The wonderful and terrifying implications of computers that can learn | Jeremy Howard

譯者: Sharon Loh 審譯者: Yamei Huang

It used to be that if you wanted to get a computer to do something new,

過去如果想用電腦來作點新東西，

you would have to program it.

你需要設計程式。

Now, programming, for those of you here that haven't done it yourself,

而現在， 你們可能沒做過程式設計這件事，

requires laying out in excruciating detail

它需要規劃相當詳細的細節

every single step that you want the computer to do

那些你想讓電腦執行的每一個步驟

in order to achieve your goal.

以達到你的目的。

Now, if you want to do something that you don't know how to do yourself,

如果你沒有概念要怎麼做的話

then this is going to be a great challenge.

那會是個很大的挑戰。

So this was the challenge faced by this man, Arthur Samuel.

亞瑟·撒姆爾也曾面對這種挑戰。

In 1956, he wanted to get this computer

他在 1956 年便想到用這台電腦

to be able to beat him at checkers.

能夠在西洋跳棋棋賽打敗他。

How can you write a program,

要如何設計這樣的程式？

lay out in excruciating detail, how to be better than you at checkers?

把細節通通寫出來， 如何讓電腦比你還會下棋？

So he came up with an idea:

於是他想出了一個點子：

he had the computer play against itself thousands of times

他讓電腦與電腦本身對弈數千次

and learn how to play checkers.

以學習如何玩西洋棋。

And indeed it worked, and in fact, by 1962,

然而，在 1962 年做到了，

this computer had beaten the Connecticut state champion.

電腦打敗了康乃狄克州的冠軍。

So Arthur Samuel was the father of machine learning,

於是亞瑟·撒姆爾 成為了機器學習之父，

and I have a great debt to him,

我尊敬他，

because I am a machine learning practitioner.

因為我也是個機器學習實踐者，

I was the president of Kaggle,

我曾是 Kaggle 的會長，

a community of over 200,000 machine learning practictioners.

Kaggle 是個超過 20 萬人的 機器學習實踐者的社群。

Kaggle puts up competitions

Kaggle 設立了一些比賽

to try and get them to solve previously unsolved problems,

讓他們參與解決 過去無法解決的問題，

and it's been successful hundreds of times.

而有上百的成功個案。

So from this vantage point, I was able to find out

從這有利的環境中， 我發現

a lot about what machine learning can do in the past, can do today,

很多機器學習在 過去和現在可以做到的事情，

and what it could do in the future.

還有未來可以做到的事。

Perhaps the first big success of machine learning commercially was Google.

第一個機器學習的 商業成功案例是谷歌。

Google showed that it is possible to find information

谷歌展示找尋資料的方法

by using a computer algorithm,

是使用計算機演算法，

and this algorithm is based on machine learning.

而這演算法是以機器學習為基礎。

Since that time, there have been many commercial successes of machine learning.

自此，機器學習 有很多的商業成功例子，

Companies like Amazon and Netflix

譬如亞馬遜和奈飛公司

use machine learning to suggest products that you might like to buy,

用機器學習會向你推薦 你可能想買的商品，

movies that you might like to watch.

你可能想看的影片。

Sometimes, it's almost creepy.

有時，你可能會很訝異。

Companies like LinkedIn and Facebook

像領英和臉書等公司

sometimes will tell you about who your friends might be

有些時候會告訴你 誰會是你的朋友

and you have no idea how it did it,

而你根本不知道他們是如何做到的，

and this is because it's using the power of machine learning.

因為他們用了 機器學習這強大的功能。

These are algorithms that have learned how to do this from data

演算法從資料去學習這類事情

rather than being programmed by hand.

不需要動手去編寫程式。

This is also how IBM was successful

這也是 IBM 過去能成功的原因

in getting Watson to beat the two world champions at "Jeopardy,"

讓超級電腦「華生」在「危機遊戲」中 打敗兩屆世界冠軍。

answering incredibly subtle and complex questions like this one.

回答一些細碎和複雜的問題，像是

["The ancient 'Lion of Nimrud' went missing from this city's national museum in 2003 (along with a lot of other stuff)"]

「2003年，古獅像在這城市的 國家博物館消失了（連同其他物品）」

This is also why we are now able to see the first self-driving cars.

這也是我們現在能看到第一部 自行駕駛汽車的原因。

If you want to be able to tell the difference between, say,

如果你能說出不同點，像是

a tree and a pedestrian, well, that's pretty important.

一棵樹和一條行人道， 那顯得非常重要。

We don't know how to write those programs by hand,

我們不知道如何設計這樣的程式，

but with machine learning, this is now possible.

不過通過機器，這就成為可能。

And in fact, this car has driven over a million miles

事實上， 這部汽車已經行駛一百萬英哩

without any accidents on regular roads.

在正常路面沒有發生事故。

So we now know that computers can learn,

我們現在都知道電腦能夠學習，

and computers can learn to do things

學習做一些

that we actually sometimes don't know how to do ourselves,

有時我們自己也不知道怎麼做的事，

or maybe can do them better than us.

還可能比我們做得更好。

One of the most amazing examples I've seen of machine learning

其中一個機器學習的經典例子

happened on a project that I ran at Kaggle

是我在 Kaggle 所做的一個專案

where a team run by a guy called Geoffrey Hinton

由傑佛里·辛頓帶領的團隊

from the University of Toronto

他是多倫多大學的教授

won a competition for automatic drug discovery.

他們贏了新藥研發的比賽。

Now, what was extraordinary here is not just that they beat

他們出色地方 不只打敗了

all of the algorithms developed by Merck or the international academic community,

默克藥廠或國際學術社群 所研發的演算法，

but nobody on the team had any background in chemistry or biology or life sciences,

他們的團隊沒有化學 生物或生命科學的背景，

and they did it in two weeks.

而且只花了兩個星期就完成。

How did they do this?

他們怎麼做到的？

They used an extraordinary algorithm called deep learning.

他們用了一個很出色的演算法 叫做「深度學習」。

So important was this that in fact the success was covered

這是重要且成功的事情

in The New York Times in a front page article a few weeks later.

在數星期後 被刊登在紐約時報頭版。

This is Geoffrey Hinton here on the left-hand side.

左手邊那位是傑佛里·辛頓。

Deep learning is an algorithm inspired by how the human brain works,

深度學習是一種 受到人類大腦啟發的演算法，

and as a result it's an algorithm

它是一種演算法

which has no theoretical limitations on what it can do.

做法不受理論限制的演算法。

The more data you give it and the more computation time you give it,

你給它越多的資料和 運算時間，

the better it gets.

會得到更好的結果。

The New York Times also showed in this article

紐約時報的文章裡

another extraordinary result of deep learning

也介紹到深度學習的非凡成就

which I'm going to show you now.

我現在要展示給你們看。

It shows that computers can listen and understand.

它顯示電腦能聽懂和理解資料的能力。

(Video) Richard Rashid: Now, the last step

（影片）理察·拉希德: 現在，最後一步是

that I want to be able to take in this process

我能夠理解這個程序

is to actually speak to you in Chinese.

我能夠跟你說中文。

Now the key thing there is,

現在關鍵的是，

we've been able to take a large amount of information from many Chinese speakers

我們從很多講中文的人士中 收集大量的資訊

and produce a text-to-speech system

然後產生文字轉化語言的系統

that takes Chinese text and converts it into Chinese language,

將中文文字轉化成中文語言，

and then we've taken an hour or so of my own voice

然後錄一個小時我自己的聲音

and we've used that to modulate

我們使用它去調變

the standard text-to-speech system so that it would sound like me.

使標準文字轉化語音系統的聲音 聽起來像我的聲音。

Again, the result's not perfect.

再一次，雖然結果沒有很完美，

There are in fact quite a few errors.

裡面還有一些錯誤。

(In Chinese)

（中文）

(Applause)

（掌聲）

There's much work to be done in this area.

在這個領域還有很多工作要做。

(In Chinese)

（中文）

(Applause)

（掌聲）

Jeremy Howard: Well, that was at a machine learning conference in China.

傑里米·霍華德：那是在中國舉行的 機器學習研討會。

It's not often, actually, at academic conferences

那不常有，事實上， 在學術會議上

that you do hear spontaneous applause,

聽到熱烈的掌聲，

although of course sometimes at TEDx conferences, feel free.

雖然有些時候 TEDx 講座不拘泥形式。

Everything you saw there was happening with deep learning.

你所看到的都是出於深度學習

(Applause) Thank you.

（掌聲）謝謝。

The transcription in English was deep learning.

英文文字翻譯由深度學習完成的。

The translation to Chinese and the text in the top right, deep learning,

翻譯成中文和右上角的文稿 也是出於深度學習，

and the construction of the voice was deep learning as well.

連創建聲音也都是深度學習。

So deep learning is this extraordinary thing.

深度學習是如此的神奇。

It's a single algorithm that can seem to do almost anything,

它是個單一的演算法 似乎可以完成任何事情，

and I discovered that a year earlier, it had also learned to see.

我一年前還發現它可以學會看

In this obscure competition from Germany

這個德國遊戲的比賽

called the German Traffic Sign Recognition Benchmark,

叫德國交通標誌確認基準，

deep learning had learned to recognize traffic signs like this one.

深度學習能認出這個交通標誌。

Not only could it recognize the traffic signs

它不只確認交通標誌的能力

better than any other algorithm,

比其他的演算法好，

the leaderboard actually showed it was better than people,

在排行榜上更顯示它做得比人類好，

about twice as good as people.

正確性是人類的兩倍。

So by 2011, we had the first example

2011 以前，我們有了第一個例子

of computers that can see better than people.

視力高於人類的電腦。

Since that time, a lot has happened.

從那時開始，許多電腦也可以做到。

In 2012, Google announced that they had a deep learning algorithm

2012 年谷歌宣佈 使用深度學習演算法

watch YouTube videos

來監看 Youtube 影片

and crunched the data on 16,000 computers for a month,

收集一個月 1,600 台電電腦的資料，

and the computer independently learned about concepts such as people and cats

電腦獨立識別 人或貓的概念

just by watching the videos.

僅透過觀看影片。

This is much like the way that humans learn.

這樣更像人類的學習方式。

Humans don't learn by being told what they see,

人類並非通過別人的指示來學習，

but by learning for themselves what these things are.

而是從自己搞懂事情來學習。

Also in 2012, Geoffrey Hinton, who we saw earlier,

在 2012 年傑佛里·辛頓 我們之前看到的人，

won the very popular ImageNet competition,

贏了很有名的映像網路比賽，

looking to try to figure out from one and a half million images

嘗試從 150 萬的圖像中找出

what they're pictures of.

想要的圖像。

As of 2014, we're now down to a six percent error rate

2014 年, 我們現在 圖像辨識的錯誤率

in image recognition.

降到 6％ 以下。

This is better than people, again.

這再次證明它比人類優秀。

So machines really are doing an extraordinarily good job of this,

可見機器 真可以做到如此非凡的成就，

and it is now being used in industry.

它現在已經用在產業上了。

For example, Google announced last year

比如說，谷歌去年宣佈

that they had mapped every single location in France in two hours,

他們可以在兩小時内把 法國每一個位置繪成地圖，

and the way they did it was that they fed street view images

他們用的方式是 把街景圖像

into a deep learning algorithm to recognize and read street numbers.

輸入深度學習演算法 來辨認和讀取街道號碼。

Imagine how long it would have taken before:

想想我們以前需要花多少時間？

dozens of people, many years.

至少好幾十人加上好幾年呢。

This is also happening in China.

同樣的情況也發生在中國。

Baidu is kind of the Chinese Google, I guess,

我想「百度」類似中國的谷歌，

and what you see here in the top left

在左上角你會看見

is an example of a picture that I uploaded to Baidu's deep learning system,

一張我上傳到 百度深度學習系統的圖片，

and underneath you can see that the system has understood what that picture is

下方你可以看到 系統可以理解這張圖片

and found similar images.

而且能找到相似的圖像。

The similar images actually have similar backgrounds,

類似的圖像 也就是有相似的背景，

similar directions of the faces,

相似面孔的角度，

even some with their tongue out.

有的圖像甚至有伸出舌頭。

This is not clearly looking at the text of a web page.

這個網頁的文字看不大清楚，

All I uploaded was an image.

因為我上傳的都是圖像。

So we now have computers which really understand what they see

這顯示了電腦能明白他們所看到的

and can therefore search databases

電腦能夠搜尋資料庫

of hundreds of millions of images in real time.

以即時的方式從億萬張圖片中搜尋。

So what does it mean now that computers can see?

現在的電腦能夠去看 是表示什麼意思呢？

Well, it's not just that computers can see.

其實電腦不只能看見。

In fact, deep learning has done more than that.

事實上深度學習可以做得更多。

Complex, nuanced sentences like this one

像這個樣複雜，僅有小小差別的句子

are now understandable with deep learning algorithms.

現在的深度學習演算法能夠理解。

As you can see here,

你可以看到，

this Stanford-based system showing the red dot at the top

這以史丹福為基礎的系統 顯示上面的紅點

has figured out that this sentence is expressing negative sentiment.

指這句子是在表達負面的情緒。

Deep learning now in fact is near human performance

深度學習現在已經接近人類的行為

at understanding what sentences are about and what it is saying about those things.

能理解句子是要表達什麼。

Also, deep learning has been used to read Chinese,

同時，深度學習也能用以閱讀中文，

again at about native Chinese speaker level.

程度相當於以中文為母語的水平。

This algorithm developed out of Switzerland

這演算法發展於瑞士

by people, none of whom speak or understand any Chinese.

沒有一個會說中文的團隊。

As I say, using deep learning

像我說的，深度學習

is about the best system in the world for this,

是一個最好的系統 對完成這任務來說，

even compared to native human understanding.

甚至比人類還要好。

This is a system that we put together at my company

這個系統是我公司建立的

which shows putting all this stuff together.

要把這些東西都集中在一起。

These are pictures which have no text attached,

這是一些沒有文字描述的圖片，

and as I'm typing in here sentences,

我在這裡輸入句子，

in real time it's understanding these pictures

它在同步理解這些照片

and figuring out what they're about

找出它們是有關什麼的照片

and finding pictures that are similar to the text that I'm writing.

也找出跟我句子相關類似的圖片。

So you can see, it's actually understanding my sentences

所以你看， 它真的能理解我的句子。

and actually understanding these pictures.

也完全的理解這些圖片。

I know that you've seen something like this on Google,

你在谷歌上也看過類似的，

where you can type in things and it will show you pictures,

你可以輸入文字 而它會顯示圖片，

but actually what it's doing is it's searching the webpage for the text.

但事實上，它在尋索網頁上的文字。

This is very different from actually understanding the images.

這跟理解圖片有很大的不同。

This is something that computers have only been able to do

理解圖片只有電腦可以做

for the first time in the last few months.

電腦在過去幾個月才會做的事。

So we can see now that computers can not only see but they can also read,

電腦不單能看見 也能閱讀，

and, of course, we've shown that they can understand what they hear.

而且我們顯示了電腦能理解所聽到的。

Perhaps not surprising now that I'm going to tell you they can write.

或許不意外地， 我要告訴你們電腦也能書寫。

Here is some text that I generated using a deep learning algorithm yesterday.

這是我昨天用深度學習演算法 所產生的文字。

And here is some text that an algorithm out of Stanford generated.

這裡有一些非史丹佛演算法 所產生的文字。

Each of these sentences was generated

這些句子的產生

by a deep learning algorithm to describe each of those pictures.

是透過深度學習演算法 對圖片進行描述。

This algorithm before has never seen a man in a black shirt playing a guitar.

這演算法是電腦從來沒有看見過 一個穿黑襯衫的男子彈吉他。

It's seen a man before, it's seen black before,

電腦見過男人， 看過黑色，

it's seen a guitar before,

見過吉他，

but it has independently generated this novel description of this picture.

它自己便對圖片做出描述。

We're still not quite at human performance here, but we're close.

雖然還沒有超越人類， 不過很接近了。

In tests, humans prefer the computer-generated caption

依據統計，人們較喜歡 電腦的圖片說明

one out of four times.

有四分之一的人會做這樣的選擇。

Now this system is now only two weeks old,

這系統在兩個星期前開發完成，

so probably within the next year,

估計在明年，

the computer algorithm will be well past human performance

電腦演算法將會超越人類

at the rate things are going.

如果依照這樣的速度發展下的話。

So computers can also write.

到時候電腦也會書寫了。

So we put all this together and it leads to very exciting opportunities.

我們把這些都放在一起， 讓它來引導到一個令人振奮的時機。

For example, in medicine,

像在藥物方面，

a team in Boston announced that they had discovered

一個波士頓的團隊 宣佈他們發現了

dozens of new clinically relevant features

數十種腫瘤的臨床特徵

of tumors which help doctors make a prognosis of a cancer.

幫助醫生預測癌症。

Very similarly, in Stanford,

同樣的，在史丹佛，

a group there announced that, looking at tissues under magnification,

一個組織宣佈 在放大鏡下觀察組織，

they've developed a machine learning-based system

他們開發 一個以機器學習為基礎的系統

which in fact is better than human pathologists

比人類病理學家更有效地

at predicting survival rates for cancer sufferers.

預測癌症病患的生存率。

In both of these cases, not only were the predictions more accurate,

這些例子， 不但能更準確地預測，

but they generated new insightful science.

而且也能帶來更多科技上的洞見。

In the radiology case,

在放射學的個案中，

they were new clinical indicators that humans can understand.

他們是人類所能理解的新臨床指標。

In this pathology case,

在這病理學個案，

the computer system actually discovered that the cells around the cancer

電腦系統發現癌症周圍的細胞

are as important as the cancer cells themselves

在診斷的時候

in making a diagnosis.

是跟癌細胞一樣重要。

This is the opposite of what pathologists had been taught for decades.

這跟病理學家 10 年來的說法相反。

In each of those two cases, they were systems developed

在這兩個個案， 系統的開發人員

by a combination of medical experts and machine learning experts,

是由醫學專家 和機器學習專家所組成，

but as of last year, we're now beyond that too.

但自去年開始， 我們也超越了這些。

This is an example of identifying cancerous areas

這是確認癌症範圍的例子

of human tissue under a microscope.

是在顯微鏡下的人類組織。

The system being shown here can identify those areas more accurately,

系統顯示可以更準確地確認範圍，

or about as accurately, as human pathologists,

如病理學家般準確，

but was built entirely with deep learning using no medical expertise

不過沒有藥物專家 來建構整套深度學習系統

by people who have no background in the field.

系統是由一些 沒有專業背景的人完成。

Similarly, here, this neuron segmentation.

同樣地，從是細胞分裂。

We can now segment neurons about as accurately as humans can,

我們的系統可以像人類般 精確地分裂神經細胞，

but this system was developed with deep learning

不過開發這套深度學習系統

using people with no previous background in medicine.

沒有一個人來自醫學背景。

So myself, as somebody with no previous background in medicine,

就是我和一些沒有醫學背景的人，

I seem to be entirely well qualified to start a new medical company,

看來我頗有資格開一家醫藥公司。

which I did.

我確實這麼做了。

I was kind of terrified of doing it,

我是以戒慎恐懼的心情開始做，

but the theory seemed to suggest that it ought to be possible

不過理論顯示 這是可行的

to do very useful medicine using just these data analytic techniques.

用這些資料分析技術來 製作有效的藥物。

And thankfully, the feedback has been fantastic,

感恩的是 回應也挺不錯，

not just from the media but from the medical community,

這回應不只是來自媒體， 而且還有醫藥社群，

who have been very supportive.

他們都很支持。

The theory is that we can take the middle part of the medical process

理論上我們能在醫務過程中

and turn that into data analysis as much as possible,

盡量轉換成資料分析，

leaving doctors to do what they're best at.

讓醫生去做他們擅長的。

I want to give you an example.

我舉一個例子。

It now takes us about 15 minutes to generate a new medical diagnostic test

我們現在花 15 分鐘 來創造一項新的醫學診斷測試

and I'll show you that in real time now,

我會讓你同步看到過程，

but I've compressed it down to three minutes by cutting some pieces out.

不過我已刪除部分資料 壓縮成三分鐘。

Rather than showing you creating a medical diagnostic test,

我不會向你們展示 創造出來的醫學診斷測試，

I'm going to show you a diagnostic test of car images,

我要向你們展示 一項汽車圖片的診斷測試，

because that's something we can all understand.

因為這個我們都能理解。

So here we're starting with about 1.5 million car images,

我們從 150 萬張 的汽車圖片開始，

and I want to create something that can split them into the angle

我希望創造一些東西 把圖片分類

of the photo that's being taken.

而且依圖片拍攝的角度來分類。

So these images are entirely unlabeled, so I have to start from scratch.

這些圖片完全沒有標題， 我必需從零開始。

With our deep learning algorithm,

深度學習演算法，

it can automatically identify areas of structure in these images.

它能自動確認 這些圖片的結構。

So the nice thing is that the human and the computer can now work together.

美好的是 人和電腦可以合作

So the human, as you can see here,

看看這裡，這個人，

is telling the computer about areas of interest

正在告訴電腦 關於感興趣的範圍

which it wants the computer then to try and use to improve its algorithm.

而電腦會嘗試用它 來改善電腦的演算法。

Now, these deep learning systems actually are in 16,000-dimensional space,

這些深度學習系統 有 16,000 個立體空間，

so you can see here the computer rotating this through that space,

你可以看見電腦 讓他們在這空間旋轉，

trying to find new areas of structure.

嘗試找出新的區域結構。

And when it does so successfully,

當它成功時，

the human who is driving it can then point out the areas that are interesting.

在開車的人能夠 指出有興趣的地方。

So here, the computer has successfully found areas,

這裡，電腦成功的找到了那地區，

for example, angles.

再舉例，角度，

So as we go through this process,

通過這個過程，

we're gradually telling the computer more and more

我們漸漸地告訴電腦更多

about the kinds of structures we're looking for.

關於我們在找的結構類型。

You can imagine in a diagnostic test

你可以想像一個診斷測試

this would be a pathologist identifying areas of pathosis, for example,

像是一個病理學家辨認 病症的範圍，

or a radiologist indicating potentially troublesome nodules.

或是放射治療師界定 潛在的腫瘤。

And sometimes it can be difficult for the algorithm.

有些時候對演算法來說 是有些困難。

In this case, it got kind of confused.

在我們這個例子，它會出現混亂。

The fronts and the backs of the cars are all mixed up.

汽車的正面和背面 都混淆不清了。

So here we have to be a bit more careful,

我們需要更小心，

manually selecting these fronts as opposed to the backs,

手動選出正面 跟背面有相反效果的文字，

then telling the computer that this is a type of group

然後告知電腦 這是一種

that we're interested in.

我們有興趣的一類。

So we do that for a while, we skip over a little bit,

這要花了一些時間來做， 所以我們跳過，

and then we train the machine learning algorithm

然後我們訓練 機器學習演算法

based on these couple of hundred things,

以好幾百張圖片去訓練它，

and we hope that it's gotten a lot better.

我們希望它會做得更好。

You can see, it's now started to fade some of these pictures out,

你可以看見，它開始 刪除一些圖片，

showing us that it already is recognizing how to understand some of these itself.

顯示它已經知道 可以自己理解這些圖片。

We can then use this concept of similar images,

我們運用相似圖片的概念，

and using similar images, you can now see,

用類似的圖片，你可以看到，

the computer at this point is able to entirely find just the fronts of cars.

電腦現在可以 完全找到正面的汽車。

So at this point, the human can tell the computer,

這時， 人類可以告訴電腦，

okay, yes, you've done a good job of that.

對，你做的很好。

Sometimes, of course, even at this point

當然，有些時候，即使在這個階段

it's still difficult to separate out groups.

分組仍然是困難的。

In this case, even after we let the computer try to rotate this for a while,

在這情況，儘管我們讓 電腦嘗試旋轉圖片一陣子，

we still find that the left sides and the right sides pictures

我們還是發現左邊 和右邊的圖片

are all mixed up together.

是混淆在一起的。

So we can again give the computer some hints,

於是我們再次 給電腦一些提示，

and we say, okay, try and find a projection that separates out

像是嘗試去發現一個計畫可以

the left sides and the right sides as much as possible

儘量區分出左邊和右邊的圖片

using this deep learning algorithm.

是透過使用深度學習演算法。

And giving it that hint -- ah, okay, it's been successful.

給予提示後， 好，它已經完成了。

It's managed to find a way of thinking about these objects

它找到一個方法 想像這些目標

that's separated out these together.

來分別這些分類。

So you get the idea here.

你現在知道了。

This is a case not where the human is being replaced by a computer,

這並不是電腦取代人類，

but where they're working together.

而是兩者一起合作。

What we're doing here is we're replacing something that used to take a team

我們在做的事情是 在過去需要

of five or six people about seven years

5 或 6 個人 花 7 年時間完成的事情

and replacing it with something that takes 15 minutes

現在只需一個人

for one person acting alone.

15 分鐘來完成。

So this process takes about four or five iterations.

這個過程需要重覆 4 或 5 次。

You can see we now have 62 percent

你現在可以看到

of our 1.5 million images classified correctly.

我們在 150 萬的圖片中 有 62％ 是正確分類。

And at this point, we can start to quite quickly

現在，可見我們可以迅速地

grab whole big sections,

掌握整個大部分資料，

check through them to make sure that there's no mistakes.

再檢查以確定沒有錯誤。

Where there are mistakes, we can let the computer know about them.

有錯誤，我們可以 讓電腦知道錯誤的地方。

And using this kind of process for each of the different groups,

每一個不同的分類 我們都使用這種程序來做，

we are now up to an 80 percent success rate

我們現在 在分辨 150 萬張的圖片時

in classifying the 1.5 million images.

有超過 80％ 的成功率，

And at this point, it's just a case

現在，在這個案例

of finding the small number that aren't classified correctly,

找到少數幾個不正確的分類，

and trying to understand why.

讓電腦了解原因。

And using that approach,

用這種方法，

by 15 minutes we get to 97 percent classification rates.

15 分鐘就有 97％ 的分辨率。

So this kind of technique could allow us to fix a major problem,

這種技術可以幫助 解決一個重要的問題，

which is that there's a lack of medical expertise in the world.

醫療專家不足的問題。

The World Economic Forum says that there's between a 10x and a 20x

世界經濟論壇表示

shortage of physicians in the developing world,

在發展中國家，內科醫生 有 10 倍到 20 倍的短缺。

and it would take about 300 years

這要三百年的時間

to train enough people to fix that problem.

才能訓練足夠的人 來處理這個問題。

So imagine if we can help enhance their efficiency

想像一下， 我們是否可以幫助提高效率

using these deep learning approaches?

是使用深度學習這個方法來提升？

So I'm very excited about the opportunities.

我對這個機會感到很興奮。

I'm also concerned about the problems.

我也關注這些問題。

The problem here is that every area in blue on this map

問題是在這地圖上每個藍色的地方

is somewhere where services are over 80 percent of employment.

那裡都有 80％ 的服務人員。

What are services?

什麼是服務？

These are services.

這些就是服務。

These are also the exact things that computers have just learned how to do.

電腦剛學會如何去做是確實的事。

So 80 percent of the world's employment in the developed world

發展中國家 80％ 的僱員工作

is stuff that computers have just learned how to do.

電腦已開始學習如何做。

What does that mean?

這意味什麼？

Well, it'll be fine. They'll be replaced by other jobs.

那可好。 他們將會被其他的職業取代。

For example, there will be more jobs for data scientists.

舉例：需要更多科學家來工作。

Well, not really.

不過，這不完全正確。

It doesn't take data scientists very long to build these things.

數據科學家 不需要花很久的時間去做這些事情。

For example, these four algorithms were all built by the same guy.

例如，這四個演算法是同一個人設計的。

So if you think, oh, it's all happened before,

若你認為這些 以前都發生過，

we've seen the results in the past of when new things come along

過去我們看過 新事物出現的結果

and they get replaced by new jobs,

他們被新的職務所取替，

what are these new jobs going to be?

那些新的職業會是什麼呢？

It's very hard for us to estimate this,

我們很難去判斷，

because human performance grows at this gradual rate,

因為人類的能力 以這個速度逐漸成長，

but we now have a system, deep learning,

我們現在有了深度學習系統，

that we know actually grows in capability exponentially.

我們知道 以指數的方式增長。

And we're here.

我們在這裡。

So currently, we see the things around us

最近，我們看周圍的事物

and we say, "Oh, computers are still pretty dumb." Right?

會說：電腦還是很笨，不是嗎？

But in five years' time, computers will be off this chart.

但是在五年內， 電腦將會超越這張圖表。

So we need to be starting to think about this capability right now.

我們需要開始思考這個能力。

We have seen this once before, of course.

當然，我們曾經看過這個。

In the Industrial Revolution,

在工業革命時期，

we saw a step change in capability thanks to engines.

發動機讓生產力往前跨一大步。

The thing is, though, that after a while, things flattened out.

雖然，一段時間之後， 事情轉為平靜。

There was social disruption,

那時社會混亂，

but once engines were used to generate power in all the situations,

發動機被普遍使用 產生動力，

things really settled down.

事情就能真正得到解決。

The Machine Learning Revolution

機器學習革命

is going to be very different from the Industrial Revolution,

與工業革命大不相同，

because the Machine Learning Revolution, it never settles down.

因為機器學習革命， 永遠不會停下來。

The better computers get at intellectual activities,

電腦更具智力活動，

the more they can build better computers to be better at intellectual capabilities,

他們能製造更好的電腦 去運作更好的智能活動，

so this is going to be a kind of change

這是一種改變

that the world has actually never experienced before,

從未經歷過的改變，

so your previous understanding of what's possible is different.

你之前的理解的可能性是不同的。

This is already impacting us.

這已經影響我們。

In the last 25 years, as capital productivity has increased,

過去 25 年， 資本生產力一直在增長，

labor productivity has been flat, in fact even a little bit down.

勞動生產力已經放緩， 事實上已有一點點下降。

So I want us to start having this discussion now.

我想我們開始討論這個議題。

I know that when I often tell people about this situation,

我知道當我告訴別人這種情況時，

people can be quite dismissive.

人們可以不以為然。

Well, computers can't really think,

電腦不會思考，

they don't emote, they don't understand poetry,

它們沒有感情， 也不了解詩，

we don't really understand how they work.

我們不真正理解它們怎麼運作。

So what?

可是，哪又如何？

Computers right now can do the things

電腦現在可以作

that humans spend most of their time being paid to do,

人們花大部分時間 得到報酬所做的事情，

so now's the time to start thinking

所以我們該是思考的時候

about how we're going to adjust our social structures and economic structures

我們如何調整我們的社會和經濟結構

to be aware of this new reality.

請關注這些新的改變。

Thank you.

謝謝

(Applause)

（掌聲）