How to Get Inside the "Brain" of AI | Alona Fyshe | TED

翻译人员: Weihao Sheng 校对人员: Yip Yan Yeung

People are funny.

人是非常有趣的。

We're constantly trying to understand and interpret

我们一直在试图理解和解释

the world around us.

我们周围的世界。

I live in a house with two black cats, and let me tell you,

我家里有两只黑猫， 我可以告诉你，

every time I see a black, bunched up sweater out of the corner of my eye,

每当我的余光瞥见 一件打结的黑色毛衣，

I think it's a cat.

我都会认为那是一只猫。

It's not just the things we see.

不只是我们看到的东西。

Sometimes we attribute more intelligence than might actually be there.

我们有时以为一些东西有超常的智慧， 但实际上未必有。

Maybe you've seen the dogs on TikTok.

比如，你也许在TikTok上 看到过狗狗的视频。

They have these little buttons that say things like "walk" or "treat."

上面有一些小按钮， 写着“遛遛”或是“要吃的”。

They can push them to communicate some things with their owners,

这些狗狗能用这些按钮 和它们的主人交流，

and their owners think they use them

它们的主人也以为用这些按钮就能

to communicate some pretty impressive things.

让狗狗做一些令人惊奇的事情。

But do the dogs know what they're saying?

但狗狗知道它们在说什么吗？

Or perhaps you've heard the story of Clever Hans the horse,

或许你听过《聪明的汉斯》的故事，

and he could do math.

这匹马居然能做数学题。

And not just like, simple math problems, really complicated ones,

不仅仅是简单的数学计算， 而是非常复杂的问题，

like, if the eighth day of the month falls on a Tuesday,

比如，如果一个月的第八天是星期二，

what's the date of the following Friday?

那么下一个星期五的日期是什么？

It's like, pretty impressive for a horse.

对于一匹马来说，这真是令人惊叹。

Unfortunately, Hans wasn't doing math,

不幸的是，汉斯并不是在做数学题，

but what he was doing was equally impressive.

但它学会的事情也很了不起。

Hans had learned to watch the people in the room

汉斯学会了观察房间里的人，

to tell when he should tap his hoof.

来判断它该在什么时候敲蹄子。

So he communicated his answers by tapping his hoof.

它通过敲蹄子来“说出”它的答案。

It turns out that if you know the answer

真实的情况是，如果你知道答案，

to "if the eighth day of the month falls on a Tuesday,

就是“如果每个月的第八天是星期二，

what's the date of the following Friday,"

那么下一个星期五是什么日子”的答案，

you will subconsciously change your posture

你会在汉斯正确地敲打 18 下的时候

once the horse has given the correct 18 taps.

下意识地改变你的姿势。

So Hans couldn't do math,

所以汉斯不会做数学，

but he had learned to watch the people in the room who could do math,

但它学会了观察房间里会做数学的人，

which, I mean, still pretty impressive for a horse.

这对一匹马来说还是很了不起的。

But this is an old picture,

但这只是一个古老的故事了，

and we would not fall for Clever Hans today.

今天的我们已经不会 再被聪明的汉斯骗到了。

Or would we?

会吗？

Well, I work in AI,

我在人工智能领域工作，

and let me tell you, things are wild.

我可以告诉你，事情很疯狂。

There have been multiple examples of people being completely convinced

有许多例子表明，人们完全相信

that AI understands them.

人工智能能理解他们。

In 2022,

2022 年，

a Google engineer thought that Google’s AI was sentient.

一位谷歌工程师认为 谷歌的人工智能有自我意识。

And you may have had a really human-like conversation

你也可能试过与ChatGPT

with something like ChatGPT.

进行类似人类的对话。

But models we're training today are so much better

我们今天训练的模型比我们

than the models we had even five years ago.

仅五年前的模型都要好得多。

It really is remarkable.

这真的很了不起。

So at this super crazy moment in time,

所以在这个疯狂的时刻，

let’s ask the super crazy question:

让我们问一个疯狂的问题：

Does AI understand us,

人工智能是真的理解我们，

or are we having our own Clever Hans moment?

还是我们又遇到了一匹聪明的汉斯？

Some philosophers think that computers will never understand language.

一些哲学家认为， 计算机永远不会理解语言。

To illustrate this, they developed something they call

为了说明这一点，他们设计了一段

the Chinese room argument.

被称为“中文房间”的论证。

In the Chinese room, there is a person, hypothetical person,

在中文房间里， 有一个人，虚构的人，

who does not understand Chinese,

他/她不懂中文，

but he has along with him a set of instructions

但他/她眼前有一套指令，

that tell him how to respond in Chinese to any Chinese sentence.

告诉他/她如何用中文 回应任何中文语句。

Here's how the Chinese room works.

中文房间中会进行以下流程。

A piece of paper comes in through a slot in the door,

门上的缝隙里递来了一张纸条，

has something written in Chinese on it.

上面写着一些中文。

The person uses their instructions to figure out how to respond.

这个人要用面前的指示想出回应的内容。

They write the response down on a piece of paper

他/她把回答写在一张纸上，

and then send it back out through the door.

再通过门上的缝隙传出去。

To somebody who speaks Chinese,

对于说中文的人来说，

standing outside this room,

站在这个房间外面，

it might seem like the person inside the room speaks Chinese.

可能会觉得房间里的人会说中文。

But we know they do not,

但我们知道他/她不会，

because no knowledge of Chinese is required to follow the instructions.

因为遵循指示不需要学会中文。

Performance on this task does not show that you know Chinese.

这项任务的表现并不能说明你懂中文。

So what does that tell us about AI?

这个论证和AI有什么关系？

Well, when you and I stand outside of the room,

当你我站在房间外面，

when we speak to one of these AIs like ChatGPT,

与像ChatGPT这样的AI说话时，

we are the person standing outside the room.

我们就是站在房间外面的人。

We're feeding in English sentences,

我们输入的是英语句子，

we're getting English sentences back.

得到的是英语句子的反馈。

It really looks like the models understand us.

看起来这些模型真的能理解我们。

It really looks like they know English.

看起来它们真的像懂英语。

But under the hood,

但在系统的底层，

these models are just following a set of instructions, albeit complex.

这些模型只是遵循一套指令， 尽管是很复杂的指令。

How do we know if AI understands us?

我们如何知道AI能否理解我们？

To answer that question, let's go back to the Chinese room again.

为了回答这个问题， 让我们再回到“中文房间”的例子。

Let's say we have two Chinese rooms.

假设我们有两个中文房间。

In one Chinese room is somebody who actually speaks Chinese,

其中一个房间里 是真正会说中文的人，

and in the other room is our impostor.

而另一个房间里的是个冒牌货。

When the person who actually speaks Chinese gets a piece of paper

当真正讲中文的人拿到一张

that says something in Chinese in it, they can read it, no problem.

写有中文的纸时， 他/她当然可以读懂。

But when our imposter gets it again,

但是，当这位冒牌货拿到纸条时，

he has to use his set of instructions to figure out how to respond.

他/她必须使用那套指令来作出回应。

From the outside, it might be impossible to distinguish these two rooms,

从外面看，可能无法区分这两个房间，

but we know inside something really different is happening.

但我们知道房间里面 发生着一些根本不同的事情。

To illustrate that,

为了说明这一点，

let's say inside the minds of our two people,

假设在两个房间里的两个人，

inside of our two rooms,

两个人的头脑中

is a little scratch pad.

各有一个小草稿本。

And everything they have to remember in order to do this task

完成这项任务需要的所有记忆

has to be written on that little scratch pad.

都必须记在小草稿本上。

If we could see what was written on that scratch pad,

如果我们能看到写在草稿本上的东西，

we would be able to tell how different their approach to the task is.

我们就能知道他们 完成任务的方法有什么不同。

So though the input and the output of these two rooms

因此，尽管这两个房间的输入和输出

might be exactly the same,

可能完全相同，

the process of getting from input to output -- completely different.

但从输入转化为输出的过程完全不同。

So again, what does that tell us about AI?

这跟人工智能又有什么关系呢？

Again, if AI, even if it generates completely plausible dialogue,

即使人工智能产生了完全合理的对话，

answers questions just like we would expect,

像我们期望的那样回答问题，

it may still be an imposter of sorts.

它也仍然可能是某种程度上的冒牌货。

If we want to know if AI understands language like we do,

如果我们想知道人工智能 能否像我们一样理解语言，

we need to know what it's doing.

我们需要知道它在做什么。

We need to get inside to see what it's doing.

我们需要深入内部，看看它在做什么。

Is it an imposter or not?

它到底是不是一个冒牌货？

We need to see its scratch pad,

我们需要看到它的草稿本，

and we need to be able to compare it

并将其与

to the scratch pad of somebody who actually understands language.

真正理解语言的人类的草稿本进行比较。

But like scratch pads in brains,

但是，大脑中的“草稿本”

that's not something we can actually see, right?

不是我们能随便看到的东西，对吧？

Well, it turns out that we can kind of see scratch pads in brains.

实际上，我们可以在一定程度上 “看到”大脑中的草稿。

Using something like fMRI or EEG,

用fMRI或EEG这样的技术，

we can take what are like little snapshots of the brain while it’s reading.

我们可以在人阅读时拍下大脑的快照。

So have people read words or stories and then take pictures of their brain.

在人们在读单词或读故事的时候， 拍摄他们大脑的状态。

And those brain images are like fuzzy,

这些脑成像的图片就像是

out-of-focus pictures of the scratch pad of the brain.

模糊、失焦的草稿本照片，

They tell us a little bit about how the brain is processing

它们能告诉我们一些信息，

and representing information while you read.

阅读时的大脑是如何处理、表现信息的。

So here are three brain images taken while a person read the word "apartment,"

这里有三张大脑图像，对应一个人 读到三个词时的情况：“公寓”、

"house" and "celery."

“房子”和“芹菜”。

You can see just with your naked eye

你一眼就能看出，

that the brain image for "apartment" and "house"

“公寓”和“房子”的脑图像

are more similar to each other

比“芹菜”的脑图像

than they are to the brain image for "celery."

更为相似。

And you know, of course that apartments and houses are more similar

当然，公寓和房子就词意来说 本来就更相似，

than they are to celery, just the words.

相比于芹菜来说。

So said another way,

换一种说法，

the brain uses its scratchpad when reading the words "apartment" and "house"

大脑在读到“公寓”和“房子”两个词时 在草稿本上记录下的内容

in a way that's more similar than when you read the word "celery."

比读到“芹菜”时的草稿本更相似。

The scratch pad tells us a little bit

这些草稿本向我们透露了一些

about how the brain represents the language.

大脑表示语言的方法。

It's not a perfect picture of what the brain's doing,

这种方法并不能完美展现 大脑中发生的情况，

but it's good enough.

但是足够用了。

OK, so we have scratch pads for the brain.

现在我们有了大脑的草稿纸，

Now we need a scratch pad for AI.

我们需要拿到AI的草稿纸。

So inside a lot of AIs is a neural network.

许多AI的内部是一个神经网络，

And inside of a neural network is a bunch of these little neurons.

而神经网络是由 一个个小的神经元组成的。

So here the neurons are like these little gray circles.

神经元就像这些灰色的小圆圈。

And we would like to know

我们想要知道

what is the scratch pad of a neural network?

神经网络的草稿纸长什么样？

Well, when we feed in a word into a neural network,

当我们给神经网络输入一个词时，

each of the little neurons computes a number.

每个小神经元都会计算一个数字。

Those little numbers I'm representing here with colors.

我用颜色来表示这些数字。

So every neuron computes this little number,

每个神经元会计算一个数字，

and those numbers tell us something

这些数字给我们展现了

about how the neural network is processing language.

神经网络是如何处理语言的。

Taken together,

总结起来，

all of those little circles paint us a picture

所有这些小圆圈为我们描述了

of how the neural network is representing language,

神经网络表示语言的方法，

and they give us the scratch pad of the neural network.

展现了神经网络的草稿本。

OK, great.

太好了。

Now we have two scratch pads, one from the brain and one from AI.

现在我们有了两张草稿纸， 一张来自大脑，一张来自人工智能。

And we want to know: Is AI doing something like what the brain is doing?

我们想知道的是： AI做的事情是否和大脑类似？

How can we test that?

我们要怎么判断呢？

Here's what researchers have come up with.

这是研究人员想出的办法。

We're going to train a new model.

我们要训练一个新的模型。

That new model is going to look at neural network scratch pad

新的模型将检查神经网络

for a particular word

对某个单词的“草稿”，

and try to predict the brain scratch pad for the same word.

并试图预测同一个单词的大脑草稿。

We can do it, by the way, around two.

顺便一提，这个过程也可以反过来。

So let's train a new model.

我们来训练一个新的模型，

It’s going to look at the neural network scratch pad for a particular word

它会检查神经网络 对特定单词的草稿，

and try to predict the brain scratchpad.

并预测大脑的草稿。

If the brain and AI are doing nothing alike,

如果大脑和AI所做的事情 没有任何相似之处，

have nothing in common,

没有任何共同之处，

we won't be able to do this prediction task.

这项预测任务将无法完成。

It won't be possible to predict one from the other.

两者中的任何一个都无法预测另一个。

So we've reached a fork in the road

现在我们到了一个岔路口，

and you can probably tell I'm about to tell you one of two things.

答案只会是以下两者之一：

I’m going to tell you AI is amazing,

要么AI是非常惊人的；

or I'm going to tell you AI is an imposter.

要么AI只是一个冒牌货。

Researchers like me love to remind you

像我这样的研究人员 特别喜欢说，

that AI is nothing like the brain.

人工智能与大脑完全不同。

And that is true.

这是事实。

But could it also be the AI and the brain share something in common?

但AI和大脑有没有相似点呢？

So we’ve done this scratch pad prediction task,

我们进行了这项预测草稿的任务，

and it turns out, 75 percent of the time

结果发现，有 75% 的概率

the predicted neural network scratchpad for a particular word

针对某一特定词语的 神经网络草稿的预测结果

is more similar to the true neural network scratchpad for that word

会更类似针对这一词语的 真实大脑神经网络草稿，

than it is to the neural network scratch pad

而不是更接近于 针对其他随机词语的

for some other randomly chosen word --

大脑神经网络草稿。

75 percent is much better than chance.

75% 要远高于随机水平。

What about for more complicated things,

那么对于更复杂的事物，

not just words, but sentences, even stories?

不只是单词， 还有句子，甚至故事呢？

Again, this scratch pad prediction task works.

这个草稿预测任务得到了同样的结果。

We’re able to predict the neural network scratch pad from the brain and vice versa.

我们可以从大脑图像预测神经网络， 反过来也可以。

Amazing.

太有意思了。

So does that mean

那么，这是否意味着

that neural networks and AI understand language just like we do?

神经网络和人工智能可以 像我们人类一样理解语言呢？

Well, truthfully, no.

说实话，并不是。

Though these scratch pad prediction tasks show above-chance accuracy,

尽管这些草稿预测任务 表现出高于随机的准确率，

the underlying correlations are still pretty weak.

两者底层的相关性仍然非常弱。

And though neural networks are inspired by the brain,

尽管神经网络的灵感来自于大脑，

they don't have the same kind of structure and complexity

它们并不具备大脑呈现的

that we see in the brain.

结构和复杂性。

Neural networks also don't exist in the world.

神经网络也不存在于真实世界中。

A neural network has never opened a door

从来没有一个神经网络打开过门，

or seen a sunset, heard a baby cry.

看到过日落，听到过婴儿的哭声。

Can a neural network that doesn't actually exist in the world,

一个并不真实存在于世界上、

hasn't really experienced the world,

没有真正体验过世界的神经网络，

really understand language about the world?

能真正理解描述世界的语言吗？

Still, these scratch pad prediction experiments have held up --

尽管如此，这些草稿预测实验 仍然站得住脚——

multiple brain imaging experiments,

多个大脑成像结果，

multiple neural networks.

多个神经网络模型。

We've also found that as the neural networks get more accurate,

我们还发现， 随着神经网络变得更加准确，

they also start to use their scratch pad in a way

它们也以一种更像大脑的方式

that becomes more brain-like.

使用着草稿纸。

And it's not just language.

这不仅仅是语言方面。

We've seen similar results in navigation and vision.

我们在导航任务和视觉任务上 也看到了相似的结果。

So AI is not doing exactly what the brain is doing,

人工智能所做的并不完全和大脑相同，

but it's not completely random either.

但也不是完全随机。

So from where I sit,

从我的角度看来，

if we want to know if AI really understands language like we do,

如果我们想真正知道 AI能否像我们这样理解语言，

we need to get inside of the Chinese room.

我们需要进入那个“中文房间”，

We need to know what the AI is doing,

需要知道AI到底在做什么，

and we need to be able to compare that to what people are doing

需要能将AI的行为与人类

when they understand language.

理解语言的行为比较。

AI is moving so fast.

人工智能发展得太快了。

Today, I'm asking you, does AI understand language

今天我还在问大家， 人工智能能否理解语言，

that might seem like a silly question in ten years.

可能十年以后， 这个问题就会看起来很“傻”。

Or ten months.

也可能十个月。

(Laughter)

（笑）

But one thing will remain true.

但有一件事不会变化。

We are meaning-making humans,

我们是创造意义的人类。

and we are going to continue to look for meaning

我们将继续寻找意义，

and interpret the world around us.

解释我们周围的世界。

And we will need to remember

我们需要记住，

that if we only look at the input and output of AI,

如果我们只看AI的输入和输出，

it's very easy to be fooled.

我们就容易被骗到。

We need to get inside of the metaphorical room of AI

我们需要真正深入 人工智能里的那个“房间”，

in order to see what's happening.

看到真正在发生的事情。

It's what's inside the counts.

房间里发生了什么才是最重要的。

Thank you.

谢谢。

(Applause)

（掌声）