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

譯者: C Leung 審譯者: 麗玲 辛

Let me show you something.

讓我展示一些東西給你看。

To be precise,

準確來說，

I'm going to show you nothing.

我要展示給你們的是 「什麼都看不到」。

This was the world 540 million years ago.

這是 540 億年前的世界，

Pure, endless darkness.

純粹、無盡的黑暗。

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

不是因為沒有光，

It was dark because of a lack of sight.

而是因為沒有視覺。

Although sunshine did filter 1,000 meters

雖然陽光確實能透射到

海平面下 1000 米。

beneath the surface of ocean,

a light permeated from hydrothermal vents to seafloor,

光線穿透到海底熱泉、到海床，

brimming with life,

充滿著生機，

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

這些古老水域卻連一隻眼睛都找不到，

No retinas, no corneas, no lenses.

沒有視網膜，沒有角膜，沒有晶狀體。

So all this light, all this life went unseen.

所以這些光，這些生命， 從未被看見。

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

曾經有一段時間， 「視覺」這個概念根本不存在。

It [had] simply never been done before.

這在以前從未發生過，

Until it was.

直到它發生。

So for reasons we're only beginning to understand,

因此，由於我們才剛開始了解的原因，

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

三葉蟲，第一種能夠 感知光的生物出現了，

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

它們是我們視為理所當然的 這個現實世界的第一批居民。

First to discover that there is something other than oneself.

首次發現除了自己之外， 還有其他東西存在，

A world of many selves.

一個有多個我的世界。

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

一般認為，視覺能力 促成寒武紀的大爆發，

a period in which a huge variety of animal species

這時期，種類繁多的動物物種

entered fossil records.

進入了化石記錄。

What began as a passive experience,

最初，視覺是一種被動體驗，

the simple act of letting light in,

單純讓光進入，

soon became far more active.

但很快就變得更加主動。

The nervous system began to evolve.

神經系統開始進化。

Sight turning to insight.

視力轉變成洞察力。

Seeing became understanding.

看見變成了理解。

Understanding led to actions.

理解導致行動。

And all these gave rise to intelligence.

這一切都促成了智能的誕生。

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

今天，我們不再滿足於 僅僅擁有自然賦予的視覺智能，

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

好奇心促使我們創造出 跟我們同樣有視覺智能的機器，

if not better.

甚至更好。

Nine years ago, on this stage,

九年前，在這個講臺，

I delivered an early progress report on computer vision,

我發表了一份有關 電腦視覺的早期進度報告，

a subfield of artificial intelligence.

這是人工智能的一個分支。 (Intelligence一字譯法以講者為準)

Three powerful forces converged for the first time.

三種強大力量第一次匯集。

Aa family of algorithms called neural networks.

稱為神經網路的一系列演算法，

Fast, specialized hardware called graphic processing units,

快速、專門化的硬體，稱為圖形處理器

or GPUs.

或 GPU，

And big data.

以及大數據。

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

就像我的實驗室花費多年時間整理

名為 ImageNet 的 1500 萬張圖像，

Together, they ushered in the age of modern AI.

這些力量共同開啟了 現代 AI 時代。

We've come a long way.

我們已有很大的進展，

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

當時，只在圖像上加入標籤 就是個重大突破。

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

但這些演算法的速度 和準確性迅速提高。

The annual ImageNet challenge, led by my lab,

由我實驗室主導的 年度 ImageNet 挑戰賽

gauged the performance of this progress.

評估了這項進展的表現。

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

在這張圖上，你會看到每年的進步

and milestone models.

和里程碑模型。

We went a step further

我們更進一步，

and created algorithms that can segment objects

創建了可以分割物件

or predict the dynamic relationships among them

或預測物件之間動態關係的演算法。

in these works done by my students and collaborators.

這些都是我的學生們 和合作者完成的工作。

And there's more.

還有更多。

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

請回想一下，我上次 向你展示的首個電腦視覺演算法，

that can describe a photo in human natural language.

它可以用人類自然語言描述照片。

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

那是我之前優秀的學生 安德烈·卡帕蒂跟我一起完成。

At that time, I pushed my luck and said,

當時我碰碰運氣跟他說：

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

「安德烈，我們可以用電腦 來做相反的事嗎？」

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

他說：「哈哈，這是不可能的。」

Well, as you can see from this post,

嗯，正如你從這篇文章中看到，

recently the impossible has become possible.

最近，這件不可能的事已成為可能。

That's thanks to a family of diffusion models

這要歸功於一組稱為擴散模型的技術，

that powers today's generative AI algorithm,

驅動了當今的生成式 AI 演算法，

which can take human-prompted sentences

可以將人類提示的句子

and turn them into photos and videos

變成全新創作的照片和影片。

of something that's entirely new.

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

在座許多人都看過 OpenAI 最近推出令人驚豔的 Sora 成果。

But even without the enormous number of GPUs,

但即使沒有大量 GPU，

my student and our collaborators

我的學生和合作夥伴

have developed a generative video model called Walt

也已開發了一個名為 Walt 的生成式影片模型，

months before Sora.

而且比 Sora 還早幾個月。

And you're seeing some of these results.

你們現在看到的是一些成果。

There is room for improvement.

還有改進空間。

I mean, look at that cat's eye

看看那隻貓的眼睛，

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

還有它在浪潮下卻不會被弄濕。

What a cat-astrophe.

真是場災難。

(Laughter)

（笑聲）

And if past is prologue,

若過去是序幕，

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

我們將從這些錯誤中學習， 創造我們所想像的未來。

And in this future,

而在未來，

we want AI to do everything it can for us,

我們希望 AI 能為我們 做一切它能做的事，

or to help us.

或協助我們。

For years I have been saying

多年來，我一直在說，

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

拍照並不等於看到和理解。

Today, I would like to add to that.

今天，我想再補充一點，

Simply seeing is not enough.

只是看到還不夠，

Seeing is for doing and learning.

看見是為了行動和學習。

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

當我們在三維空間和時間中行動時，

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

我們會學習，並學會去看和做得更好。

Nature has created this virtuous cycle of seeing and doing

大自然創造了由「空間智能」 驅動的良性循環，

powered by “spatial intelligence.”

觀察和行動。

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

為了向你說明你的空間智能 如何不斷地運作，

look at this picture.

請看這張圖片。

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

如果你覺得自己想採取行動，請舉手。

(Laughter)

（笑聲）

In the last split of a second,

在剛剛極短的時間裏，

your brain looked at the geometry of this glass,

你的大腦觀察了 這個玻璃杯的幾何形狀，

its place in 3D space,

在三維空間中的位置，

its relationship with the table, the cat

它與桌子、貓

and everything else.

和其他物品的關係。

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

你可以預測接下來會發生什麼事。

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

行動的衝動是所有 具有空間智能的生物的天性，

which links perception with action.

它將感知與行動聯繫起來。

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

如果我們想讓 AI 超越目前的能力，

we want more than AI that can see and talk.

我們需要的不僅僅是 能看見和交談的 AI，

We want AI that can do.

我們需要的是能夠做到 實際行動的 AI。

Indeed, we're making exciting progress.

的確，我們正取得令人興奮的進展。

The recent milestones in spatial intelligence

空間智能最近的里程碑

is teaching computers to see, learn, do

是教電腦去看見、學習、行動，

and learn to see and do better.

並且學會更好地看見和行動。

This is not easy.

這並不容易。

It took nature millions of years to evolve spatial intelligence,

大自然花了數百萬年的時間 才進化出空間智能，

which depends on the eye taking light,

這取決於眼睛接收光線，

project 2D images on the retina

將二維影像投射到視網膜上，

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

大腦再將這些數據轉換為三維信息。

Only recently, a group of researchers from Google

就最近，一群來自谷歌的研究人員

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

開發出一種演算法，

可以將一組照片轉換為三維空間，

and translate that into 3D space,

like the examples we're showing here.

就像我們在這裡展示的例子。

My student and our collaborators have taken a step further

我的學生和我們的合作夥伴更進一步，

and created an algorithm that takes one input image

創建了一個演算法，

可以從一張輸入圖片生成三維形狀。

and turn that into 3D shape.

Here are more examples.

以下是更多的例子。

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

回想一下，我們剛提到一個電腦程式，

可以將人類句子轉換為影片，

and turn it into videos.

A group of researchers in University of Michigan

密歇根大學的一群研究人員 已經找到了一種方法，

have figured out a way to translate that line of sentence

將這些句子轉化成三維的 房間佈局，如影片所示。

into 3D room layout, like shown here.

And my colleagues at Stanford and their students

我在史丹福的同事和他們的學生

have developed an algorithm that takes one image

開發了一種演算法，

可以從一張圖像生成無限可能的空間，

and generates infinitely plausible spaces

for viewers to explore.

供觀眾探索。

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

這些是未來可能性的 最初萌芽跡象的原型。

One in which the human race can take our entire world

在這個未來，人類可以將整個世界

and translate it into digital forms

轉化為數位形式，

and model the richness and nuances.

並模擬其豐富性和細微差別。

What nature did to us implicitly in our individual minds,

大自然在我們的個體大腦中 隱含完成的事情，

spatial intelligence technology can hope to do

空間智能技術可以為 我們的集體意識做到。

for our collective consciousness.

As the progress of spatial intelligence accelerates,

隨著空間智能的進展加速，

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

一個新的良性循環時代 正在我們眼前展開。

This back and forth is catalyzing robotic learning,

這一來一回促進了機器學習，

a key component for any embodied intelligence system

這是任何實體智能系統的 關鍵組成部份，

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

因其需要理解三維世界並與之互動。

A decade ago,

十年前，

ImageNet from my lab

我實驗室的 ImageNet

enabled a database of millions of high-quality photos

啟用了包含數百萬張 高品質照片的資料庫，

to help train computers to see.

以幫助訓練電腦去看。

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

今天，我們同樣用行為和行動

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

訓練電腦和機器人在三維世界中行動。

But instead of collecting static images,

我們不是收集靜態影像，

we develop simulation environments powered by 3D spatial models

而是開發由三維空間模型 驅動的模擬環境，

so that the computers can have infinite varieties of possibilities

這樣電腦就可以有無限種可能性

to learn to act.

來學習行動。

And you're just seeing a small number of examples

而你只是看到了少量例子，

to teach our robots

在教導我們的機器人，

in a project led by my lab called Behavior.

來自我實驗室主導 一個叫「行為」的專案。

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

我們也在機器人語言智能方面 取得令人興奮的進展。

Using large language model-based input,

使用大型語言模型輸入，

my students and our collaborators are among the first teams

我的學生和我們的合作夥伴 是眾多團隊中第一批

that can show a robotic arm performing a variety of tasks

能夠展示機械臂

根據口頭指令，執行各種任務，

based on verbal instructions,

like opening this drawer or unplugging a charged phone.

如打開這個抽屜 或拔下充好電的手機，

Or making sandwiches, using bread, lettuce, tomatoes

或者製作三明治， 用麵包，生菜，蕃茄，

and even putting a napkin for the user.

甚至為用戶放張餐巾紙。

Typically I would like a little more for my sandwich,

通常我想我的三明治要更多配料，

but this is a good start.

但這是個好開始。

(Laughter)

（笑聲）

In that primordial ocean, in our ancient times,

在那遠古時代的原始海洋中，

the ability to see and perceive one's environment

觀察和感知環境的能力

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

引發了寒武紀大爆發， 各種生命形式的相互作用。

Today, that light is reaching the digital minds.

如今，這種光芒正照亮數位思維。

Spatial intelligence is allowing machines

空間智能不僅讓機器能彼此互動，

to interact not only with one another,

but with humans, and with 3D worlds,

還可以與人類以及三維世界互動，

real or virtual.

無論是真實還是虛擬世界。

And as that future is taking shape,

隨著未來漸漸成形，

it will have a profound impact to many lives.

它將對許多人的生活產生深遠的影響。

Let's take health care as an example.

我們以醫療保健為例。

For the past decade,

過去十年來，

my lab has been taking some of the first steps

我的實驗室已經邁出了第一步，

in applying AI to tackle challenges that impact patient outcome

應用 AI 來應對影響患者治療結果

and medical staff burnout.

和醫護人員過勞的挑戰。

Together with our collaborators from Stanford School of Medicine

我們與史丹福醫學院的合作夥伴

and partnering hospitals,

和合作醫院

we're piloting smart sensors

一起試行智能感測器，

that can detect clinicians going into patient rooms

檢測醫護人員是否 在未正確洗手的情況下進入病房。

without properly washing their hands.

Or keep track of surgical instruments.

或追蹤手術器械的位置。

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

或者，當患者面臨跌倒等 風險時，提醒護理團隊。

such as falling.

We consider these techniques a form of ambient intelligence,

我們認為這些技術是 環境智能的一種形式，

like extra pairs of eyes that do make a difference.

就像多一雙眼睛，確實能帶來改變。

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

但我希望為我們的患者、臨床醫生 和護理人員提供更多互動幫助，

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

他們也迫切需要額外的協助。

Imagine an autonomous robot transporting medical supplies

想像一個自主機器人運送醫療用品，

while caretakers focus on our patients

與此同時，護理人員專注於照顧患者，

or augmented reality, guiding surgeons to do safer, faster

或擴增實境指導外科醫生 進行更安全、更快速、

and less invasive operations.

更少侵入性的手術。

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

或者想像嚴重癱瘓的患者 用他們的思想控制機器人。

That's right, brainwaves, to perform everyday tasks

沒錯，用腦電波來執行

你我都認為理所當然的日常任務。

that you and I take for granted.

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

你可以在我實驗室最近進行的 試點研究中看到這一未來。

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

在這部影片中， 機械臂正在烹調日本壽喜燒，

controlled only by the brain electrical signal,

它僅由大腦電信號控制，

non-invasively collected through an EEG cap.

信號是由非侵入性 腦電圖（EEG）帽收集的。

(Applause)

（掌聲）

Thank you.

謝謝

The emergence of vision half a billion years ago

五億年前，視覺的出現

turned a world of darkness upside down.

顛覆了黑暗世界。

It set off the most profound evolutionary process:

它引發了最深遠的進化過程：

the development of intelligence in the animal world.

動物世界智能的發展。

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

AI 在過去十年中取得的 驚人進步同樣令人驚訝。

But I believe the full potential of this digital Cambrian explosion

但我相信，這種數位寒武紀爆發的 全部潛力不會完全實現，

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

直到我們為電腦和機器人 提供空間智能，

with spatial intelligence,

just like what nature did to all of us.

就像大自然給予我們所有人的一樣。

這是一個令人興奮的時機，

It’s an exciting time to teach our digital companion

教導我們的數位伴侶學會推理，

to learn to reason

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

並與我們這個稱為家的 美麗三維空間互動，

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

也創造許多我們可以探索的新世界。

To realize this future won't be easy.

實現這個未來並不容易。

It requires all of us to take thoughtful steps

這需要我們所有人深思熟慮的努力，

and develop technologies that always put humans in the center.

開發始終以人類為中心的技術。

But if we do this right,

但如果我們做得對，

the computers and robots powered by spatial intelligence

由空間智能驅動的電腦和機器人

will not only be useful tools

不僅會成為有用的工具，

but also trusted partners

也會是值得信賴的合作夥伴，

to enhance and augment our productivity and humanity

可以增強、提升我們的生產力和人性，

while respecting our individual dignity

同時尊重我們的個體尊嚴，

and lifting our collective prosperity.

並提升我們的集體繁榮。

What excites me the most in the future

未來最令我興奮的事

is a future in which that AI grows more perceptive,

是 AI 變得更加敏銳、

insightful and spatially aware,

有洞察力和空間感知，

and they join us on our quest

並與我們一起

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

努力追求更好的方法， 以創造更美好的世界。

Thank you.

謝謝。

(Applause)

（掌聲）