because that seems to be the key driver

因為機器學習似乎是推動人工智慧

around artificial intelligence.

的關鍵

How does machine learning work?

機器學習是如何運作的呢？

Sebastian Thrun: So, artificial intelligence and machine learning

賽巴斯・汀索朗：人工智慧和機器學習

is about 60 years old

大約有 60 年的歷史，

and has not had a great day in its past until recently.

一直到近期才達到極致。

And the reason is that today,

那是因為現在，

we have reached a scale of computing and datasets

我們的計算能力和資料庫規模已經達到

that was necessary to make machines smart.

讓機器變聰明所必須具備的條件。

So here's how it works:

它的運作方式是這樣的：

If you program a computer today, say, your phone,

如果現在你要為一台電腦寫程式，比如你的手機，

then you hire software engineers

你會僱用軟體工程師，

that write a very, very long kitchen recipe,

他們會寫一份非常非常長的指令，像廚房食譜，

like, "If the water is too hot, turn down the temperature.

像是「如果水太熱，就把溫度調低。

If it's too cold, turn up the temperature."

如果水太冷，就把溫度調高。」

The recipes are not just 10 lines long.

這樣的「食譜」並不是只有十行的長度。

They are millions of lines long.

它們長達數百萬行

A modern cell phone has 12 million lines of code.

一台現代手機有 1200 萬行的程式碼

A browser has five million lines of code.

一個瀏覽器就有五百萬行的程式碼

And each bug in this recipe can cause your computer to crash.

而且食譜中的每一個錯誤， 都會造成你的電腦當機

That's why a software engineer makes so much money.

那就是軟體工程師能賺那麼多錢的原因

The new thing now is that computers can find their own rules.

現在的新發展是，電腦能找到它們自己的規則

So instead of an expert deciphering, step by step,

所以不再需要解碼專家，去針對每個情況

a rule for every contingency,

一步一步地做理解辨識，

what you do now is you give the computer examples

現在你的做法是，給電腦一些範例，

and have it infer its own rules.

讓它推導出它自己的規則

A really good example is AlphaGo, which recently was won by Google.

最近 Google 的阿爾法圍棋贏得比賽， 就是一個很好的例子

Normally, in game playing, you would really write down all the rules,

通常，在玩遊戲時， 你會寫下所有的規則，

but in AlphaGo's case,

但在阿爾法圍棋這個例子，

the system looked over a million games

系統是去看了一百多萬場的比賽，

and was able to infer its own rules

並且推導出它自己的規則，

and then beat the world's residing Go champion.

然後打敗現在的世界棋王

That is exciting, because it relieves the software engineer

這事件振奮人心的事， 因為軟體工程師能鬆口氣了，

of the need of being super smart,

他們不需要超級聰明，

and pushes the burden towards the data.

這個重任已經落到資料上頭。

As I said, the inflection point where this has become really possible --

如我所言，這件事可能發生的轉折點在於──

very embarrassing, my thesis was about machine learning.

真是不好意思，我的論文就是寫機器學習

It was completely insignificant, don't read it,

它完全不重要，請不要去讀，

because it was 20 years ago

因為那是 20 年前寫的

and back then, the computers were as big as a cockroach brain.

那個時候，電腦不過和蟑螂大腦一樣大而已

Now they are powerful enough to really emulate

現在，電腦已經強大到能夠真正地模擬

kind of specialized human thinking.

人類的特定思想

And then the computers take advantage of the fact

而且，電腦也因為可以比人類看更多的資料

that they can look at much more data than people can.

而取得優勢

So I'd say AlphaGo looked at more than a million games.

就如同我所說的， 阿爾法圍棋已經研究過一百多萬場的比賽

No human expert can ever study a million games.

沒有任何專家能夠研究一百多萬場的比賽

Google has looked at over a hundred billion web pages.

Google 已經看過了一千多億個網頁

No person can ever study a hundred billion web pages.

從來沒有人有能力研究一千多億個網頁

So as a result, the computer can find rules

因此，電腦能夠找出

that even people can't find.

人類找不到的規則

CA: So instead of looking ahead to, "If he does that, I will do that,"

安德森：那麼，電腦應該不是： 「如果他那樣下，我就這樣下。」

it's more saying, "Here is what looks like a winning pattern,

應該比較像是：「下在這裡比較像是獲勝的模式，

here is what looks like a winning pattern."

下在那裡比較像是獲勝的模式。」

ST: Yeah. I mean, think about how you raise children.

索朗：沒錯， 想想看你如何養育你的孩子

You don't spend the first 18 years giving kids a rule for every contingency

你並不會花前 18 年的時間， 針對每種狀況給孩子一條規則，

and set them free and they have this big program.

然後放他們自由， 他們就會做出這個大程式

They stumble, fall, get up, they get slapped or spanked,

他們會摔跤，會爬起來， 他們會被賞巴掌或打屁股，

and they have a positive experience, a good grade in school,

他們會有正向的經驗， 在學校有好的成績，

and they figure it out on their own.

他們會靠自己去了解這些

That's happening with computers now,

現在的電腦也是這樣，

which makes computer programming so much easier all of a sudden.

突然間電腦寫程式變得簡單多了

Now we don't have to think anymore. We just give them lots of data.

我們不用再花腦筋思考，只要給它們大量資料即可

CA: And so, this has been key to the spectacular improvement

安德森：所以，這是車輛自動駕駛能力

in power of self-driving cars.

能夠有重大改善的關鍵

I think you gave me an example.

我想你給了我一個例子

Can you explain what's happening here?

你可以解釋一下這裡發生了什麼事嗎？

ST: This is a drive of a self-driving car

索朗：這是自動駕駛車輛，

that we happened to have at Udacity

我們優達學城（Udacity）剛好有，

and recently made into a spin-off called Voyage.

最近成為 Voyage 的副產品

We have used this thing called deep learning

我們用所謂的「深度學習」

to train a car to drive itself,

來訓練汽車自動駕駛

and this is driving from Mountain View, California,

這趟行程從加州的山景城出發，

to San Francisco

前往舊金山，

on El Camino Real on a rainy day,

在雨天行駛在 El Camino Real 路上，

with bicyclists and pedestrians and 133 traffic lights.

路上有腳踏車騎士及行人， 並且經過了 133 個紅綠燈

And the novel thing here is,

新奇的事情是，

many, many moons ago, I started the Google self-driving car team.

許多個月前，我成立了 Google 自動駕駛汽車團隊

And back in the day, I hired the world's best software engineers

那時，我僱用世界上最厲害的軟體工程師，

to find the world's best rules.

來找出世界上最好的規則。

This is just trained.

這是訓練出來的

We drive this road 20 times,

這條路我們開了 20 次

we put all this data into the computer brain,

我們把所有資料放到電腦中，

and after a few hours of processing,

經過幾個小時的處理，

it comes up with behavior that often surpasses human agility.

電腦所找出的行為，通常都超越人類的機敏。

So it's become really easy to program it.

所以電腦很容易為它寫程式

This is 100 percent autonomous, about 33 miles, an hour and a half.

這是 100% 自動化的，大約 33 英哩，一個半小時

CA: So, explain it -- on the big part of this program on the left,

安德生：來解釋一下──這程式左邊大部分，

you're seeing basically what the computer sees as trucks and cars

我們可以看到電腦所看到的卡車和汽車，

and those dots overtaking it and so forth.

還有那些超越它的點。

ST: On the right side, you see the camera image, which is the main input here,

索朗：右邊可以看到攝影機的影像，也就是主要輸入，

and it's used to find lanes, other cars, traffic lights.

這個用來找車道、其它車輛和紅綠燈

The vehicle has a radar to do distance estimation.

這個車用雷達用來估算距離

This is very commonly used in these kind of systems.

這是這類系統常用的方式

On the left side you see a laser diagram,

左邊的是雷射圖，

where you see obstacles like trees and so on depicted by the laser.

我們可以看到雷射槍描繪出來的障礙物，如樹木等等

But almost all the interesting work is centering on the camera image now.

但幾乎所有有趣的部份，都是著重在攝影機影像上

We're really shifting over from precision sensors like radars and lasers

我們其實從精準的感測器，像是雷達和雷射，

into very cheap, commoditized sensors.

轉換到更便宜的一般感測器

A camera costs less than eight dollars.

一台攝影機的成本不到 $8

CA: And that green dot on the left thing, what is that?

安德森：左邊的綠點是什麼？

Is that anything meaningful?

是有意義的嗎？

ST: This is a look-ahead point for your adaptive cruise control,

索朗：這是「向前看」的點，提供自動調整航程控制用，

so it helps us understand how to regulate velocity

它會根據前車的距離，

based on how far the cars in front of you are.

幫助我們調整速度

CA: And so, you've also got an example, I think,

安德森：這樣的話，我認為，你也有個例子

of how the actual learning part takes place.

說明真正的學習是如何進行的

Maybe we can see that. Talk about this.

也許我們可以邊看那個畫面，邊談這個。

ST: This is an example where we posed a challenge to Udacity students

索朗：這是我們挑戰 Udacity 學生的一個例子，

to take what we call a self-driving car Nanodegree.

是取得「自駕車奈米學位」的挑戰

We gave them this dataset

我們提供他們這個資料庫，

and said "Hey, can you guys figure out how to steer this car?"

並且告訴他們：「你們能不能想出要如何駕駛這台車？」

And if you look at the images,

如果從影像來看，

it's, even for humans, quite impossible to get the steering right.

即使是人類也很難駕駛正確。

And we ran a competition and said, "It's a deep learning competition,

我們進行了一項競賽，並說： 「這是場深度學習競賽，

AI competition,"

這是人工智慧競賽。」

and we gave the students 48 hours.

我們給學生 48 小時

So if you are a software house like Google or Facebook,

如果你是間軟體公司，如 Google 或臉書，

something like this costs you at least six months of work.

像這樣的東西會花你至少六個月的時間

So we figured 48 hours is great.

所以我們認為 48 小時是很棒的

And within 48 hours, we got about 100 submissions from students,

在 48 小時內，我們得到了約 100 件學生提交的結果，

and the top four got it perfectly right.

前四名完全正確。

It drives better than I could drive on this imagery,

和我在這影像上用深度學習相比，

using deep learning.

它駕駛得更好

And again, it's the same methodology.

再一次，同樣的方法，

It's this magical thing.

這是件很神奇的事

When you give enough data to a computer now,

當你提供電腦足夠的資料，

and give enough time to comprehend the data,

並給它足夠時間來理解這些資料，

it finds its own rules.

它就會自己找到規則

CA: And so that has led to the development of powerful applications

安德森：所以，那就導致了在各種領域

in all sorts of areas.

應用程式的強大發展

You were talking to me the other day about cancer.

之前你有和我談過癌症的事

Can I show this video?

我能播那段影片嗎？

ST: Yeah, absolutely, please. CA: This is cool.

索朗：當然，請放。安德森：這很酷

ST: This is kind of an insight into what's happening

索朗：這有點像是對完全不同的領域

in a completely different domain.

洞察所發生的事

This is augmenting, or competing --

在旁觀者眼裡，

it's in the eye of the beholder --

這可以說是

with people who are being paid 400,000 dollars a year,

和那些年薪 $40 萬的

dermatologists,

皮膚科醫生的擴增或競爭，

highly trained specialists.

他們是訓練良好的專家，

It takes more than a decade of training to be a good dermatologist.

要受十年以上的訓練才可能成為好的皮膚科醫生

What you see here is the machine learning version of it.

這裡所看到的是它的機器學習版本，

It's called a neural network.

稱為「神經網路」

"Neural networks" is the technical term for these machine learning algorithms.

「神經網路」是機器學習演算法的專有名詞，

They've been around since the 1980s.

大約出自 1980 年代

This one was invented in 1988 by a Facebook Fellow called Yann LeCun,

這個是在 1988 年由臉書的研究專員揚・勒丘恩所發明的

and it propagates data stages

它透過一種你可視為是人腦的方式

through what you could think of as the human brain.

依階段傳播數據

It's not quite the same thing, but it emulates the same thing.

它不是人腦，但它模仿人腦

It goes stage after stage.

一個階段接著一個階段，

In the very first stage, it takes the visual input and extracts edges

在第一個階段取得視覺輸入，粹取出邊界、

and rods and dots.

線和點

And the next one becomes more complicated edges

下個階段就變成更複雜的邊界

and shapes like little half-moons.

以及像是半月的形狀。

And eventually, it's able to build really complicated concepts.

最後，它能建立出非常複雜的概念。

Andrew Ng has been able to show

Andrew Ng 就展示過，

that it's able to find cat faces and dog faces

它能夠在非常大量的影像中

in vast amounts of images.

找出貓和狗的臉。

What my student team at Stanford has shown is that

我在史丹佛的學生團隊也展示過，

if you train it on 129,000 images of skin conditions,

如果你用十二萬九千張皮膚症狀的影像來訓練它，

including melanoma and carcinomas,

包括黑色素瘤和癌，

you can do as good a job

你就能和最好的人類皮膚科醫生

as the best human dermatologists.

做得一樣好。

And to convince ourselves that this is the case,

為了說服我們自己事實確實是如此，

we captured an independent dataset that we presented to our network

我們取得了一個獨立的資料集，拿給我們的網路看，

and to 25 board-certified Stanford-level dermatologists,

也拿給 25 位認證過的史丹佛水準的皮膚科醫生看，

and compared those.

來做比較

And in most cases,

在大部份狀況，

they were either on par or above the performance classification accuracy

在分類正確性上，網路的表現都和人類皮膚科醫生

of human dermatologists.

並駕齊驅或者更好

CA: You were telling me an anecdote.

安德森：你跟我說過一則軼事

I think about this image right here.

我想應該是這張影像的這個地方

What happened here?

這裡發生了什麼事？

ST: This was last Thursday. That's a moving piece.

索朗：時間是上星期四，是個正在進行的故事。

What we've shown before and we published in "Nature" earlier this year

我們之前展示過，今年稍早也刊在「Nature」期刊中，

was this idea that we show dermatologists images

我們的想法是，我們讓皮膚科醫生看影像，

and our computer program images,

也讓我們的電腦程式看，

and count how often they're right.

計算它們判斷正確的頻率

But all these images are past images.

但所有影像都是過去的影像

They've all been biopsied to make sure we had the correct classification.

都已經做過切片檢查，確保分類正確

This one wasn't.

但是這一張沒有

This one was actually done at Stanford by one of our collaborators.

這張其實是史丹佛的一位合作者做的

The story goes that our collaborator,

這個故事跟我們的合作者有關，

who is a world-famous dermatologist, one of the three best, apparently,

他是世界知名的皮膚科醫生，很顯然是三位最好的皮膚科醫生之一，

looked at this mole and said, "This is not skin cancer."

他看著這個痣說：「這不是皮膚癌。」

And then he had a second moment, where he said,

他想了一下，接著說：

"Well, let me just check with the app."

「讓我用應用程式確認一下。」

So he took out his iPhone and ran our piece of software,

他拿出他的 iPhone，執行我們的軟體，

our "pocket dermatologist," so to speak,

iPhone 可說是我們的「口袋皮膚科醫生」，

and the iPhone said: cancer.

而 iPhone 說是癌症，

It said melanoma.

是黑色素瘤

And then he was confused.

他很困惑，

And he decided, "OK, maybe I trust the iPhone a little bit more than myself,"

他決定：「好吧，也許我應該相信 iPhone 比相信我自己多一點。」

and he sent it out to the lab to get it biopsied.

他把它送去實驗室做切片檢查，

And it came up as an aggressive melanoma.

結果是惡性黑色素瘤

So I think this might be the first time that we actually found,

我想，這可能是我們第一次

in the practice of using deep learning,

在深度學習上實際遇到，

an actual person whose melanoma would have gone unclassified,

如果沒有這個深度學習的機會，

had it not been for deep learning.

這個人的黑色素瘤就不會被發現

CA: I mean, that's incredible.

安德森：那真的很了不起。

It feels like there'd be an instant demand for an app like this right now,

像這樣的應用程式，現在可能已經有很迫切的需求，

that you might freak out a lot of people.

這可能會嚇壞很多人

Are you thinking of doing this, making an app that allows self-checking?

你有想過要這麼做嗎？做個自我檢測的應用程式？

ST: So my in-box is flooded about cancer apps,

索朗：我的收件匣被關於癌症應用程式的信件給淹沒了，

with heartbreaking stories of people.

那些信都是令人心碎的故事

I mean, some people have had 10, 15, 20 melanomas removed,

有些人已經移除了 10、15、20 個黑色素瘤，

and are scared that one might be overlooked, like this one,

很害怕會漏掉任何一個，就像這個例子一樣，

and also, about, I don't know,

還有些內容是，我不知道，

flying cars and speaker inquiries these days, I guess.

飛天車、這幾天的演說邀請，我猜是吧

My take is, we need more testing.

我的重點是，我們需要更多測試

I want to be very careful.

我必須非常小心，

It's very easy to give a flashy result and impress a TED audience.

畢竟 TED 的觀眾很容易會對一些出色的演說結果感到印象深刻

It's much harder to put something out that's ethical.

相對地，要端出合乎道德的東西就難很多

And if people were to use the app

如果人們要用這個應用程式，

and choose not to consult the assistance of a doctor

而選擇不去尋求醫生的協助，

because we get it wrong,

如果程式判斷錯誤的話，

I would feel really bad about it.

我就會感覺非常難過

So we're currently doing clinical tests,

所以我們目前在做臨床實驗，

and if these clinical tests commence and our data holds up,

如果這些實驗開始之後， 我們的資料站得住腳，

we might be able at some point to take this kind of technology

在某個時間點，我們或許可以把這技術

and take it out of the Stanford clinic

應用到史丹佛的臨床課程，

and bring it to the entire world,

甚至把它帶到全世界，

places where Stanford doctors never, ever set foot.

帶到史丹佛的醫生從來不會去的地方

CA: And do I hear this right,

安德森：我沒聽錯吧，

that it seemed like what you were saying,

你的意思聽起來像是

because you are working with this army of Udacity students,

因為你在和這支 Udacity 學生大軍合作，