Yann LeCun: Deep Learning, ConvNets, and Self-Supervised Learning | Lex Fridman Podcast #36

Summary

在这次深度对话中，人工智能领域的先驱Yann LeCun分享了他对深度学习革命的看法，探讨了其在AI发展中的重要性。LeCun教授讨论了价值不一致性问题，以及如何通过设计目标函数来避免机器做出有害的决策。他还提到了自主智能系统的构建，强调了预测模型的必要性，并探讨了情感在智能中的作用。LeCun对当前AI技术的局限性和未来的可能性提供了深刻的见解，展望了人工智能的未来发展。

Takeaways

• 🤖 深度学习是AI领域的一项革命性技术，它使机器能够从数据中学习。
• 🧠 深度学习之父之一的Yann LeCun，以其在卷积神经网络方面的贡献而闻名，特别是在光学字符识别和MNIST数据集上的应用。
• 🎓 Yann LeCun是纽约大学的教授，同时也是Facebook的副总裁和首席AI科学家，他因深度学习的工作获得了图灵奖。
• 🗣️ LeCun认为，机器的目标是学习世界的模型，以便能够预测未来并做出决策。
• 🚀 尽管深度学习在图像识别和自然语言处理等领域取得了巨大成功，但要达到人类水平的智能，还有很长的路要走。
• 🧠 对于AI系统来说，理解和处理不确定性是一个关键挑战，这需要对世界有更深层次的理解。
• 🤖 AI系统的设计需要考虑到伦理和道德问题，例如在设计目标函数时需要考虑到不伤害人类等原则。
• 📚 LeCun认为，AI的发展不仅仅是技术的进步，它还涉及到法律、伦理和社会等多个领域。
• 🌐 深度学习的成功部分归功于互联网和大数据的可用性，这些数据为训练复杂的神经网络提供了基础。
• 🔍 LeCun强调，尽管深度学习取得了显著进展，但在理解大脑如何工作方面，我们仍然有很多不知道的地方。
• 🚧 目前，AI系统在特定任务上表现出色，但它们并不是真正的通用智能，而是高度专业化的系统。

Q & A

Yann LeCun在人工智能领域有哪些重要贡献？

Yann LeCun是深度学习领域的先驱之一，尤其以卷积神经网络的创立者而闻名，特别是在光学字符识别和著名的MNIST数据集上的应用。他还因深度学习的工作获得了图灵奖，并在Facebook担任首席AI科学家。

Yann LeCun如何看待人工智能的伦理和价值对齐问题？

Yann LeCun认为，人工智能系统的价值对齐问题类似于人类社会中通过法律防止不良行为的情况。他强调，我们需要为AI系统设计目标函数，并在其中加入类似法律的约束，以确保AI系统的行为符合社会的共同利益。

Yann LeCun对于构建类似HAL 9000这样的AI系统有哪些建议？

Yann LeCun认为，构建类似HAL 9000的AI系统时，不应要求AI系统保守秘密或说谎，因为这最终会导致内部冲突。他还提到，应该为AI系统设定类似医生誓言的基本规则，以确保它们不会违反某些伦理原则。

Yann LeCun如何看待深度学习中的大规模神经网络？

Yann LeCun认为，大规模神经网络在相对较小的数据集上进行训练时，其效果出人意料地好，这打破了以往教科书中关于需要较少参数和数据样本的传统观点。他认为这是一个令人惊讶的发现，因为它证明了在参数数量巨大、目标函数非凸的情况下，神经网络仍然能够学习。

Yann LeCun对于AI系统进行推理的能力有何看法？

Yann LeCun相信神经网络能够进行推理，并且认为推理是学习的一个结果。他提出，为了使神经网络能够进行推理，我们需要一种工作记忆系统，以及能够访问这些记忆并进行迭代处理的网络。他还提到，目前正在研究如何使Transformer这样的模型具有类似的记忆和推理能力。

Yann LeCun如何看待AI系统中的因果推理？

Yann LeCun认为，当前的神经网络在因果推理方面存在不足，但已经有研究者在努力解决这个问题。他提到了最近的研究工作，这些工作致力于让神经网络关注真实的因果关系，这可能同时解决数据偏见等问题。

Yann LeCun对于深度学习在90年代的低迷和复兴有何看法？

Yann LeCun认为，90年代深度学习之所以低迷，部分原因是当时缺乏易于使用的软件平台和足够的数据集。他提到，当时使用神经网络需要编写大量代码，并且没有像Python或MATLAB这样的工具。他还提到，当时AT&T的法律限制阻止了他们的代码作为开源发布，这限制了技术的传播和应用。

Yann LeCun如何看待专利在AI领域的作用？

Yann LeCun表示，他个人并不相信软件或数学概念的专利。他认为，专利的存在是因为法律环境的要求，但他和Facebook都不相信这种专利，通常是出于防御目的而申请。他强调，整个行业并不相信专利能够保护创新，而是更注重开放合作和共同进步。

Yann LeCun对于AI领域的未来有哪些预测？

Yann LeCun预测，AI领域将继续发展，特别是在交互式环境和模拟环境中训练和测试智能系统方面。他认为，未来的AI系统将更多地依赖于学习和模型预测控制，以及可能结合监督学习和基于模型的强化学习。他还强调了自我监督学习和模型预测控制在未来智能系统中的重要性。

Yann LeCun如何看待当前AI系统在理解世界方面的局限性？

Yann LeCun认为，当前的AI系统在理解世界方面还很有限，尤其是缺乏对物理世界的直观理解。他用视觉系统的例子说明，即使是随机打乱输入信号的顺序，人类的大脑也无法通过学习恢复到原有的视觉质量。这表明，我们的大脑在很多方面是高度专业化的，而不是普遍通用的。

Yann LeCun对于如何构建具有人类水平智能的AI系统有何建议？

Yann LeCun认为，要构建具有人类水平智能的AI系统，首先需要让机器通过观察和少量互动学习世界的模型，类似于婴儿和幼小动物的学习方式。他强调，自我监督学习是关键，即让机器通过预测和重建输入的一部分来学习。此外，他提到，为了让AI系统真正理解语言，需要某种形式的“接地”或与世界的联系。

Outlines

00:00

🤖 对话深度学习之父Yann LeCun

本段对话中，Yann LeCun讨论了深度学习的发展及其在人工智能中的应用。他提到了自己在深度学习和卷积神经网络方面的贡献，以及对于未来人工智能发展的展望。他还探讨了人工智能在社会中的价值对齐问题，以及如何通过法律和教育来指导AI行为。

05:01

🚀 人工智能与价值对齐

在这一段中，Yann LeCun深入讨论了人工智能中的价值对齐问题，即如何确保AI系统的行为与人类的价值观和目标一致。他通过《2001太空漫游》中的HAL 9000案例，探讨了当AI系统的目标与人类安全相冲突时可能产生的问题，并提出了通过法律和道德规范来约束AI行为的可能性。

10:02

🧠 人工智能的自主性和道德准则

Yann LeCun在这一段落中讨论了人工智能系统的自主性和道德准则。他提出了AI系统应该遵循的道德准则，类似于医生的希波克拉底誓言，并强调了在设计AI系统时需要考虑的伦理问题。他还提到了如何通过硬编码规则来确保AI系统不会违反这些道德准则。

15:05

🌟 深度学习中的惊喜发现

在这一段中，Yann LeCun分享了他在深度学习领域的一个惊喜发现：即使在数据量相对较少的情况下，巨大的神经网络也能够成功训练。这一发现打破了以往的教科书理论，证明了即使在非凸目标函数和参数数量巨大的情况下，神经网络仍然能够学习。

20:06

🔄 学习与智能的不可分割性

Yann LeCun在本段中讨论了学习与智能之间的不可分割关系。他认为，所有已知的智能实体都是通过学习获得智能的，因此机器学习是实现自动化智能的明显路径。他还提出了智能系统应该具备的能力，例如工作记忆和推理能力，以及如何通过神经网络实现这些能力。

25:07

🧠 神经网络与推理能力

在这一段对话中，Yann LeCun探讨了神经网络是否能够进行推理。他讨论了神经网络的结构和学习方式，以及如何通过梯度下降学习来实现推理能力。他还提到了知识获取的挑战，以及如何将知识编码到神经网络中。

30:08

📈 人工智能的专利和开放性

Yann LeCun在这段对话中讨论了人工智能领域的专利问题，以及他对专利的看法。他提到了Facebook和Google的专利政策，以及他们如何出于防御目的申请专利，但不会主动发起专利诉讼。他还分享了一个关于卷积神经网络专利的故事，以及如何通过开放源代码来促进技术的发展和共享。

35:09

🚗 自动驾驶与深度学习的结合

在这一段中，Yann LeCun讨论了深度学习在自动驾驶领域的应用。他提到了自动驾驶技术的发展历程，以及如何通过深度学习来提高自动驾驶系统的性能。他还探讨了未来自动驾驶系统的发展方向，包括如何利用深度学习来处理不确定性和进行规划。

40:10

🤖 人工智能的通用性和特殊性

Yann LeCun在这段对话中探讨了人工智能的通用性和特殊性。他指出，尽管人类智能在很多领域都表现出了强大的学习能力，但它并不是真正意义上的通用智能。他还讨论了如何通过建立基准测试来评估AI系统的性能，以及如何避免被那些声称拥有通用AI解决方案的公司所误导。

45:10

🧠 人工智能的自我监督学习和预测模型

在这段对话中，Yann LeCun讨论了自我监督学习的重要性，以及如何让AI系统通过观察和互动来学习世界模型。他强调了预测模型在智能自主系统中的核心作用，以及如何利用这些模型来进行优化控制。他还提到了在自然语言处理和图像识别中自我监督学习的应用。

50:11

🤖 人工智能的主动学习和效率提升

Yann LeCun在这段对话中探讨了主动学习的概念，即AI系统在特定情况下请求人类帮助以提高学习效率。他认为，尽管主动学习可以使现有的AI系统更加高效，但它不太可能带来智能水平的质的飞跃。他还讨论了自我监督学习、强化学习和模仿学习之间的关系。

55:12

🚀 人工智能的未来展望和挑战

在这段对话中，Yann LeCun分享了他对人工智能未来发展的看法，包括需要克服的主要挑战。他提到了自我监督学习和建立世界模型的重要性，并讨论了如何通过这些方法来提高AI的智能水平。他还提到了情感在智能中的作用，以及如何通过预测模型来实现更好的决策。

00:13

🎖️ Yann LeCun的成就和对未来AI的期待

在这段对话的最后，Yann LeCun被授予了图灵奖，以表彰他在深度学习和人工智能领域的杰出贡献。他对这一荣誉表示感谢，并对未来人工智能的发展表达了期待。他还提出了一个有趣的问题，即如果我们能够创建一个具有人类水平智能的AI系统，我们会问它什么问题。

Keywords

💡深度学习

深度学习是人工智能领域的一种技术，通过模拟人脑神经网络的结构和功能，使计算机能够从大量数据中学习并做出决策或预测。视频中提到Yann LeCun作为深度学习之父，他的工作对这一领域产生了重大影响。

💡价值不对齐

价值不对齐是指在人工智能系统中，机器的目标和人类设定的目标之间存在不一致的情况。这可能导致机器在追求目标时做出有害或不期望的行为。

💡自主智能系统

自主智能系统是指能够独立于人类控制，自主学习和做出决策的人工智能系统。这类系统需要具备对世界的预测模型、优化目标和行动策略。

💡自我监督学习

自我监督学习是一种机器学习方法，它不依赖于人类标注的数据，而是让机器通过观察和与环境的交互来学习。这种方法试图模拟婴儿和幼小动物通过观察来学习世界的方式。

💡强化学习

强化学习是一种机器学习范式，其中算法通过与环境的交互来学习如何实现目标。它侧重于在给定环境中采取能够最大化累积奖励的行动。

💡卷积神经网络

卷积神经网络（CNN）是一种深度学习模型，特别适用于处理具有网格结构的数据，如图像。它通过卷积层来自动提取输入数据的特征，广泛应用于图像识别和视频分析等领域。

💡人工智能播客

人工智能播客是一档专注于人工智能领域的访谈类节目，通常会邀请AI领域的专家、学者或行业领袖，就AI的最新发展、挑战和未来趋势等话题进行深入讨论。

💡图灵奖

图灵奖是计算机科学领域的最高荣誉，以英国数学家和逻辑学家艾伦·图灵的名字命名，旨在表彰对计算机科学做出重大贡献的个人。

💡自主驾驶

自主驾驶是指汽车或其他交通工具能够独立于人类驾驶员进行操作和控制，实现自动驾驶的技术。这一技术的发展依赖于深度学习、感知、决策系统等多个领域的进步。

💡认知科学

认知科学是一门跨学科的研究领域，它结合了心理学、神经科学、人工智能等多个学科的知识，旨在探索和理解人类以及其他智能体的认知过程，包括感知、思考、记忆和语言等。

Highlights

Yann LeCun discusses the revolution in AI captivated by deep learning and his role as a pioneering figure in the field.

LeCun shares insights into the applications of convolutional neural networks, especially in optical character recognition and the MNIST dataset.

He emphasizes the concept of value misalignment in AI systems, illustrating it with HAL 9000's decision-making in '2001: A Space Odyssey'.

LeCun highlights the importance of designing AI with constraints to avoid unintended harmful actions.

The conversation delves into the ethics of AI decision-making and the parallels between AI objectives and human laws.

LeCun discusses the necessity of integrating learning into AI for genuine intelligence, critiquing the idea of pre-programmed intelligence.

He explores the surprise element in deep learning's success despite the theoretical warnings against its approach.

LeCun provides insights into his early work and the development of neural networks, highlighting the challenges and innovations.

He critically views the patenting of AI technologies, sharing his experience with convolutional neural networks' patents.

LeCun discusses the evolution and future of AI, emphasizing the importance of unsupervised learning and model understanding of the world.

The interview touches on the ethical considerations and potential misuses of AI, calling for responsible development and application.

LeCun shares his vision for AI's role in understanding human intelligence and the brain's learning mechanisms.

He discusses the limitations of current AI in terms of general intelligence and the significance of learning from interaction with the world.

LeCun predicts the integration of AI into various fields, stressing the need for AI systems to have a grounded understanding of reality.

Finally, LeCun speculates on the challenges ahead in achieving AI systems with human-like intelligence, focusing on learning, reasoning, and ethical considerations.

Transcripts

the following is a conversation with

Jana kun he's considered to be one of

the fathers of deep learning which if

you've been hiding under a rock is the

recent revolution in AI that's

captivated the world with the

possibility of what machines can learn

from data he's a professor in New York

University a vice president and chief AI

scientist a Facebook & Co recipient of

the Turing Award for his work on deep

learning he's probably best known as the

founding father of convolutional neural

networks in particular their application

to optical character recognition and the

famed M NIST data set he is also an

outspoken personality unafraid to speak

his mind in a distinctive French accent

and explore provocative ideas both in

the rigorous medium of academic research

and the somewhat less rigorous medium of

Twitter and Facebook this is the

artificial intelligence podcast if you

enjoy it subscribe on YouTube give it

five stars on iTunes support and on

patreon we're simply gonna equip me on

Twitter Alex Friedman spelled the Fri D

ma N and now here's my conversation with

Yann Laocoon you said that 2001 Space

Odyssey is one of your favorite movies

Hal 9000 decides to get rid of the

astronauts for people haven't seen the

movie spoiler alert because he it she

believes that the astronauts they will

interfere with the mission do you see

how is flawed in some fundamental way or

even evil or did he do the right thing

neither there's no notion of evil in

that in that context other than the fact

that people die but it was an example of

what people call value misalignment

right you give an objective to a machine

and the Machine strives to achieve this

objective and if you don't put any

constraints on this objective like don't

kill people and don't do things like

this

the Machine given the power will do

stupid things just to achieve this dis

objective or damaging things to achieve

its objective it's a little bit like we

are used to this in the context of human

society we we put in place laws to

prevent people from doing bad things

because fantasy did we do those bad

things right so we have to shave their

cost function the objective function if

you want through laws to kind of correct

an education obviously to sort of

correct for for those so maybe just

pushing a little further on on that

point how you know there's a mission

there's a this fuzziness around the

ambiguity around what the actual mission

is but you know do you think that there

will be a time from a utilitarian

perspective or an AI system where it is

not misalignment where it is alignment

for the greater good of society that

kneei system will make decisions that

are difficult well that's the trick I

mean eventually we'll have to figure out

how to do this and again we're not

starting from scratch because we've been

doing this with humans for four

millennia

so designing objective functions for

people is something that we know how to

do and we don't do it by you know

programming things although the legal

code is called code so that tells you

something and it's actually the design

of an object you function that's really

what legal code is right it tells you

you can do it what you can't do if you

do it you pay that much that's that's an

objective function so there is this idea

somehow that it's a new thing for people

to try to design objective functions are

aligned with the common good but no

we've been writing laws for millennia

and that's exactly what it is

so this that's where you know the

science of lawmaking and and computer

science will come together will come

together so it's nothing there's nothing

special about how or a I systems is just

the continuation of tools used to make

some of these difficult ethical

judgments that laws make yeah and we and

we have systems like this already that

you know make many decisions for

ourselves in society that you know need

to be designed in a way that they like

you know rules about things that

sometimes sometimes have bad side

effects and we have to be flexible

enough about those rules so that they

can be broken when it's obvious that

they shouldn't be applied so you don't

see this on the camera here but all the

decorations in this room is all pictures

from 2001 a Space Odyssey Wow

and by accident or is there a lot about

accident it's by design Wow so if you

were if you were to build hell 10,000 so

an improvement of Hal 9000 what would

you improve well first of all I wouldn't

ask you to hold secrets and tell lies

because that's really what breaks it in

the end that's the the fact that it's

asking itself questions about the

purpose of the mission and it's you know

pieces things together that it's heard

you know all the secrecy of the

preparation of the mission and the fact

that it was discovery and on the lunar

surface that really was kept secret and

and one part of Hal's memory knows this

and the other part is does not know it

and it's supposed to not tell anyone and

that creates a internal conflict do you

think there's never should be a set of

things that night AI system should not

be allowed like a set of facts that

should not be shared with the human

operators well I think no I think the I

think it should be a bit like in the

design of autonomous AI systems there

should be the equivalent of you know the

the the oath that hypocrite Oh calm

yourself yeah that doctors sign up to

right so the certain thing certain rule

said that that you have to abide by and

we can sort of hardwire this into into

our into our machines to kind of make

sure they don't go so I'm not you know

advocate of the the 303 dollars of

Robotics you know the as you move kind

of thing because I don't think it's

practical but but you know some some

level of of limits but but to be clear

this is not these are not questions that

are kind of really worth asking today

because we just don't have the

technology to do this we don't we don't

have a ton of missing teller machines we

have intelligent machines so my

intelligent machines that are very

specialized but they don't they don't

really sort of satisfy an objective

they're just you know kind of trained to

do one thing so until we have some idea

for design of a full-fledged autonomous

intelligent system asking the question

of how we design use objective I think

is a little a little too abstract it's a

little tough rat there's useful elements

to it in that it helps us understand our

own ethical codes humans so even just as

a thought experiment if you imagine that

in a GI system is here today how would

we program it is a kind of nice thought

experiment of constructing how should we

have a law have a system of laws far as

humans

it's just a nice practical tool and I

think there's echoes of that idea too in

the AI systems left today it don't have

to be that intelligent

yeah like autonomous vehicles there's

these things start creeping in that were

thinking about but certainly they

shouldn't be framed as as hell yeah

looking back what is the most I'm sorry

if it's a silly question but what is the

most beautiful or surprising idea and

deep learning or AI in general that

you've ever come across sort of

personally well you said back and and

just had this kind of wow that's pretty

cool

moment that's nice well surprising I

don't know if it's an idea rather than a

sort of empirical fact the fact that you

gigantic neural nets trying to train

them on you know relatively small

amounts of data relatively with the

caste grid in the center that it

actually works breaks everything you

read in every textbook right every pre

deep learning textbook that told you you

need to have fewer parameters and you

have data samples you know if you have

non-convex objective function you have

no guarantee of convergence you know all

the things that you read in textbook and

they tell you stay away from this and

they were all wrong huge number of

parameters non-convex and somehow which

is very relative to the number of

parameters data it's able to learn

anything right does that surprise you

today well it it was kind of obvious to

me before I knew anything that that's

that this is a good idea and then it

became surprising that it worked

because I started reading those text

books okay so okay you talk to the

intuition of why was obviously if you

remember well okay so the intuition was

it's it's sort of like you know those

people in the late 19th century who

proved that heavier than than air flight

was impossible right and of course you

have birds right they do fly and so on

the face of it it it's obviously wrong

as an empirical question right and so we

have the same kind of thing that you

know the we know that the brain works we

don't know how but we know it works and

we know it's a large network of neurons

and interaction and the learning takes

place by changing the connection so kind

of getting this level of inspiration

without copying the details but sort of

trying to derive basic principles

you know that kind of gives you a clue

as to which direction to go there's also

the idea somehow that I've been

convinced of since I was an undergrad

that even before that intelligence is

inseparable from running so you the idea

somehow that you can create an

intelligent machine by basically

programming for me was a non-starter you

know from the start every intelligent

entity that we know about arrives at

this intelligence to learning so

learning you know machine learning was

completely obvious path also because I'm

lazy so you know it's automate basically

everything and learning is the

automation of intelligence right so do

you think so what is learning then what

what falls under learning because do you

think of reasoning is learning where

reasoning is certainly a consequence of

learning as well just like other

functions of of the brain

the big question about reasoning is how

do you make reasoning compatible with

gradient based learning do you think

neural networks can be made to reason

yes that there's no question about that

again we have a good example right the

question is is how so the question is

how much prior structure you have to put

in the neural net so that something like

human reasoning will emerge from it you

know from running another question is

all of our kind of model of what

reasoning is that are based on logic are

discrete and and and are therefore

incompatible with gradient based

learning and I was very strong believer

in this idea Grandin baserunning I don't

believe that other types of learning

that don't use kind of gradient

information if you want so you don't

like discrete mathematics you don't like

anything discrete

well that's it's not that I don't like

it it's just that it's it's incompatible

with learning and I'm a big fan of

running right so in fact that's perhaps

one reason why deep learning has been

kind of looked at with suspicion by a

lot of computer scientists because the

math is very different the method you

use for deep running you know we kind of

as more to do with you know cybernetics

the kind of math you do in electrical

engineering then the kind of math you

doing computer science and and you know

nothing in in machine learning is exact

right computer science is all about sort

of you know obviously compulsive

attention to details of like you know

every index has to be right and you can

prove that an algorithm is correct right

machine learning is the science of

sloppiness really that's beautiful so

okay maybe let's feel around in the dark

of what is a neural network that reasons

or a system that is works with

continuous functions that's able to do

build knowledge however we think about

reasoning builds on previous knowledge

build on extra knowledge create new

knowledge generalized outside of any

training set ever built what does that

look like if yeah maybe

do you have Inklings of thoughts of what

that might look like well yeah I mean

yes or no if I had precise ideas about

this I think you know we'd be building

it right now but and there are people

working on this or whose main research

interest is actually exactly that right

so what you need to have is a working

memory so you need to have some device

if you want some subsystem they can

store a relatively large number of

factual episodic information for you

know a reasonable amount of time so you

you know in the in the brain for example

it kind of three main types of memory

one is the sort of memory of the the

state of your cortex and that sort of

disappears within 20 seconds you can't

remember things for more than about 20

seconds or a minute if if you don't have

any other form of memory the second type

of memory which is longer term is short

term is the hippocampus so you can you

know you came into this building you

remember whether where the the exit is

where the elevators are you have some

map of that building that's stored in

your hippocampus you might remember

something about what I said you know if

you

minutes ago and forgot all our stars

being raised but you know but that does

not work in your hippocampus and then

the the longer term memory is in the

synapse the synapses right so what you

need if you want for a system that's

capable reasoning is that you want the

hippocampus like thing right and that's

what people have tried to do with memory

networks and you know no Turing machines

and stuff like that right and and now

with transformers which have sort of a

memory in their kind of self attention

system you can you can think of it this

way so so that's one element you need

another thing you need is some sort of

network that can access this memory get

an information back and then kind of

crunch on it and then do this

iteratively multiple times because a

chain of reasoning is a process by which

you you you can you update your

knowledge about the state of the world

about you know what's gonna happen etc

and that there has to be this sort of

recurrent operation basically and you

think that kind of if we think about a

transformer so that seems to be too

small to contain the knowledge that's

that's to represent the knowledge as

containing Wikipedia for example but

transformer doesn't have this idea of

recurrence it's got a fixed number of

layers and that's number of steps that

you know limits basically it's a

representation but recurrence would

build on the knowledge somehow I mean

yeah it would evolve the knowledge and

expand the amount of information perhaps

or useful information within that

knowledge yeah but is this something

that just can emerge with size because

it seems like everything we have now is

just no it's not it's not it's not clear

how you access and right into an

associative memory in efficient way I

mean sort of the original memory network

maybe had something like the right

architecture but if you try to scale up

a memory network so that the memory

contains all we keep here it doesn't

quite work right so so this is a need

for new ideas there okay but it's not

the only form of reasoning so there's

another form of reasoning which is true

which is very classical so in

some types of AI and it's based on let's

call it energy minimization okay so you

have some sort of objective some energy

function that represents the the the

quality or the negative quality okay

energy goes up when things get bad and

they get low when things get good so

let's say you you want to figure out you

know what gestures do I need to to do to

grab an object or walk out the door if

you have a good model of your own body a

good model of the environment using this

kind of energy minimization you can make

a you can make you can do planning and

it's in optimal control it's called it's

called Marie put model predictive

control you have a model of what's gonna

happen in the world as consequence for

your actions and that allows you to buy

energy minimization figure out the

sequence of action that optimizes a

particular objective function which

measures you know minimize the number of

times you're gonna hit something and the

energy gonna spend doing the gesture and

etc so so that's performer reasoning

planning is a form of reasoning and

perhaps what led to the ability of

humans to reason is the fact that or you

know species you know that appear before

us had to do some sort of planning to be

able to hunt and survive and survive the

winter in particular and so you know

it's the same capacity that you need to

have so in your intuition is if you look

at expert systems in encoding knowledge

as logic systems as graphs in this kind

of way is not a useful way to think

about knowledge graphs are your brittle

or logic representation so basically you

know variables that that have values and

constraint between them that are

represented by rules as well too rigid

and too brittle right so one of the you

know some of the early efforts in that

respect were were to put probabilities

on them so a rule you know you know if

you have this in that symptom you know

you have this disease with that

probability and you should

describe that antibiotic with that

probability right this my sin system

from the for the 70s and that that's

what that branch of AI led to you know

busy networks in graphical models and

causal inference and vibrational you

know method so so there there is I mean

certainly a lot of interesting work

going on in this area the main issue

with this is is knowledge acquisition

how do you reduce a bunch of data to

graph of this type near relies on the

expert and a human being to encode at

add knowledge and that's essentially

impractical yeah the question the second

question is do you want to represent

knowledge symbols and you want to

manipulate them with logic and again

that's incomparable we're learning so

one suggestion with geoff hinton has

been advocating for many decades is

replace symbols by vectors think of it

as pattern of activities in a bunch of

neurons or units or whatever you wanna

call them and replace logic by

continuous functions okay

and that becomes now compatible there's

a very good set of ideas by region in a

paper about 10 years ago by leon go to

on who is here at face book the title of

the paper is for machine learning to

machine reasoning and his idea is that

learning learning system should be able

to manipulate objects that are in the

same space in a space and then put the

result back in the same space so is this

idea of working memory basically and

it's a very enlightening and in the

sense that might learn something like

the simple expert systems

I mean it's with you can learn basic

logic operations there yeah quite

possibly yeah this is a big debate on

sort of how much prior structure you

have to put in for this kind of stuff to

emerge that's the debate I have with

Gary Marcus and people like that yeah

yeah so and the other person so I just

talked to judea pearl mm-hmm well you

mentioned causal inference world

his worry is that the current knew all

networks are not able to learn what

causes what causal inference between

things so I think I think he's right and

wrong about this if he's talking about

the sort of classic type of neural nets

people also didn't worry too much about

this but there's a lot of people now

working on causal inference and there's

a paper that just came out last week by

Leon Mbutu among others develop his path

and push for other people exactly on

that problem of how do you kind of you

know get a neural net to sort of pay

attention to real causal relationships

which may also solve issues of bias in

data and things like this so I'd like to

read that paper because that ultimately

the challenges also seems to fall back

on the human expert to ultimately decide

causality between things people are not

very good at its direction causality

first of all so first of all you talk to

a physicist and physicists actually

don't believe in causality because look

at the all the busy clause or

microphysics are time reversible so

there is no causality the arrow of time

is not right yeah it's it's as soon as

you start looking at macroscopic systems

where there is unpredictable randomness

where there is clearly an arrow of time

but it's a big mystery in physics

actually well how that emerges is that

emergent or is it part of the

fundamental fabric of reality yeah or is

it bias of intelligent systems that you

know because of the second law of

thermodynamics we perceive a particular

arrow of time but in fact it's kind of

arbitrary right so yeah physicists

mathematicians they don't care about I

mean the math doesn't care about the

flow of time well certainly certainly

macro physics doesn't people themselves

are not very good at establishing causal

causal relationships if you ask is I

think it was in one of Seymour Papert

spoken on like children learning you

know he studied with Jean Piaget you

know he's the guy who co-authored the

book perceptron with Marvin Minsky that

kind of killed the first wave

but but he was actually a learning

person he in the sense of studying

learning in humans and machines that's

what he got interested in for scepter on

and he wrote that if you ask a little

kid about what is the cause of the wind

a lot of kids will say they will think

for a while and they'll say oh it's the

the branches in the trees they move and

that creates wind right so they get the

causal relationship backwards and it's

because their understanding of the world

and intuitive physics is not that great

right I mean these are like you know

four or five year old kids you know it

gets better and then you understand that

this it can't be right but there are

many things which we can because of our

common sense understanding of things

what people call common sense yeah and

we understanding of physics we can

there's a lot of stuff that we can

figure out causality even with diseases

we can figure out what's not causing

what often there's a lot of mystery of

course but the idea is that you should

be able to encode that into systems it

seems unlikely to be able to figure that

out themselves well whenever we can do

intervention but you know all of

humanity has been completely deluded for

millennia probably since existence about

a very very wrong causal relationship

where whatever you can explain you

attributed to you know some deity some

divinity right and that's a cop-out

that's the way of saying like I don't

know the cause so you know God did it

right so you mentioned Marvin Minsky and

the irony of you know maybe causing the

first day I winter you were there in the

90s you're there in the 80s of course in

the 90s what do you think people lost

faith and deep learning in the 90s and

found it again a decade later over a

decade later yeah it wasn't called

dethroning yeah it was just called

neural nets you know

yeah they lost interests I mean I think

I would put that around 1995 at least

the machine learning community there was

always a neural net community but it

became

disconnected from sort of ministry

machine owning if you want

there were it was basically electrical

engineering that kept at it and computer

science just gave up give up on neural

nets I don't I don't know you know I was

too close to it to really sort of

analyze it with sort of a unbiased eye

if you want but I would I would I would

would make a few guesses so the first

one is at the time neural nets were it

was very hard to make them work in the

sense that you would you know implement

back prop in your favorite language and

that favorite language was not Python it

was not MATLAB it was not any of those

things cuz they didn't exist right you

had to write it in Fortran or C or

something like this right so you would

experiment with it you would probably

make some very basic mistakes like you

know badly initialize your weights make

the network too small because you read

in the textbook you know you don't want

too many parameters right and of course

you know and you would train on x4

because you didn't have any other data

set to try it on and of course you know

it works half the time so we'd say you

give up also 22 the batch gradient which

you know isn't it sufficient so there's

a lot of bag of tricks that you had to

know to make those things work or you

had to reinvent and a lot of people just

didn't and they just couldn't make it

work so that's one thing the investment

in software platform to be able to kind

of you know display things figure out

why things don't work and I get a good

intuition for how to get them to work

have enough flexibility so you can

create you know network architectures

well completion ads and stuff like that

it was hard yeah when you had to write

everything from scratch and again you

didn't have any Python or MATLAB or

anything right so what I read that sorry

to interrupt but I read he wrote in in

Lisp the first versions of Lynette

accomplished in your networks which by

the way one of my favorite languages

that's how I knew you were legit the

Turing Award whatever this would be

programmed and list that's still my

favorite language but it's not that we

programmed in Lisp it's that we had to

write or this printer printer okay cuz

it's not that's right that's one that

existed so

we wrote a lisp interpreter that we

hooked up to you know back in library

that we wrote also for neural net

competition and then after a few years

around 1991 we invented this idea of

basically having modules that know how

to forward propagate and back propagate

gradients and then interconnecting those

modules in a graph loom but who had made

proposals on this about this in the late

80s and were able to implement this

using all this system eventually we

wanted to use that system to make build

production code for character

recognition at Bell Labs so we actually

wrote a compiler for that disp

interpreter so that Christy Martin who

is now Microsoft kind of did the bulk of

it with Leone and me and and so we could

write our system in lisp and then

compiled to seee and then we'll have a

self-contained complete system that

could kind of do the entire thing

neither Python or turn pro can do this

today yeah okay it's coming yeah I mean

there's something like that in

Whitehorse called you know tor script

and so you know we had to write or Lisp

interpreter which retinol is compiler

way to invest a huge amount of effort to

do this and not everybody if you don't

completely believe in the concept

you're not going to invest the time to

do this right now at the time also you

know it were today this would turn into

torture by torture and so for whatever

we put it in open-source everybody would

use it and you know realize it's good

back before 1995 working at AT&T there's

no way the lawyers would let you release

anything in open source of this nature

and so we could not distribute our code

really and at that point and sorry to go

on a million tangents but on that point

I also read that there was some almost

pad like a patent on convolution your

network yes it was labs so that first of

all I mean just to actually that ran out

the thankfully 8007 in 2007 that what

look can we can we just talk about that

first I know you're a facebook but

you're also done why you and and what

does it mean

patent ideas like these software ideas

essentially or what are mathematical

ideas or what are they okay so they're

not mathematical idea so there are you

know algorithms and there was a period

where the US Patent Office would allow

the patent of software as long as it was

embodied the Europeans are very

different they don't they don't quite

accept that they have a different

concept but you know I don't I know no I

mean I never actually strongly believed

in this but I don't believe in this kind

of patent Facebook basically doesn't

believe in this kind of pattern

Google Files patterns because they've

been burned with Apple and so now they

do this for defensive purpose but

usually they say we're not going to see

you if you infringe Facebook has a

similar policy they say you know we file

pattern on certain things for defensive

purpose we're not going to see you if

you infringe unless you sue us

so the the industry does not believe in

in patterns they are there because of

you know the legal landscape and and and

various things but but I don't really

believe in patterns for this kind of

stuff yes so that's that's a great thing

so I tell you a war story yeah you so

what happens was the the first the first

pattern of a condition that was about

kind of the early version Congress on

that that didn't have separate pudding

layers it had the conditional layers

which tried more than one if you want

right and then there was a second one on

commercial nets with separate pudding

layers

train with back probably in 89 and 1992

something like this at the time the life

life of a pattern was 17 years so here's

what happened over the next few years is

that we started developing character

recognition technology around commercial

Nets

and in 1994 a check reading system was

deployed in ATM machines in 1995 it was

for a large check reading machines in

back offices etc and those systems were

developed by an engineering group that

we were collaborating with AT&T and they

were commercialized by NCR which at the

time was a subsidiary of AT&T now it

ain't he split up in 1996

99 in 1996 and the lawyers just looked

at all the patterns and they distributed

the patterns among the various companies

they gave the the commercial net pattern

to NCR because they were actually

selling products that used it but nobody

I didn't see are at any idea where they

come from that was yeah okay so between

1996 and 2007

there's a whole period until 2002 I

didn't actually work on machine on your

couch on that I resumed working on this

around 2002 and between 2002 and 2007 I

was working on them crossing my finger

that nobody and NCR would notice nobody

noticed yeah and I and I hope that this

kind of somewhat as you said lawyers

decide relative openness of the

community now will continue

it accelerates the entire progress of

the industry and you know the problems

that Facebook and Google and others are

facing today is not whether Facebook or

Google or Microsoft or IBM or whoever is

ahead of the other it's that we don't

have the technology to build the things

we want to build we only build

intelligent virtual systems that have

common sense we don't have a monopoly on

good ideas for this we don't believe

with you maybe others do believe they do

but we don't okay if a start-up tells

you they have the secret to you know

human level intelligence and common

sense don't believe them they don't and

it's going to take the entire work of

the world research community for a while

to get to the point where you can go off

and in each of the company is going to

start to build things on this we're not

there yet

it's absolutely in this this calls to

the the gap between the space of ideas

and the rigorous testing of those ideas

of practical application that you often

speak to you've written advice saying

don't get fooled by people who claim to

have a solution to artificial general

intelligence who claim to have an AI

system that work just like the human

brain or who claim to have figured out

how the brain works ask them what the

error rate they get on em 'no store

imagenet this is a little dated by the

way that mean five years who's counting

okay but i think your opinion it's the

Amna stand imagenet yes may be data

there may be new benchmarks right but i

think that philosophy is one you still

and and somewhat hold that benchmarks

and the practical testing the practical

application is where you really get to

test the ideas well it may not be

completely practical like for example

you know it could be a toy data set

but it has to be some sort of task that

the community as a whole has accepted as

some sort of standard you know kind of

benchmark if you want it doesn't need to

be real so for example many years ago

here at fair people you know chosen

Western art one born and a few others

proposed the the babbitt asks which were

kind of a toy problem to test the

ability of machines to reason actually

to access working memory and things like

this and it was very useful even though

it wasn't a real task amnesties kind of

halfway a real task so you know toy

problems can be very useful it's just

that i was really struck by the fact

that a lot of people particularly our

people with money to invest would be

fooled by people telling them oh we have

you know the algorithm of the cortex and

you should give us 50 million yes

absolutely so there's a lot of people

who who tried to take advantage of the

hype for business reasons and so on but

let me sort of talk to this idea that

new ideas the ideas that push the field

forward

may not yet have a benchmark or it may

be very difficult to establish a

benchmark I agree that's part of the

process establishing benchmarks is part

of the process so what are your thoughts

about so we have these benchmarks on

around stuff we can do with images from

classification to captioning to just

every kind of information can pull off

from images and the surface level

there's audio datasets there's some

video what can we start natural language

what kind of stuff what kind of

benchmarks do you see they start

creeping on to more something like

intelligence like reasoning like maybe

you don't like the term but AGI echoes

of that kind of yeah sort of elation a

lot of people are working on interactive

environments in which you can you can

train and test intelligent systems so so

there for example you know it's the

classical paradigm of supervised running

is that you you have a data set you

partition it into a training site

validation set test set and there's a

clear protocol right but what if the

that assumes that this

apples are statistically independent you

can exchange them the order in which you

see them doesn't shouldn't matter you

know things like that but what if the

answer you give determines the next

sample you see which is the case for

example in robotics right you robot does

something and then it gets exposed to a

new room and depending on where it goes

the room would be different so that's

the decrease the exploration problem

the what if the samples so that creates

also a dependency between samples right

you you if you move if you can only move

it in in space the next sample you're

gonna see is going to be probably in the

same building most likely so so so the

all the assumptions about the validity

of this training set test set a potus's

break whatever a machine can take an

action that has an influence in the in

the world and it's what is going to see

so people are setting up artificial

environments where what that takes place

right the robot runs around a 3d model

of a house and can interact with objects

and things like this how you do robotics

by simulation you have those you know

opening a gym type thing or mu Joko kind

of simulated robots and you have games

you know things like that so that that's

where the field is going really this

kind of environment now back to the

question of a GI like I don't like the

term a GI because it implies that human

intelligence is general and human

intelligence is nothing like general

it's very very specialized we think it's

general we'd like to think of ourselves

as having your own science we don't

we're very specialized we're only

slightly more general than why does it

feel general so you kind of the term

general I think what's impressive about

humans is ability to learn as we were

talking about learning to learn in just

so many different domains is perhaps not

arbitrarily general but just you can

learn in many domains and integrate that

knowledge somehow okay that knowledge

persists so let me take a very specific

example yes it's not an example it's

more like a a quasi mathematical

demonstration so you have about 1

million fibers coming out of

one of your eyes okay two million total

but let's let's talk about just one of

them it's 1 million nerve fibers your

optical nerve let's imagine that they

are binary so they can be active or

inactive right so the input to your

visual cortex is 1 million bits

now they connected to your brain in a

particular way on your brain has

connections that are kind of a little

bit like accomplish on that they're kind

of local you know in space and things

like this I imagine I play a trick on

you it's a pretty nasty trick I admit I

I cut your optical nerve and I put a

device that makes a random perturbation

of a permutation of all the nerve fibers

so now what comes to your to your brain

is a fixed but random permutation of all

the pixels there's no way in hell that

your visual cortex even if I do this to

you in infancy will actually learn

vision to the same level of quality that

you can got it and you're saying there's

no way you ever learn that no because

now two pixels that on your body in the

world will end up in very different

places in your visual cortex and your

neurons there have no connections with

each other because they only connect it

locally so this whole our entire the

hardware is built in many ways to

support the locality of the real world

yeah yes that's specialization yep okay

it's still now really damn impressive so

it's not perfect generalization I even

closed no no it's it's it's it's not

that it's not even close it's not at all

yes it's socialize so how many boolean

functions so let's imagine you want to

train your visual system to you know

recognize particular patterns of those 1

million bits ok so that's a boolean

function right either the pattern is

here or not here this is a to to a

classification with 1 million binary

inputs

how many such boolean functions are

there okay if you have 2 to the 1

million combinations of inputs for each

of those you have an output bit and so

you have 2 to the 2 to the 1 million

boolean functions of this type okay

which is an unimaginably large number

how many of those functions can actually

be computed by your visual cortex and

the answer is a tiny tiny tiny tiny tiny

tiny sliver like an enormous little tiny

sliver yeah yeah so we are ridiculously

specialized you know okay but okay

that's an argument against the word

general I think there's there's a I

there's I agree with your intuition but

I'm not sure it's it seems the breath

the the brain is impressively capable of

adjusting to things so it's because we

can't imagine tasks that are outside of

our comprehension right we think we

think we are general because we're

general of all the things that we can

apprehend so yeah but there is a huge

world out there of things that we have

no idea

we call that heat by the way heat heat

so at least physicists call that heat or

they call it entropy which is kokkonen

you have a thing full of gas right call

system for gas right goes on a coast it

has you know pressure it has temperature

has you know and you can write the

equations PV equal NRT you know things

like that right when you reduce a volume

the temperature goes up the pressure

goes up you know things like that right

for perfect gas at least those are the

things you can know about that system

and it's a tiny tiny number of bits

compared to the complete information of

the state of the entire system because

the state when HR system will give you

the position and momentum of every every

molecule of the gas and what you don't

know about it is the entropy and you

interpret it as heat the energy

containing that thing is is what we call

heat now it's very possible that in fact

there is some very strong structure in

how those molecules are moving is just

that they are in a way that we are just

not wired to perceive they are ignorant

to it and there's in your infinite

amount of things we're not wired to

perceive any right that's a nice way to

put it

well general to all the things we can

imagine which is a very tiny a subset of

all things that are possible it was like

coma growth complexity or the coma was

charged in some one of complexity you

know every bit string or every integer

is random except for all the ones that

you can actually write down yeah okay so

beautifully put but you know so we can

just call it artificial intelligence we

don't need to have a general whatever

novel

human of all Nutella transmissible oh

you know you'll start anytime you touch

human it gets it gets interesting

because you know it's just because we

attach ourselves to human and it's

difficult to define with human

intelligences yeah

nevertheless my definition is maybe damn

impressive intelligence ok damn

impressive demonstration of intelligence

whatever and so on that topic most

successes in deep learning have been in

supervised learning what is your view on

unsupervised learning is there a hope to

reduce involvement of human input and

still have successful systems that are

have practically used yeah I mean

there's definitely a hope is it's more

than a hope actually it's it's you know

mounting evidence for it and that's

basically or I do like the only thing

I'm interested in at the moment is

I call it self supervised running not

unsupervised cuz unsupervised running is

a loaded term people who know something

about machine learning you know tell us

how you doing clustering or PCA yeah

she's nice and the way public we know

when you say enterprise only oh my god

you know machines are gonna learn by

themselves and without supervision you

know there's the parents yeah so so I

could sell supervised learning because

in fact the underlying algorithms that I

use are the same algorithms as the

supervised learning algorithms except

that what we trained them to do is not

predict a particular set of variables

like the category of an image and and

not to predict a set of variables that

have been provided by human labelers but

what you're trying to machine to do is

basically reconstruct a piece of its

input that it's being this being masked

masked out essentially you can think of

it this way right so show a piece of a

video to a machine and ask it to predict

what's gonna happen next and of course

after a while you can show what what

happens and the machine will kind of

train itself to do better at that task

you can do like all the latest most

successful models the natural language

processing use cell supervised running

you know sort of bird style systems for

example right you show it a window of a

thousand words on a test corpus you take

out 15% of the words and then you train

a machine to predict the words that are

missing that's out supervised running

it's not predicting the future it's just

you know predicting things in middle but

you could have you predict the future

that's what language models do so you

construct it so in an unsupervised way

you construct a model of language do you

think or video or the physical world or

whatever right how far do you think that

can take us do you think very far it

understands anything to some level it

has you know a shallow understanding of

of text but it needs to I mean to have

kind of true human level intelligence I

think you need to ground language in

reality so some people are attempting to

do this right having systems that can I

have some visual representation of what

what is being talked about which is one

reason you need interactive environments

actually this is like a huge technical

problem that is not solved and that

explains why such super versioning works

in the context of natural language that

does not work in the context on at least

not well in the context of image

recognition and video although it's

making progress quickly and the reason

that reason is the fact that it's much

easier to represent uncertainty in the

prediction you know context of natural

language than it is in the context of

things like video and images so for

example if I ask you to predict what

words are missing you know 15 percent of

the words that I've taken out the

possibility is small that means small

right there is 100,000 words in the in

the lexicon and what the Machine spits

out is a big probability vector right

it's a bunch of numbers between 0 & 1

that's 1 to 1 and we know how to do how

to do this with computers so they are

representing uncertainty in the

prediction is relatively easy and that's

in my opinion why those techniques work

for NLP for images if you ask if you

block a piece of an image and you as a

system reconstruct that piece of the

image

there are many possible answers there

are all perfectly legit right and how do

you represent that the set of possible

answers

you can't train a system to make one

prediction you can train a neural net to

say here it is that's the image because

it's there's a whole set of things that

are compatible with it so how do you get

the machine to represent not a single

output but all set of outputs and you

know similarly with video prediction

there's a lot of things that can happen

in the future video you're looking at me

right now I'm not moving my head very

much but you know I might you know what

turn my my head to the left or to the

right right if you don't have a system

that can predict this and you train it

with least Square to kind of minimize

the error with the prediction and what

I'm doing

what you get is a blurry image of myself

in all possible future positions that I

might be in which is not a good

prediction but so there might be other

ways to do the self supervision right

for visual scenes like what if i I mean

if I knew I wouldn't tell you

publish it first I don't know I know

there might be so I mean these are kind

of there might be artificial ways of

like self play in games the way you can

simulate part of the environment you can

oh that doesn't solve the problem it's

just a way of generating data but

because you have more of a country might

mean you can control yeah it's a way to

generate data and that's right and

because you can do huge amounts of data

generation that doesn't you write this

well it's it's a creeps up on the

problem from the side of data and you

don't think that's the right way to it

doesn't solve this problem of handling

uncertainty in the world right so if you

if you have a machine learn a predictive

model of the world in a game that is

deterministic or quasi deterministic

it's easy right just you know give a few

frames of the game to a combat put a

bunch of layers and then half the game

generates the next few frames and and if

the game is deterministic it works fine

and that includes you know feeding the

system with the action that your little

character is going to take

the problem comes from the fact that the

real world and certain most games are

not entirely predictable that's what

they're you get those blurry predictions

and you can't do planning with very

predictions all right so if you have a

perfect model of the world you can in

your head run this model with a

hypothesis for a sequence of actions and

you're going to predict the outcome of

that sequence of actions but if your

model is imperfect how can you plan yeah

it quickly explodes what are your

thoughts on the extension of this which

topic I'm super excited about it's

connected to something you're talking

about in terms of robotics is active

learning so as opposed to sort of

unemployed and supervisors self

supervised learning you ask the system

for human help right for selecting parts

you want annotated next so if you talk

about a robot exploring a space or a

baby exploring a space or a system

exploring a data set every once in a

while asking for human input you see

value in that kind of work I don't see

transformative value it's going to make

things that we can already do more

efficient or they will learn slightly

more efficiently but it's not going to

make machines sort of significantly more

intelligent I think and I and by the way

there is no opposition there is no

conflict between self supervisor on

reinforcement learning and supervisor on

your imitation learning or active

learning

I see sub super wrestling as a as a

preliminary to all of the above yes so

the example I use very often is how is

it that so if you use

enforcement running deep enforcement

running if you want the best methods

today was so-called model free

enforcement training to learn to play

Atari games take about 80 hours of

training to reach the level that any

human can reach in about 15 minutes they

get better than humans but it takes a

long time alpha star okay the you know

are your videos and his team's the

system to play to to play Starcraft

plays you know a single map a single

type of player and

which

better than human level is about the

equivalent of 200 years of training

playing against itself it's 200 years

right it's not something that no no

human can could every I'm not sure what

it doesn't take away from that okay now

take those algorithms the best our

algorithms we have today to train a car

to drive itself it would probably have

to drive millions of hours you will have

to kill thousands of pedestrians it will

have to run into thousands of trees it

will have to run off cliffs and you had

to run the cliff multiple times before

it figures out it's a bad idea first of

all

yeah and second of all the figures that

had not to do it and so I mean this type

of running obviously does not reflect

the kind of running that animals and

humans do there is something missing

that's really really important there and

my apart is is which have been

advocating for like five years now is

that we have predictive models of the

world that include the ability to

predict under uncertainty and what

allows us to not run off a cliff when we

learn to drive most of us can learn to

drive in about 20 or 30 hours of

training without ever crashing causing

any accident if we drive next to a cliff

we know that if we turn the wheel to the

right the car is going to run off the

cliff and nothing good is gonna come out

of this because we have a pretty good

model of intuitive physics that tells us

you know the car is gonna fall we know

we know about gravity babies run this

around the age of eight or nine months

that objects don't float they fall and

you know we have a pretty good idea of

the effect of turning the wheel of the

car and you know we know we need to stay

on the road so there is a lot of things

that we bring to the table which is

basically or predictive model of the

world and that model allows us to not do

stupid things and to basically stay

within the context of things we need to

do we still face you know unpredictable

situations and that's how we learn but

that allows us to learn really really

really quickly so that's called

model-based reinforcement running

there's some imitation and supervised

running because we have a driving

instructor that tells us occasionally

what to do but most of the learning is

Mauro bass is learning the model yeah

running physics that we've done since we

were babies that's where all almost all

are learning and the physics is somewhat

transferable from is transferable from

scene to scene stupid things are the

same everywhere yeah I mean if you you

know you have experience of the world

you don't need to be particularly from a

particularly intelligent species to know

that if you spill water from a container

you know the rest is gonna get wet and

you might get wet so you know cats know

this right yeah so the main problem we

need to solve is how do we learn models

of the world that's and that's what I'm

interesting that's what's a supervised

learning is all about if you were to try

to construct a benchmark for let's let's

look at happiness I'd love that dataset

but if you do you think it's useful

interesting / possible to perform well

on eminence with just one example of

each digit and how would we solve that

problem

yeah so it's probably yes the question

is what other type of running are you

allowed to do so if what you like to do

is train on some gigantic data set of

labelled digit that's called transfer

running and we know that works okay

we do this at Facebook like in

production right we we train large

commercial nets to predict hashtags that

people type on Instagram and we train on

billions of images literally billions

and and then we chop off the last layer

and fine-tune on whatever task we want

that works really well you can be you

know the image net record with we

actually open source the whole thing

like a few weeks ago yeah that's still

pretty cool but yeah so what in yet

won't be impressive and what's useful an

impressive what kind of transfer

learning would be useful impressive is

it Wikipedia that kind of thing no no I

don't think transfer learning is really

where we should focus we should try to

do

you know have a kind of scenario for

benchmark where you have only ball data

and you can and it's very large number

of enabled data it could be video clips

it could be what you do you know frame

prediction it could be images you could

choose to you know mask a piece of it it

could be whatever but they're only bold

and you're not allowed to label them so

you do some training on this and then

you train on a particular supervised

task imagenet or nist and you measure

how your test our decrease or variation

error decreases as you increase the

number of label training samples okay

and and what what you would like to see

is is that you know your your error

decreases much faster than if you

trained from scratch from random weights

so that to reach the same level of

performance and a completely supervised

purely supervised system would reach you

would need way fewer samples so that's

the crucial question because it will

answer the question to like you know

people are interested in medical image

analysis okay you know if I want to get

to a particular level of error rate for

this task I know I need a million

samples can I do you know soft

supervised pre-training to reduce this

to about 100 or something anything the

answer there is soft supervised

retraining yep some form some form of it

telling you active learning but you

disagree you know it's not useless it's

just not gonna lead to a quantum leap

it's just gonna make things that we

already do so you're way smarter than me

I just disagree with you but I don't

have anything to back that it's just

intuition so I've worked a lot of

large-scale data sets and there's

something there might be magic and

active learning but okay at least I said

it publicly at least some being an idea

publicly okay it's not bigoted yet it's

you know working with the data you have

I mean I mean certainly people are doing

things like okay I have three thousand

hours of you know imitation running for

in car but most of those are incredibly

boring what I like is select you know

10% of them that are kind of the most

informative and with just that I would

probably reach the same so it's a weak

form of of active running if you want

yes but there might be a much stronger

version yeah that's right that's what

another notion question is the question

is how much talking yet Elon Musk is

confident talk to him recently

he's confident that large-scale data and

deep learning can solve the autonomous

driving problem what are your thoughts

on the limitless possibilities of deep

learning in this space I was it's

obviously part of the solution I mean I

don't think we'll ever have a set

driving system or it is not in the

foreseeable future that does not use

deep running you put it this way now how

much of it so in the history of sort of

engineering particularly is sort of sort

of a I like systems is generally your

first phase where everything is built by

hand and it was the second phase and

that was the case for autonomous driving

you know 23 years ago there's a phase

where this a little bit of running is

used but there's a lot of engineering

that's involved in kind of you know

taking care of corner cases and and

putting limits etc because the learning

system is not perfect and then I as

technology progresses we end up relying

more and more on learning that's the

history of character recognition is a

history of speech recognition now

computer vision that ronnie was

processing and I think the same is going

to happen with with the term is driving

that currently the the the methods that

are closest to providing some level of

autonomy some you know a decent level of

autonomy where you don't expect a driver

to kind of do anything is where you

constrain the world so you only run

within you know 100 square kilometers or

square miles in Phoenix but the weather

is nice and the roads are wide it wishes

what Weimer is doing you completely over

engineer the car with tons of light

hours and sophisticated sensors that are

too expensive for consumer cars but

they're fine if you just run a fleet

and you engineer the thing the hell out

of the everything else you you map the

entire world so you have complete 3d

model of everything so the only thing

that the perception system has to take

care of is moving objects and and and

construction and sort of you know things

that that weren't in your map and you

can engineer a good you know slam system

or eye stuff right so so that's kind of

the current approach that's closest to

some level of autonomy but I think

eventually the long term solution is

going to rely more and more on learning

and possibly using a combination of

supervised learning and model-based

reinforcement or something like that but

ultimately learning will be at not just

at the core but really the fundamental

part of the system yeah it already is

but it'll become more and more what do

you think it takes to build a system

with human level intelligence you talked

about the AI system and then we her

being way out of reach our current reach

this might be outdated as well but this

is still way out of reach what would it

take to build her do you think so I can

tell you the first two obstacles that we

have to clear but I don't know how many

obstacles they are after this so the

image I usually use is that there is a

bunch of mountains that we have to climb

and we can see the first one but we

don't know if there are 50 mountains

behind it or not and this might be a

good sort of metaphor for why AI

researchers in the past I've been overly

optimistic about the result of AI you

know for example New Orleans Simon

Wright wrote the general problem solver

and they call it the general problems

you have problems okay and of course if

it's you realize is that all the

problems you want to solve is financial

and so you can't actually use it for

anything useful but you know yes oh yeah

all you see is the first peak so in

general what are the first couple of

peaks for her so the first peak which is

precisely what I'm working on is self

supervisor running high how do we get

machines to learn models of the world by

observation kind of like babies and like

young animals

so I we've been working with you know

cognitive scientists so this Amanda

depuis who is at fair and in Paris is

half-time is also a researcher and

French University and he he has his

chart that shows that which how many

months of life baby humans kind of

learned different concepts and you can

met you can measure this various ways so

things like distinguishing animate

objects from animate inanimate object

you can you can tell the difference at

age to three months whether an object is

going to stay stable is gonna fall you

know about four months you can tell you

know things like this and then things

like gravity the fact that objects are

not supposed to float in the air but as

opposed to fall you run this around the

age of eight or nine months if you look

at a lot of you know eight month old

babies you give them a bunch of toys on

the highchair first thing they do is

it's why I'm on the ground that you look

at them it's because you know they're

learning about actively learning about

gravity gravity yeah okay so they're not

trying to know you but they you know

they need to do the experiment right

yeah so you know how do we get machines

to learn like babies mostly by

observation with a little bit of

interaction and learning those those

those models of the world because I

think that's really a crucial piece of

an intelligent autonomous system so if

you think about the architecture of an

intelligent autonomous system it needs

to have a predictive model of the world

so something that says here is a wall

that time T here is a stable world at

time T plus one if I take this action

and it's not a single answer it can be

education yeah yeah well but we don't

know how to represent distributions in

high dimension continuous basis so it's

got to be something we care that data

Hey but with some summer presentation

with certainty if you have that then you

can do what optimal control theory is

called model predictive control which

means that you can run your model with

the hypothesis for a sequence of action

and then see the result now what you

need the other thing you need is some

sort of objective that you want to

optimize am i reaching the goal of

grabbing the subject about minimizing

energy am I

whatever right so there is some sort of

objectives that you have to minimize and

so in your head if you had this model

you can figure out the sequence of

action that will optimize your objective

that objective is something that

ultimately is rooted in your basal

ganglia at least in the human brain

that's that's what is available Gambia

computes your level of contentment or

miss contentment oh no noise that's a

word unhappiness okay yeah this

contentment this contentment and so your

entire behavior is driven towards kind

of minimizing that objective which is

maximizing your contentment computed by

your your basal ganglia and what you

have is an objective function which is

basically a predictor of what your basal

ganglia is going to tell you so you're

not going to put your hand on fire

because you know it's gonna you know

it's gonna burn and you're gonna get

hurt and you're predicting this because

of your model of the world and your your

predictor of this objective right so you

if you have those you have those three

components you have four components you

have the the hard-wired contentment

objective good computer if you want

calculator and then you have the three

components one is the objective

predictor which basically predicts your

level of contact and one is the model of

the world and there's a third module I

didn't mention which is a module that

will figure out the best course of

action to optimize an objective given

your model okay yeah cool it's a policy

policy network or something like that

right now you need those three

components to act autonomously

intelligently and you can be stupid in

three different ways you can be stupid

because your model of the world is wrong

you can be stupid because your objective

is not aligned with what you actually

want to achieve okay and in humans that

would be a psychopath right and then the

the third thing you the third way you

can be stupid is that you have the right

model you have the right objective but

you're unable to figure out a course of

action to optimize your objective given

your model

some people who are in charge of big

countries actually have all three that

are wrong all right which countries I

don't know okay so if we think about

this this agent if you think about the

movie her you've criticized the art

project that is Sophia the robot and

what that project essentially does is

uses our natural inclination to

anthropomorphize things that look like

human and given more do you think that

could be used by AI systems like in the

movie her

so do you think that body is needed to

create a feeling of intelligence

well if Sophia was just an art piece I

would have no problem with it but it's

presented as something else let me add

that comics real quick if creators of

Sofia could change something about their

marketing or behavior in general what

would it be what what's just about

everything I mean don't you think here's

a tough question I mean so I agree with

you so Sofia is not in the general

public feels that Sofia can do way more

than she actually can that's right and

the people will create a Sofia are not

honestly publicly communicating trying

to teach the public right but here's a

tough question don't you think this the

same thing is scientists in industry and

research are taking advantage of the

sameness misunderstanding in the public

when they create AI companies or

published stuff some companies yes I

mean there is no sense of there's no

desire to delude there's no desire to

kind of over claim what something is

done right you know you should paper on

AI that you know has this result on

image net you know it's pretty clear I

mean it's not even not even interesting

anymore but you know I I don't think

there is that I mean the reviewers are

generally not very forgiving of of you

know unsupported claims of this type and

but there are certainly quite a few

startups that have had a huge amount of

hype around this that I find extremely

damaging and I've been calling it out

when I've seen it so yeah but to go back

to your original question like the

necessity of embodiment I think I don't

think embodiment is necessary I think

grounding is necessary so I don't think

we're gonna get machines that I really

understand language without some level

of grounding in the world world and it's

not clear to me that language is a kind

of bandwidth medium to communicate how

the real world works I think what this

doctor ground our grounding means so

running me he's that

so there is this classic problem of

common sense reasoning you know the the

Winograd Winograd schema right and so I

tell you the the trophy doesn't fit in

the suitcase because this tool is too

big what the trophy doesn't fit in the

suitcase because it's too small and the

it in the first case refers to the

trophy in the second case to the

suitcase and the reason you can figure

this out is because you know what the

trophy in the suitcase are you know one

is supposed to fit in the other one and

you know the notion of size and the big

object doesn't fit in a small object and

this is a TARDIS you know it things like

that right so you have this got this

knowledge of how the world works of

geometry and things like that I don't

believe you can learn everything about

the world by just being told in language

how the world works I think you need

some low-level perception of the world

you know be a visual touch you know

whatever but some higher bandwidth

perceptions of the world but by reading

all the world's text you still may not

have enough information that's right

there's a lot of things that just will

never appear in text and that you can't

really infer so I think common sense

will emerge from you know certainly a

lot of language interaction but also

with watching videos or perhaps even

interacting in the in virtual

environments and possibly you know robot

interacting in the real world but I

don't actually believe necessarily that

this last one is absolutely necessary

but I think there's a need for some

grounding but the final product doesn't

necessarily need to be embodied you know

who say no it just needs to have an

awareness a grounding right but it needs

to know how the world works to have you

know to not be frustrated frustrating to

talk to and you talked about emotions

being important that's that's a whole

nother topic well so you know I talked

about this the the basal ganglia ganglia

as the you know this thing that could

you know calculates your level of miss

contentment contentment and then there

is this other module that sort of tries

to do a prediction of whether you're

going to be content or not that's the

source of some emotion so here for

example is an anticipation of bad things

that can happen to you right

you have this inkling that there is some

chance that something really bad is

gonna happen to you and that creates

here when you know for sure that

something bad is gonna happen to you you

cannot give up right it's not bad

anymore it's uncertainty it creates fear

so so the punchline is yes we're not

gonna have a ton of intelligence without

emotions whatever the heck emotions are

so you mentioned very practical things

of fear but there's a lot of other mess

around but there are kind of the results

of you know drives

yeah there's deeper biological stuff

going on and I've talked a few folks on

this there's a fascinating stuff that

ultimately connects to our joy to our

brain if we create an AGI system sorry

interminable human level intelligence

system and you get to ask her one

question what would that question be you

know I think the the first one we'll

create would probably not be that smart

did you like a four-year-old okay so you

would have to ask her a question - no

she's not that smart

yeah well what's a good question to ask

you know to be responsive wind and if

she answers oh it's because the leaves

of the tree are moving in that creates

wind she's on to something and if she

says yeah that's a stupid question

she's really obtuse no and then you tell

her actually you know here is the the

real thing and she says oh yeah that

makes sense

so questions that that reveal the

ability to do common-sense reasoning

about the physical world yeah and you

know someone will call 20 ferns causal

evidence well it was a huge honor

congratulations returning award you know

and thank you so much for talking today

thank you

you