Are Insect Brains the Secret to Great AI? | Frances S. Chance | TED

譯者: Lilian Chiu 審譯者: Helen Chang

Creating intelligence on a computer.

在電腦上創造智慧。

This has been the Holy Grail for artificial intelligence

從好一段時間之前，這就 一直是人工智慧的聖杯。

for quite some time.

But how do we get there?

但我們要如何做到？

So we view ourselves as highly intelligent beings.

我們自認是具有高度智慧的生物。

So it's logical to study our own brains,

所以很合邏輯的做法 就是研究我們自己的大腦，

the substrate of our cognition, for creating artificial intelligence.

我們認知的基質，

來創造人工智慧。

Imagine if we could replicate how our own brains work on a computer.

想像一下如果我們能把我們 大腦運作的方式複製到電腦上。

But now consider the journey that would be required.

但，再想想會需要 什麼樣的過程才能達成。

The human brain contains 86 billion neurons.

人類的大腦有八百六十億個神經元。

Each is constantly communicating with thousands of others,

每個神經元都經常在 和數千個其他神經元溝通，

and each has individual characteristics of its own.

每個神經元也都有它自己的特徵。

Capturing the human brain on a computer

把人類大腦放到 電腦上，簡單來說就是

may simply be too big and too complex a problem

太大、太複雜的問題，

to tackle with the technology and the knowledge that we have today.

用現今的科技和知識無法辦到。

I believe that we can capture a brain on a computer,

我相信我們可以把大腦放到電腦上，

but we have to start smaller.

但我們得從比較小的大腦著手。

Much smaller.

小很多。

These insects have three of the most fascinating brains in the world to me.

對我來說，這三種昆蟲 有世界上最炫的三種大腦。

While they do not possess human-level intelligence,

雖然牠們沒有人類等級的智慧，

each is remarkable at a particular task.

但每一種都非常擅長 一項特定的工作任務。

Think of them as highly trained specialists.

把牠們想成受過高度訓練的專家。

African dung beetles are really good at rolling large balls in straight lines.

非洲蜣螂非常擅長沿直線滾出大球。

(Laughter)

（笑聲）

Now, if you've ever made a snowman,

如果你曾經做過雪人，

you know that rolling a large ball is not easy.

就會知道要滾出大球並不容易。

Now picture trying to make that snowman

想像一下，你要用跟你 一樣大的雪球來做的雪人，

when the ball of snow is as big as you are

and you're standing on your head.

你還呈倒立狀態。

(Laughter)

（笑聲）

Sahara desert ants are navigation specialists.

撒哈拉銀蟻是導航專家。

They might have to wander a considerable distance to forage for food.

牠們可能得要四處遊走 相當遠的距離去找食物。

But once they do find sustenance,

但，一旦找到食物，

they know how to calculate the straightest path home.

牠們知道如何計算出 回到家最直的路線。

And the dragonfly is a hunting specialist.

而蜻蜓則是打獵專家。

In the wild, dragonflies capture approximately 95 percent

在野外，被蜻蜓選中的獵物 有近 95% 會被牠們捉到。

of the prey they choose to go after.

These insects are so good at their specialties

這些昆蟲在牠們專長的領域很出色，

that neuroscientists such as myself study them as model systems

因此像我這樣的神經科學家

會把牠們當作模範系統來研究，

to understand how animal nervous systems solve particular problems.

以了解動物的神經系統 如何解決特定的問題。

And in my own research, I study brains to bring these solutions,

在我自己的研究中，我鑽研大腦，

將生物學所能提供 最棒的解決方案帶給電腦。

the best that biology has to offer, to computers.

So consider the dragonfly brain.

以蜻蜓的大腦為例。

It has only on the order of one million neurons.

它只有大約一百萬個神經元。

Now, it's still not easy to unravel a circuit of even one million neurons.

就算只有一百萬個神經元， 也仍然不容易搞懂一個迴路。

But given the choice

但如果要選擇去試著破解

between trying to tease apart the one-million-neuron brain

一百萬個神經元的大腦， 或八百六十億個神經元的大腦，

versus the 86-billion-neuron brain,

which would you choose to try first?

你會選擇先試哪一個？

(Laughter)

（笑聲）

When studying these smaller insect brains,

研究這些較小的昆蟲大腦時，

the immediate goal is not human intelligence.

當前的目標並不是人類智慧。

We study these brains for what the insects do well.

我們針對這些昆蟲的專長 來研究這些大腦。

And in the case of the dragonfly, that's interception.

以蜻蜓來說，專長就是攔截。

So when dragonflies are hunting,

蜻蜓在打獵時，

they do more than just fly straight at the prey.

牠們不會只是直直飛向獵物。

They fly in such a way that they will intercept it.

牠們的飛行方式 讓牠們能攔截該獵物。

They aim for where the prey is going to be.

牠們會瞄準該獵物將會到達的地方。

Much like a soccer player, running to intercept a pass.

很像是足球員跑去攔截對手傳球。

To do this correctly,

要正確做好攔截，

dragonflies need to perform what is known as a coordinate transformation,

蜻蜓需要做我們所知的座標轉換，

going from the eye’s frame of reference, or what the dragonfly sees,

從眼睛的參照座標系， 也就是蜻蜓看見的，

to the body's frame of reference,

轉到身體的參照座標系，

or how the dragonfly needs to turn its body to intercept.

即蜻蜓要怎麼做才能 把身體轉對方向去做攔截。

Coordinate transformations are a basic calculation

動物必須要會座標轉換 這種基本計算，

that animals need to perform to interact with the world.

才能夠跟世界互動。

We do them instinctively every time we reach for something.

每當我們伸手拿東西時 就是直覺地在做座標轉換。

When I reach for an object straight in front of me,

當我伸手去拿我前面的物體時，

my arm takes a very different trajectory than if I turn my head,

我的手臂移動的軌跡 完全不同於我轉動我的頭

look at that same object when it is off to one side

去看在旁邊的同一個物體

and reach for it there.

並伸手去拿它的軌跡。

In both cases, my eyes see the same image of that object,

在這兩個例子中，我的眼睛 都看到同樣的物體影像。

but my brain is sending my arm on a very different trajectory

但我的大腦會指示我的手臂 沿不同的軌跡移動，

based on the position of my neck.

其依據是脖子的位置。

And dragonflies are fast.

蜻蜓的速度很快。

This means they calculate fast.

這就表示牠們的計算很快。

The latency, or the time it takes for a dragonfly to respond

「延遲時間」，也就是 蜻蜓看到獵物轉彎之後

once it sees the prey turn,

做出反應需要的時間，

is about 50 milliseconds.

是大約五十毫秒。

This latency is remarkable.

這延遲時間短得不可思議。

For one thing, it's only half the time of a human eye blink.

一則，它僅是人類眨眼一次 所需要的時間的一半。

But for another thing,

但，二則，

it suggests that dragonflies capture how to intercept

這也表示蜻蜓要知道 如何攔截，只需要用

in only relatively or surprisingly few computational steps.

相對比較少的計算步驟， 而且少得驚人。

So in the brain,

在大腦中，

a computational step is a single neuron

一個計算步驟就是單一個神經元

or a layer of neurons working in parallel.

或者是一層神經元平行運作。

It takes a single neuron about 10 milliseconds

一個神經元需要大約十毫秒

to add up all its inputs and respond.

把所有的輸入資訊 加總起來做出反應。

The 50-millisecond response time

反應時間只要五十毫秒，

means that once the dragonfly sees its prey turn,

意味著一旦蜻蜓看到牠的獵物轉向，

there's only time for maybe four of these computational steps

也許時間只夠進行 四個這種計算步驟，

or four layers of neurons, working in sequence, one after the other,

或者四層神經元， 依序運作，一層接著一層，

to calculate how the dragonfly needs to turn.

來計算蜻蜓需要如何轉向。

In other words, if I want to study

換言之，如果我想要研究蜻蜓如何做

how the dragonfly does coordinate transformations,

座標轉換，

the neural circuit that I need to understand,

我需要了解的神經迴路，

the neural circuit that I need to study,

我需要研究的神經迴路，

can have at most four layers of neurons.

最多只有四層神經元。

Each layer may have many neurons,

一層當中可能有許多神經元，

but this is a small neural circuit.

這算是很小的神經迴路。

Small enough that we can identify it

小到我們可以用現今 既有的工具來辨識、研究它。

and study it with the tools that are available today.

And this is what I'm trying to do.

這就是我在試圖做的事。

I have built a model of what I believe is the neural circuit

我建立了一個模型， 我相信它就是用來計算

that calculates how the dragonfly should turn.

蜻蜓應該如何轉向的神經迴路。

And here is the cool result.

很酷的結果如下。

In the model,

在模型中，

dragonflies do coordinate transformations in only one computational step,

蜻蜓只用一個計算步驟 就完成座標轉換，

one layer of neurons.

只用一層神經元。

This is something we can test and understand.

這是我們能測試並了解的現象。

In a computer simulation,

在電腦模擬中，

I can predict the activities of individual neurons

我可以預測蜻蜓在打獵時， 每個個別神經元的活動。

while the dragonfly is hunting.

For example, here I am predicting the action potentials, or the spikes,

比如，在這裡我可以預測 當蜻蜓看到獵物移動時，

that are fired by one of these neurons

這些神經元的任何一個會 產生的動作電位，或峰電位。

when the dragonfly sees the prey move.

To test the model, my collaborators and I

為了測試這個模型， 我和我的合作夥伴現在

are now comparing these predicted neural responses

正在將這些預測的神經反應拿來比對

with responses of neurons recorded in living dragonfly brains.

從活的蜻蜓大腦 所記錄下來的神經元反應。

These are ongoing experiments

這些實驗還在進行中，

in which we put living dragonflies in virtual reality.

在實驗中，我們將活的 蜻蜓放到虛擬實境中。

(Laughter)

（笑聲）

Now, it's not practical to put VR goggles on a dragonfly.

讓蜻蜓戴 VR 頭戴式 顯示器不太實際。

So instead, we show movies of moving targets to the dragonfly,

所以，我們改成讓蜻蜓 看有移動目標的電影，

while an electrode records activity patterns of individual neurons

同時用電極來記錄大腦中 個別神經元的活動模式。

in the brain.

Yeah, he likes the movies.

是的，牠喜歡看電影。

If the responses that we record in the brain

如果我們從大腦中記錄到的反應

match those predicted by the model,

符合模型所做的預測，

we will have identified which neurons are responsible

就表示我們找出了哪些神經元 負責做座標轉換的工作。

for coordinate transformations.

The next step will be to understand the specifics

下一步是要去了解明確細節，

of how these neurons work together to do the calculation.

搞懂這些神經元如何合作來做計算。

But this is how we begin to understand how brains do basic

但，我們就是用這種方式 開始了解大腦如何做基本

or primitive calculations.

或原始的計算。

Calculations that I regard as building blocks for more complex functions,

我把這些計算視為 是更複雜計算的積木，

not only for interception but also for cognition.

不只用在攔截上，也用在認知上。

The way that these neurons compute may be different from anything

這些神經元計算的方式 可能和電腦上現有的方法不同。

that exists on a computer today.

And the goal of this work is to do more than just write code

而這項研究的目標不只是要

撰寫程式來複製神經元的活動模式。

that replicates the activity patterns of neurons.

We aim to build a computer chip

我們想要打造一種電腦晶片，

that not only does the same things as biological brains

它不只能做到生物大腦會做的事，

but does them in the same way as biological brains.

連做的方式也和生物大腦一樣。

This could lead to drones driven by computers

可能可以發明出一種無人機，

控制它的電腦只有 蜻蜓的大腦那麼大，

the same size of the dragonfly's brain

that captures some targets and avoid others.

這種無人機能捕捉某些 目標物，避開其他的。

Personally, I'm hoping for a small army of these

我個人希望有一支 這種無人機的小軍隊，

to defend my backyard from mosquitoes in the summer.

在夏天時守衛我的後院，對抗蚊子。

(Laughter)

（笑聲）

The GPS on your phone could be replaced by a new navigation device

你手機中的全球定位系統也可以換成

根據蜣螂或螞蟻 做出的新的導航裝置，

based on dung beetles or ants

that could guide you to the straight or the easy path home.

能幫你找到最直或最簡單的回家路。

And what would the power requirements of these devices be like?

這些裝置的電源需求有多大？

As small as it is --

以大腦這麼小——喔， 抱歉——以它這麼大的尺寸，

Or, sorry -- as large as it is,

the human brain is estimated to have the same power requirements

估計電力需求會等同於

as a 20-watt light bulb.

一個二十瓦的燈泡。

Imagine if all brain-inspired computers

想像一下，如果所有 靈感取自大腦的電腦

had the same extremely low-power requirements.

對於電力的需求 都同樣低到這種程度。

Your smartphone or your smartwatch probably needs charging every day.

你的智慧手機或智慧手錶 可能每天都需要充電。

Your new brain-inspired device might only need charging every few months,

靈感來自大腦的裝置可能 幾個月充電一次就夠了，

or maybe even every few years.

甚至幾年充電一次。

The famous physicist, Richard Feynman, once said,

知名的物理學家 理查‧費曼曾經說過：

"What I cannot create, I do not understand."

「我創造不出來的東西， 就表示我不了解它。」

What I see in insect nervous systems

我在昆蟲的神經系統中看到的，

is an opportunity to understand brains

是一個了解大腦的機會，

through the creation of computers that work as brains do.

透過創造模仿大腦運作方式的 電腦來了解大腦 。

And creation of these computers will not just be for knowledge.

創造這些電腦的目的 不只是為了知識。

There's potential for real impact on your devices, your vehicles,

有可能可以造成真實的影響， 改變你的裝置、你的車，

maybe even artificial intelligences.

可能甚至人工智慧。

So next time you see an insect,

所以，下次當你看到昆蟲，

consider that these tiny brains can lead to remarkable computers.

想想這些微小的大腦 能夠促成非凡的電腦。

And think of the potential that they offer us for the future.

也想想它們為我們的未來 提供的可能性。

Thank you.

謝謝。

(Applause)

（掌聲）