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

翻译人员: Yip Yan Yeung 校对人员: Sue Lu

I grew up with parents who are engineers.

我在我父母身边长大， 他们都是工程师。

They were among the first to bring computerized manufacturing

他们是最早将计算机化制造 带到我印度的家乡的那批工程师。

to my hometown in India.

Growing up as a young girl,

还是个小女孩的时候，

I remember being fascinated

我记得我很着迷于

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

这些计算机程序， 它们不仅存在于计算机中，

but touched the physical world

还能接触物理世界，

and produced these beautiful and precise metal parts.

帮助生产出精美而精密的金属部件。

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

过去二十年的 AI 研究中，

this memory continued to inspire me

这段记忆一直激励着我

to connect the physical

将物理世界和数字世界联结在一起。

and digital worlds together.

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

我的 AI 研究将改变 我们从事科学和工程工作的方式。

Scientific research and engineering design

科学研究和工程设计

currently involves a lot of trial and error.

目前涉及大量的反复试错。

Many long hours are spent in the lab doing experiments.

我们在实验室里 耗费了大量时间做实验。

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

推动科学向前发展的不仅仅是好主意,

You need these experiments to validate findings

你得用实验证明你的发现， 激发新的想法。

and spark new ideas.

How can language models help here?

语言模型在这方面有什么用处呢？

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

ChatGPT 会怎么做， 如果我让它设计更好的飞机机翼，

or a drone that flies on a turbulent wind?

或是能在湍流中飞行的无人机？

It may suggest something.

它可能会提出一些建议，

It may even draw something.

甚至画点图案，

But how do we know this is any good?

但我们怎么知道能不能用呢？

We don't.

我们不知道。

Language models hallucinate because they have no physical grounding.

语言模型会产生幻觉， 因为它们没有物理基础。

While language models may help generate new ideas,

尽管语言模型 可能会帮你生成新的想法，

they cannot attack the hard part of science

但是它们无法攻克科学中的难题，

which is simulating the necessary physics

即模拟必需的物理存在 以取代实验室实验。

to replace the Nab experiments.

In order to model scientific and physical phenomena,

要模拟科学和物理现象，

text alone is not sufficient.

光靠语言文字是不够的。

To get to AI with universal physical understanding,

为了使 AI 具有普通物理理解能力，

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

我们需要用观察到的数据训练它。

And not just that, also its hidden details.

除此之外，还有数据中隐藏的细节。

From the intricacies of quantum chemistry

从量子化学的错综复杂

that happen at the smallest level

在最微小层面的影响，

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

到分子和蛋白质 对生物过程运作的影响，

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

到洋流和云层在星球 甚至更大范围内的变化，

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

我们需要 AI 来捕捉这些物理现象。

We need AI that can really zoom into the fine details

我们需要 AI 能够聚焦细枝末节，

in order to simulate these phenomena accurately.

才能准确地模拟这些现象。

To capture the cloud movements,

为了捕捉云层的运动，

and predict how clouds move and change in our atmosphere,

预测云在大气层中的移动和变化，

we need to be able to zoom into the fine details

我们需要能够放大湍流流动的细节。

of the turbulent fluid flow.

Standard deep learning uses a fixed number of pixels.

标准深度学习使用固定数量的像素，

So if you zoom in, it gets blurry

因此，一放大就会变模糊，

and not all the details are captured.

无法捕捉到所有细节。

We invented an AI technology called neural operators

我们发明了一种 名为“神经算子”的 AI 技术，

that represents the data as continuous functions or shapes,

它将数据表示为连续的函数或形状，

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

让我们能无限放大到任何分辨率或比例。

Neural operators allow us to train on data

神经算子让我们可以 在多种尺度或分辨率的数据上训练，

at multiple scales or resolutions.

And also allows us to incorporate

也让我们可以利用数学等式的知识

the knowledge of mathematical equations

to fill in the finer details

填补精密细节，

when only limited resolution data is available.

弥补分辨率数据有限的情况。

Such learning at multiple scales is essential for scientific understanding

这种多尺度的学习 对于科学理解至关重要，

and neural operators enable this.

而神经算子可以实现这一点。

With neural operators,

使用神经算子，

we can simulate physical phenomena such as fluid dynamics

我们可以模拟流体动力学等物理现象，

as much as a million times faster than traditional simulations.

速度比传统模拟快一百万倍。

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

去年，我们利用神经算子 发明了一种更好的医用导管。

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

医用导管是将液体吸出人体的管道。

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

不幸的是，细菌往往顺着导管逆流而上，

and infect the human.

感染人体。

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

每年有超过 50 万

of such healthcare-related infections,

此类医疗相关的感染病例，

and this is one of the leading causes.

而导管感染就是主要原因之一。

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

去年，我们使用神经算子将导管内部

from smooth to ridged.

从光滑改为凹凸不平的结构。

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

有了棱，液体流动时会产生涡流，

and we can hope to stop the bacteria from swimming upstream

因此有望借助涡流阻止细菌逆流而上。

because of these vortices.

But to get this correct,

但要做到这一点，

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

我们得让棱的形状完全正确。

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

以前会通过反复试错完成。

Design a version of the catheter,

先设计一款导管，

build it out, take it to the lab,

制作出来，带到实验室，

observe a hypothesis if something went wrong,

观察这个假设，如果出错，

rinse and repeat and redesign again.

洗净导管、重复试错，然后再重新设计。

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

现在，我们只需教会 AI 液体在导管内部流动的行为规律，

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

神经算子模型就能直接给出优化设计。

an optimized design.

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

我们只需 3D 打印一次这个设计， 验证它是否可行。

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

在视频中，导管正在实验室里测试。

The bacteria are not able to swim upstream,

这些细菌无法向上游动，

are instead being pushed out with the fluid flow.

而是顺着液体流向被排出体外。

In fact, we measured the reduction in bacterial contamination

我们测得细菌污染减少超过 100 倍。

by more than 100-fold.

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

在此例中，神经算子被专门用来 理解导管中的液体流动。

fluid flow in a tube.

What other applications can AI tackle

AI 还能应对哪些应用，

and help us solve such pressing problems?

帮助我们解决紧迫的问题呢？

Can deep learning beat numerical weather models?

深度学习能否超越数值天气模型？

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

2021 年 2 月， 一群顶尖气象科学家提出了这个问题，

发表在英国皇家学会的刊物中。

in a "Royal Society" publication.

They felt that AI was still in its infancy,

他们认为，AI 仍处于起步阶段，

and that a number of fundamental breakthroughs would be needed

需要一些根本性的突破

for AI to become competitive with traditional weather models,

才能与传统的天气模型媲美，

and that would take years or even decades.

而这一切将耗费数年甚至数十年。

Exactly a year later,

整整一年后，

we released FourCastNet.

我们发布了 FourCastNet。

Using neural operators,

利用神经算子，

we built the first fully AI-based weather model

我们构建了第一个完全 基于 AI 的天气模型。

that is high resolution

该模型具有高分辨率，

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

比传统的天气模型快数万倍。

What used to take a big supercomputer

过去需要一台大型超级计算机，

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

现在家里的游戏电脑就可以运行。

This model is also running

该模型还运行于欧洲中期天气预报中心，

at the European Centre for Medium-Range Weather Forecasting,

one of the premier weather agencies of the world.

该中心是世界主要气象机构之一。

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

我们的 AI 模型不仅 比传统模型快数万倍，

than traditional models.

It's also more accurate in many cases.

在许多情况下，它也更准确。

On September 16 last year,

去年 9 月 16 日，

Hurricane Lee hit the coast of Nova Scotia, Canada.

飓风李侵袭了 加拿大新斯科舍省海岸。

A full ten days earlier,

整整十天之前，

our FourCastNet model correctly predicted

我们的 FourCastNet 模型 正确地预测了飓风将登陆，

that the hurricane would make landfall,

but the traditional weather model

但传统天气模型预测飓风将绕过海岸。

predicted the hurricane would skip the coast.

Only five days later, on September 11,

仅仅五天后，即 9 月 11 日，

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

传统天气模型才将预测更正为登陆。

Extreme weather events such as Hurricane Lee

像飓风李这样的极端天气事件

will only increase further unless we take action

只会进一步增加， 除非我们对气候变化采取措施,

on climate change.

Such as finding new, clean sources of energy.

例如寻找新的清洁能源。

Nuclear fusion is one of them.

核聚变就是其中之一。

But unfortunately, there are still big challenges with it.

但不幸的是， 它仍然面临着巨大挑战。

The fusion reactor heats up the plasma

核聚变反应堆要将等离子体 加热到极高的温度才能开始聚变。

to extremely high temperatures to get fusion started.

And sometimes this hot plasma can escape confinement

有时这种高温等离子体会脱离限制，

and can damage the reactor.

从而损坏反应堆。

We train neural operators to simulate and predict

我们训练神经算子模拟和预测

the evolution of plasma inside the reactor.

反应堆内等离子体的演变。

And with it,

借此，

we can use this to predict disruptions before they occur

在干扰发生之前就能预测到，

and take corrective action in the real world.

并采取校正措施。

We are enabling the possibility of nuclear fusion

我们正在促进核聚变成为现实。

becoming a reality.

So neural operators and AI broadly

神经算子和广泛意义的 AI

are enabling us to tackle hard scientific challenges

使我们可以应对艰巨的科学挑战，

such as climate change and nuclear fusion.

比如气候变化和核聚变。

To me, this is just the beginning.

我认为这仅仅是个开始。

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

目前，这些 AI 模型仅限于 它们经过训练的狭窄领域。

What if you had an AI model

如果有个可以解决所有 科学问题的 AI 模型，会怎么样呢？

that could solve all and any scientific problem?

From designing better drones, aircrafts, rockets,

从设计更好的无人机、飞机、火箭

and even better drugs and medical devices?

到更好的药物和医疗器械？

Such an AI model would greatly benefit humanity.

这样的 AI 模型会让人类受益匪浅。

This is what we are working on.

这就是我们正在研究的问题。

We are building a generalist AI model with emergent capabilities

我们正在构建一个 具有新兴能力的通才 AI 模型，

that can simulate any physical phenomena

可以模拟任何物理现象，

and generate novel designs that were previously out of reach.

生成以前无法企及的新颖设计。

This is how we scale up neural operators

这就是我们如何拓展神经算子的应用，

to enable general intelligence with universal physical understanding.

实现具有普通物理理解能力的通用智能。

Thank you.

谢谢。

(Applause)

（掌声）