that is that once you have a self

play environment if we put compute into

it you turn it into data self play

allows you to turn compute into data and

I think you will see a lot more of that

as being an extremely important thing to

be able to turn compute into essentially

data generalization simply because the

speed of neural net processors will

increase very dramatically over the next

few years so neural net cycles will be

cheap and it will be important to make

use of this new of newly-found

overabundance of cycles

I also want to talk a little bit about

the endgame of the self approach so one

thing that we know about the human brain

is that it has increased in sized fairly

rapidly over the past two million years

my theory the reason I think it happened

is because our ancestors got to a point

where the thing that's most important

for your survival is your standing in

the tribe and less the tiger and the

lion once the most important thing is

how you deal with those other things

which have a large brain then it really

helps to have a slightly larger brain

and I think that's what happened and

there exists at least one paper from

science which supports this point of

view so apparently there has been

convergent evolution between social apps

and social Birds even though in terms of

various behaviors even though the

divergence in evolutionary timescale

between humans and birds has occurred a

very long time ago and humans and humans

apes and humans apes and birds have very

different brain structure so I think

what should happen if we succeed if we

successfully follow the path of this

approach is that you should create a

society of agents which will have

language and theory of mind negotiation

social skills trade economy politics

justice system all these things should

happen inside the multi-agent

environment and it will also be some

alignment issue of how do you make sure

that the agents we learn behave in a way

that we want now I want to make a

speculative digression here which is I

want to make the following observation

if you believe that this kind of society

of agents is a plausible place where

truly where the fuller fully general

intelligence will emerge and if you

accept that our experience with the dota

BOTS we've seen a very rapid increase in

competence will carry over once all the

details are right if you assume both of

these conditions then it should follow

that we should see a very rapid increase

in the competence of our agents as they

live in the Society of agents so now

that we've talked about a potentially

interesting way of increasing the

competence and teachings of an agent's

social skills and language and a lot of

things that actually exist in humans as

well we want to talk a little bit about

how you convey goals to agents and the

question of the main goal to eight calls

to agents is just a technical problem

but it will be important because it is a

lot more likely than not that the agents

of evil train will eventually be

dramatically smarter than us and this is

work by the opening eye safety team by

Paul Christiana at all and others so I'm

just going to show you this video which

basically explains how the whole thing

works you there is some behavior looking

for and you the human gets to see pairs

of behaviors and you simply click on the

one that looks better and after a very

modest number of clicks you can get this

little simulated leg to do back flips

and there

go picking out the back flips and in

this to get this specific behavior it

took about 500 clicks by human

annotators the way it works is that you

take all the so this is a very data

efficient reinforcement learning

algorithm but it is efficient in terms

of rewards and not in terms of the

environment interactions so what you do

here is that you take all the clicks so

you've got your here is one B here which

is better than other you fit a reward

function a numerical reward function to

those clicks so you want to fit a reward

function which satisfies those clicks

clicks and you optimize this reward

function with reinforcement learning and

it actually works so this requires 500

bits of information you've also been

able to train lots of Atari games using

several thousand bits of information so

in all these cases you had human and

human annotators or human judges just

like in the previous slide looking at

the pairs of trajectories and clicking

on the one that they thought was better

and here's an example of an unusual goal

where this is a car racing game but the

goal was to ask the the agent to train

the white car drive right behind the

orange car so it's a different goal and

it was very straightforward to

communicate this goal using this

approach so then to finish off alignment

is a technical problem it has to be

solved but of course the determination

of the correct goals we want array

assistance the systems to have will be a

very challenging political problem and

on this note I want to thank you so much

for your attention and I just want to

say that will be a happy hour at

Cambridge Brewing Company at 8:45 if you

want to chat more about AI and other

topics please come by I think that

deserves an applause

so back propagation is a or neural

networks of bio-inspired but back

propagation doesn't look as though it's

what's going on in the brain because

signals in the brain go one direction

down the axons whereas back propagation

requires the errors to be propagated

back up the the wires so can you just

talk a little bit about that whole

situation where it looks as the brain is

doing something a bit different than our

highly successful algorithms our

algorithm is going to be improved once

we figure out what the brain is doing or

is the brain really sending signals back

even though it's got no obvious way of

doing that what's what's happening in

that area so that's a great question so

first of all I'll say that the true

answer is that the honest answer is that

I don't know but I have opinions and so

I'll say two things

but first of all given that look if you

agree if we agree like so rather it is a

true fact the back propagation solves

the problem of circuit search this

problem feels like an extremely

fundamental problem and for this reason

I think that it's unlikely to go away

now you also write that the brain

doesn't obviously do back propagation

although they've been multiple proposals

of how it could be how it could be doing

them for example there's been a work by

Tim little crap and others where they've

shown that if you use that it's possible

to learn a different set of connections

but can be used for the backward pass

and that can result in successful

learning now the reason this hasn't been

like really pushed to the limit by

practitioners is because they say well I

got TF to the gradients I'm just not

going to worry about it but you are

right this is an important issue and you

know one of two things is going to

happen so my personal opinion is that

back propagation is just going to stay

with us till the very end and will

actually build fully human level and

beyond systems before we understand how

the brain does what it does so that's

what I believe but of course it is a

difference that has to be acknowledged

okay thank you do you think it was a

fair matchup for the dota bot and that

person given the constraints of the

system so I'd say that like the biggest

advantage computers have in games like

this like one of the big advantages is

that they obviously have a better

reaction time although in DotA in

particular the number of clicks per

second over the top players is fairly

small which is different from Starcraft

so in Starcraft stuff up is a very

compact mechanically heavy game because

of a large number of units and so the

top players that is click all the time

in DotA every player controls just one

hero and so that greatly reduces the

total number of actions they need to

make now still precision matters I think

that will discover that but what I think

it'll really happen is if you'll

discover that computers have the

advantage in any domain or rather every

domain not yet so do you think that the

emergent behaviors from the agent were

actually kind of directed because the

constraints already kinda in place like

so it was kind of forced discover those

or do you think that like that was

actually something quite novel that like

wow it actually discovered these on its

own like you didn't actually am biased

towards constraining it so it's

definitely discover new strategies and I

can share an anecdote where our tester

we have a probe which would test the

bots and he played against for a long

time and the bots would do all kinds of

things against the player the human

player which were effective then at some

point that Pro decided to play against

the better plot Pro and he decided to

imitate one of the things that the bot

was doing and this image but by

imitating if he was able to defeat a

better pro so I think I think the

strategy discovers are real and so like

it means that like this very real

transformative Tran you know I would say

I think what that means is that he

because the strategies discovered by the

bot of the humans it means that we like

a fundamental game plays deeply related

for a long time now I've heard that the

objective of reinforcement learning is

to determine a policy that chooses an

action to maximize the expected reward

which is what you said earlier would you

ever want to look at the standard

deviation of possible rewards does that

even make sense yeah I mean I think for

sure I think it's a really application

dependent one of the reasons to maximize

the expected reward it's because it's

easier to design algorithms for it

so you write down this equation the

formula you do a little bit of

derivation you get something which

amounts to a nice-looking algorithm now

I think there exist like really there

exist applications where you'd never

want to make mistakes and you want to

work on the standard deviation as well

but in practice it seems that the just

looking at the expected reward covers a

large fraction of the B the situation as

you'd like to apply this door Thanks

we talked last week about motivations

and that has a lot to do with the

reinforcement and some of the ideas is

that the our motivations are actually

connection with others and cooperation

and I'm wondering if they're thrown off

and I understand it's very popular to

have the computers play these

competitive games but is there any use

in like having an agent self play

collaboratively collaborative games Yeah

right that's an extremely good question

I don't think one place from which we

can get some inspiration is from the

evolution of cooperation

like I think cooperation like we

cooperate ultimately because it's much

better for you the person to be

cooperative than not and so I think what

should happen

if you have a sufficiently open-ended

game then cooperation will be the

winning strategy and so I think we will

get cooperation whether we like it or

not Hey

you mentioned the complexity of this

simulation of friction I was wondering

if you feel that there exists open

complexity theoretic problems relevant

to relevant to AI or whether it's just a

matter of finding good approximations

that humans of the types of problems

that humans tend to solve yeah so

complexity theory well like at a very

basic level we know that whatever

algorithm we gonna run is going to run

fairly efficiently on some hardware so

that puts a pretty strict upper bound

and the true complexity of the problems

we're solving but by definition we are

solving problems which aren't too hard

in a complexity theoretic sense now it

is also the case that many of the

problems so while the overall thing that

we do is not hard from a complexity

theory makes sense and indeed humans

cannot solve np-complete problems in

general it is true that many of the like

optimization problems that we pose to

our algorithms are intractable in the

general case starting from a neural net

optimization itself it is easy to create

a family of data sets for a neural

network with a very small number of

neurons such that find a global optimum

is np-complete and so how do we avoid it

well we just try gradient descent anyway

and somehow it works but without

question like we cannot we do not solve

problems which are truly intractable so

I mean I hope this answer the question

hello

it seems like an important sub-problem

on the path towards AGI will be

understanding language and the state of

generative language modeling right now

is pretty abysmal what do you think are

the most productive research

trajectories towards generative language

models so

I'll first say that you are completely

correct that the situation with language

is still far from great although

progress has been made even without any

particular innovations beyond models

that exist today simply scaling up

models that exist today on larger

datasets is going to go surprisingly far

not even large datasets but larger and

deeper models for example if you trained

a language model be the thousand layers

and it's the same layer I think it's

gonna be a pretty amazing language model

like we don't have the cycles for it yet

but to think it will change very soon

now I also agree with you that there are

some fundamental things missing in a

current understanding of deep learning

which prevent us from really solving the

problem that we want so I think one of

these problems one of the things that's

missing is that or that seems like

patently wrong is the fact that we train

a model then you stop training the model

and you freeze it even though it's the

training process where the magic really

happens but the magic is that if you

think about it like the training process

is the true general part of the whole of

the whole of the whole story because you

tends to flow code doesn't care which

data set to optimize it just says

whatever just give me the data set I

don't care which one solve I'll sew them

all

so like the ability to do that feels

really special and I think we are not

using it at test time like it's hard to

speculate about like things which you

don't know the answer but all I'll say

is that simply train bigger deeper

language models you'll go surprisingly

far scaling up but also doing things

like training a test them and inference

the test time I think would be another

important boosts the performance hi

thank you for the talk so it seems like

right now another interesting approach

to solving reinforcement learning

problems could be to go for the

evolutionary roots using evolutionary

strategies and although they have they

their cave Hut's I wanted to know if I'd

open a I particularly you're working on

something related and what are what is

your general opinion on them

so like at present I believe that

something evolutionary strategies is not

great for reinforcement learning I think

that normal reinforcement learning

algorithms especially with big policies

are better but I think if you want to

evolve a small compact object like like

a piece of code for example I think that

would be a place where this would be

seriously was considering but this all

you know evolving a beautiful piece of

code is a cool idea hasn't been done yet

so still a lot of work to be done before

we get there hi thank you so much for

coming my question is you mentioned what

is the right go is a political problem

so I'm wondering if you can elaborate a

bit on that and also what do you think

would be their approach for us to maybe

get there well I can't I can't really

comment too much because all the

thoughts that you know we have we now

have a few people who are thinking about

this full-time at opening I I don't have

enough of a super strong opinion to say

anything too definitive all I can say at

the very high level is given the size

like if you go into the future whenever

soon or late you know whenever it's

going to happen when you build a

computer which can do anything better

than a human it will happen because the

brain is physical the impact on society

is going to be completely massive and

overwhelming it's it's very difficult to

imagine even if you try really hard and

I think what it means is that people who

care a lot and that's what I was

alluding to the fact that this will be

something that many people who care

about strongly and like as the impact

increases gradually with self-driving

cars more automation I think we will see

a lot more people care do we need to

have a very accurate model of the

physical world and then simulate that in

order to have these agents that can

eventually come out into the real world

and do something approaching you know

human level intelligence tasks that's a

very good question so I think if that

were the case

be in trouble and I am very certain that

it could be avoided so specifically the

real answer has to be that look you

learn the problem so we learn to

negotiate you learn to persist you not a

lots of different useful life lessons in

the simulation and yes you learn some

physics too but then you go outside to

the real world

and you have to start over to some

extent because many of you are deeply

held assumptions will be false in one of

the goals so what was that's one reasons

I care so much about never stopping

training you've accumulated your

knowledge now we go into an environment

for some of your assumptions of valid

you continue training you try to connect

the new data to your old data and this

is an important requirement from our

algorithms which is already met to some

extent but it will have to be met a lot

more so that you can take the partial

knowledge if you've acquired then go in

a new situation learn some more

literally the example of you go to

school ballon useful things then you go

to work it's not a perfect it's not you

know you pour your four years of CS and

undergrad is not going to fully prepare

you for whatever it is you need to know

it work

it will help somewhat you'll be able to

get off the ground but it will be lots

of new things you need to learn so

that's that's the spirit of it I think

of a toes of the school one of the

things you mentioned pretty early on in