AI That Connects the Digital and Physical Worlds | Anima Anandkumar | TED

譯者: Enzo Liu 審譯者: Shelley Tsang 曾雯海

I grew up with parents who are engineers.

我和工程師的父母一起長大

They were among the first to bring computerized manufacturing

他們是第一個將電腦化製造

to my hometown in India.

帶到我的家鄉印度的人之一

Growing up as a young girl,

當我還是個年經女孩時

I remember being fascinated

我記得一個很衝擊的事

how these computer programs didn't just reside within a computer,

這些店撓程序不僅存在於計算機中

but touched the physical world

而且連接到物理世界

and produced these beautiful and precise metal parts.

並生產了這些美麗而精確的金屬零件

Over the last two decades, as I pursued AI research,

在過去的二十年中，我從事人工智能研究，

this memory continued to inspire me

這種記憶繼續激發我

to connect the physical

將物理和數位世傑連接

and digital worlds together.

I am working on AI that transforms the way we do science and engineering.

我正在研發人工智能這將改變我們從事科學和工程方式的方式。

Scientific research and engineering design

科學研究和工程設計

currently involves a lot of trial and error.

目前涉及大量的試驗和錯誤

Many long hours are spent in the lab doing experiments.

在實驗室中花了許多長時間進行實驗

So it's not just the great ideas that propel science forward.

因此推動科學向前不僅僅是偉大的想法

You need these experiments to validate findings

您需要這些實驗來驗證發

and spark new ideas.

現並激發新想法。

How can language models help here?

語言模型如何在這裡提供幫助？

What if I ask ChatGPT to come up with a better design of an aircraft wing,

如果我要求 ChatGPT為飛機翼設計更好的設計？

or a drone that flies on a turbulent wind?

或者在亂流中飛行的無人機？

It may suggest something.

它可能會告訴了我們一些東西

It may even draw something.

它甚至可能畫一些東西

But how do we know this is any good?

但是我們怎麼知道這有好處？

We don't.

可惜我們不知道

Language models hallucinate because they have no physical grounding.

語言模型會產生幻覺因為它們沒有物理基礎

While language models may help generate new ideas,

they cannot attack the hard part of science

which is simulating the necessary physics

而就是模擬必要的物理

to replace the Nab experiments.

以取代 Nab 實驗

In order to model scientific and physical phenomena,

為了模型化科學和物理現象

text alone is not sufficient.

單獨文自是不足夠的

To get to AI with universal physical understanding,

要透過普遍的物理理解來進行AI

we need to train it on the data of the world we observe.

我們需要在我們觀察的數據上訓練它

And not just that, also its hidden details.

不僅如此，還有其隱藏的細節

From the intricacies of quantum chemistry

從量子化學在最小程度上發生的複雜性，

that happen at the smallest level

to molecules and proteins that influence how all biological processes work,

到影響所有生物過程運作方式的分子和蛋白質

to ocean currents and clouds that happen at planetary scales and beyond,

再到行星規模和其他地方發生的海流和雲

we need AI that can capture these whole range of physical phenomena.

我們需要能夠捕捉所有物理現象的AI

We need AI that can really zoom into the fine details

我們更需要可以真正放大細節的AI

in order to simulate these phenomena accurately.

來準確模擬這些現象

To capture the cloud movements,

為了捕捉雲的運動

and predict how clouds move and change in our atmosphere,

並預測雲如何在我們大氣中移

we need to be able to zoom into the fine details

動和變化我們需要能夠縮放雜流體流動的細節

of the turbulent fluid flow.

Standard deep learning uses a fixed number of pixels.

標準深度學習使用固定數量的像素

So if you zoom in, it gets blurry

因此如果放大，它會變模糊

and not all the details are captured.

而且並非所有細節都能被捕捉到

We invented an AI technology called neural operators

我們發明了一種稱為神經運算子的AI 技術

that represents the data as continuous functions or shapes,

這技術會將數據表示為連續函數或形狀

and allows us to zoom in indefinitely to any resolution or scale.

並讓我們無限次放大到任何解析度或比例

Neural operators allow us to train on data

神經運算子允許我們以多種規模或分辨率進行數據進行訓練。

at multiple scales or resolutions.

And also allows us to incorporate

並且還允許我們將數學方程式的知識結

the knowledge of mathematical equations

to fill in the finer details

合在僅有限的分辨率數據

when only limited resolution data is available.

時填寫更細緻的細節

Such learning at multiple scales is essential for scientific understanding

這種多個規模的學習對於科學理解至關重要

and neural operators enable this.

神經運算子可以實現這一點

With neural operators,

使用神經運算子

we can simulate physical phenomena such as fluid dynamics

我們可以模擬流體動力學

as much as a million times faster than traditional simulations.

等物理現象，比傳統模擬更快一百萬倍

Last year, we used neural operators to invent a better medical catheter.

去年我們使用神經操作員發明更好的醫療導管。

A medical catheter is a tube that draws fluids out of the human body.

醫療導管是從人體中抽出液體的管道。

Unfortunately, the bacteria tend to swim upstream against the fluid flow

可是細菌傾向於向上游，抵抗液體流動

and infect the human.

並感染人類

In fact, annually there is more than half a million cases

事實上每年有超過半百萬這種

of such healthcare-related infections,

與醫療保健相關的感染個案

and this is one of the leading causes.

這也是主要原因之一

Last year, we used neural operators to change the inside of the catheter

去年我們使用神經操作員將導管的內部

from smooth to ridged.

從平滑改為凹痕

With ridges, now we have vortices created as the fluid flows,

現在我們在流體流動時產生了渦流

and we can hope to stop the bacteria from swimming upstream

我們可以希望阻止細菌

because of these vortices.

因為這些渦流而阻止細菌在上游。

But to get this correct,

但是為了這個問題

we need the shape of the ridges to be exactly right.

我們需要脊柱的形狀完全正確

In the past, this would have been done by trial and error.

這也是通過嘗試和錯誤來完成的

Design a version of the catheter,

設計導管的一個版本構建它將其

build it out, take it to the lab,

帶到實驗室建它

observe a hypothesis if something went wrong,

出現問題時觀察假設

rinse and repeat and redesign again.

並重複並重新設計。

But instead, we taught AI the behavior of the fluid flow inside the tube,

但相反我們教導 AI管內流體的行為

and with it, our neural operator model was able to directly propose

並且通過它我們的神經運算子模型能夠直接提出優化

an optimized design.

的設計

We 3D-printed the design only once to verify that it worked.

我們只進行 3D列印一次設計以驗證它是否有效

In the video, you're seeing our catheter being tested in the lab.

在影片中，您看到我們的導管在實驗室中進行測試

The bacteria are not able to swim upstream,

細菌無法向上游

are instead being pushed out with the fluid flow.

而是隨著流體流動並被推出

In fact, we measured the reduction in bacterial contamination

事實上我們測量了細菌污染的

by more than 100-fold.

減少了至少 100 倍以上

So in this case, the neural operators were specialized to understand

因此在這種情況下神經運算子

fluid flow in a tube.

專門了解管中的流體流量

What other applications can AI tackle

AI 可以解決哪些其他應用程序

and help us solve such pressing problems?

並幫助我們解決這樣迫切的問題？

Can deep learning beat numerical weather models?

深度學習能否擊敗 數字天氣模型？

A group of leading weather scientists asked this question in February 2021,

一群領先的天氣科學家 在 2021年 2 月在「皇家協會」

in a "Royal Society" publication.

出版物中提出了這個問題。

They felt that AI was still in its infancy,

他們認為人工智慧仍處於初級階段

and that a number of fundamental breakthroughs would be needed

若要讓人工智慧與傳統天氣模式競爭

for AI to become competitive with traditional weather models,

還需要一系列的基礎性突破

and that would take years or even decades.

而這需要數年甚至數十年

Exactly a year later,

一年後

we released FourCastNet.

我們發布了FourCastNet

Using neural operators,

使用神經運算子

we built the first fully AI-based weather model

我們構建了第一個完全基於AI 的

that is high resolution

天氣模型

and is tens of thousands of times faster than traditional weather models.

該模型具有高解析度而且比傳統天氣模型快數千倍

What used to take a big supercomputer

以前使用大型超級計算機的東西

can now run on a gaming PC that you may have at home.

現在可以在您家中上電腦上運行了

This model is also running

此模型還在歐洲等

at the European Centre for Medium-Range Weather Forecasting,

天氣預測中心運行

one of the premier weather agencies of the world.

該機構是世界頂尖的天氣機構之一

And our AI model is not just tens of thousands of times faster

而我們的AI模型不僅比傳統型號快數千倍

than traditional models.

It's also more accurate in many cases.

在許多情況下它也更準確

On September 16 last year,

去年 9 月 16 日

Hurricane Lee hit the coast of Nova Scotia, Canada.

颶風“Lee”襲擊了加拿大新斯科舍省的海岸

A full ten days earlier,

十天前

our FourCastNet model correctly predicted

我們的 FourCastNet 模型正確預測了颶

that the hurricane would make landfall,

風將進入陸地

but the traditional weather model

但傳統天氣模型

predicted the hurricane would skip the coast.

預測颱風將跨過海岸

Only five days later, on September 11,

僅僅五天後9 月 11 日

did the traditional weather model correct its forecast to predict landfall.

傳統天氣模型也預測陸地

Extreme weather events such as Hurricane Lee

除非我們對氣候變化採取行動

will only increase further unless we take action

否則颶風“Lee”等極端天氣事件

on climate change.

只會進一步增加

Such as finding new, clean sources of energy.

例如尋找新的清潔能源

Nuclear fusion is one of them.

核融合也是其中之一

But unfortunately, there are still big challenges with it.

但不幸的是它仍然存在很大的挑戰

The fusion reactor heats up the plasma

融合反應器將等離子體

to extremely high temperatures to get fusion started.

加熱至極高溫以開始融合

And sometimes this hot plasma can escape confinement

有時這種熱的等離子可以逃離監禁

and can damage the reactor.

並可能會損壞反應器

We train neural operators to simulate and predict

我們訓練神經操作員以模擬

the evolution of plasma inside the reactor.

和預測反應器內等離子體的演變

And with it,

有了它

we can use this to predict disruptions before they occur

我們可以利用它在它們發生之前預測中斷

and take corrective action in the real world.

並在現實世界中採取糾正措施

We are enabling the possibility of nuclear fusion

我們正在使核融合成為現實的可能性

becoming a reality.

So neural operators and AI broadly

因此神經運算子和 AI 廣泛

are enabling us to tackle hard scientific challenges

使我們能夠應對艱難的科學挑戰

such as climate change and nuclear fusion.

例如氣候變化和核融合

To me, this is just the beginning.

對我來說這只是個開始

So far, these AI models are limited to the narrow domains they're trained on.

到目前為止這些AI模型僅限於它們受到訓練的狹窄領域

What if you had an AI model

如果你有一個可以解決

that could solve all and any scientific problem?

所有科學問題的AI模型怎麼辦？

From designing better drones, aircrafts, rockets,

從設計更好的無人機，飛機，火箭

and even better drugs and medical devices?

甚至更好的藥物 和醫療設備？

Such an AI model would greatly benefit humanity.

這樣的AI模型將對人類有很多的好處

This is what we are working on.

這也就是我們正在努力的

We are building a generalist AI model with emergent capabilities

我們正在構建一個具有新興功能的通用AI模型

that can simulate any physical phenomena

可以模擬任何物理現象

and generate novel designs that were previously out of reach.

並產生以前無法觸及的新型設計

This is how we scale up neural operators

這就是我們擴展神經運算子的方式

to enable general intelligence with universal physical understanding.

以實現通用物理理解的通用智能。

Thank you.

謝謝大家

(Applause)

（掌聲）