Jeff Dean: AI isn't as smart as you think -- but it could be | TED

譯者: Yiann C. 審譯者: Helen Chang

Hi, I'm Jeff.

嗨，我是傑夫

I lead AI Research and Health at Google.

我在 Google 帶領人工智慧研究與健康

I joined Google more than 20 years ago,

我在 20 多年前加入 Google

when we were all wedged into a tiny office space,

那時我們還擠在一個小辦公室

above what's now a T-Mobile store in downtown Palo Alto.

在帕羅奧圖市中心 現在一家 T-Mobile 樓上

I've seen a lot of computing transformations in that time,

我見證了很多當時電腦技術的轉變

and in the last decade, we've seen AI be able to do tremendous things.

過去十年，我們也看到 人工智慧做到了許多厲害的事

But we're still doing it all wrong in many ways.

但我們還是在許多方面都做錯了

That's what I want to talk to you about today.

這就是我今天想要和你們說的

But first, let's talk about what AI can do.

不過，我們先來說說 人工智慧能做些什麼

So in the last decade, we've seen tremendous progress

過去十年，人工智慧有很大的進展

in how AI can help computers see, understand language,

能幫助電腦更懂得去看、理解語言

understand speech better than ever before.

聽懂人們說的話

Things that we couldn't do before, now we can do.

以前我們做不到的事，現在能做到了

If you think about computer vision alone,

單就電腦視覺來說

just in the last 10 years,

光最近十年

computers have effectively developed the ability to see;

電腦已經有效地發展出了視覺

10 years ago, they couldn't see, now they can see.

十年前它們看不見，現在可以了

You can imagine this has had a transformative effect

這對於我們能如何應用電腦 有轉變性的影響

on what we can do with computers.

我們來看看幾個 這些功能所帶來的應用

So let's look at a couple of the great applications

enabled by these capabilities.

我們能運用機器學習 更準確地預測洪水

We can better predict flooding, keep everyone safe,

確保大家的安全

using machine learning.

我們可以翻譯超過一百種語言 讓大家更無礙地溝通

We can translate over 100 languages so we all can communicate better,

也能更準確地預測和診斷疾病

and better predict and diagnose disease,

where everyone gets the treatment that they need.

讓大家獲得所需的治療

So let's look at two key components

我們來看看兩個關鍵的因素

that underlie the progress in AI systems today.

它們促成了現今人工智慧的發展

The first is neural networks,

第一個是類神經網路

a breakthrough approach to solving some of these difficult problems

是個突破性的方法 去解決許多困難的問題

that has really shone in the last 15 years.

最近 15 年廣受運用

But they're not a new idea.

但它們不是新的發想

And the second is computational power.

第二個是計算能力

It actually takes a lot of computational power

事實上，讓類神經網路 發揮它真正的潛力

to make neural networks able to really sing,

需要很多計算能力

and in the last 15 years, we’ve been able to halve that,

過去 15 年 我們成功減半所需的計算能力

and that's partly what's enabled all this progress.

這是造就了目前的發展 部分的原因

But at the same time, I think we're doing several things wrong,

但同時，我認為我們做錯了幾件事

and that's what I want to talk to you about

這是我在後半段想跟大家說的

at the end of the talk.

First, a bit of a history lesson.

先來上一下歷史課

So for decades,

幾十年以來

almost since the very beginning of computing,

幾乎從計算科學的一開始

people have wanted to be able to build computers

大家就想建造出看得見、懂得語言、 聽得懂人們說話的電腦

that could see, understand language, understand speech.

The earliest approaches to this, generally,

最早的方法

people were trying to hand-code all the algorithms

基本上是將所需的演算法 全部人工編碼

that you need to accomplish those difficult tasks,

來完成那些困難的任務

and it just turned out to not work very well.

結果這個方法不怎麼好用

But in the last 15 years, a single approach

但最近 15 年

出乎意料，有一個方法 同時促進了所有方面的發展

unexpectedly advanced all these different problem spaces all at once:

neural networks.

類神經網路

So neural networks are not a new idea.

類神經網路不是全新的概念

They're kind of loosely based

是基於一些真的神經系統的特性

on some of the properties that are in real neural systems.

And many of the ideas behind neural networks

很多關於類神經網路的想法

have been around since the 1960s and 70s.

早在 1960、70 年代就出現了

A neural network is what it sounds like,

類神經網路就如其名

a series of interconnected artificial neurons

是一連串互相連結的人工神經元

that loosely emulate the properties of your real neurons.

大致上模仿人類神經元的特性

An individual neuron in one of these systems

系統裡每個神經元

has a set of inputs,

有一組輸入值

each with an associated weight,

它們有各自的權重

and the output of a neuron

神經元的輸出

is a function of those inputs multiplied by those weights.

則是將這些輸入值乘以各自的權重 所組成的函數

So pretty simple,

很簡單

and lots and lots of these work together to learn complicated things.

透過很多這樣的神經元一起運作 來學習複雜的事情

So how do we actually learn in a neural network?

類神經網路到底如何學習的呢？

It turns out the learning process

其實學習的過程

consists of repeatedly making tiny little adjustments

是不斷地微調權重的分配

to the weight values,

strengthening the influence of some things,

透過加強某些輸入對輸出的影響

weakening the influence of others.

和減弱其他輸入的權重

By driving the overall system towards desired behaviors,

朝著我們想要的結果 來調整整體系統

these systems can be trained to do really complicated things,

我們可以訓練它們 達成非常複雜的任務

like translate from one language to another,

像是翻譯語言

detect what kind of objects are in a photo,

偵測相片裡的物體

all kinds of complicated things.

各種複雜的事

I first got interested in neural networks

我對類神經網路感興趣的契機

when I took a class on them as an undergraduate in 1990.

是 1990 年在大學修的一門課

At that time,

當時

neural networks showed impressive results on tiny problems,

類神經網路在小問題上有顯著的表現

but they really couldn't scale to do real-world important tasks.

但無法擴展到實際生活中 重要的任務上

But I was super excited.

但我超級興奮

(Laughter)

（笑聲）

I felt maybe we just needed more compute power.

我覺得我們只是需要更多的計算能力

And the University of Minnesota had a 32-processor machine.

明尼蘇達大學有台機器 有 32 個處理器

I thought, "With more compute power,

我就想，有了更強大的計算能力

boy, we could really make neural networks really sing."

天啊，我們就可以 發揮類神經網路的極致

So I decided to do a senior thesis on parallel training of neural networks,

所以我決定以類神經網路的平行訓練 為畢業論文的主題

the idea of using processors in a computer or in a computer system

概念是在一台電腦或一個系統中 運用多個處理器

to all work toward the same task,

一起完成同一項任務

that of training neural networks.

來訓練類神經網路

32 processors, wow,

32 個處理器，哇

we’ve got to be able to do great things with this.

我們一定可以達成很棒的事

But I was wrong.

但我錯了

Turns out we needed about a million times as much computational power

結果發現我們需要 1990 年的百萬倍的計算能力

as we had in 1990

before we could actually get neural networks to do impressive things.

才能讓類神經網路做厲害的事

But starting around 2005,

不過 2005 年左右開始

thanks to the computing progress of Moore's law,

多虧摩爾定律

we actually started to have that much computing power,

我們開始有足夠的電腦效能

and researchers in a few universities around the world started to see success

世界各地幾所大學的研究員開始發現

in using neural networks for a wide variety of different kinds of tasks.

類神經網路能成功做到 各式各樣不同的任務

I and a few others at Google heard about some of these successes,

我和其他幾位 Google 的同事 聽聞這些消息

and we decided to start a project to train very large neural networks.

就決定開始一個新專案 來訓練大規模的類神經網路

One system that we trained,

我們訓練了一個系統

we trained with 10 million randomly selected frames

用一千萬張 從 Youtube 影片裡隨機選出的畫面

from YouTube videos.

The system developed the capability

這個系統發展出

to recognize all kinds of different objects.

識別各種不同物體的能力

And it being YouTube, of course,

因為是 Youtube

it developed the ability to recognize cats.

這個系統發展出識別貓的能力

YouTube is full of cats.

Youtube 上充滿著貓咪

(Laughter)

（笑聲）

But what made that so remarkable

但驚人的是

is that the system was never told what a cat was.

我們從來沒有告訴這個系統 貓是什麼

So using just patterns in data,

所以只靠數據裡的規律

the system honed in on the concept of a cat all on its own.

系統自己整理出貓的概念

All of this occurred at the beginning of a decade-long string of successes,

這是十年間 一連串成功案例的開端

of using neural networks for a huge variety of tasks,

Google 和其他人成功運用類神經網路 在很多不同的事情上

at Google and elsewhere.

Many of the things you use every day,

很多每天用到的東西

things like better speech recognition for your phone,

像是手機上更準確的語音辨識

improved understanding of queries and documents

改良對查詢和文件的理解

for better search quality,

提升搜尋的品質

better understanding of geographic information to improve maps,

對地理資訊有更好的理解 來改良地圖

and so on.

等等

Around that time,

那時候

we also got excited about how we could build hardware that was better tailored

我們也很興奮地想知道

如何製造出 更適合類神經網路的計算的硬體

to the kinds of computations neural networks wanted to do.

類神經網路計算有兩個特性

Neural network computations have two special properties.

第一是它們對準確度有很大的包容性

The first is they're very tolerant of reduced precision.

Couple of significant digits, you don't need six or seven.

只要幾個有效數字，不需要六個或七個

And the second is that all the algorithms are generally composed

第二，基本上所有演算法

都是以矩陣和向量所組成的不同序列

of different sequences of matrix and vector operations.

So if you can build a computer

所以如果你做了一台電腦

that is really good at low-precision matrix and vector operations

有不錯的低精確度 矩陣和向量運算能力

but can't do much else,

但不太能做其他的

that's going to be great for neural-network computation,

這會很適合類神經網路運算

儘管不能拿它來做別的事

even though you can't use it for a lot of other things.

如果你做出這樣的東西 人們會找到它的用處

And if you build such things, people will find amazing uses for them.

這是我們做出來的第一版，TPU v1

This is the first one we built, TPU v1.

"TPU" stands for Tensor Processing Unit.

TPU 是張量處理器的縮寫

These have been used for many years behind every Google search,

它們已經被廣為應用好幾年

包括 Google 搜尋、翻譯

for translation,

in the DeepMind AlphaGo matches,

還有 DeepMind AlphaGo 圍棋比賽

so Lee Sedol and Ke Jie maybe didn't realize,

所以李世乭和柯潔大概不知道

but they were competing against racks of TPU cards.

他們是跟一整架的 TPU 卡在比賽

And we've built a bunch of subsequent versions of TPUs

我們接著做了好幾個新的版本

that are even better and more exciting.

更好更棒

But despite all these successes,

不過儘管有這麼多成功案例

I think we're still doing many things wrong,

我認為我們還是做錯了很多事情

and I'll tell you about three key things we're doing wrong,

讓我說明三件我們做錯的事

以及我們該怎麼修正

and how we'll fix them.

首先，目前大部分的類神經網路

The first is that most neural networks today

are trained to do one thing, and one thing only.

都是訓練來做一件事，只有一件事

You train it for a particular task that you might care deeply about,

訓練它來做某一件你特別關心的事

but it's a pretty heavyweight activity.

其實很麻煩的

You need to curate a data set,

必須準備一個資料庫

you need to decide what network architecture you'll use

必須決定針對這個問題 要用什麼網路架構

for this problem,

you need to initialize the weights with random values,

必須用隨機值來初始化權重分配

apply lots of computation to make adjustments to the weights.

再用各種運算來調整

And at the end, if you’re lucky, you end up with a model

幸運的話，最後 會得到一個很好的模型

that is really good at that task you care about.

能做好你在乎的那件事

But if you do this over and over,

但如果你不斷重複這些步驟

you end up with thousands of separate models,

最後會有上千個獨立的模型

each perhaps very capable,

可能每個都表現得很好

but separate for all the different tasks you care about.

但各自解決不同的問題

But think about how people learn.

想想人類是怎麼學習的

In the last year, many of us have picked up a bunch of new skills.

過去一年，相信很多人學了很多新技能

I've been honing my gardening skills,

我更精進了我的園藝技巧

experimenting with vertical hydroponic gardening.

試驗了垂直水耕栽培

To do that, I didn't need to relearn everything I already knew about plants.

在那之前，我不需要重新學習 對植物既有的知識

I was able to know how to put a plant in a hole,

我已經知道如何把植物放進洞裡

how to pour water, that plants need sun,

如何澆水、植物需要陽光

and leverage that in learning this new skill.

然後在這個基礎上學習新的技能

Computers can work the same way, but they don’t today.

電腦可以用同樣的方式學習， 但目前並非如此

If you train a neural network from scratch,

如果你從頭訓練一個類神經網路

it's effectively like forgetting your entire education

就像是每當你嘗試一件新的事

every time you try to do something new.

就得忘掉一生所受過的教育

很荒唐吧？

That’s crazy, right?

So instead, I think we can and should be training

所以我認為我們可以，也應該要

multitask models that can do thousands or millions of different tasks.

訓練能多工處理上千 或上百萬不同任務的模型

Each part of that model would specialize in different kinds of things.

模型裡各個部分專精於不同的事

And then, if we have a model that can do a thousand things,

如果我們有個模型可以做一千件事

and the thousand and first thing comes along,

然後第 1001 件事出現了

we can leverage the expertise we already have

我們可以利用和它相關的既有知識

in the related kinds of things

so that we can more quickly be able to do this new task,

更快地達成這個新任務

just like you, if you're confronted with some new problem,

就像你一樣，如果遇到了一個新問題

你會快速地想到 17 件已經知道的事

you quickly identify the 17 things you already know

that are going to be helpful in solving that problem.

來幫助你解決問題

Second problem is that most of our models today

第二個問題是，目前大部分的模型

deal with only a single modality of data --

只處理單一一種型態的資料

with images, or text or speech,

影像，或是文字，或是語音

but not all of these all at once.

不是同時處理全部

But think about how you go about the world.

但想想你是怎麼感受周遭的

You're continuously using all your senses

你不斷地用所有的感官

to learn from, react to,

去學習、去反應

figure out what actions you want to take in the world.

來決定你該做出什麼行為

這樣做很合理

Makes a lot more sense to do that,

and we can build models in the same way.

我們也可以用同樣的道理來建模型

We can build models that take in these different modalities of input data,

我們可以建立 接收不同資料型態的模型

文字、影像、語音

text, images, speech,

but then fuse them together,

然後把它們融合在一起

so that regardless of whether the model sees the word "leopard,"

所以不管模型是看到「豹」這個字

sees a video of a leopard or hears someone say the word "leopard,"

看到一部豹的影片 還是聽到某個人說「豹」這個字

the same response is triggered inside the model:

都會產生相同的反應

the concept of a leopard

就是豹的概念

can deal with different kinds of input data,

也可以處理其他不同種的資料

even nonhuman inputs, like genetic sequences,

甚至是非人為產出的資料 例如基因序列

3D clouds of points, as well as images, text and video.

3D 點雲，和影像、文字、影片

The third problem is that today's models are dense.

第三個問題是，現在的模型是密集的

There's a single model,

單一一個模型

the model is fully activated for every task,

完全啟動來處理每一個任務

for every example that we want to accomplish,

每一件我們想完成的事

whether that's a really simple or a really complicated thing.

無論是非常簡單還是非常複雜的事

This, too, is unlike how our own brains work.

這也不像我們腦袋運作的方式

Different parts of our brains are good at different things,

腦袋不同的部位負責不同的功能

and we're continuously calling upon the pieces of them

我們不斷運用各個 與眼前任務有關的部份

that are relevant for the task at hand.

For example, nervously watching a garbage truck

例如，當你緊張地看著垃圾車倒車

back up towards your car,

越來越靠近你的車

the part of your brain that thinks about Shakespearean sonnets

腦中負責思考 莎士比亞十四行詩的部位

is probably inactive.

大概是沒有活動的

(Laughter)

（笑聲）

AI models can work the same way.

人工智慧模型也可以這樣運作

Instead of a dense model,

與其用一個密集的模型

we can have one that is sparsely activated.

我們可以用一個零星式啟動的模型

So for particular different tasks, we call upon different parts of the model.

所以不同的任務 會啟動模型裡不同的部分

During training, the model can also learn which parts are good at which things,

訓練時，模型也可以學到 哪部分擅長哪些事

to continuously identify what parts it wants to call upon

來持續判斷它需要利用哪些部分

in order to accomplish a new task.

去完成一項新的任務

The advantage of this is we can have a very high-capacity model,

這個方法的優點是 模型的適應能力會很高

也非常的有效率

but it's very efficient,

因為我們只用各項任務所需的部分

because we're only calling upon the parts that we need

for any given task.

修正這三個問題之後，我認為

So fixing these three things, I think,

will lead to a more powerful AI system:

會做出更強大的人工智慧系統

instead of thousands of separate models,

與其用上千個獨立的模型

train a handful of general-purpose models

不如訓練一些通用於 上千或上百萬件事的模型

that can do thousands or millions of things.

Instead of dealing with single modalities,

與其只處理單一資料型態

deal with all modalities,

不如同時處理所有型態的資料

and be able to fuse them together.

然後把它們融合在一起

And instead of dense models, use sparse, high-capacity models,

與其用密集的模型 不如用零星、高適應力的模型

where we call upon the relevant bits as we need them.

只在需要的時候用所需的部分

We've been building a system that enables these kinds of approaches,

我們已經在建立運用這些方法的系統了

and we’ve been calling the system “Pathways.”

我們叫它「 Pathways 」

So the idea is this model will be able to do

概念是這個模型可以做 上千、上百萬不同的任務

thousands or millions of different tasks,

and then, we can incrementally add new tasks,

當我們逐漸增加新的任務

and it can deal with all modalities at once,

它可以同時處理所有資料型態

and then incrementally learn new tasks as needed

並且必要時，會漸漸學習新的任務

and call upon the relevant bits of the model

然後針對不同的內容

for different examples or tasks.

運用與其相關的部分

And we're pretty excited about this,

我們很興奮

相信這會帶我們邁向 理想中的人工智慧系統

we think this is going to be a step forward

in how we build AI systems.

But I also wanted to touch on responsible AI.

不過我也想提提 負責任的人工智慧

We clearly need to make sure that this vision of powerful AI systems

很顯然地，我們需要保證 這個強大的人工智慧系統遠景

benefits everyone.

能讓每個人都受惠

These kinds of models raise important new questions

這些模型衍生出許多重要的問題

about how do we build them with fairness,

如何在建立的過程中 考量對所有使用者的公平性

interpretability, privacy and security,

可解釋性、隱私與安全

for all users in mind.

For example, if we're going to train these models

例如，如果我們訓練這些模型

on thousands or millions of tasks,

來做上千、上百萬種任務

we'll need to be able to train them on large amounts of data.

我們得用非常大量的資料來訓練它們

And we need to make sure that data is thoughtfully collected

我們必須確保這些資料 經過謹慎的蒐集

and is representative of different communities and situations

並且對世界上不同的群體和情況

all around the world.

都具有代表性

And data concerns are only one aspect of responsible AI.

對資料的顧慮只是 負責任的人工智慧的一部分

We have a lot of work to do here.

我們有很多工作要做

So in 2018, Google published this set of AI principles

2018 年，Google 發表了這些 人工智慧原則

by which we think about developing these kinds of technology.

適用於開發這類的科技

And these have helped guide us in how we do research in this space,

這些原則引導我們 如何在這個領域做研究

how we use AI in our products.

如用在產品中應用人工智慧

And I think it's a really helpful and important framing

我認為這是非常有用，而且重要的架構

for how to think about these deep and complex questions

來幫助我們思考這些 關於如何在社會中應用人工智慧

about how we should be using AI in society.

深入又複雜的問題

We continue to update these as we learn more.

隨著我們知道更多 我們會持續更新這些原則

Many of these kinds of principles are active areas of research --

很多是正在研究中的領域

super important area.

超級重要的領域

Moving from single-purpose systems that kind of recognize patterns in data

從辨識資料中的規律的 單一目的系統

to these kinds of general-purpose intelligent systems

到對世界有更深的理解的 多方通用智慧系統

that have a deeper understanding of the world

will really enable us to tackle

可以真正讓我們解決

some of the greatest problems humanity faces.

人類長久以來面臨的大問題

For example,

例如，我們可以診斷出更多疾病

we’ll be able to diagnose more disease;

we'll be able to engineer better medicines

我們可以製造出更好的藥

by infusing these models with knowledge of chemistry and physics;

透過注入化學和物理的知識 到這些模型裡

we'll be able to advance educational systems

我們可以提供更多個人化家教

by providing more individualized tutoring

來提升教育系統

to help people learn in new and better ways;

讓大家用全新更好的方法來學習

we’ll be able to tackle really complicated issues,

我們能處理非常複雜的問題 像是氣候變遷

like climate change,

and perhaps engineering of clean energy solutions.

或者可能可以實現 乾淨能源的解決方案

So really, all of these kinds of systems

所以，這些系統

are going to be requiring the multidisciplinary expertise

將會需要全世界所有人

of people all over the world.

跨領域的專業

So connecting AI with whatever field you are in,

所以為了有所進展，不論你在什麼領域

in order to make progress.

應用人工智慧吧

So I've seen a lot of advances in computing,

我見證了許多電腦運算科學上的進展

and how computing, over the past decades,

在過去的幾十年

has really helped millions of people better understand the world around them.

運算科學幫助了上百萬的人 理解周遭的世界

And AI today has the potential to help billions of people.

如今，人工智慧有能夠 幫助數十億人類的潛力

We truly live in exciting times.

我們真的身處於一個令人興奮的時代

Thank you.

謝謝

(Applause)

（掌聲）

Chris Anderson: Thank you so much.

克里斯·安德森：非常謝謝你

I want to follow up on a couple things.

我想多問幾個問題

This is what I heard.

這是我聽到的

Most people's traditional picture of AI

大部分人對傳統人工智慧的理解

is that computers recognize a pattern of information,

是電腦辨識資訊中的規律

and with a bit of machine learning,

利用一些機器學習

they can get really good at that, better than humans.

電腦會表現得比人類還好

What you're saying is those patterns

你說那些規律

are no longer the atoms that AI is working with,

不再是人工智慧處理的元素

that it's much richer-layered concepts

而是一種多層的概念

that can include all manners of types of things

包含了，像是各式各樣

that go to make up a leopard, for example.

和豹相關的東西

So what could that lead to?

所以，這會往什麼方向發展？

Give me an example of when that AI is working,

舉個例子，當這樣的人工智慧實現時

what do you picture happening in the world

未來的五或十年裡 你期待什麼樣的事情會發生？

in the next five or 10 years that excites you?

Jeff Dean: I think the grand challenge in AI

傑夫：我認為人工智慧很大的一個挑戰

is how do you generalize from a set of tasks

是如何概括一組已經知道如何做的事

you already know how to do

to new tasks,

到新的任務上

as easily and effortlessly as possible.

而且要盡量輕鬆容易

And the current approach of training separate models for everything

目前，為所有事 訓練獨立模型的方法

means you need lots of data about that particular problem,

意味著你需要有 針對每個問題的大量數據

because you're effectively trying to learn everything

因為你試著從零來學習

about the world and that problem, from nothing.

關於這個世界和那個問題的所有事

But if you can build these systems

但如果你可以建造一些系統

that already are infused with how to do thousands and millions of tasks,

它們已經知道如何做 上千或上百萬個任務

then you can effectively teach them to do a new thing

那你就可以有效地用相對少的例子 來教它們新的事

with relatively few examples.

So I think that's the real hope,

所以我認為這是我們的希望

that you could then have a system where you just give it five examples

你只需要提供這個系統 關於某件你在乎的事的

of something you care about,

五個例子

and it learns to do that new task.

它就會學習怎麼做新的任務

CA: You can do a form of self-supervised learning

克里斯：你可以提供極少的原料

that is based on remarkably little seeding.

來做自監督式的學習

JD: Yeah, as opposed to needing 10,000 or 100,000 examples

傑夫：對，不需要一萬或十萬個例子

to figure everything in the world out.

去理解世界上所有的事

克里斯：難道不會有預期之外

CA: Aren't there kind of terrifying unintended consequences

令人害怕的後果嗎？

possible, from that?

JD: I think it depends on how you apply these systems.

傑夫：這就看你如何應用這些系統

It's very clear that AI can be a powerful system for good,

人工智慧用在好的地方 很顯然會是很厲害的系統

or if you apply it in ways that are not so great,

如果用在不好的地方

it can be a negative consequence.

會帶來負面的影響

So I think that's why it's important to have a set of principles

所以我認為這就是為什麼原則很重要

by which you look at potential uses of AI

按照原則去衡量人工智慧可以用在哪

and really are careful and thoughtful about how you consider applications.

非常小心、周全地思考如何應用

CA: One of the things people worry most about

克里斯：大家擔心的事之一是

is that, if AI is so good at learning from the world as it is,

假如人工智慧這麼擅長 學習社會現在的樣貌

it's going to carry forward into the future

它將會把一些其實不對的事

aspects of the world as it is that actually aren't right, right now.

原封不動地延續到未來

And there's obviously been a huge controversy about that

最近 Google 在這個議題上

recently at Google.

有個很大的爭議

Some of those principles of AI development,

那些你所說的人工智慧開發原則

you've been challenged that you're not actually holding to them.

有人質疑你們自己並沒有遵守

Not really interested to hear about comments on a specific case,

我並不是想聽對於特定事件的回應

but ... are you really committed?

但是...你們真的有決心嗎？

How do we know that you are committed to these principles?

我們怎麼知道 你們真的承諾遵守這些原則？

這只是個宣傳嗎？還是 真的有實踐在你們每天的工作中？

Is that just PR, or is that real, at the heart of your day-to-day?

JD: No, that is absolutely real.

傑夫：不是宣傳，是真的

Like, we have literally hundreds of people

我們真的有上百位研究員

working on many of these related research issues,

在研究這些相關議題

because many of those things are research topics

因為這些議題本身就是個研究主題

in their own right.

如何從現實世界取得資料

How do you take data from the real world,

that is the world as it is, not as we would like it to be,

但它們反映出現在社會真實的樣貌 而不是我們想要的模樣

and how do you then use that to train a machine-learning model

然後如何用這些資料 來訓練一個機器學習模型

and adapt the data bit of the scene

並針對情況去調整資料

or augment the data with additional data

或是加上額外的資料來擴增

so that it can better reflect the values we want the system to have,

讓模型更能反映出 我們想要系統擁有的價值

not the values that it sees in the world?

而不是它所看到社會目前的樣子

CA: But you work for Google,

克里斯： 但是你為 Google 工作

Google is funding the research.

Google 為研究提供資金

How do we know that the main values that this AI will build

我們怎麼知道這個人工智慧建立的價值

are for the world,

是為了這個世界

and not, for example, to maximize the profitability of an ad model?

例如，不是為了 一個下廣告的模型最大化利潤

When you know everything there is to know about human attention,

當你知道所有關於 人類注意力的運作模式

你也會知道很多

you're going to know so much

about the little wriggly, weird, dark parts of us.

我們見不得人、奇怪、黑暗的部分

你的團隊裡，有沒有規定相關的界線

In your group, are there rules about how you hold off,

church-state wall between a sort of commercial push,

防止商業利益干擾，明確區隔

"You must do it for this purpose,"

「為了這個使命，一定要這麼做」

so that you can inspire your engineers and so forth,

所以你可以激勵工程師等等

to do this for the world, for all of us.

為了整個社會、整體人類而做？

JD: Yeah, our research group does collaborate

傑夫：有，我們的研究團隊

和 Google 各個不同的團隊合作

with a number of groups across Google,

including the Ads group, the Search group, the Maps group,

包括廣告組、搜尋組、地圖組

所以我們內部有互相合作 但也有很多基本的研究

so we do have some collaboration, but also a lot of basic research

是對外公開的

that we publish openly.

去年，我們在不同主題上 發表了一千多篇論文

We've published more than 1,000 papers last year

in different topics, including the ones you discussed,

包括你提到的

關於機器學習模型的 公平性、可解釋性

about fairness, interpretability of the machine-learning models,

things that are super important,

非常重要的主題

and we need to advance the state of the art in this

我們須要提升當前的最佳技術

in order to continue to make progress

才能確保

在模型的安全性和開發責任方面 持續有所進展

to make sure these models are developed safely and responsibly.

CA: It feels like we're at a time when people are concerned

克里斯：感覺現在大家 都很擔心科技巨頭的力量

about the power of the big tech companies,

and it's almost, if there was ever a moment to really show the world

如果有機會展現給世界看

that this is being done to make a better future,

說為了更美好的未來，這已經做到了

that is actually key to Google's future,

對 Google 的未來會很關鍵

as well as all of ours.

對我們的未來也是

JD: Indeed.

傑夫：沒錯

CA: It's very good to hear you come and say that, Jeff.

克里斯：非常高興聽到你這麼說，傑夫

Thank you so much for coming here to TED.

非常謝謝你來到 TED

JD: Thank you.

傑夫：謝謝

(Applause)

（掌聲）