With Spatial Intelligence, AI Will Understand the Real World | Fei-Fei Li | TED

翻译人员: Yip Yan Yeung 校对人员: Lening Xu

Let me show you something.

给大家展示一下。

To be precise,

确切地说，

I'm going to show you nothing.

我什么都不会展示。

This was the world 540 million years ago.

这是 5.4 亿年前的世界。

Pure, endless darkness.

纯粹、无尽的黑暗。

It wasn't dark due to a lack of light.

黑暗并非因为缺乏光线。

It was dark because of a lack of sight.

而是因为缺乏观察的眼睛。

Although sunshine did filter 1,000 meters

虽然阳光穿透了海洋表面，

beneath the surface of ocean,

深入到 1000 米以下，

a light permeated from hydrothermal vents to seafloor,

光线通过热液喷口照射着 充满着生命的海底，

brimming with life,

there was not a single eye to be found in these ancient waters.

但在这古老的水域中并没有一只眼睛。

No retinas, no corneas, no lenses.

没有视网膜，没有角膜，没有晶状体。

So all this light, all this life went unseen.

每一束光、每一个生命都不为人知。

There was a time that the very idea of seeing didn't exist.

曾经有一段时间， “看”这一概念并不存在。

It [had] simply never been done before.

以前就是没有这种做法。

Until it was.

直到看见的这一刻。

So for reasons we're only beginning to understand,

出于我们才摸到门槛的一些原因，

trilobites, the first organisms that could sense light, emerged.

三叶虫——第一种能感光的生物，出现了。

They're the first inhabitants of this reality that we take for granted.

它们是我们现在理所当然 身处的环境里的第一批栖息动物。

First to discover that there is something other than oneself.

它们第一个发现了 除了自己还有别的东西。

A world of many selves.

一个由许多“自我”组成的世界。

The ability to see is thought to have ushered in Cambrian explosion,

视觉被认为推动了寒武纪生命大爆发，

a period in which a huge variety of animal species

在此期间， 各种各样的动物物种

entered fossil records.

有了化石记录。

What began as a passive experience,

最初的被动体验， 即 让光透进来简单动作，

the simple act of letting light in,

soon became far more active.

很快就变得更加主动了。

The nervous system began to evolve.

神经系统开始进化。

Sight turning to insight.

观察变成了洞察。

Seeing became understanding.

看见变成了理解。

Understanding led to actions.

理解带来了行动。

And all these gave rise to intelligence.

它们促成了智能。

Today, we're no longer satisfied with just nature's gift of visual intelligence.

如今，我们不再满足于 大自然赋予的视觉智能。

Curiosity urges us to create machines to see just as intelligently as we can,

好奇心促使我们创造出那些机器， 使其尽可能智能，

if not better.

甚至更好。

Nine years ago, on this stage,

九年前，在这个舞台上，

I delivered an early progress report on computer vision,

我发表了一场有关计算机视觉的 早期进展报告。

a subfield of artificial intelligence.

这是人工智能的一个细分领域。

Three powerful forces converged for the first time.

三股强大的力量首次 汇聚在一起。

Aa family of algorithms called neural networks.

一种称为神经网络的算法。

Fast, specialized hardware called graphic processing units,

称为“图形处理器”或 GPU的快速、专业的硬件，

or GPUs.

And big data.

还有大数据。

Like the 15 million images that my lab spent years curating called ImageNet.

就像我的实验室花了多年时间整理的

名为 ImageNet 的 1500 万张图像一样。

Together, they ushered in the age of modern AI.

它们共同开创了现代 AI 时代。

We've come a long way.

我们已经走了很长一段路。

Back then, just putting labels on images was a big breakthrough.

当时，单单是给图片打标签 都是一个重大突破。

But the speed and accuracy of these algorithms just improved rapidly.

但是这些算法的速度和准确性 迅速提高了。

The annual ImageNet challenge, led by my lab,

由我的实验室领导的 一年一度 ImageNet 挑战赛

gauged the performance of this progress.

评测了这一进展的表现。

And on this plot, you're seeing the annual improvement

在这张图上，可以看到每年的改进

and milestone models.

和里程碑式模型。

We went a step further

我们更进一步，

and created algorithms that can segment objects

创造了算法切分物体

or predict the dynamic relationships among them

或预测它们之间的动态关系，

in these works done by my students and collaborators.

由我的学生和合作者们 在这些项目中共同完成。

And there's more.

还有更多。

Recall last time I showed you the first computer-vision algorithm

回想一下上一次我展示了 第一个计算机视觉算法，

that can describe a photo in human natural language.

使用人类自然语言描述照片。

That was work done with my brilliant former student, Andrej Karpathy.

那是我与我以前的优秀学生 安德烈·卡帕蒂 （Andrej Karpathy）

合作完成的。

At that time, I pushed my luck and said,

当时，我想碰碰运气地说：

"Andrej, can we make computers to do the reverse?"

“安德烈， 我们能让计算机反向操作吗？”

And Andrej said, "Ha ha, that's impossible."

安德烈说： “哈哈，不可能。”

Well, as you can see from this post,

正如这篇推文所示，

recently the impossible has become possible.

最近，不可能变成了可能。

That's thanks to a family of diffusion models

这要归功于一系列扩散模型，

that powers today's generative AI algorithm,

是它们驱动了如今的生成式 AI 算法，

which can take human-prompted sentences

将人为输入的句子

and turn them into photos and videos

转换为闻所未闻事物的照片和视频。

of something that's entirely new.

Many of you have seen the recent impressive results of Sora by OpenAI.

很多人已经看到了 OpenAI 的 Sora

最近展示出的惊人成果。

But even without the enormous number of GPUs,

但是，即使没有大量的 GPU，

my student and our collaborators

我的学生和合作者们

have developed a generative video model called Walt

开发了一个生成式视频模型， 名为“Walt”，

months before Sora.

领先 Sora 数月。

And you're seeing some of these results.

来看一些结果。

There is room for improvement.

还有改进的余地。

I mean, look at that cat's eye

看看那只猫的眼睛，

and the way it goes under the wave without ever getting wet.

还有它怎么能在水下 却没有被弄湿。

What a cat-astrophe.

真是个大“猫”病。

(Laughter)

（笑声）

And if past is prologue,

如果过去是序言，

we will learn from these mistakes and create a future we imagine.

我们会从这些错误中吸取教训， 创造一个我们想象的未来。

And in this future,

在未来，

we want AI to do everything it can for us,

我们希望 AI 为我们尽其所能，

or to help us.

或者帮助我们。

For years I have been saying

多年来，我一直在说

that taking a picture is not the same as seeing and understanding.

拍照不等于看见和理解。

Today, I would like to add to that.

今天，我想补充一下。

Simply seeing is not enough.

仅仅看见是不够的。

Seeing is for doing and learning.

看见是为了做和学。

When we act upon this world in 3D space and time,

当我们在三维空间和时间中 对这个世界采取行动时，

we learn, and we learn to see and do better.

我们会学习， 学会观察并做得更好。

Nature has created this virtuous cycle of seeing and doing

大自然创造了这种视与行的良性循环，

powered by “spatial intelligence.”

由“空间智能”所驱动。

To illustrate to you what your spatial intelligence is doing constantly,

为了说明空间智能 一直在做些什么，

look at this picture.

请看这张图片。

Raise your hand if you feel like you want to do something.

如果你想做点什么，请举手。

(Laughter)

（笑声）

In the last split of a second,

在刚才的一秒钟里，

your brain looked at the geometry of this glass,

你的大脑观察了 这个杯子的几何形状、

its place in 3D space,

它在三维空间中的位置、

its relationship with the table, the cat

它与桌子、猫 以及其他一切的关系。

and everything else.

And you can predict what's going to happen next.

而且你可以预测接下来会发生什么。

The urge to act is innate to all beings with spatial intelligence,

采取行动的冲动是所有 拥有空间智能的生物与生俱来的，

which links perception with action.

空间智能将感知与行动联系起来。

And if we want to advance AI beyond its current capabilities,

如果我们想推动 AI 超越其现有能力，

we want more than AI that can see and talk.

我们需要的不仅仅 是能看见和说话的 AI。

We want AI that can do.

我们要能做事的 AI。

Indeed, we're making exciting progress.

我们确实正在取得 一些令人兴奋的进展。

The recent milestones in spatial intelligence

空间智能领域最近的里程碑

is teaching computers to see, learn, do

是教会计算机看见、学习、做事，

and learn to see and do better.

学会看见，做得更好。

This is not easy.

这并不容易。

It took nature millions of years to evolve spatial intelligence,

大自然花了数百万年 才发展出空间智能，

which depends on the eye taking light,

由眼睛捕捉光线，

project 2D images on the retina

将二维图像投射到视网膜

and the brain to translate these data into 3D information.

和大脑上， 将这些数据转化为三维信息。

Only recently, a group of researchers from Google

就在最近，谷歌的一组研究人员

are able to develop an algorithm to take a bunch of photos

开发出了一种算法， 拍摄大量照片

and translate that into 3D space,

并将其转换到三维空间，

like the examples we're showing here.

就像我们在这里展示的例子一样。

My student and our collaborators have taken a step further

我的学生和我们的合作者更进一步，

and created an algorithm that takes one input image

创建了一种算法， 采集一张输入图像

and turn that into 3D shape.

并将其转换为三维形状。

Here are more examples.

以下是更多示例。

Recall, we talked about computer programs that can take a human sentence

回想一下，我们谈到过 可以将人类语句

转换为视频的计算机程序。

and turn it into videos.

A group of researchers in University of Michigan

密歇根大学的一组研究人员

have figured out a way to translate that line of sentence

研究出了将这行句子转换成

into 3D room layout, like shown here.

三维房间布局的方法， 如下所示。

And my colleagues at Stanford and their students

我在斯坦福大学的同事 和他们的学生

have developed an algorithm that takes one image

开发了一种算法， 可以通过一张图像

and generates infinitely plausible spaces

生成无限可能的空间

for viewers to explore.

供观众探索。

These are prototypes of the first budding signs of a future possibility.

这是未来可能性的 第一个萌芽的雏形。

One in which the human race can take our entire world

在这个可能性里， 人类采集了整个世界，

and translate it into digital forms

将其转化为数字形式，

and model the richness and nuances.

模拟其丰富性和细微的参差。

What nature did to us implicitly in our individual minds,

正如大自然暗暗 对我们每个人的大脑所做的，

spatial intelligence technology can hope to do

空间智能技术有望 提高我们的集体意识。

for our collective consciousness.

As the progress of spatial intelligence accelerates,

随着空间智能进步的加速，

a new era in this virtuous cycle is taking place in front of our eyes.

这种良性循环的新时代 正在我们眼前发生。

This back and forth is catalyzing robotic learning,

这种往复的交流 正在加速机器人学习，

a key component for any embodied intelligence system

它是任何需要理解 并与三维世界互动的

that needs to understand and interact with the 3D world.

具身智能系统的关键组成部分。

A decade ago,

十年前，

ImageNet from my lab

我实验室的 ImageNet

enabled a database of millions of high-quality photos

启用了一个包含数百万张 高质量照片的数据库，

to help train computers to see.

帮助训练计算机“看见”。

Today, we're doing the same with behaviors and actions

如今，我们也在用行为和动作

to train computers and robots how to act in the 3D world.

训练计算机和机器人 在三维世界中采取行动。

But instead of collecting static images,

但是，我们没有收集静态图像，

we develop simulation environments powered by 3D spatial models

而是开发了 基于三维空间模型的仿真环境，

so that the computers can have infinite varieties of possibilities

这样计算机就可以有无限的可能性

to learn to act.

来学会行动。

And you're just seeing a small number of examples

你刚看到了一小部分示例，

to teach our robots

是在指导我们的机器人，

in a project led by my lab called Behavior.

来自我的实验室领导的 一个名为 Behavior 的项目。

We’re also making exciting progress in robotic language intelligence.

我们还在机器人语言智能方面 取得了令人兴奋的进展。

Using large language model-based input,

通过基于大语言模型的输入，

my students and our collaborators are among the first teams

我的学生和合作者们是第一批

that can show a robotic arm performing a variety of tasks

能够展示机械臂根据口头指示

执行各种任务的团队之一，

based on verbal instructions,

like opening this drawer or unplugging a charged phone.

例如打开抽屉 或拔掉充满电的手机。

Or making sandwiches, using bread, lettuce, tomatoes

或者用面包、生菜、西红柿 制作三明治，

and even putting a napkin for the user.

甚至会为用户放上一张纸巾。

Typically I would like a little more for my sandwich,

通常我会想在三明治里多放点料，

but this is a good start.

但这是一个不错的开始。

(Laughter)

（笑声）

In that primordial ocean, in our ancient times,

在我们远古时代的那片原始海洋中，

the ability to see and perceive one's environment

看见和感知环境的能力

kicked off the Cambrian explosion of interactions with other life forms.

拉开了与其他生命形式互动的 寒武纪生命大爆发的序幕。

Today, that light is reaching the digital minds.

此刻，这道光射向了数字大脑。

Spatial intelligence is allowing machines

空间智能使机器

to interact not only with one another,

不仅可以相互交互，

but with humans, and with 3D worlds,

还可以与人类、 真实或虚拟的三维世界进行交互。

real or virtual.

And as that future is taking shape,

随着未来逐渐成形，

it will have a profound impact to many lives.

它将对许多生命产生深远的影响。

Let's take health care as an example.

我们以医疗保健为例。

For the past decade,

在过去的十年中，

my lab has been taking some of the first steps

我的实验室一直在 进行一些初步研究，

in applying AI to tackle challenges that impact patient outcome

将 AI 应用于应对影响患者预后

and medical staff burnout.

和医务人员倦怠的挑战。

Together with our collaborators from Stanford School of Medicine

我们正在与斯坦福医学院

and partnering hospitals,

和合作医院的合作者们一起

we're piloting smart sensors

试点使用智能传感器，

that can detect clinicians going into patient rooms

检测临床医生在 未正确洗手的情况下进入病房。

without properly washing their hands.

Or keep track of surgical instruments.

或者追踪手术器械。

Or alert care teams when a patient is at physical risk,

或者，当患者面临身体危险， 例如跌倒时，提醒护理团队。

such as falling.

We consider these techniques a form of ambient intelligence,

我们将这些技术 视为环境智能的一种形式，

like extra pairs of eyes that do make a difference.

就像多几双帮得上忙的眼睛。

But I would like more interactive help for our patients, clinicians

但我希望为我们的患者、临床医生

和看护人员提供更多的互动式帮助， 他们也迫切需要帮助。

and caretakers, who desperately also need an extra pair of hands.

Imagine an autonomous robot transporting medical supplies

想象一下， 一个自主机器人运送医疗用品，

while caretakers focus on our patients

看护人员则专注于我们的患者，

or augmented reality, guiding surgeons to do safer, faster

或用增强现实 引导外科医生更安全、更快地做手术

and less invasive operations.

或者让手术的侵入性更低。

Or imagine patients with severe paralysis controlling robots with their thoughts.

或者想象一下严重瘫痪的患者 用自己的意念控制机器人。

That's right, brainwaves, to perform everyday tasks

没错，脑电波，可以执行 你我司空见惯的日常任务。

that you and I take for granted.

You're seeing a glimpse of that future in this pilot study from my lab recently.

在我实验室最近进行的这项试点研究中， 你可以一窥那样的未来。

In this video, the robotic arm is cooking a Japanese sukiyaki meal

在这段视频中，机械臂 正在烹饪一顿日本寿喜烧，

controlled only by the brain electrical signal,

仅由大脑电信号控制，

non-invasively collected through an EEG cap.

通过脑电帽非侵入性收集。

(Applause)

（掌声）

Thank you.

谢谢。

The emergence of vision half a billion years ago

五亿年前视觉的出现

turned a world of darkness upside down.

颠覆了黑暗的世界。

It set off the most profound evolutionary process:

它开启了最重大的进化过程：

the development of intelligence in the animal world.

动物世界智力的发展。

AI's breathtaking progress in the last decade is just as astounding.

AI 在过去十年中取得的惊人进步 同样令人震惊。

But I believe the full potential of this digital Cambrian explosion

但我相信，这场 数字寒武纪大爆发的全部潜力

won't be fully realized until we power our computers and robots

不会被完全发挥出来， 除非我们用空间智能

with spatial intelligence,

赋能我们的计算机和机器人，

just like what nature did to all of us.

正如大自然对我们所有人做的那样。

It’s an exciting time to teach our digital companion

这是一个激动人心的时刻， 可以教导我们的数字伙伴

to learn to reason

学会推理

and to interact with this beautiful 3D space we call home,

并与这个我们称为“家”的 美丽三维空间互动，

and also create many more new worlds that we can all explore.

同时创造更多我们可以探索的新世界。

To realize this future won't be easy.

要实现这个未来并不容易。

It requires all of us to take thoughtful steps

需要我们所有人每一步深思熟虑，

and develop technologies that always put humans in the center.

开发始终以人为本的技术。

But if we do this right,

但是，如果我们做对了，

the computers and robots powered by spatial intelligence

由空间智能驱动的计算机和机器人

will not only be useful tools

将不仅是实用的工具，

but also trusted partners

而且是值得信赖的合作伙伴，

to enhance and augment our productivity and humanity

可以提高、增强 我们的生产力和人性，

while respecting our individual dignity

同时，不损害我们每个人的尊严，

and lifting our collective prosperity.

促进我们的共同繁荣。

What excites me the most in the future

未来最令我兴奋的是，

is a future in which that AI grows more perceptive,

在未来，AI 有更强的理解能力、

insightful and spatially aware,

洞察力和空间感知能力，

and they join us on our quest

它们将和我们一起追求

to always pursue a better way to make a better world.

更好的方式， 创造更美好的世界。

Thank you.

谢谢。

(Applause)

（掌声）