How computers are learning to be creative | Blaise Agüera y Arcas

譯者: 易帆 余

So, I lead a team at Google that works on machine intelligence;

我在 Google 帶領 一個團隊做機械智慧；

in other words, the engineering discipline of making computers and devices

換句話說，就是制定一些訓練方法，

able to do some of the things that brains do.

讓電腦和裝置能做些大腦做的事。

And this makes us interested in real brains

而這也讓我們對真實的大腦

and neuroscience as well,

以及神經科學產生了興趣，

and especially interested in the things that our brains do

特別是一些我們大腦能做

但電腦仍無法呈現出來的事。

that are still far superior to the performance of computers.

Historically, one of those areas has been perception,

長期以來，機械智慧的 其中一個領域談的就是機械感知，

the process by which things out there in the world --

它是一種轉化的過程——

sounds and images --

像是把聲音和影像——

can turn into concepts in the mind.

轉化成心智上的概念。

This is essential for our own brains,

這是我們大腦必備的能力，

and it's also pretty useful on a computer.

這個能力對電腦來說也很有用。

The machine perception algorithms, for example, that our team makes,

所謂的機械感知演算法， 像是我們團隊做的，

能讓你 Google 相簿裡的照片

are what enable your pictures on Google Photos to become searchable,

based on what's in them.

根據照片裡的東西 把它們變成可以被搜尋的資料。

The flip side of perception is creativity:

感知的另一面是創意：

把概念轉化成另一種東西。

turning a concept into something out there into the world.

所以過去幾年， 我們團隊在機器感知上的努力，

So over the past year, our work on machine perception

has also unexpectedly connected with the world of machine creativity

已經可以把創意與

機器藝術結合在一起。

and machine art.

I think Michelangelo had a penetrating insight

我覺得米開朗基羅對「感知」 與「創意」這兩者之間的關係

into to this dual relationship between perception and creativity.

有一種很透析的看法。

他有一句名言：

This is a famous quote of his:

「每一塊石頭裡都藏著一座雕像，

"Every block of stone has a statue inside of it,

等待雕刻家將它雕塑出來。」

and the job of the sculptor is to discover it."

所以我覺得米開朗基羅 當時的體悟是：

So I think that what Michelangelo was getting at

is that we create by perceiving,

我們的「創意」來自「感知」，

and that perception itself is an act of imagination

而感知本身就是一個想像行為

and is the stuff of creativity.

及創意的來源。

The organ that does all the thinking and perceiving and imagining,

人體中有一個器官 能做出思考、感受和想像，

of course, is the brain.

當然，那就是我們的大腦。

我想先簡單地來談一談

And I'd like to begin with a brief bit of history

about what we know about brains.

我們對大腦認知的歷史。

Because unlike, say, the heart or the intestines,

因為大腦不像我們的心臟或腸道，

you really can't say very much about a brain by just looking at it,

你不能光用看的來瞭解大腦，

光靠肉眼根本看不出個所以然來。

at least with the naked eye.

The early anatomists who looked at brains

早期研究大腦的解剖學家，

gave the superficial structures of this thing all kinds of fanciful names,

在大腦表皮結構上 取了許多稀奇古怪的名字，

like hippocampus, meaning "little shrimp."

例如海馬體，意思是「小蝦子」。

But of course that sort of thing doesn't tell us very much

當然，這樣的命名方式

並沒有讓我們對 大腦的認識有太多的幫助。

about what's actually going on inside.

The first person who, I think, really developed some kind of insight

我認為，第一個有真正深入了解

into what was going on in the brain

大腦如何運作的，

was the great Spanish neuroanatomist, Santiago Ramón y Cajal,

是偉大的西班牙神經解剖學家 桑地牙哥·拉蒙卡哈，

in the 19th century,

他在十九世紀，

who used microscopy and special stains

就已經開始用顯微鏡和特殊染劑

that could selectively fill in or render in very high contrast

把大腦裡的特定細胞篩選出來染色，

the individual cells in the brain,

或以強烈的對比色來觀察細胞，

in order to start to understand their morphologies.

這樣做，是為了瞭解 它們的形態結構。

And these are the kinds of drawings that he made of neurons

這些是他在十九世紀時

in the 19th century.

畫的神經細胞圖，

這一張是鳥的大腦。

This is from a bird brain.

但當時已經可以看到 各式各樣不同的細胞圖片，

And you see this incredible variety of different sorts of cells,

即使細胞的原理 在當時是個相當新穎的概念。

even the cellular theory itself was quite new at this point.

And these structures,

這些結構，

these cells that have these arborizations,

這些樹枝狀的細胞結構，

可以延伸到相當相當長──

these branches that can go very, very long distances --

this was very novel at the time.

在當時來講， 這樣的發現算是相當神奇了。

They're reminiscent, of course, of wires.

當然，它們也會讓人聯想到電線，

That might have been obvious to some people in the 19th century;

這對 19 世紀的人來說， 這樣的比喻可能比較恰當，

the revolutions of wiring and electricity were just getting underway.

因為當時電線和電力的變革 正如火如荼的進行。

But in many ways,

但就很多方面來說，

像拉蒙卡哈這樣的顯微鏡解剖圖

these microanatomical drawings of Ramón y Cajal's, like this one,

they're still in some ways unsurpassed.

現在看來還是很厲害。

We're still more than a century later,

但我們卻在一個世紀後，

trying to finish the job that Ramón y Cajal started.

才想試著去完成 當年拉蒙卡哈的研究。

These are raw data from our collaborators

這些原始資料，來自我們

at the Max Planck Institute of Neuroscience.

馬克斯·普朗克 神經科學機構的合作夥伴。

And what our collaborators have done

而我們的合作夥伴的工作就是

is to image little pieces of brain tissue.

把大腦組織切成 一小片一小片的圖像。

The entire sample here is about one cubic millimeter in size,

整個樣本的大小 大約只有 1 立方毫米，

and I'm showing you a very, very small piece of it here.

我展示給各位看的只有小小的一片。

That bar on the left is about one micron.

你可以看到， 左邊的長度標誌僅有一微米。

The structures you see are mitochondria

各位現在看到的結構是粒線體，

that are the size of bacteria.

大小跟細菌一樣。

And these are consecutive slices

這些連續切片圖，

through this very, very tiny block of tissue.

是由一塊很小的組織中 一片片切出來的。

舉個例子做比較，

Just for comparison's sake,

the diameter of an average strand of hair is about 100 microns.

一根頭髮的直徑 大約有 100 微米。

So we're looking at something much, much smaller

我們在研究的

是比一根頭髮還更細更小的東西。

than a single strand of hair.

而這一系列的電子顯微鏡切片圖像，

And from these kinds of serial electron microscopy slices,

可以組成像這樣的 神經元 3D 立體成像。

one can start to make reconstructions in 3D of neurons that look like these.

這些和拉蒙卡哈 當年的研究相去不遠。

So these are sort of in the same style as Ramón y Cajal.

Only a few neurons lit up,

但只有幾個神經元可以打光，

because otherwise we wouldn't be able to see anything here.

否則我們會看不到東西。

It would be so crowded,

因為空間太壅擠、

so full of structure,

結構太複雜了，

of wiring all connecting one neuron to another.

神經元蜿蜒地一個接著一個。

So Ramón y Cajal was a little bit ahead of his time,

所以，拉蒙卡哈在當時 也算是走在時代的尖端，

但在那之後的幾十年，

and progress on understanding the brain

proceeded slowly over the next few decades.

人類對大腦的認識卻相當緩慢。

But we knew that neurons used electricity,

但我們已經知道 神經元是利用電子傳遞訊號，

and by World War II, our technology was advanced enough

到第二次世界大戰前， 我們的科技已經進步到

to start doing real electrical experiments on live neurons

可以在活體神經元上做電子實驗，

用來更好地理解它們是如何運作的。

to better understand how they worked.

This was the very same time when computers were being invented,

這也是電腦被發明出來的時間，

當初有一個模擬人腦的基礎想法——

very much based on the idea of modeling the brain --

是由艾倫·圖靈所提出， 他稱之為「智能機械」，

of "intelligent machinery," as Alan Turing called it,

one of the fathers of computer science.

他是計算機科學之父之一。

Warren McCulloch and Walter Pitts looked at Ramón y Cajal's drawing

當時沃倫麥卡洛克和華特彼特斯 （人工神經科學家）

of visual cortex,

看到的視覺皮質圖，

which I'm showing here.

就是上面這張拉蒙卡哈的圖片。

This is the cortex that processes imagery that comes from the eye.

這個皮質層是負責把 眼睛傳來的訊號轉換成圖像。

And for them, this looked like a circuit diagram.

他們當時發現， 它看起來像是一張電路圖。

So there are a lot of details in McCulloch and Pitts's circuit diagram

雖然麥卡洛克和彼特斯

在電路圖上有很多細節不太正確，

that are not quite right.

But this basic idea

但這樣的基礎概念，

that visual cortex works like a series of computational elements

視覺皮層的工作原理

that pass information one to the next in a cascade,

像一系列的計算子 在串聯的電路圖上傳遞著資訊，

is essentially correct.

這樣的概念卻是相當正確的。

Let's talk for a moment

我們稍微聊一下，

about what a model for processing visual information would need to do.

產生視覺資訊的模型， 需要做哪些事情。

The basic task of perception

覺察力的基本任務就是

is to take an image like this one and say,

比如說，看到這一張圖片，

就要會判斷出，「這是一隻鳥」，

"That's a bird,"

which is a very simple thing for us to do with our brains.

這對我們大腦來說是很簡單的任務。

But you should all understand that for a computer,

但各位要知道，這對電腦來說

this was pretty much impossible just a few years ago.

在幾年前根本是不可能的事。

The classical computing paradigm

傳統的計算模式

is not one in which this task is easy to do.

根本不太容易跑出來這樣的任務。

So what's going on between the pixels,

所以，像素、

between the image of the bird and the word "bird,"

鳥圖與文字之間，

is essentially a set of neurons connected to each other

一定要有一組彼此連結的神經元

in a neural network,

在神經網路內相互作用著，

就像我這張示意圖。

as I'm diagramming here.

This neural network could be biological, inside our visual cortices,

這張神經網路圖 就像我們的視覺皮質運作原理。

or, nowadays, we start to have the capability

如今，我們已經有能力

to model such neural networks on the computer.

用電腦來模擬這樣的神經網路。

And I'll show you what that actually looks like.

接下來我向各位展示一下， 實際的操作大概是怎樣。

So the pixels you can think about as a first layer of neurons,

圖片的像素你可以把它想像成是 第一層的神經元，

and that's, in fact, how it works in the eye --

實際上，就是眼睛裡面 像素的呈現方式，

that's the neurons in the retina.

像素是透過 視網膜上的神經元做傳遞。

And those feed forward

而這些前饋資訊

會一層一層地傳遞到下一層神經元，

into one layer after another layer, after another layer of neurons,

all connected by synapses of different weights.

全部由不同的「突觸權重」所連結。

The behavior of this network

神經網路的行為

is characterized by the strengths of all of those synapses.

全都由這些突觸的強度所控制。

Those characterize the computational properties of this network.

它們決定了神經網路的計算模式。

And at the end of the day,

最後，

會有一個或一小群的 神經元發出訊號，

you have a neuron or a small group of neurons

that light up, saying, "bird."

辨識出該圖片就是，「鳥」。

Now I'm going to represent those three things --

我現在要來解釋一下這三個元素——

the input pixels and the synapses in the neural network,

輸入的「像素」、 神經網路裡的「突觸」、

and bird, the output --

還有「鳥」這個輸出的字元—— 它們是如何運作的。

by three variables: x, w and y.

它們是由三種變數所組成， x、w 和 y。

There are maybe a million or so x's --

圖片中可能有一百多萬個 x ——

a million pixels in that image.

100 多萬個像素。

There are billions or trillions of w's,

而 w 可能有數十億或好幾兆個，

它們代表著神經網路中 各個突觸的權重。

which represent the weights of all these synapses in the neural network.

And there's a very small number of y's,

而這個網路能輸出的 y

of outputs that that network has.

只有少數幾個。

"Bird" is only four letters, right?

「bird」只有四個字母，對吧?

我們假設它的原理是 一個簡單的公式，

So let's pretend that this is just a simple formula,

x "x" w = y.

x 「乘以」 w = y

I'm putting the times in scare quotes

我把乘法符號用引號標示起來

because what's really going on there, of course,

因為它其實是一個

非常複雜的數學運算概念。

is a very complicated series of mathematical operations.

這個方程式

That's one equation.

There are three variables.

有三個變數，

我們都知道，如果你想要 解開這個方程式，

And we all know that if you have one equation,

you can solve one variable by knowing the other two things.

可以從兩個已知數 交叉算出未知的數。

所以要推斷出

So the problem of inference,

that is, figuring out that the picture of a bird is a bird,

圖片中的影像是一隻鳥，

is this one:

可以用這種方式得知：

it's where y is the unknown and w and x are known.

y 是未知數，而 w 和 x 是已知數。

You know the neural network, you know the pixels.

已知神經網路和圖片像素，

As you can see, that's actually a relatively straightforward problem.

其實可以很直接的就得到答案，

2x3=6，就做完了。

You multiply two times three and you're done.

I'll show you an artificial neural network

我向各位展示一個

我們最近做的人工神經網路，

that we've built recently, doing exactly that.

This is running in real time on a mobile phone,

它可以在手機上做及時的操作，

and that's, of course, amazing in its own right,

當然，手機的運算能力相當驚人，

that mobile phones can do so many billions and trillions of operations

手機每秒

可以做出數十億至上兆次的運算。

per second.

What you're looking at is a phone

你現在看到的是一隻手機

looking at one after another picture of a bird,

正對著一張張的鳥圖拍照，

and actually not only saying, "Yes, it's a bird,"

手機不但可以正確的說出， 「是的，這是一隻鳥。」

but identifying the species of bird with a network of this sort.

還能透過神經網路分類 分辨出這是哪一種鳥。

So in that picture,

所以，在這些圖片上，

the x and the w are known, and the y is the unknown.

x 和 w 是已知，而 y 是未知。

I'm glossing over the very difficult part, of course,

我現在來解釋一下這個 最困難的 「w」，

我們到底是如何算出來的？

which is how on earth do we figure out the w,

the brain that can do such a thing?

為什麼大腦可以做出這樣的判斷？

How would we ever learn such a model?

我們到底是如何學到 這樣的認知模式的？

So this process of learning, of solving for w,

這個學習的過程， 是一個求解 w 的過程，

if we were doing this with the simple equation

如果我們要解這個一次方程式，

in which we think about these as numbers,

當它們都是數字時，

we know exactly how to do that: 6 = 2 x w,

我們都知道如何解 6=2 x w，

我們只要把 6 除以 2 就可以得到答案。

well, we divide by two and we're done.

問題在於這個運算符號，

The problem is with this operator.

So, division --

除法這個符號——

we've used division because it's the inverse to multiplication,

我們會用除法的方式求解， 是因為它跟乘法相反，

但就如同我剛剛提到的，

but as I've just said,

the multiplication is a bit of a lie here.

乘法在這裡有點像是個幌子。

這是非常非常複雜的概念， 它們是「非線性運算」的概念；

This is a very, very complicated, very non-linear operation;

it has no inverse.

無法直接用除的求解。

所以，我們要另外 找個方法來解方程式，

So we have to figure out a way to solve the equation

without a division operator.

而不能直接用除的。

And the way to do that is fairly straightforward.

方法相當簡單，

You just say, let's play a little algebra trick,

可以說，我們只用了點 代數的小技巧，

and move the six over to the right-hand side of the equation.

將 6 移動到等號的右邊。

Now, we're still using multiplication.

如此我們就可以繼續用乘法來運算。

And that zero -- let's think about it as an error.

而等號左邊的零—— 我們把它想像成是誤差。

In other words, if we've solved for w the right way,

換言之，如果要解出 w，

then the error will be zero.

誤差就要變成 0。

And if we haven't gotten it quite right,

如果我們沒找到答案

the error will be greater than zero.

誤差會永遠大於 0。

So now we can just take guesses to minimize the error,

所以，我們現在 只能用猜的來縮小誤差，

and that's the sort of thing computers are very good at.

而這就是電腦非常擅長的地方。

So you've taken an initial guess:

所以，你會從頭開始猜：

what if w = 0?

假設 w=0

那誤差會等於6

Well, then the error is 6.

What if w = 1? The error is 4.

但假如 w=1 呢？誤差等於 4。

And then the computer can sort of play Marco Polo,

接下來電腦有點像是在玩 馬可波羅探索遊戲，

and drive down the error close to zero.

探索到誤差接近零為止。

As it does that, it's getting successive approximations to w.

當它一直探索到零， 那麼 w 就解出來了。

原則上，它會不停探索直到接近零， 但大約經過多次步驟後，

Typically, it never quite gets there, but after about a dozen steps,

we're up to w = 2.999, which is close enough.

我們就能得出 w=2.999， 相當接近了。

And this is the learning process.

這就是電腦學習的過程。

回想一下剛剛發生了什麼事情，

So remember that what's been going on here

is that we've been taking a lot of known x's and known y's

我們有很多已知的 x 和 y，

and solving for the w in the middle through an iterative process.

透過重複迭代的過程解出了 w。

It's exactly the same way that we do our own learning.

而這就是我們人類學習的過程，

We have many, many images as babies

我們從小看了很多圖片

and we get told, "This is a bird; this is not a bird."

被告知「這是鳥」，「這不是鳥」；

And over time, through iteration,

經過了一段時間，不停地重複，

we solve for w, we solve for those neural connections.

我們解出了 w， 產生了神經元的連結關係。

So now, we've held x and w fixed to solve for y;

所以現在，我們的 x 和 w 是固定數，可以解出 y；

that's everyday, fast perception.

這就是我們人類每天 經常性的快速直覺判斷。

We figure out how we can solve for w,

我們搞懂了如何解出 w，

that's learning, which is a lot harder,

而學習本身是一條相當艱辛的路程，

因為為了讓誤差最小化，

because we need to do error minimization,

我們必須使用很多的訓練樣本。

using a lot of training examples.

And about a year ago, Alex Mordvintsev, on our team,

約一年前，我們團隊的 艾力克斯摩文斯夫

決定做個實驗，

decided to experiment with what happens if we try solving for x,

看看如果我們試著給出了 w 和 y， 解出來的 x 會變什麼樣。

given a known w and a known y.

換句話說，

In other words,

you know that it's a bird,

電腦知道它是一隻鳥，

and you already have your neural network that you've trained on birds,

電腦有你給它訓練出來 辨識鳥圖片的神經網路，

但對電腦而言，鳥是怎樣的圖像？

but what is the picture of a bird?

原來，使用一模一樣的 「誤差最小化」程序

It turns out that by using exactly the same error-minimization procedure,

以及訓練出來 用來辨識鳥的神經網路，

one can do that with the network trained to recognize birds,

and the result turns out to be ...

你就能辨識出……

a picture of birds.

這是一張鳥圖，

So this is a picture of birds generated entirely by a neural network

所以，這是一張完全由

訓練辨認鳥的神經網路 自行創造出來的鳥圖，

that was trained to recognize birds,

just by solving for x rather than solving for y,

只要透過不斷地重複解出 x，

and doing that iteratively.

而不是解 y 就可以了。

Here's another fun example.

這裡有另一個有趣的範例。

This was a work made by Mike Tyka in our group,

我們團隊裡的 另外一位組員麥克泰卡，

他稱這些畫為《動物大遊行》。

which he calls "Animal Parade."

It reminds me a little bit of William Kentridge's artworks,

這讓我有點回想起了 威廉肯特基的作品，

in which he makes sketches, rubs them out,

他畫好素描後，擦掉它，

makes sketches, rubs them out,

然後反覆地畫、反覆地擦

and creates a movie this way.

透過這樣的方式， 創造出了一部影片。

In this case,

在這個展示裡，

what Mike is doing is varying y over the space of different animals,

麥可做的就是把不同動物的 y ，

in a network designed to recognize and distinguish

透過設計好的神經網路，

different animals from each other.

彼此辨認並分別出不一樣的動物。

And you get this strange, Escher-like morph from one animal to another.

如此，你就能得到一張像艾雪一樣的 不同動物的變體圖像。

Here he and Alex together have tried reducing

這一張是他和艾力克斯一起完成的，

the y's to a space of only two dimensions,

他們試著減少 y 的數量， 將這些圖案丟到一個 2D 平面上，

thereby making a map out of the space of all things

透過這個網路的辨識，

創造出了這一張有各種動物的地圖。

recognized by this network.

Doing this kind of synthesis

要做出這樣的綜合體，

or generation of imagery over that entire surface,

或透過整張圖面產出圖像，

varying y over the surface, you make a kind of map --

你只要在圖面上給出各式各樣的 y ， 你就能做出一張地圖來——

一張由神經網路辨識出的視覺地圖。

a visual map of all the things the network knows how to recognize.

The animals are all here; "armadillo" is right in that spot.

所有動物都會在這上面， 犰狳就在圖上這個點。

You can do this with other kinds of networks as well.

你也可以透過不同的神經網路， 做出類似這樣的作品，

This is a network designed to recognize faces,

這一張由辨識臉的神經網路

to distinguish one face from another.

所做出來的作品，

And here, we're putting in a y that says, "me,"

這一張是用「我」當作 y ， 所做出來的圖畫，

my own face parameters.

用我的臉當參數。

And when this thing solves for x,

當電腦解出 x 後，

it generates this rather crazy,

它就畫出了這一張相當瘋狂、

kind of cubist, surreal, psychedelic picture of me

有點像立體派藝術、 超現實、迷幻效果的我，

同一張圖卻有不同的視角。

from multiple points of view at once.

The reason it looks like multiple points of view at once

而會有這種「同一張圖 不同視角」的感覺，

is because that network is designed to get rid of the ambiguity

是因為這個神經網路的設計，

of a face being in one pose or another pose,

可以將不同姿勢臉之間的 模糊地帶移除掉，

being looked at with one kind of lighting, another kind of lighting.

透過觀察不同的光源就可以做到。

So when you do this sort of reconstruction,

所以，當你重新製作圖像時，

if you don't use some sort of guide image

如果你沒有使用指導圖，

or guide statistics,

或特定的統計資料，

then you'll get a sort of confusion of different points of view,

那你就能得到來自 不同角度的混合體圖像，

because it's ambiguous.

因為它是模糊的。

This is what happens if Alex uses his own face as a guide image

所以如果艾力克斯 用他自己的臉當作指導圖

在優化過程中重新建造我的臉， 就會產生這樣的圖像。

during that optimization process to reconstruct my own face.

So you can see it's not perfect.

各位可以看到， 這作品還不是很完美，

There's still quite a lot of work to do

在圖像優化的過程方面，

on how we optimize that optimization process.

還有很多工作要做。

但如果用我的臉當指導圖，

But you start to get something more like a coherent face,

rendered using my own face as a guide.

就能漸漸地顯現出比較 條理分明的臉。

You don't have to start with a blank canvas

你不需要從一張空白的畫布

or with white noise.

或用白雜訊畫起。

When you're solving for x,

當你解出 x 後，

you can begin with an x, that is itself already some other image.

你就可以從 x 開始畫起， 因為它本身就有一些圖像。

That's what this little demonstration is.

這個小小的展示 說明了它的運作原理。

This is a network that is designed to categorize

這個網路是設計用來 分辨各種不同的物體，

all sorts of different objects -- man-made structures, animals ...

像是人造結構、動物……等。

Here we're starting with just a picture of clouds,

這一張畫我們是從 雲朵的圖像開始畫起的，

and as we optimize,

當我們把它優化後，

基本上，這個神經網路 正在搞懂它在雲朵中看見了什麼。

basically, this network is figuring out what it sees in the clouds.

And the more time you spend looking at this,

當你看得越久，

the more things you also will see in the clouds.

你就能在雲層中看得越多。

你也可以運用人臉網路 讓它產生幻覺，

You could also use the face network to hallucinate into this,

and you get some pretty crazy stuff.

然後就會跑出相當瘋狂的畫作。

(Laughter)

（笑聲）

Or, Mike has done some other experiments

或者，麥可已經有作出 一些其它的實驗，

他用那張雲朵的圖像，

in which he takes that cloud image,

使電腦產生幻覺、然後放大、 產生幻覺、再放大。

hallucinates, zooms, hallucinates, zooms hallucinates, zooms.

And in this way,

用這樣的方式，

you can get a sort of fugue state of the network, I suppose,

我在想，你就能得到一種 像是在神遊狀態的網路，

or a sort of free association,

或者像是一種無拘束的聯想，

in which the network is eating its own tail.

彷彿神經網路正在吃著自己的尾巴。

So every image is now the basis for,

所以每一張圖像基本上像是正在想：

"What do I think I see next?

「我接下來會看到什麼？

What do I think I see next? What do I think I see next?"

接下來會看到什麼？ 接下來會看到什麼？」

I showed this for the first time in public

我第一次在一個 公眾場合上展示這個影片，

to a group at a lecture in Seattle called "Higher Education" --

是在西雅圖的「高等教育」 機構做演說時展示的，

this was right after marijuana was legalized.

當時剛好是大麻剛合法化的時候。

(Laughter)

（笑聲）

So I'd like to finish up quickly

所以，我快速總結一下，

by just noting that this technology is not constrained.

這項技術並不會受到約束。

我剛剛展示的是純粹的視覺範例， 因為觀察它的變化，真的很好玩。

I've shown you purely visual examples because they're really fun to look at.

It's not a purely visual technology.

它不單只有視覺科技。

Our artist collaborator, Ross Goodwin,

我們的藝術合作者，羅斯谷穎 已經做了一些實驗，

has done experiments involving a camera that takes a picture,

他用相機拍了一張照片，

and then a computer in his backpack writes a poem using neural networks,

然後他背包裡的電腦 會根據圖片上的內容，

透過神經網路，創作出一首詩。

based on the contents of the image.

這個會作詩的神經網路

And that poetry neural network has been trained

是透過大量 20 世紀的詩集 所訓練出來的，

on a large corpus of 20th-century poetry.

And the poetry is, you know,

而做出來的詩，

I think, kind of not bad, actually.

實際上，我覺得還得不錯。

(Laughter)

（笑聲）

In closing,

整體而言，

I think that per Michelangelo,

我在想，米開朗基羅，

I think he was right;

他是對的；

perception and creativity are very intimately connected.

感知和創意的關係是相當緊密的。

What we've just seen are neural networks

我們剛剛看的神經網路，

that are entirely trained to discriminate,

它們是被訓練出來分辯

or to recognize different things in the world,

或辨認世界上不同的東西，

able to be run in reverse, to generate.

也可以反過來，自行創作出東西來。

而我從中所得到的

One of the things that suggests to me

is not only that Michelangelo really did see

不僅有米開朗基羅的啟發：

the sculpture in the blocks of stone,

「看見石頭裡的雕像」，

but that any creature, any being, any alien

還有任何能做出感知活動的 生物、生命、外來物種

that is able to do perceptual acts of that sort

都能透過這樣的方式

被呈現並創造出來，

is also able to create

because it's exactly the same machinery that's used in both cases.

因為這兩者與剛才舉的例子 都有著相同的機制。

Also, I think that perception and creativity are by no means

我也認為，感知及創意

不是只有我們人類獨有。

uniquely human.

We start to have computer models that can do exactly these sorts of things.

我們已經有電腦模式 可以做出相當類似的事。

And that ought to be unsurprising; the brain is computational.

所以不需要感到驚訝； 因為大腦是會運算的。

And finally,

最後，我要說的是，

computing began as an exercise in designing intelligent machinery.

設計智能機器已經開始成為 電腦界的活動。

It was very much modeled after the idea

在如何讓機器更智能的領域方面，

of how could we make machines intelligent.

已經有很多的模式產生。

And we finally are starting to fulfill now

我們終於開始

some of the promises of those early pioneers,

完成一些早期前輩們

像是圖靈、馮諾伊曼、

of Turing and von Neumann

and McCulloch and Pitts.

馬庫洛奇和皮斯的期望。

而我也認為電腦不是只有拿來計算

And I think that computing is not just about accounting

or playing Candy Crush or something.

或玩玩 Candy Crush 而已，

From the beginning, we modeled them after our minds.

回到初衷，我們想要的 是讓電腦能仿效人腦。

它不僅讓我們更了解了人類的心智，

And they give us both the ability to understand our own minds better

並讓我們獲得延伸發展心智的能力。

and to extend them.

Thank you very much.

非常感謝大家。

(Applause)

（掌聲）