Season 2 Ep 22 Geoff Hinton on revolutionizing artificial intelligence... again

Summary

在这段深入的访谈中，人工智能领域的先驱杰夫·辛顿（Geoff Hinton）分享了他对深度学习和神经网络的深刻见解。辛顿教授讨论了深度学习的起源、发展以及它如何成为当今最突出的人工智能方法。他提到了自己在深度学习领域的工作，包括他在图像识别方面的突破，这一成就被称为“ImageNet时刻”，极大地推动了整个AI领域的发展。此外，辛顿还探讨了当前AI技术的局限性，特别是与人类大脑的工作方式相比，他认为现有的深度学习技术如反向传播可能与大脑的处理机制大相径庭。他提出了未来可能的研究方向，包括无监督学习、局部目标函数以及模拟大脑功能的新型学习算法。辛顿的工作不仅对理解大脑的工作原理具有重要意义，也为开发更高效、更强大的AI系统指明了方向。

Takeaways

• 📈 过去十年，人工智能在多个领域取得了突破性进展，包括计算机视觉、语音识别、机器翻译、机器人技术、医学、计算生物学等，这些进展直接推动了数万亿美元公司和许多新创公司的业务发展。
• 🧠 深度学习作为人工智能的一个子领域，是这些突破的基础。杰夫·辛顿（Jeff Hinton）是深度学习领域的先驱，他的工作被引用超过五十万次，对整个领域产生了深远影响。
• 🏆 杰夫·辛顿因其在深度学习领域的贡献，被授予相当于计算机科学领域的诺贝尔奖，并且至今仍在领导该领域的研究。
• 🔬 辛顿在2012年展示了深度学习在图像识别方面比其他方法更优越，这一成果被称为ImageNet时刻，它改变了整个人工智能领域的研究方向。
• 🤖 辛顿认为，尽管现有的人工智能系统在某些方面非常高效，但它们与大脑的工作方式存在根本差异，特别是在反向传播机制方面。
• 🌟 辛顿提出了关于大脑如何使用局部目标函数来学习的想法，这可能与大脑的学习机制更为接近，而与现有的深度学习算法不同。
• 📚 辛顿和他的同事们提出了一种新的学习算法，称为SimCLR，它通过自我监督学习来提高性能，这与大脑可能使用的学习机制相似。
• 🔬 辛顿讨论了大脑如何处理大量参数并从中学习，以及这与现有神经网络的对比，特别是在能量效率和计算方式上的差异。
• 💡 辛顿提出了“死亡计算”（mortal computing）的概念，这是一种新的计算方式，它允许计算机通过学习获得知识，而不是通过编程，并且当计算机停止工作时，其知识也随之消失。
• 🧵 辛顿还探讨了关于睡眠的计算功能，提出了睡眠可能与神经网络中的负相学习（negative phase of learning）有关的理论。
• 👁️‍🗨️ 最后，辛顿讨论了如何通过可视化技术，如t-SNE，来理解神经网络在高维空间中的数据表示，以及这些技术如何帮助我们更好地理解机器学习模型的内部工作机制。

Q & A

深度学习在哪些领域取得了突破性进展？

深度学习在计算机视觉、语音识别、机器翻译、机器人技术、医学、计算生物学、蛋白质折叠预测等多个领域取得了突破性进展。

深度学习为何能驱动万亿级公司和众多初创公司的业务发展？

深度学习通过提供先进的算法和技术，使得公司能够在图像识别、数据处理、自动化决策等方面取得显著进步，从而推动了业务的创新和增长。

杰弗里·辛顿（Geoffrey Hinton）在人工智能历史上的地位如何？

杰弗里·辛顿被认为是人工智能历史上最重要的人物之一，他不仅是深度学习领域的先驱，而且至今仍在领导该领域的研究。

什么是神经网络，我们为什么需要关注它？

神经网络是一种模仿人脑神经元工作方式的计算模型，它通过调整神经元之间的权重来学习和处理信息。我们关注神经网络是因为它们在模式识别、数据处理和自动化决策等方面展现出强大的能力。

杰弗里·辛顿对于大脑工作原理的理解有何看法？

辛顿认为我们对大脑工作原理的理解还很有限，但他相信在未来五年内会取得重大突破。他提出现有的人工智能并非完全模仿大脑的工作方式，尤其是在反向传播算法方面。

为什么说反向传播算法可能与大脑的学习机制不同？

辛顿认为反向传播算法在参数调整上可能与大脑的机制不同。大脑可能使用了一种不同的方式获取梯度，这种方式可能更适合从较少的数据中抽象出结构。

自监督学习和无监督学习在深度学习中扮演什么角色？

自监督学习和无监督学习是深度学习中的关键部分，它们允许模型从未标记的数据中学习，这减少了对大量标记数据的依赖，并且能够从数据中提取更多的结构化信息。

为什么说大脑可能使用了多个局部目标函数来学习？

辛顿提出大脑可能使用了多个局部目标函数来学习，这样可以通过局部的不一致性来学习更多的信息。这种学习方式可能比单一的端到端系统更为高效和灵活。

在深度学习中，如何处理从不同图像块中提取的特征之间的一致性？

通过设计目标函数，使得不同图像块提取的特征之间存在一致性。如果这些特征在给定的上下文中预测一致，那么它们通常会达成一致，如果不一致，则可以作为学习信号。

为什么说大规模的神经网络模型（如GPT-3）可能并不完全理解它们所生成的内容？

尽管大规模神经网络模型能够生成连贯且看似有逻辑的内容，但它们可能只是在统计层面上理解了训练数据中的模式，而并没有真正理解内容的深层含义。

在深度学习中，如何处理模型的泛化能力，特别是在面对对抗性样本时？

对抗性样本揭示了模型可能依赖于纹理等浅层特征进行识别，而非深层的语义理解。提高模型的泛化能力需要更深入地研究其决策过程，并可能需要开发新的算法或架构。

Outlines

00:00

😀 深度学习的起源与影响

本段落介绍了深度学习作为人工智能的一个子领域，在过去十年中在多个领域取得的突破性进展。特别提到了杰弗里·辛顿（Geoffrey Hinton）的贡献，他被视为AI历史上最重要的人物之一，其研究工作被引用超过五十万次。辛顿在2012年的ImageNet竞赛中展示了深度学习在图像识别上的显著优势，这一成就促使整个AI领域转向深度学习。

05:01

🧠 大脑工作机制的探索

辛顿讨论了大脑的工作方式，特别是神经元如何通过调整权重来响应输入信号。他提出了关于大脑如何调整这些权重的疑问，并认为这是理解大脑工作的关键。辛顿还表达了对未来几年内破解大脑工作机制的乐观态度，尽管他怀疑目前深度学习中使用的反向传播（back propagation）算法与大脑的实际工作机制大相径庭。

10:01

🔄 反向传播的效率与局限性

辛顿纠正了关于他发明反向传播的说法，他实际上与同事一起展示了反向传播能够学习有趣的表示，如词嵌入。他认为反向传播比大脑中的学习过程更高效，但可能不如大脑在从少量数据中抽象出大量结构方面做得好。辛顿提出了无监督目标函数的概念，认为这是关键的学习方式，并且大脑可能使用许多局部目标函数。

15:02

🤖 自监督学习与大脑的相似性

辛顿讨论了自监督学习的概念，特别是通过比较图像的小区域来提取表示，并将其与基于其他区域表示的上下文预测进行比较。他认为这种方法可能更接近大脑的工作方式，并且可以通过局部不同意来学习。他还提到了SimCLR论文和他早期关于通过图像不同区域之间的表示达成一致来进行自监督学习的论文。

20:02

🧬 学习算法与大脑的相似性

辛顿讨论了当前的学习方法，如端到端学习和反向传播，并提出了如何从更少的数据中提取更多信息以改进学习方法的问题。他提出了使用许多小的局部目标函数来提高学习带宽的想法，并讨论了大脑如何使用这些局部目标函数。他还提到了大脑可能不使用与当前神经网络相同的函数来进行局部处理。

25:03

🌱 知识的转移与共享

辛顿提出了通过局部区域之间的相互教学来实现知识转移的想法，这与权重共享不同，但提供了一种更灵活且更符合生物神经网络的方式。他认为这种相互教学的机制可能是大脑如何共享知识的方式，并且这种方法比权重共享更有效。

30:07

🚀 从学术界到工业界的转变

辛顿讲述了他从学术界转向工业界的经历，包括他在多伦多大学的不愉快经历，以及他如何通过Coursera课程和随后的公司拍卖来寻找赚取资金的其他方式。他详细描述了拍卖的过程，以及他们如何决定在Google工作，因为他喜欢那里的研究环境和团队。

35:08

🏆 学术成就与职业转变

辛顿分享了他的学术背景，包括他在剑桥的学习经历，以及后来成为木工的时期。他讨论了他对木工的热爱，以及他如何意识到自己在木工方面的不足，这促使他回到学术界。他还提到了他在爱丁堡大学攻读神经网络博士学位的经历，以及他早期关于神经网络的研究工作。

40:09

🤔 深度学习的未来与大脑的比较

辛顿反思了他对深度学习的看法，包括他所说的“深度学习会做到一切”的评论。他解释了他对深度学习的定义，以及他如何看待大脑使用局部目标函数的方式。他还讨论了计算机的当前设计，以及它们与大脑的工作方式的差异。

45:10

🧵 制造与生长计算机的未来

辛顿提出了一个关于计算机设计的未来愿景，其中包括放弃程序的不朽性以换取低能耗计算和便宜的制造过程。他讨论了使用纳米技术“生长”计算机的想法，这些计算机通过学习获得所有知识，并在它们死亡时失去所有知识。

50:13

🌟 神经网络的规模与理解

辛顿讨论了大型神经网络，如大型语言模型，以及它们在处理语言和图像识别方面的成就。他提出了关于这些模型是否真的理解它们所处理的信息的问题，并将它们的能力与昆虫的视觉系统进行了比较。

55:14

💤 睡眠的计算功能

辛顿探讨了睡眠的计算功能，提出了睡眠可能与学习过程的负面阶段有关的理论。他认为睡眠期间大脑可能在进行一种“非学习”或“非优化”的过程，这有助于巩固记忆并避免过度拟合。

00:16

📉 学生超越老师的现象

辛顿描述了即使在标签错误的情况下，神经网络（学生）如何能够超越其训练数据（老师）。他讨论了通过使用大量有噪声的标签数据，神经网络如何学习并最终提供比训练数据更准确的输出。

05:18

📈 t-SNE的发明与高维数据可视化

辛顿讲述了他如何发明t-SNE（t分布随机邻域嵌入）算法，这是一种用于高维数据可视化的技术。他解释了t-SNE如何通过使用高斯混合模型来改进早期的SNE算法，并且如何通过区分不同的尺度来同时显示数据的粗略结构和细节。

Keywords

💡深度学习

深度学习是一种机器学习技术，它通过模拟人脑中的神经网络来识别模式，广泛应用于图像识别、语音识别、自然语言处理等领域。在视频中，深度学习是讨论的核心，因为它是实现AI突破的关键技术之一。

💡神经网络

神经网络是深度学习的基础，由相互连接的神经元组成，模仿人脑的处理方式。视频提到，理解神经网络的工作原理对于揭示大脑的工作机制至关重要。

💡反向传播

反向传播是深度学习中用于训练神经网络的一种算法，通过调整网络中的权重来最小化误差。视频中提到反向传播可能与大脑的学习机制不同，这表明未来可能需要新的学习算法。

💡自监督学习

自监督学习是一种无监督的机器学习方法，它使模型能够从未标记的数据中学习。视频中讨论了自监督学习的重要性，以及它是如何通过比较不同图像块的表示来实现的。

💡梯度下降

梯度下降是一种优化算法，用于最小化损失函数，通过迭代调整参数来寻找函数的最小值。在深度学习中，梯度下降用于训练神经网络，调整网络权重。

💡ImageNet时刻

ImageNet时刻是指2012年深度学习在ImageNet竞赛中取得显著成绩，大幅超越其他方法，这一事件标志着深度学习时代的到来。视频中提到这一时刻对整个AI领域产生了深远影响。

💡局部目标函数

局部目标函数是针对数据的一小部分定义的优化标准，与全局目标函数相对。视频中提到，大脑可能使用许多局部目标函数来进行学习，这与当前深度学习使用的端到端学习策略不同。

💡能量函数

在机器学习中，能量函数用于表示一个系统的能量状态，通常用于指导模型学习。视频中提到，通过调整神经元的权重来改变能量函数，可以帮助神经网络学习。

💡dropout

dropout是一种正则化技术，通过在训练过程中随机关闭神经网络中的一些神经元来防止过拟合。视频中提到dropout作为一种技术，帮助了深度学习模型在ImageNet竞赛中取得成功。

💡稀疏编码

稀疏编码是一种表示数据的方式，其中大部分元素是零或接近零。在深度学习中，稀疏编码有助于提取更有意义的特征。视频中提到稀疏编码作为一种技术，可以提高神经网络的性能。

💡玻尔兹曼机

玻尔兹曼机是一种随机神经网络，能够学习输入数据的概率分布。视频中提到玻尔兹曼机在早期深度学习研究中的重要性，以及它是如何启发了后续的深度学习模型。

Highlights

过去10年中，人工智能在计算机视觉、语音识别、机器翻译、机器人技术、医学、计算生物学等多个领域取得了突破性进展。

深度学习作为AI的一个子领域，对上述突破起到了基础性作用。

杰弗里·辛顿（Geoffrey Hinton）被认为是AI历史上最重要的人物之一，对深度学习的起源和发展有着深远的影响。

辛顿的工作被引用超过五十万次，意味着有大量研究论文在其研究基础上进行。

2012年，辛顿展示了深度学习在图像识别方面比其他方法更优越，这一时刻被称为ImageNet时刻，引发了AI领域的转变。

辛顿认为，尽管现有的AI主要基于反向传播算法，但人脑可能使用了一种不同的机制来调整神经网络的权重。

辛顿提出了无监督目标函数的概念，强调通过观察世界来学习一个好的世界模型，从而基于模型而非原始数据采取行动。

辛顿认为大脑可能使用许多局部目标函数，而非单一的端到端系统。

辛顿提出了一种基于图像局部区域表示的潜在目标函数，通过比较局部提取的特征与基于上下文的预测来学习。

辛顿讨论了自监督学习的重要性，以及它与大脑学习方式的相似性。

辛顿指出，大脑的神经元通过尖峰信号工作，这与当前人工神经网络的工作原理有显著不同。

辛顿认为，为了更好地模拟大脑，未来的学习算法可能需要放弃程序的“不朽性”，转而设计能够通过学习获得知识的“有死亡率的计算机”。

辛顿提出了“学生超越老师”的概念，说明了即使在标签错误的数据集上，深度学习模型仍然能够学习并超越其训练数据的准确性。

辛顿讨论了睡眠的计算功能，提出了睡眠可能与负面学习或记忆整合有关的理论。

辛顿介绍了t-SNE（t分布随机邻域嵌入）算法，这是一种用于高维数据可视化的技术，能够揭示数据点之间的相似性。

辛顿强调了在构建模型时理解模型学习内容的重要性，并讨论了可视化技术在理解学习过程中的应用。

Transcripts

[Music]

[Music]

over the past 10 years ai has

experienced breakthrough after

breakthrough after breakthrough in

computer vision in speech recognition in

machine translation in robotics in

medicine in computational biology

protein folding prediction and the list

goes on and on and on and the

breakthroughs aren't showing any signs

of stopping not to mention these ai

breakthroughs are directly driving the

business of trillion dollar companies

and many many new startups

underneath all of these breakthroughs is

one single subfield of ai

deep learning

so

when and where did deep learning

originate

and when did it become the most

prominent ai

approach today's guest has everything to

do with this

today's guest is arguably the single

most important person in ai history and

continues to lead the charge today

award the equivalent of the nobel prize

for computer science

today's guest has their work cited over

half a million times

that means there is half a million and

counting other research papers out there

that build on top of his work

today's guest has worked on deep

learning for about half a century

and most of the time

in relative obscurity

but that all changed in 2012

when he showed deep learning is better

at image recognition than any other

approaches to computer vision and by a

very large margin

that result that moment

known as the imagenet moment

changed the whole ai field

pretty much everyone dropped what they

had been doing and switched to deep

learning

former students of today's cast include

vlognee who put deep mind on the map

with their first major result on

learning to play atari games

and includes our season one finale guest

eliza skiver founder and research

director of openai

in fact every single guest in our

podcast has built on top of the work

done by today's guest

i am of course

talking about no one less than jeff

hinton

chef welcome to the show so happy to

have you here

well thank you very much for inviting me

well so glad to get to talk with you on

the show here and i'd say let's dive

right in with maybe the you know the

highest level question

i can ask you um

what are neural nets and why should we

care

okay if you already know a lot about

neural nets please forgive the

simplifications

um

here's how your brain works

it has lots of little processing

elements called neurons

and every so often a neuron goes ping

and what makes it go ping is that it's

hearing pings from other neurons and

each time it hears a ping from another

neuron

it adds a little weight to some store of

um input that it's got and when it gets

it when it's got enough input it goes

ping

and so if you want to know how the brain

works all you need to know

is how the neurons decide to adjust

those weights

that they add when a ping arrives um

that's all you need to know this there's

got to be some procedure used for

adjusting those weights and if we could

figure it out we'd know how the brain

works and that's been your quest for a

long time now figuring out how the brain

might work and

what's the status do you do we as a

field understand how the brain works

okay i always think we're going to crack

it in the next five years since that's

quite a productive thing to think

um but i actually do and i think we're

going to crack it in the next five years

um i think we're getting closer

i'm fairly confident now that it's not

back propagation

so all of existing ai i think

is built on something that's quite

different from what the brain's doing

at a high level it's got to be the same

that is you have a lot of parameters

these weights between neurons

and you adjust those parameters

um on the basis of lots of training

examples

and that causes wonderful things to

happen if you have billions of

parameters

the brain's like that and deep learning

is like that

the question is

how do you

um get the gradient for adjusting those

parameters so what you want is some

some measure of how well you're doing

and then you want to adjust the

parameters so they improve that measure

of how well you're doing

um

but my belief currently is that

back propagation which is the way deep

learning works at present

is quite different from what the brain's

doing the brain's getting gradients in a

different way now that's interesting

you're the one saying that jeff because

you actually

you

wrote a paper on back recreation for

training neural networks and it's

powering everything everybody's doing

today

and now here you are saying actually

it's time probably time for us to figure

out how do you think we should change it

close to what the brain is doing or do

you think maybe

back repetition could be better than

what the brain is doing

let me first correct you

um yes we did write the most cited paper

on back propagation on

ronald hunt and williams and me um

back propagation was already known um

to a number of different authors what we

really did was showed that it could

learn interesting representations so it

wasn't that we invented back propagation

we

ronald hart reinvented back propagation

we showed that it could learn

interesting representations like for

example word embeddings so

i think back propagation is probably

much more efficient

than what we have in the brain that's

squeezing a lot of information into a

few connections

whereby a few connections i mean only a

few billion

so the problem the brain has is that

connections are very cheap

um we've got

hundreds of trillions of them

um

experience is very expensive

and so we are willing to throw lots and

lots of parameters

at a small amount of experience

whereas the neural nets we're using are

basically the other way around

they have lots and lots of experience

and they're trying to get the

information

about what relates the input to the

output into the parameters

and i think back propagation is much

more efficient than what brain's using

at doing that

but maybe not as good

at from not much data

abstracting a lot of structure

and well this begs the question of

course do you have any hypothesis on

approaches

that might get

better performance in that regard i have

a sort of general view which i've had

for a long long time

which is that we need unsupervised

objective functions so i'm talking

mainly about perceptual learning

um

which i think is the sort of key if you

can learn a good model of the world by

looking at it

um

then you can base your actions on that

model rather than on the raw data

and that's going to make

doing the right things much easier

i'm convinced that the brain

is using lots of little local objective

functions

so rather than being a kind of

end-to-end system chain trained to

optimize one objective function

i think it's using lots of little local

ones um so as an example

the kind of thing i think will make a

good objective function though it's hard

to make it work

is

if you look at a small patch of an image

and try and extract some representation

of what you think is there

you can now compare the representation

you got from that small patch of the

image

with a contextual bet that was got by

taking the representations of other

nearby patches and based on those

predicting what that patch of the image

should have in it

and obviously

um

once you're very familiar with the

domain

those predictions from context and

locally extracted features will agree

generally agree and you'll be very

surprised when they don't and you can

learn an awful lot on one trial if they

disagree radically

so that's an example of where i think

the brain could learn a lot from this

local disagreement

um

it's hard to get that to work but i'm

convinced something like that is going

to be the objective function and if you

think of a big image

and lots of little local patches in the

image

that means you get lots and lots of

feedback in terms of the agreement of

what was extracted locally and what was

predicted contextually

um

all over the image

and at many different levels of

representation

and so we can get a much much richer

feedback

from these agreements with contextual

predictions but making all that work is

difficult

but i think it's going to be along those

lines

now

what you're describing strikes me as

part of what people are trying to do in

self-supervised and unsupervised

learning and in fact you wrote one of

the breakthrough papers the sim clear

paper with with a couple of

collaborators of course

in this space what do you think about

the simclear work and contrast of

learning more generally

and what do you think about the recent

mast auto encoders and how does that

relate to what you just described it

relates quite closely to what i've it's

evidence that that kind of objective

function is good

um

i didn't write the sim clear paper um

team chain written simply a paper um

with help from the major co-authors i

was my name was on the paper for general

inspiration but

i did write a paper a long time ago with

sue becker

on the idea of getting agreement

between representations you got from two

different patches of the image

so that was i think of that as the

origin of this idea

of doing self-supervised learning by

having agreement between representations

from two patches of the same image

um

the method that sue and i used

didn't work very well

because of a subtle thing that we didn't

understand at the time but i now do

understand

um

and i could explain that if you like but

i'll lose most of the audience

well i'm curious i think it'd be great

great great to hear it but maybe we can

zoom out for a moment before zooming

back in you talk about

current methods use end-to-end learning

back propagation

to power the end-to-end learning

and you're saying switch to learn from

less data and extract more from less

data is going to be key as as as a way

to make progress to get closer to how

the brain learns

um yes so you get much bigger bandwidth

for learning by having many many little

local objective functions

and when when we look at these local

objective functions like filling in a

blanked out part of an image or maybe

filling back in a word

if we look at today's technologies in

fact this is the current frontier you've

contributed a lot of people are working

exactly on on that problem of

learning from unlabeled data effectively

because it requires a lot less

human labor but they still use back

propagation

the same mechanism so what i do

what i don't like about the mast auto

encoder is

you have

your input patches

and then you go through many layers of

representation

and at the output of the net you try to

reconstruct the missing input patches

i think

the brain you have these levels of

representation but at each

level you're trying to reconstruct

what's at the level below

um so it's not like you go through this

many many layers and then come back out

again

um it's that you have all these levels

each of which is trying to reconstruct

what's at the level below

um

so i think that's much more brainlike

and the question is can you do that

without using back propagation obviously

if you go through many many levels and

then reconstruct the missing patches of

the output you need to get information

back through all those levels

and since we have back propagation it's

built into all the simulators you might

as well do it that way but i don't think

that's how the brain's doing it

and now imagine the brain is doing with

all these local objectives

do you think for for our engineered

systems

will it matter whether

in some sense there are three choices to

make it seems one choice is

what are the objectives what are those

local objectives that we want to

optimize

a second choice is

what's the algorithm

to use

to optimize it and then a third choice

is

what's the architecture of how do we

wire the neurons together that are

doing this this learning

and

among those three it seems like all

three

could be the missing piece that we're

not getting right or what do you think i

if you're interested in perceptual

learning

i think it's fairly clear you want

retinotopic maps a hierarchy of written

topic maps

so the architecture is local

connectivity

um

and the point about that is

you can solve a lot of the credit

assignment problem by just assuming that

something in one locality in a

retrotronpic map is going to be

determined

by the corresponding locality in the

retinotopic map that feeds into it

so

you're not trying to low down in the

system

um figure out how

pixels

determine what's going on a long

distance away in the image

you're going to just use local

interactions and that gives you a lot of

locality

um

and you'd be crazy not to lose not to

use that

one thing neural nets do at present is

they assume you're going to be using the

same functions at every locality so

convolutional let's do that and

transformers do that too

um

i don't think the brain can do that

because that would involve weight

sharing

and it would involve doing exactly the

same computation at each locality so you

can use the same weights

i think it's most unlikely the brain

does that

but actually

there's a way to achieve what weight

sharing does what convolutional

nets do in the brain in a much more

plausible way than i think people have

suggested before

which is

if you do have

contextual predictions trying to agree

with locally extracted things

then imagine a whole bunch of columns

that are making local predictions and

looking at nearby columns to get their

contextual prediction

you can think of the context as a

teacher for the local thing

but also vice versa

but think of the context as a teacher

for what you're attracting locally

so you can think of the information

that's in the context as being distilled

into the local extractor but that's true

for all the local extractors

so what you've got is mutual

distillation

where they're all providing teaching

signals for each other

and what that means is knowledge about

what you should extract in one location

is getting transferred

into other locations

if they're trying to agree if you're

trying to get different locations to

agree on something if for example you

find a nose and you find a mouth and you

want them both to agree that they're

part of the same face

so they should both give rise to the

same representation

then the fact that you're trying to get

the same representation at different

locations allows knowledge to be

distilled from one location to another

and there's a big advantage of that over

actual weight sharing

obviously biologically one advantage is

that the detailed architecture in these

different locations doesn't need to be

identical

but the other advantage is the front-end

processing doesn't need to be the same

so if you take your retina

different parts of the retina have

different size receptive fields

and convolutional nets try to ignore

that they sometimes have multiple

different resolutions and do convolution

at each resolution but they just can't

deal with different front-end processing

whereas if you're distilling knowledge

from one location to another

what you're trying to do is get the same

function

from the optic array

to the representation in these different

locations

and

it's fine if you pre-process the optic

array differently in the two different

locations

you can still distill the knowledge

across the function from the optic array

to the representation even though the

frontend processing is different

and so

although distillation is less efficient

than actually showing the weights it's

much more flexible and it's much more

neurally plausible

so for me that was a kind of big insight

i had about a year ago that

we have to have something like weight

sharing to be efficient

but local distillation will work if

you're trying to get neighboring things

to agree on a representation

but that idea of trying to get them to

agree gives you the signal you need for

knowledge in one location to supervise

knowledge in another location

and jeff do you think so what you're

describing one way to think of it is to

say hey we're cheering

is clever because it's something the

brain kind of does too it just does it

differently so we should continue to do

weight sharing another way to think of

it is that actually we shouldn't

continue to do weight sharing because

the brain does it somewhat differently

and it might be be a reason to do it

differently

what's your thinking i think the brain

doesn't do weight sharing because it's

hard for it to

ship cement strengths about the place

it's very easy if they're all sitting in

ram

so i think we should continue to do

convolutional things in convnets and in

transformers we should share weights

um we should share knowledge by sharing

weights

but just bear in mind that the brain is

going to share knowledge not by sharing

weights but by sharing the function from

input to output

and using distillation to transfer

knowledge

now there's the other topic that is

talked about quite a bit where the brain

is drastically different

from our current neural nets and is the

fact that

neurons are work with spiking signals

and that's very different from our

artificial neurons in our gpus

and so i'm very curious on your thinking

on that is is that just an engineering

difference or do you think there could

be

more to it that we need to understand

better and benefits to spiking

i think it's not just an engineering

difference i think once we understand

why that hardware is so good

why you can do so much in such an energy

efficient way with that kind of hardware

thing

we'll see that um it's sensible for the

brain geospiking units the retina for

example doesn't use spiking neurons the

retina does lots of processing with

non-spiking nerves

so

once we understand why cortex is using

those

um

we'll see that it was the right thing

for biology to do and i think that's

going to hinge on what the learning

algorithm is how you get gradients

for networks of spiking neurons and at

present nobody really knows the present

what people do is say

you see the problem with the spiking

urine is

there's two diff quite different kinds

of decision

one is exactly when does it spike and

the other is does it or doesn't it spike

so there's this discrete decision should

the neuron spike or not and then this

continuous variable of exactly when it

should spike

and people trying to optimize a system

like that have come up with various kind

of surrogate functions which sort of

smooth things a bit so you can get

continuous functions they didn't seem

quite right

um

it'd be really nice to have a learning

algorithm and in fact

in nips in about 2000 andy brown and i

had a paper on trying to learn spiking

boltzmann machines um

but it'd be really nice to

get a learning algorithm that's good for

spiking yards and i think that's the

main thing that's holding up spiking

neuron hardware so people like steve

ferber in manchester have realized that

many other people have realized that um

you can make more energy efficient

hardware this way and they built great

big systems what they don't have is a

good learning outcome for it and i think

until we've got a good learning

algorithm for it we won't really be able

to exploit what we can do with spiking

neurons and there's one obvious thing

you can do with them that isn't easy in

conventional neural nets

and that's

agreement

so if you take a standard artificial

neuron

then you simply ask the question can it

tell if it's two inputs have the same

value

well it can't it's not an easy thing for

a standard neuron to do the standard

artificial one

um if you use spiking neurons it's very

easy to build a system where if the two

spikes arrive at the same time they'll

make when you're on fire and if they're

over different times they won't

so using the time of a spike

seems like a very good way of measuring

agreement we know the biological system

does that

so

you can see the direction a sound is

coming from or rather here the direct

sound is coming from by the time delay

in the signals reaching the two

ears and

if you take a foot

that's about a nanosecond for light

and it's about a millisecond first sound

and the point is if i move something

sideways in front of you by a few inches

the difference

in

the time delay to the two ears the

length of the path to the two ears

is only a small fraction of an inch

and so

it's only a small fraction of a

millisecond difference

in the time the signal gets to the two

is and we can deal with that and owls

can deal with it even better um

and so we're measuring we're sensitive

to times of like um 30 milliseconds 30

microseconds

in order to get stereo from sound

um i can't remember what i was sensitive

to but it's i think it's a lot better

than 30 microseconds

and we do that by having

um two axons with spikes traveling in

different directions one from one and

one from the other ear and then you have

cells that fire if the spikes get at the

same time

that's the simplification but roughly

that um

so we know that spike timing can be used

for exquisitely sensitive things like

that

and it would sort of be very surprising

if the precise times the spike wasn't

being used but we really don't know how

and for a long time i thought

it'd be really nice if you could use

spike times to detect agreement for

things like self-supervised learning

or for things like

um

if i've extracted your mouth and i've

extracted your nose or other

representations of them

and

i

from your mouth i can now predict

something about your whole face and from

your nose i can break something about

your whole face

and if your mouth and nose are in the

right relationship to make a face those

predictions will agree

and it'd be really nice to use spike

timing to see that those predictions

agree

um

but it's hard to make that work and one

of the reasons it's hard to make that

work is because we don't know we don't

have a good algorithm for training

networks just like in neurons so that's

one of the things

i'm focused on now how can we get a good

training out of the networks of spiking

yours and i think that'll have a big

impact on hardware

that's a really interesting question

you're putting forward there because i

doubt too many people are working on

that compared to let's say the number of

people working on

large language models or

other problems that are much more i

guess

visible in terms of progress recently

um

i think

yeah it's always a good idea to figure

out what huge numbers of very smart

people are working on and to work on

something else

yeah i think the challenge of course for

most people

i'd say including myself but i

definitely hear the question from many

students too is that

it's easy to work on something else than

everybody else but it's hard to make

sure that something else is actually

relevant

because there's many other things

out there that are not not very relevant

you could possibly spend time on yeah

that involves having good intuitions

yeah

listening to you for example could help

um so i've actually a follow-up question

something you just said jeff which is um

that

the retina

doesn't use all spiking neurons are you

saying that the brain has two types of

neurons some that are more like our

artificial neurons and some that are

spiking neurons

i'm not sure the retina is more like

artificial neurons but um

certainly the cortex has

the neocortex has spiking neurons

um and that's its primary mode of

communication is by sending spikes

to from one parameter to another

parameter cell

um and i don't think we're gonna

understand the brain until we understand

why he chooses to send spikes

for a while i thought i had a good

argument that didn't involve

the precise time to spikes and the

argument went like this

the brains in the regime where it's got

lots and lots of parameters and not much

data

relative to

the typical neural nets we use

and

there's a potential overfitting in that

regime unless you use very strong

regularization

and a good regularization technique is

dropout where each time you use a neural

net you ignore a whole bunch of the

units

and so maybe

the fact that the neurons are sending

spikes

what they're really communicating

is the underlying poisson rate

so let's assume it's pass on but she's

close enough for this argument um

there's a price on process which spends

sends spike stochastically

but the rate of that process varies and

that's determined by the input to the

neuron

and you might think you'd like to send

the real valued rate from one urine to

another

um

but if you want to do lots and lots of

regularization

you could send the real valued rate with

some noise added

and one way to add noise is to just use

spikes that'll add lots of noise

and so

this was the motivation for dropout that

the

most of the times most of the neurons

aren't involved in things if you look at

any fine time window

um

and you can think of

spikes

as a representational underlying

personal rate it's just a very very

noisy representation

which sounds like a very very bad idea

because it's very very noisy but

actually once you understand about

regularization we have too many

parameters it's a very very good idea

so

i still have a

lingering fondness for the idea but

actually we're not using spike timing at

all um it's just about

using very noisy representations of

personal rates to be a good regularizer

and i sort of flip between i think it's

very important when you do science

not to totally commit to one idea and

ignore all the evidence for other ideas

but if you do that you end up

um

flipping between ideas every few years

so

some years i think neural nets are

deterministic i mean we should have

deterministic neural nets and that's

what backwards using

another years i think it's about a

five-year cycle i think no no it's very

important the best stochastic

um and that

changes the play for everything so

boltzmann machines were intrinsically

stochastic and that was very important

to them

um

but the main thing is not to fully

commit to either of those but to be open

to both

now one thing if we think more about

what you just said the importance of

spiking neurons and figuring out how to

train a spiking neuron network

effectively

what if we for now just say let's not

worry about the training part

given that

seemingly it's

far more power efficient

um wouldn't people want to distribute

pure inference chips that are you you

pre-trained effectively separately and

then you compile it onto a spiking

neuron

chip to have very low power inference

capabilities what about that

yeah so lots of people have thought of

that and um it's a very sensible idea

and it's probably on the evolutionary

path to getting to use backing neural

nets

because once you're using them for

inference

um

and it works and it's all people are

already doing that and it's already

working being shown to be more power

efficient

and various companies have produced

these big spiking systems um

once you're doing them for inference

anyway you'll get more and more

interested in

how you could learn in a way that makes

more use of the available power in these

spike times

so

you can imagine a system where you learn

using backprop

um

but not on

not on analog hardware for example not

on the this low energy hardware

um

and then you transfer it to the lower

energy hardware and that's fine

um

but we'd really like to learn directly

in the hardware

now one thing that really strikes me

jeff is when i think about

your talks back around

2005 6 7 8 when i was a phd student

essentially pre-alex ned talks

those talks i think topically have a lot

of resemblance to what you're excited

about now

and it almost feels like alexnet is an

outlier in in your path

um how did you go from thinking so

closely how the brain might work

to deciding that you know maybe you can

first explain what was alex net and but

also how did it come about and what was

that path to go from working on

restricted boltzmann machines trying to

see how the brain works too

i would say that

the more traditional approach to neural

nets that you all of a sudden show it

can actually work

well um if you're an academic you have

to raise grant money

and

it's convenient to have things that

actually work

even if they don't work the way you're

interested in

um so part of it's that just go with the

flow and

um

if you can make back prop work well

and back then in about 2006 2005 i got

fascinated by the idea you could use

stacks of restricted voltage machines

to pre-train feature detectors and then

it would be much easier to get backdrop

to work

it turned out with enough data

which is what you had in speech

recognition

um and later on because of faith ali and

her team

in image recognition with enough data

you don't need the pre-training although

pre-training is coming back i mean gpt-3

has pre-training

and pre-training is a thoroughly good

idea

um

but

once we

discovered that you can pre-train and

that will make backdrop work better and

that did great things for speech

which george john and abdul rahman

muhammad did um

in 2009

then alex

who was a graduate student in my group

thing

um

started uh

applying the same ideas to

vision

um

and pretty soon we discovered that you

didn't actually need this pre-training

especially if you had the imagenet

data

and in fact that project

um

was partly due to india's persistence so

i remember ilya coming into the lab one

day and saying look we now that we've

got speech recognition working this

stuff really works we've got to do

imagenet before anybody else does

and retrospectively learned that janella

come was going into the lab and saying

look we've got to do imagenet with

compliments before anybody else does

and

jan's students also and postdoc said oh

but i'm busy doing something else so

well

he he couldn't actually get someone to

commit to it

and

yeah initially couldn't get people to

commit to it

and so he persuaded alex to commit to it

by pre-processing the data for him so he

didn't have to pre-process the data the

data was all pre-processed to be just

what he needed

and then alex really went to china and

alex is just a superb programmer

and it was

alex was able to make a couple of gpus

really sing he made them work together

in his bedroom at home

um i don't think his parents realized

that they were paying most of the cost

because that was the electricity um

but

he did a superb job of programming

convolutional nets on them

um

so

he said we've got to do this

and helped alex with the design and so

on alex did the really intricate

programming

and i provided support um and a few

ideas like using dropout

i also did some good management i'm not

often very good at management but i'm

very proud of the management idea which

is alex kruszewski had to write a depth

or to show that he was sort of capable

of understanding research literature

which is what you have to do after a

couple of years to stay in the phd

program

and he doesn't really like writing um

and he didn't really want to do the

depth of it but it was way past the

deadline and the development was

hassling us

so i said to him

um each

each time you can improve the

performance by one percent on imagenet

um

you can delay your depth order by

another week

and alex delayed his death roll by a

whole lot of weeks

yeah and just for context for i mean a

lot of researchers know this of course

but maybe not everybody alex's result

with you and ilya

cut the error rate in half compared to

prior work on the imagenet image

recognition competition which was just

more or less i

i used to be a professor so it wasn't

quite in half close it cut it from about

a whole available well that's why

everybody switched from what they were

doing which was hand engineered

approaches to computer vision try to

program directly

how can a computer understand what's an

image to to deep learning i should say

one thing that's important to say here

um

jalakhar spent many years um developing

convolutional neural nets

um

and it really should have been him

his lab that developed that system we

had a few little extra tricks but they

weren't the important thing the

important thing was to apply

convolutional nets

using gpus to a big data set

um

so yam was kind of unlucky in that um he

didn't get the win on that

but it was using many of the techniques

that he developed it didn't have the the

russian immigrants that uh toronto and

eu had been able to attract to make it

happen

well once russian one's ukrainian and

it's important to confuse those even

though the ukraine is a russian-speaking

ukrainian don't confuse russian

absolutely

it's a it's a different country

so

now jeff that moment actually also

marked a big change in your career

because as far as i understand you've

never been involved in

corporate

work

but

it marked a transition for you soon

thereafter from being a pure academic to

being ending up at google actually

uh can you see a bit about that how was

that for you like

did you have any internal resistance i

can say why that transition happened

what triggered i'm curious

so

um i have a lonely disabled son who

needs

um

future provisions so i needed to get a

lump of money and i thought one way i

might get a lump of money was by

teaching a coursera course

and so i did a coursera course on neural

networks in

2012. and it was one of the early

coursera courses so their software

wasn't very good so it's extremely

irritating to do

um

it really was very irritating then

i'm not very good on software so i

didn't like that

and

from my point of view it amounted to

you agree to supply a chapter of a

textbook one chapter every week

um

so

you had to give them these videos and

then a whole bunch of people are going

to watch the videos

like sometimes the next day yoshi or

banjo would say why did you say that

um

so you know that it's going to be people

who know very little but also people

know a whole lot

and so it's stressful you know that if

you make mistakes they're going to be

caught not like a normal lecture where

you can just sort of

press on the sustaining pedal and sort

of

blow your way through it if you get some

slightly confused about something um

here you have to get it straight

and

the deal with the university of toronto

originally was that

um

if any money was made from these courses

which i was hoping there would be

um

the money that came to the university

would be split with the professor they

didn't specify exactly what the split

would be

but one assumed it would be like 50 50

or something like that

and i was okay with that

the university didn't provide any

support in preparing the videos

and

then after i started the course and when

i could no longer back out of it

the provost made a unilateral decision

without consulting

me or anybody else um that actually if

money came from coursera the university

would take all the money and the

professor would get zero

which is exactly the opposite of what

happens with textbooks

and the process was very like bringing

textbooks

i actually asked the university to help

me prepare the videos

and the av people came back to me and

said do you have any idea how expensive

it is to make videos

and i actually did have an idea because

i've been doing this

so i got really pissed off with my

university because they unilaterally

sort of canceled the idea i get any

immunization for this they said it was

part of my teaching well actually it

wasn't part of my teaching it was

clearly based on lectures i given as

part of my teaching but i was doing my

teaching as well as that and that i

wasn't using that course for my teaching

and that got me pissed off enough

that i was willing to consider

alternatives to being a professor

um and at that time

we then

suddenly got interest from all sorts of

companies

in the

in recruiting us

um either in funding giving big grants

or in funding a startup

it was clear that a number of big

companies were just very interesting

getting in on the act

and so

normally i would have just said no i'm i

get paid

by the state we're doing research

um i don't want to try and make extra

money from my research i'd rather get on

with the research

but because that particular experience

with

the university

um cheating me out of the money no it

turned out they didn't shoot me out of

anything because

uh no money came from of course anyway

um

but that pushed me over the edge into

thinking well

okay i'm going to find some other way to

make some money that was the end of my

princess oh no

well

but the result is that these companies

are and in fact if you read the the

genius makers book by kid metz which i

reread last week in preparation for this

conversation um if you read the book it

starts off with actually you running an

auction for these companies to try to

acquire your company which is quite the

start of a book

um very intriguing but how is it for you

oh when it was happening it was at nips

and terry had organized nips in a casino

um

at lake tahoe

um and so in the basement of the hotel

there were these smoke-filled rooms full

of people pulling one arm bandits and

big lights flashing saying you won 25

000 and all that stuff and people

gambling um in other ways

and upstairs we were running this

auction

and

we felt like we were in a movie we felt

like this was like being in

that movie the social network it sort of

felt like that it was great

the reason we did it was we had

absolutely no idea how much we were

worth

and i consulted

a lawyer who an ip lawyer who said

there's two ways to go about this

you could hire a professional negotiator

um

in that case um

you'll end up working for a company but

they'll be pissed off with you

or you could just run an auction

um

as far as i know this was the first time

a small group like that just ran an

auction we ran it on gmail

i've worked at google over the summer so

i knew enough about google to know that

they wouldn't read our gmail

um

and

i'm still pretty confident they didn't

read our gmail

microsoft wasn't so confident

and we just ran this auction where

people had to gmail me their bids

and we then immediately mailed them out

to everybody else with the timestamp of

the gmail

and

um

he just kept going up by a half million

dollars to and he was

half a million dollars to begin with and

then a million dollars after that

um

and

yeah it was pretty exciting

and we discovered we were worth a lot

more than we thought

retrospectively

we could probably have got more but we

we got to an amount that we thought was

astronomical

and then

basically

we wanted to work for google so we

stopped the auction so we could be sure

of working for you and as i understand

it you're still at google today

i'm still at google today

i'm nine years later i'm in my tenth

year there

i think i'll get some kind of award when

i beat them for 10 years because it's so

rare um

although people tend to stay in google

longer than other companies

yeah i like it there the

the main reason i like it is because

the brain team's a very nice team

and i get along very well with jeff dean

um

he's kind of

very smart but i'm very straightforward

to deal with

and what he wants me to do is do

what i want to do which is basic

research um he thinks what i should be

doing is trying to come up with

radically new algorithms and that's what

i want to do anyway

so it's just a

very nice fit i'm no good at managing a

big team to improve speech recognition

by one percent i'd be happy

well it's better to just revolutionize

the field again right yeah

i would like to do it one more time but

i'm looking forward to it i wouldn't be

surprised at all

now when when i look at your career

and some of this information actually

comes from the book as i didn't notice

before

i had read the book the first time

i mean

you are

you were a computer science professor at

the university of toronto emeritus now i

believe but computer science but you

never got a computer science

degree you got a psychology degree

and

you actually at some point were a

carpenter

how does it come about how do you go

from yes studying psychology to becoming

a carpenter to getting into ai what's

the path for you there how do you look

at that

in my last year at cambridge

i had a very difficult time and got very

unhappy and i dropped out

just after the exams i dropped out

and

became a carpenter

and

i'd always enjoyed carpentry more than

anything else

so at high school

um to be sort of all the classes and

then you could stay in the evenings and

do carpentry that's what i really look

forward to

and

so i became a carpenter and then after

i've been a carpenter for about six

months you couldn't actually make a

living as a carpenter um so i was a

carpenter and decorator and i made the

money during decorating but i had the

fun doing carpentry

um

and

the point is carpentry is more work than

it looks and decorating is less work

than it looks

um so you can you can charge more per

hour for decorating

um

unless you're a very good card

and then i met a real carpenter

and i realized i was completely hopeless

at carpentry

and so he

he's making a door

for a basement for a

coal seller under the sidewalk that was

very damp

and he was taking pieces of wood and

arranging them

so that they would walk in opposite

directions so that it would cancel out

and that was kind of a level of kind of

understanding and thought about the

process that never occurred to me he

could also take a piece of wood and just

cut it exactly square with a hand saw

and he explained something useful to me

he said if you want to cut a piece of

wood square

you have to

line the saw bench up with the room and

you have to line the piece of wood up

with the room

you can't cut it square if it's not

aligned with the room

which is very interesting in terms of

coordinate frames

so

anyway because i was so hopeless

compared with him i decided i might as

well go back into am

now when you say get back into ai

as i understand this was at the

university of edinburgh where you went

for your phd

yeah

i went to do a phd there and i went to a

phd on neural networks

with

an eminent professor called christopher

longa higgins

who was really very brilliant um

he almost got a nobel prize when he was

in his 30s to figuring out something

about the structure of boron hydride

um

and i i still don't understand what it

is because it'll do with quantum

mechanics but it hinged on the fact that

360 degree rotation is not the identity

operator it's 720 degrees

um

there's a thing you want to find one's

books about it

um anyway

he was interested in neural nets and the

relation to holograms and about the day

i arrived in edinburgh he lost interest

in neural nets

because he read winograd's thesis and he

became completely converted

um

he thought neural nets was the wrong way

to think about it we should do symbolic

ai he was very impressed by one great

thesis

um

and so we had

he had a lot of integrity

so even though he completely disagreed

with what he was i was doing he didn't

stop me doing it

he kept trying to get me to do

stuff more like winner thesis but he let

me carry on doing what i was doing

um

and

yeah i was a bit of a loner everybody

else back then in the early 70s was

saying minsky and papua shown that

neural nets are nonsense why are you

doing this stuff it's crazy

and in fact

the first talk i ever gave to that group

was about how to do true recursion with

neural networks um

so this was talking 1973 so 49 years ago

and so my one of my first projects i

discovered a write-up of it recently

was

um

you want a neural network

that will

be able to draw a shape

and

you want it to pass the shape into

parts

and you want it to be possible for a

part of the shape

to be done drawn by the same neural

hardware as the whole shape's being

drawn by

so

the neural hub where they're storing the

whole shape

has to remember where it's got to in the

whole shape and what the

orientation and position size is for the

whole shape

but now it has to go off and you want to

use the very same neurons

for drawing a part of the shape

so you need somewhere to remember what

the whole shape was and how far you got

in it

so you can pop back to that once you've

finished doing this subroutine this part

of the shape

and the question is how is the neural

network going to remember that because

obviously you can't just copy the

neurons

and so i managed to get a system working

where the neural network remembered it

by having fast heavy and weights

that were just adapting all the time

and were adapting so that

any state that it had been in recently

could be retrieved by giving it part of

that state and then say fill in the rest

and so i had a neural net that was doing

true recursion reusing the same neurons

and the same weights

to do the recursive call as it used for

the high level call

and that was in 1973 and

the

i think people didn't understand the

talk because i wasn't very good at

giving talks but they also said why

would you want to do recursion within

your match you can do recursion with

lisp

um

they didn't understand the point which

is that

unless we get neural nets to do

something like recursion we're never

going to be able to explain a whole

bunch of things

and

now that's become sort of an interesting

question again

so i'm going to wait one more year until

that idea is an antique a genuine

antique it'll be 50 years old and then

i'm gonna sort of write up the research

i did then and it was all about fast

weights for as a member so

i have many questions here jeff one the

first one is

you're standing in this room

where everybody's you're a phd student

or maybe fresh out of phd

you're standing in a room with

essentially everybody telling you what

you're working on

is a waste of time

and

you were convinced somehow was not where

do you get that conviction from

i think a large part of it was my

schooling

so

my father was a communist

but he sent me to an expensive private

school because they had good science

education

and i was there from the age of seven

[Music]

while they had a preschool

and it was a christian school

and all the other kids believed in god

and it was just

at home i was taught that that was

nonsense

and it did seem to me that

it was nonsense

um

[Music]

and so

i was used to

just having

everybody else being wrong and obviously

wrong

and i think that's important

i think you need you need

i was back to say you need the faith

which is funny in this

situation

um you need the faith in science

to um

be willing to work on stuff just because

it's obviously right

even though everybody else says it's

nonsense

and in fact it wasn't everybody else it

was everybody else in the early 70s

doing ai said it was nonsense or nearly

everybody else

um

but if you look a bit early if you look

in the 50s

both von neumann and turing believed in

neural nets

turing in particular believed in neural

nets training with reinforcement um

so

if i i still believe if they hadn't both

died early the whole history of ai might

have been very different

because they were sort of powerful

enough intellects to have swayed a field

and

they were very interested in sort of how

does the brain work

so i think it was just bad luck we both

died early

well british intelligence might have

come into it but

now

you go from believing in this

well

at the time many people didn't

getting the big breaks resulted in that

power almost everything

that's being done today and now there is

this in some sense the next question

right

is

it's not just that deep learning works

and works great the question becomes

is it all we need or

will we need other things

and you've said things

maybe i'm not literally quoting you but

to the extent of deep learning will do

everything what i really meant by that i

i i sometimes say things without

thinking without being accurate enough

and then people call me

like saying we won't need radiologists

um

so

what i really meant was um

using stochastic gradient to send to it

just a whole bunch of parameters that's

what i sort of had in mind when i said

deep learning

the way you get the gradient might not

be back propagation

and the thing you get the gradient of

might not be some final performance

measure but rather these lots of local

objective functions

but i think that's how the brain works

and i think that's gonna explain

everything yes well nice nice to see it

confirmed um

so one other thing i want to say is the

kind of computers we have now

um

are very good for

um doing banking

because they can remember how much you

have in your account it wouldn't be so

good if you went in and they said well

you got roughly this much we're not

really sure because we don't do it to

that precision but roughly this much

um we don't want that in a computer

doing banking

um

or in a computer guiding the space

shuttle or something we would really

rather it got the answer exactly right

um

and they're very different from us

and i think

people aren't sufficiently aware

that we made a decision

about how computing would be

um which is that

um

our computer our

knowledge will be immortal

so if you look at

existing computers you have a computer

program

or maybe you just have a lot of weights

for a neural net that's a different kind

of program

um but if your hardware dies you can run

the same program on another piece of

hardware

and so that makes the knowledge immortal

it doesn't hinge on that particular

piece of hardware surviving now the cost

of the immortality is huge

because it means the two-bit different

bits of hardware have to do exactly the

same thing obviously zero correction all

that but after you've done all the error

correction they have to do exactly the

same thing

which means there better be digital or

mostly digital

um

[Music]

and they're probably gonna do things

like multiplying numbers together which

involves

using lots and lots of energy to make

things very discreet

which is not what hardware really wants

to be

and so as soon as you

commit yourself to the immortality of

your program or your neural net

you're committed to

um very

expensive computations and also to very

expensive manufacturing processes you

need to manufacture these things

accurately and probably in 2d and then

put lots of 2d things together

um if you're just willing to give up on

immortality

sort of in fiction normally what you get

in return is love

um

but if if we're willing to give up

immortality what we'll get in return is

very low energy computation and very

cheap manufacturing

so

instead of manufacturing computers

what we should do is grow them

um we should use nanotechnology to just

grow the things in 3d

and

each one will be slightly different

so the image i have is if you take a pot

plant

and you sort of pull it out of its pot

there's a root ball and it's the shape

of the pot right

and so all the different pot plants have

the same shaped root ball but the

details of the roots are all different

but they're all doing the same thing

they're extracting nutrients from the

soil and they got the same function and

they're pretty much the same

but the details are all very different

um

so that's what real brains are like

and i think that's what

what i call mortal computers will be

like

so these are computers that are grown

rather than manufactured

you can't program them they just learn

they obviously have to have a learning

algorithm sort of built into them

they learn they can do most of their

computation in analog

because analog is very good for doing

things like taking a voltage times a

resistance and turning it into a charge

and then adding up the charge and

already chips that do things like that

the problem is what do you do next

um and how do you learn in those chips

and at present people have suggested

back propagation or various versions of

boxing machines

um

i think we're going to need something

else

but i think

sometime in the not too distant future

we're going to see mortal computers

which are very cheap to create

have to get all their knowledge in there

by learning

and are very low energy

and

these mortal computers when they die

they die and their knowledge dies with

them

and so and it's no use looking at the

weights because those weights only work

for that hardware

um

so what you have to do is distill the

knowledge into other computers so when

these multiple computers get old

they're gonna have to do lots of

podcasts to try and get the knowledge

into younger modern english first one

you build i'll happily have that one on

let me know

so jeff this reminds me of another uh

question that's been on my mind

for you which is

when you think about today's neural nets

the ones that

grab the headlines

are very very large i mean not as large

as as the brain maybe but

in some sense starting to get in that

way

size right the large language models

um

but and and the results look very

very impressive

um

so one i'm i'm curious

about your take on those kinds of models

and what you see in them and we see as

limitations but two i'm also curious

about

what do you think about working on the

other end of the spectrum for example

ants have much smaller brains

obviously than humans

yet

it's fair to say that our

visual motor systems that we have

developed artificially are not yet at

the level of what ants can can pull off

or b's and so forth and so i'm curious

about that spectrum as well as the the

recent big advances in language models

where you think about those

so b's they may look small to you but i

think a b has about a million neurons

so

i think a b is like closer to gpg3 um

certainly closer than romantics um but a

b is actually quite a big neural net

um

my belief is that

um

if you take a system with lots of

parameters and they're tuned sensibly

using some kind of gradient descent in

some kind of sensible objective function

then you'll get wonderful properties out

of it and you'll get all these emerging

properties

um

like you do with gpg3 and

also the

the the google equivalents that i've

talked about so much

that doesn't sort of

settle the issue of whether they're

doing the same way as us

and i think

um

we're doing

a lot more things like recursion which i

think we do in neural nets

and i tried to address some of these

issues in a paper i put on the web last

year called glom

um well i call it glom it's how you do

part hole hierarchies in neural nets so

you definitely have to have structure

and if what you mean by symbolic

computation is just that you have part

whole structure

then we do symbolic computation that's

not normally what people meant by

symbolic computation

the sort of hardline symbolic

computation means

you're using

symbols and you're operating on symbols

using rules that just depend on the form

of the symbol string you're processing

and that a symbol

the only property a symbol has is that

it's either identical or not identical

to some other symbol

and perhaps that it points to something

it can be used as a pointer to get

something

um the neural nets are very different

from that

so

the sort of hard-line symbol processing

i don't think we do that but we

certainly

deal with

pothole hierarchies

but i think we do it in great big neural

nets

and i'm sort of up in the air at present

as to

to what extent does gpt3 really

understand what it's saying

i think it's fairly clear

it's not just like the old eliza program

which just rearranges strings of symbols

and had no clue what it was talking

about

um and the reason for believing that is

you say you say in english show me a

picture of a hamster wearing a red hat

and it draws a picture of a hamster

wearing a red hat

um

and you're fairly sure it never got that

pair before

so it has to understand the relationship

between the english string and the

picture

and

before it had done that if you'd asked

any of these um doubters

these neural net skeptics

um neural net deniers let's call them

neural net deniers um

if you'd ask them well how would you

show that it understands

i think they'd have accepted that well

if you asked to draw a picture something

draws a picture of that thing then it

understood

just as with winograd's thesis you ask

it to put the blue

the blue block in the green box and it

puts the blue block in the green box and

so that's pretty good evidence it

understood what you said

um

but now that it does it of course the

skeptics then say well you know that

doesn't really count

there's nothing that was satisfied

basically yeah

the goal line's always moving uh

for true skeptics

yeah

now there's the recent one um

the google won the paul model that uh in

in the paper showed how it was

explaining

effectively how jokes work that was

extraordinary that just seemed a very

deep understanding of language no it was

just rearranging the words it had in its

training you think so no

no it had i i didn't see how it could

generate those explanations without sort

of understanding what's going on now i'm

still open to the idea that

because it was framed with back

propagation it's going to end up with a

very different sort of understanding

from us

and obviously

adversarial images um

tell you a lot that you can recognize

objects by using their textures

and you can be correct about it in the

sense that it'll generalize to other

instances of those objects

but it's a completely different way of

doing it from what we do

and i like to think of the example of

insects and flowers so insects seeing

the ultraviolet

so two flowers that look the same to us

can look completely different to insects

and now

because the flowers look the same to us

do we say the insects are getting it

wrong

um because these flowers evolved with

the insects to give signals to the

insects in the ultraviolet to tell them

which flower it is so it's clear the

insects are getting it right and we just

can't see the difference

and that's another way of thinking about

adversarial examples

um

it looks you know this this thing that

it says is an ostrich looks like a looks

like a school bus to us

but actually if you look in the texture

domain then it's actually an ostrich so

um

the question is who's right in the case

of the insects

um just because two flowers look

identical to us it doesn't mean they're

really the same the insects are right

about them being very different in that

case it's different parts of the

electromagnetic spectrum that are

indicating the difference that we can

pick up on but it could be in the case

of image recognition for our current

neural nets so you could argue maybe

that

um

since we build them and we want them to

do things for us in our world then

we really

don't want to

just say okay they got it right and we

got it wrong i mean they need to

recognize the car and the pedestrian

yeah i agree i just want to show it's

not as simple as you might think about

who's right and who's wrong

and

part of the point of my glom paper was

to try and build

perceptual systems that work more like

us

so they're much more likely to make the

same kinds of mistakes as us

and not make

very different kinds of mistakes and

obviously

um

if you've got a self-driving car for

example if it makes a mistake that any

normal human driver would have made

that seems much more acceptable than

making a really dumb mistake

so jeff as i understand it sleep is

something you also think about can you

say a bit more

yes i often think about it when i'm not

sleeping at night

um so

there's something funny about sleep

which is um

animals do it fruit flies sleep

and it may just be to stop them flying

around in the dark but um

if you deprive people of sleep

then they go really weird

like if you surprise someone for three

days they'll start hallucinating if you

surprise someone for a week they'll go

psychotic and they never recover

um

these are nice experiments done by the

cia i think um

and

the question is why

why do we what what is the computational

function of sleep there's presumably

some pretty important question for it if

depriving you of it makes you just

completely fall apart

and so current theories are things like

it's for consolidating memories or maybe

for downloading things from hippocampus

into cortex which is a bit odd since i

had to come through court exactly on the

campus in the first place

um

so a long time ago in the early 80s

terry sanofsky and i had this theory

called baltimore machines

and it was partly based on an insight of

francis crick

um when he was thinking about hopfield

nets princess quickly rare mitchelson

had a paper about sleep and the idea

that

um

you would hit the net with random things

and tell it not to be happy with random

things so in a hot field next you give

it something you wanted to memorize and

it changes the weights so the energy of

that vector is lower

and the idea is if you also give it

random vectors and say make the energy

higher the whole thing works better

and that led to boltzmann machines where

we figured out that um

if you instead of giving it random

things you get things generated from a

markov chain the model's own markov

chain and you say make those less likely

and make the data more likely that is

actually a maximum likelihood learning

and so we got very excited about that

because we thought okay that's what

sleep is for sleep is this negative

phase of learning

it comes up again now in contrasting

learning

where you have

two patches from the same image you try

and get get them to have similar

representations

and two patches from different images

you try and get them to have

representations that are sufficiently

different once they're different you

don't make them any more different

but you stop them being too similar and

that's how contrastive learning works

now with boltzmann machines

you couldn't actually separate the

positive face from the negative face you

had to interleave positive examples and

negative examples

otherwise the whole thing would go wrong

and i went to i tried a lot not

interleaving them it's quite hard to do

a lot of positive examples followed by a

lot of negative examples

what i discovered a couple of years ago

they got me very excited and caused me

to agree to give lots of talks that i

then cancelled when i couldn't make it

work better um

was that

with contrastive learning

you can actually separate the positive

and negative phases

so you can do

lots of examples of positive pairs

followed by lots of examples in negative

pairs

and that's great because what that means

is

you can have something like a video

pipeline

where you're just trying to make things

similar while you're awake

and trying to make things dissimilar

while you're asleep

um

if you can figure out how sleep can

generate video for you um

so it makes

it makes a contrastive learning area

much more plausible if you can separate

the positive and negative phases and do

them at different times and do a whole

bunch of positive updates followed by a

whole bunch of negative updates

and

even for the standard contrastive

learning you can do that moderately well

you have to use lots of momentum and

stuff like that there's all sorts of

little tricks to make it work but you

can make it work

um

so i think it's quite likely that the

function of sleep

is to do unlearning or negative examples

and that's why you don't remember your

dreams you don't want to remember them

you're unlearning them

qrik pointed this out you'll remember

the ones that are in the fast weights

when you wake up

um

because the fast weights are a temporary

store so

that's not unrunning that still works

the same way

but the long-term memory um

the whole point is to get rid of those

things and that's why you dream for many

hours a night but when you wake up you

can just remember the last minute of the

dream you're having when you woke up

um

and i think this is a much more

plausible theory of sleep than any other

i've seen because it explains why if you

got rid of it the whole system would

just fall apart

you'll go disastrously wrong and start

hallucinating and doing all sorts of

weird things

and let me say a little bit more about

the need for negative examples that

you're having a trust of learning if

you've got a neural net

and it's trying to optimize some

internal objective function something

about the kinds of representations it

has or something about the agreement

between contextual predictions and local

predictions

it wants this agreement to be

a property of the real data

and the problem inside a neural net is

that

you might get all sorts of correlations

in your inputs i'm a neuron right so i

get all sorts of correlations in my

inputs and those correlations have

nothing to do with the real data they're

caused by the wiring of the network and

the way it's in the network

if these two neurons are both looking at

the same pixel they'll have a

correlation but that doesn't tell you

anything about the data

and so the question is how do you learn

to extract structure that's about

the real the real data

and not about the wiring of your network

and the way to do that is to feed it

positive examples and say find structure

in the positive examples

that isn't in the negative examples

because the negative examples are going

to go through exactly the same wiring

and if the structure is not in the

negative examples but it is in the

positive examples

then the structure is about the

difference between the positive and

negative examples not about your wiring

so as people don't think about this much

but

if you have

powerful learning algorithms they you

better not make them learn about the

neural network's own weights and wine

that's not what's interesting

now when you think about people who

don't get sleep then and start

hallucinating is hallucinating just

effectively trying to do the same thing

you're just doing it while you're awake

obviously you can have little naps and

that's very helpful and maybe

hallucinating when you're awake is

serving the same function in sleep

and it's i mean all the experiments i've

done say it's better to not have

16 hours awake and eight hours of sleep

it's better to have a few hours awake in

a few hours of sleep so

and a lot of people have discovered that

little naps help einstein used to

take little naps all the time and he did

okay

yeah he did very well no

for sure

now

there's this other thing

you uh you've brought up this notion of

student beats teacher

what does that refer to

okay so

um a long time ago ago i didn't

experiment on mnist

which is a standard digit database for

every 900 digits

where

um

you take the data the training data

and you corrupt it and you corrupt it by

substituting

the wrong label one of the other nine

labels

eighty percent of the time

so now you've got a data set in which

the labels are correct

um 20 of the time

and wrong 80 of the time

and the question is

um can you learn from that

and how well do you learn from that

and the answer is you can learn to get

like 95 correct on that

so now you've got a teacher who's wrong

80 of the time

then the student

is right 95 of the time

so the student is much much better than

the teacher

and this isn't

um each time you get an example you

corrupted you take the training examples

you can wrap them once and for all so

you can't average away the corruption

over different

you might be able to average it away

over different training cases that

happen to have similar images but

and if you ask well how many training

cases do you need if you have corrupted

ones

and this was a great interest because of

the tiny images data set some time ago

where they had 80 million tiny images

with a lot of wrong labels in

and the question is

would you rather have a million

things that are flakily labeled or would

you rather have 10 000 things with

accurate labels

and i had a hypothesis that what counts

is the amount of mutual information

between the label and the truth

so if the labels are correct corrupted

ninety percent of the time there's no

mutual information between the labels

and the truth

if they corrupt eighty percent of the

time there's only a small amount of

mutual information is that's what you

think is i think it's about my memory is

it's 0.06 bits per case

whereas if it's uncorrected it's about

3.3 bits per case

um so it's only a tiny amount and then

the question is well

suppose i balance the size of the

training set by putting as much mutual

information in there

um

so if if there's like a 50th of the

mutual information i have 50 times as

many examples do i now get the same

performance

and the answer is yes you do to within a

factor of two i mean the training set

actually needs to be twice that big but

roughly speaking

you can see how useful a training

example is by the amount of mutual

information between the label and the

truth

and i noticed recently you have

something for doing sim to real

where you're labeling real data using a

neural net and those labels aren't

perfect

and then you take the student that

learned from those labels and the

student is better than the teacher it

learned from

and people are always puzzled by how

could the student be better than the

teacher

um

but in neural nets it's very easy

the student will be better than the

teacher

if there's enough training data

even if the teachers are very flaky and

i have a paper

a few years ago with melody guan about

this for some medical data

the first part of paper talks about this

but the the rule of thumb is basically

what counts is the mutual information

between

the assigned label and the truth

and that tells you how valuable a

training example is

and so you can make do with lots of

flaky ones

that's so interesting now in the work we

did that you just referenced javen and

the work i've seen

quite popular recently

usually the teacher

provides

noisy labels

but then

not all the noise labels are used

there's a notion that only look at the

ones where the teacher is more

confident

your description doesn't matter that's

obviously a good hatch

yeah you don't need to do that you don't

need to do that it's a good hack

and it probably helps to only look at

the ones where you have reason to

believe the teacher got it right but

it'll work even if you just look at them

all and there's a phase transition

so with with mnist

melody plotted a graph and as soon as

you get like 20 of the labels right

your student will get like 95 correct

wow

but as you get down to about 15 right

you suddenly get a phase transition

where you don't do any better than

chance

because somehow the student has to get

it the teacher is saying these labels

and the student has to in some sense

understand

which cases are right and which case is

wrong and sort of see the relationship

between the labels and the inputs

and then once the student's seen that

relationship a wrongly labeled thing is

just very obviously wrong

um so it's fine if it's randomly wrongly

enabled

but there is a phase transition where

you have to have it good enough so the

students sort of get the idea

but that explains how our students are

all smarter than us

we all need to get it right a small

fraction of the time

right and i'm sure the students do some

of this data curation where you say

something and the student thinks oh

that's rubbish i'm not going to listen

to that

those are the very best students you

know

yeah those are the ones that can

surprise us um

now one of the things that

is really important in

neural net learning and especially when

you're building models is to get an

understanding of what is it what is it

learning and often people

try to somehow visualize what's

happening during learning and one of the

most prevalent visualization techniques

is called disney

which is something you invented jeff so

i'm curious how do you come up with that

what if maybe first describe what it

does and then what's the story behind it

so if you have some high dimensional

data

and you

try and draw a 2d or a 3d map of it

you could take the first two principal

components and just plot the first two

principal components

but what principal components cares

about is getting the big distances right

so if two things are very different

principal components is very concerned

to get them very different

in the 2d space

it doesn't care at all about the small

differences because it's it's sort of

operating on the squares of the big

differences

so it won't preserve similarity very

well

high dimensional similarity

and

you're often interested in just the

opposite you've got some data you're

interested in what's very similar to

what when you don't care if it gets the

big distances a bit wrong as long as it

gets the small distances right

so i had the idea

a long time ago that

what if we took the

distances and we turned them into

probabilities of pairs

there's various versions of tc's but

suppose we turned them into the

probability of a pair

such that we say

pairs with a small distance are probable

and players with a big distance are

improbable

so we're converting distances into

probabilities in such a way that small

distances correspond to big

probabilities

and we do that by putting a gaussian

around a point a data point

and computing the density of the other

data point under this gaussian

and that's an unnormalized probability

then you normalize these things

and then you try and lay the points out

in 2d so as to preserve those

probabilities

and so it won't care much if two points

are far apart they'll have a very low

pairwise probability and it doesn't care

the relative positions of those two

points but it cares about the relative

positions of the ones with

hyperventilators

and that produced quite nice maps and

that was called stochastic neighbour

embedding

because we thought of this you put a

gaussian and you stochastically pick a

neighbor according to the density under

the gaussian

um and i did that work with samurais and

it had very nice simple derivatives

which convinced me that we were onto

something and we got nice maps but they

tended to crowd things together

and

there's obviously a basic problem

in converting high dimensional data into

low dimensional data

so sneak tends to crowd things together

stochastic in everybody and that's

because of the nature of high

dimensional space and low dimension

spaces

in a high dimensional space a data point

can be close to lots of other points

without them all being too close to each

other

in a low dimensional space they all have

to be close to each other if they're all

close to this data point

so

you've got a problem in embedding

closenesses from high dimensions to low

dimensions

and

i had the idea when i was doing snee

that since i was using probabilities as

this kind of intermediate currency

there should be a mixture model it

should be a mixture version where you

say in high dimensions the probability

of a pair is proportional to

e to the minus s squared distance

on my gaussian um

and in low dimensions suppose you have

two different maps

the probability of a pair is the sum of

e to the minus the distance in the first

2d map

and e to the minus the squared distance

in the second 2d studio

and that way if we have a word like bank

and we're trying to put similar words

near one another bank can be close to

greed in one map

and can be close to river in the other

map

without river ever being close to greed

so

i really pushed that idea because i

thought this was a really neat idea and

you could have a mixture of maps

and we managed to get to where elio was

one of the first people to work on that

and james cook worked on it a lot

and several other students worked on it

and we never really got it to work well

um

and i was very disappointed that someone

hadn't been able to make use of the

mixture idea

and then i went to a simpler version

which i called unicene

which was a mixture of a gaussian and a

uniform

and that worked much better

um

so the idea is in one map

all pairs

are equally probable

and that gives you a sort of background

probability which goes through the big

distances

a small background probability

and then in the other map

you contribute

um

a probability proportional to your

squared distance in this other map

but it means in this other map

things can be very far apart if they

want to be

because the fact that then

they need some probability is taken care

of by the uniform

and

then i got a review paper from a plumber

called lawrence van der marton

which i thought was actually a published

paper because of the form it arrived in

but wasn't actually a published paper

and he wanted to come do research with

me and i thought he had this published

paper so i invited him to come do

research um it turned out he was

extremely good and it's lucky i've been

mistaken in thinking it was a published

paper um

and we started on unisoning

and then i realized that actually eunice

me

is a special case of using a mixture of

a gaussian

and a very very broad gaussian which is

a uniform

so what if we used a whole hierarchy of

gaussians

many many gaussians with different

widths

and that's called a t distribution um

and that led to t-sne and t c works much

better

and tc has a very nice property

that

um it can show you things at multiple

scales

because it's got a kind of one over d

squared property

that

um

once distances get big

it behaves just like gravity and

clusters of galaxies and things you have

pluses of galaxies and galaxies and

clusters of stars and so on

and you get structured many different

levels in it you get the course

structure and the fine structure all

showing up

now the

objective function used for all this

which was the sort of relative densities

under a gaussian

came from other work i did with alberto

pacinero earlier

um

that we found hard to get published

i got a review saying

yeah i got a review of that work when it

was rejected by some conference

saying hinton's been working on this

idea for seven years and nobody's

interested

um

i take those reviews as telling me i'm

on to something very original

um

and that actually had the function in it

that's now used i think it's called nce

it's using these contrastive methods

um and t-sne is actually a version of

that function

um

but it's being used for making maps

so it's a very long history of tc of

getting the original sne and then trying

to make a mixture version and it's just

not working and not working not working

and then eventually getting the

coincidence of figuring out it was a

t-distribution of what you wanted to use

that was the kind of mixture

and

lauren's arriving and laurence was very

smart in a very good program really made

it all work beautifully

this is really interesting because it

seems a lot of the um

a lot of the progress these days the the

bigger idea place plays a big role but

here it seems it was really getting the

details right

was the only way to get it to fully work

you typically need both

you have to have a big idea for it to be

interesting original stuff but you also

have to get the details right

and that's what graduate students are

for

so

jeff thank you thank you for such a

wonderful uh conversation for our part

one of our season finale

you