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

譯者: Lilian Chiu 審譯者: Helen Chang

People are funny.

人很有趣。

We're constantly trying to understand and interpret

我們經常在試著了解 和詮釋我們周遭的世界。

the world around us.

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

我住的房子裡有兩隻黑貓，

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

讓我告訴各位，每當我看到黑色、

皺成一團的毛衣出現在我的 眼角餘光，就以為是貓。

I think it's a cat.

It's not just the things we see.

不只是我們看到的東西， 有時，我們會認為

Sometimes we attribute more intelligence than might actually be there.

它們有更高的智慧，實際上卻沒有。

Maybe you've seen the dogs on TikTok.

也許各位看過抖音上的狗，

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

說「散步」或「吃點心」 這些詞就能觸發牠們，

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

這讓牠們能和牠們的 主人做一些溝通，

and their owners think they use them

而主人認為他們用這些詞 做了些很不簡單的溝通。

to communicate some pretty impressive things.

But do the dogs know what they're saying?

但狗知道主人在說什麼嗎？

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

也許各位聽過「聰明的漢斯」 這匹馬的故事。

and he could do math.

牠會算數學。

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

且不只是解簡單的數學題， 還能解很複雜的，比如，

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

若這個月的八號是星期二， 下一個星期五是幾號？

what's the date of the following Friday?

It's like, pretty impressive for a horse.

以馬來說，這很不簡單。

Unfortunately, Hans wasn't doing math,

不幸的是，漢斯並不是在算數學，

but what he was doing was equally impressive.

但牠所做的也同樣不簡單。

Hans had learned to watch the people in the room

漢斯學會觀察房間中的人，

to tell when he should tap his hoof.

讓他們告訴牠何時該點蹄，

So he communicated his answers by tapping his hoof.

牠用點蹄的方式來表達牠的答案。

It turns out that if you know the answer

結果發現是，如果你知道

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

「若這個月的八號是星期二， 下一個星期五是幾號？」

what's the date of the following Friday,"

當這匹馬正確地點了十八次蹄時，

you will subconsciously change your posture

once the horse has given the correct 18 taps.

你就會下意識改變你的姿勢。

So Hans couldn't do math,

所以漢斯不會算數學，但牠 學會觀察房間中會算數學的人。

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

對馬來說，這樣也是相當不簡單。

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

But this is an old picture,

但，這張照片很古老，現今我們 不會上「聰明的漢斯」的當了。

and we would not fall for Clever Hans today.

Or would we?

或者，還是會？

Well, I work in AI,

我做的是人工智慧的工作，

and let me tell you, things are wild.

讓我告訴各位，這個領域很瘋狂。

There have been multiple examples of people being completely convinced

有不少例子都是有人完全被說服，

that AI understands them.

認為人工智慧懂他們。

In 2022,

2022 年，

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

一名 Google 工程師認為 Google 的人工智慧有感情。

And you may have had a really human-like conversation

各位可能也曾和 ChatGPT 有過非常像人類的對談。

with something like ChatGPT.

But models we're training today are so much better

但我們現今訓練的模型 比五年前的好太多了，

than the models we had even five years ago.

真的非常驚人。

It really is remarkable.

So at this super crazy moment in time,

那麼，在這超瘋狂的時刻，

let’s ask the super crazy question:

咱們來問這個超瘋狂的問題：

Does AI understand us,

人工智慧真的懂我們嗎？或者，

or are we having our own Clever Hans moment?

我們也遇到了「聰明的漢斯」的狀況？

Some philosophers think that computers will never understand language.

有些哲學家認為 電腦永遠不可能懂語言。

To illustrate this, they developed something they call

為了闡明這一點， 他們發展出了所謂的

the Chinese room argument.

中文房間論證。

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

在中文房間中，

有一個人，假設的人，他不懂中文。

who does not understand Chinese,

but he has along with him a set of instructions

但他手上有一組指示教他如何用中文

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

回應任何中文的句子。

Here's how the Chinese room works.

這個論證是這樣的： 一張紙被從門縫塞進來，

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

上面寫的是中文。

has something written in Chinese on it.

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

房間中的人要用他們 手上的指示來想辦法回應，

They write the response down on a piece of paper

把回應寫在紙上， 再把紙從門縫送回去。

and then send it back out through the door.

To somebody who speaks Chinese,

站在房間外面且會說中文的人

standing outside this room,

會以為房間中的人會說中文。

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

But we know they do not,

但我們知道他們不會說中文，

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

因為不需要對中文的知識， 就能遵照指示回應。

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

執行這項工作任務 並不能展現出你懂中文。

So what does that tell us about AI?

那跟人工智慧有什麼關係？

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

當你我站在房間的外面，

跟 ChatGPT 這類人工智慧說話，

when we speak to one of these AIs like ChatGPT,

we are the person standing outside the room.

我們就是站在房間外面的人。

We're feeding in English sentences,

我們輸入英文句子， 得到的回應也是英文句子。

we're getting English sentences back.

It really looks like the models understand us.

感覺真的像是這個模型懂我們。

It really looks like they know English.

真的像是它們懂英文。

But under the hood,

但在背後，這些模型也只是 遵循一組更複雜的指示。

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

How do we know if AI understands us?

我們要怎麼知道 人工智慧是否懂我們？

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

為了回答這個問題， 咱們再回到中文房間。

Let's say we have two Chinese rooms.

假設我們有兩間中文房間。

In one Chinese room is somebody who actually speaks Chinese,

在一間中文房間裡的人 真的會說中文，

and in the other room is our impostor.

另一間則是冒牌的。

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

真正會說中文的人拿到了寫著 中文的紙，他們能讀，沒問題。

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

But when our imposter gets it again,

但當冒牌者拿到紙時，就得 再用手上的指示來想辦法回應。

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

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

從外面可能完全無法 區別這兩間房間，

but we know inside something really different is happening.

但我們知道房間內 發生的狀況很不同。

To illustrate that,

為了說明，

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

假設在兩間房間中兩個人的腦袋裡，

inside of our two rooms,

is a little scratch pad.

是一本小便條簿。

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

他們為了完成任務 所必須要記住的一切

has to be written on that little scratch pad.

都得要寫在那本小便條簿上。

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

如果我們能看到 小便條簿上寫了什麼，

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

我們就能夠知道他們完成 任務的方法有多麼不同。

So though the input and the output of these two rooms

所以雖然兩間房間的輸入 和輸出可能都完全一樣，

might be exactly the same,

從輸入到輸出的過程則是完全不同。

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

So again, what does that tell us about AI?

再一次，那跟人工智慧有什麼關係？

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

再一次，如果人工智慧能 產生出貌似真實的對話，

answers questions just like we would expect,

照我們預期的方式回答問題，

it may still be an imposter of sorts.

它仍然可能是冒牌的。

如果我們想知道人工智慧 是否像我們這樣懂語言，

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

we need to know what it's doing.

我們就得知道它在做什麼。

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

我們得要進去看看它在 做什麼。它是不是冒牌的？

Is it an imposter or not?

We need to see its scratch pad,

我們得看它的便條簿，

and we need to be able to compare it

且我們得把它和真正懂 語言的人的便條簿拿來比較。

to the scratch pad of somebody who actually understands language.

But like scratch pads in brains,

但，如同腦袋中的便條簿， 那是我們看不到的，對吧？

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

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

結果發現，我們算是可以 看得見腦袋裡的便條簿。

Using something like fMRI or EEG,

用比如功能性磁振造影或腦電圖，

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

我們可以在大腦讀取資料時 拍下類似大腦的照片，

所以，找人來閱讀字詞或故事，

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

接著拍下他們大腦的照片。

And those brain images are like fuzzy,

那些大腦影像就像是

模糊、沒聚焦的大腦便條簿照片。

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

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

它們能讓讓我們略知大腦如何

and representing information while you read.

在你閱讀的時候處理和表達資訊。

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

這三張大腦影像是在一個人 閱讀這三個詞時拍下的：

「公寓」、「房子」，及「芹菜」。

"house" and "celery."

You can see just with your naked eye

各位用肉眼就能看見，

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

「公寓」和「房子」的 大腦影像比較相似，

are more similar to each other

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

比較不像「芹菜」的大腦影像。

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

當然，各位知道， 公寓和房子的相似度

than they are to celery, just the words.

高於芹菜，就字面上來說。

So said another way,

換句話說，

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

在閱讀「公寓」和「房子」時， 大腦使用其便條簿的方式比較相近，

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

和閱讀「芹菜」時相差較多。

The scratch pad tells us a little bit

便條簿能讓我們略知 大腦如何表達語言。

about how the brain represents the language.

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

它無法完全呈現大腦 在做什麼，但也夠好了。

but it's good enough.

OK, so we have scratch pads for the brain.

好，我們有大腦的便條簿了，

Now we need a scratch pad for AI.

現在我們需要人工智慧的便條簿。

So inside a lot of AIs is a neural network.

許多人工智慧的內部是神經網路，

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

而神經網路的內部是一大堆神經元。

So here the neurons are like these little gray circles.

在這裡，神經元是那些灰色小圓點。

And we would like to know

我們想知道，神經網路的 便條簿是什麼？

what is the scratch pad of a neural network?

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

當我們將一個字詞 輸入到神經網路中，

each of the little neurons computes a number.

每一個小神經元都會計算一個數字。

Those little numbers I'm representing here with colors.

在這裡我用顏色來代表那些小數字。

So every neuron computes this little number,

每個神經元會計算這個小數字，

and those numbers tell us something

而那些數字能告訴我們

about how the neural network is processing language.

神經網路如何處理語言。

Taken together,

所有這些小圓點綜合起來

all of those little circles paint us a picture

就能為呈現出神經網路 如何表達語言，

of how the neural network is representing language,

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

提供我們神經網路的便條簿。

OK, great.

很好，現在我們有大腦的 便條簿和人工智慧的便條簿，

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

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

我們想知道：人工智慧 是否在做大腦做的事？

How can we test that?

我們要如何檢測這一點？

Here's what researchers have come up with.

研究者想出這個方法：

We're going to train a new model.

我們要訓練一個新模型。

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

那個新模型會參考

便條簿如何回應 某一個詞，接著試著預測

for a particular word

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

大腦的便條簿怎麼回應這個詞。

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

順道一提，也可以反向操作。

So let's train a new model.

所以，咱們來訓練一個新模型。

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

它會到神經網路的條便簿去找 某個詞，再預測大腦的條便簿。

and try to predict the brain scratchpad.

如果大腦和人工智慧 做的事完全不同，

If the brain and AI are doing nothing alike,

have nothing in common,

沒有共通點，我們就 無法成功做到預測。

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

就不可能用一本條便簿 來預測另一本。

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

So we've reached a fork in the road

我們來到了道路的分叉口， 各位應該看得出來，

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

我接著會說出這兩個 答案的其中一個：

I’m going to tell you AI is amazing,

我可能會說人工智慧很不可思議，

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

或者會說人工智慧是冒牌貨。

Researchers like me love to remind you

像我這種研究者要提醒大家，

that AI is nothing like the brain.

人工智慧完全不像大腦，那是事實。

And that is true.

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

但有沒有可能人工智慧 和大腦有些共通性？

So we’ve done this scratch pad prediction task,

我們做了這種便條簿 預測，結果發現，

and it turns out, 75 percent of the time

有 75% 的時候，

the predicted neural network scratchpad for a particular word

針對某個詞所預測出來的 神經網路便條簿

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

比較接近神經網路對於 那個詞的真正便條簿，

than it is to the neural network scratch pad

而沒那麼接近神經網路對於 其他隨機字詞的便條簿——

for some other randomly chosen word --

75 percent is much better than chance.

75% 比隨機的機率更高許多。

What about for more complicated things,

若不只是字詞，更複雜的 如句子或甚至故事呢？

not just words, but sentences, even stories?

Again, this scratch pad prediction task works.

同樣的，便條簿 預測的方式也行得通。

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

我們可以用神經網路的便條簿 來預測大腦的，反之亦然。

Amazing.

很驚人。

So does that mean

那是否意味著

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

神經網路和人工智慧懂 語言的方式和我們一樣？

Well, truthfully, no.

絕對不是。

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

雖然這些便條簿預測的 正確率比隨機猜測更佳，

the underlying correlations are still pretty weak.

背後的相關性仍然很弱。

And though neural networks are inspired by the brain,

雖然神經網路的靈感來自大腦，

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

它們並沒有大腦的 那種結構和複雜度。

that we see in the brain.

Neural networks also don't exist in the world.

神經網路並不存在於世界上。

A neural network has never opened a door

神經網路從來沒有開過門、 看過日落，或聽過寶寶哭泣。

or seen a sunset, heard a baby cry.

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

沒有真正存在於這個世界上， 沒體驗過這個世界的神經網路

hasn't really experienced the world,

really understand language about the world?

有可能了解關於這個世界的語言嗎？

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

不過，這些便條簿預測實驗 還是得到了證實——

許多大腦成像實驗，許多神經網路。

multiple brain imaging experiments,

multiple neural networks.

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

我們也發現，隨著 神經網路越來越正確，

they also start to use their scratch pad in a way

它們運用便條簿的方式 也變得越來越像大腦。

that becomes more brain-like.

And it's not just language.

不只是在語言方面，在導航 和視覺方面也有類似的結果。

We've seen similar results in navigation and vision.

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

所以，人工智慧的 運作方式和大腦不同，

but it's not completely random either.

但它也不是隨機亂猜的。

So from where I sit,

從我的角度來看，

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

若我們想知道人工智慧

能否像我們一樣真正懂語言， 我們就得進入中文房間裡，

we need to get inside of the Chinese room.

We need to know what the AI is doing,

我們得知道人工智慧在做什麼，

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

並比較看看，和懂語言的人 在做的有什麼差別。

when they understand language.

AI is moving so fast.

人工智慧進展超快。

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

今天我問各位人工智慧是否懂語言，

that might seem like a silly question in ten years.

十年後這個問題可能會顯得很蠢，

Or ten months.

或十個月後。

(Laughter)

（笑聲）

But one thing will remain true.

但有一點仍然不會變： 我們是會創造意義的人類，

We are meaning-making humans,

and we are going to continue to look for meaning

我們將會持續尋找意義，

and interpret the world around us.

並詮釋我們周遭的世界。

And we will need to remember

大家要切記，

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

如果我們只去看 人工智慧的輸入和輸出，

it's very easy to be fooled.

很容易被愚弄。

We need to get inside of the metaphorical room of AI

我們得進入人工智慧的 房間（比喻說法），

in order to see what's happening.

去看看裡面是怎麼回事。

It's what's inside the counts.

裡面的才是重要的。

Thank you.

謝謝。

(Applause)

（掌聲）