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

翻译人员: chunhua zhang 校对人员: Chen Zou

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

我在谷歌领导着一个 机器智能的项目组，

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

换句话说，利用工程学原理制造出

able to do some of the things that brains do.

能够像人脑一样 完成某些任务的电脑和设备。

And this makes us interested in real brains

这也使我们对人类的 大脑以及神经科学

and neuroscience as well,

产生了兴趣，

and especially interested in the things that our brains do

尤其在那些大脑的表现

that are still far superior to the performance of computers.

比电脑强太多的领域。

Historically, one of those areas has been perception,

长期以来，我们研究的 其中一个领域便是感知，

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

一种将外界事物——

sounds and images --

比如图像或声音—

can turn into concepts in the mind.

转化为大脑内概念的过程。

This is essential for our own brains,

这对我们的大脑很重要，

and it's also pretty useful on a computer.

对计算机的作用也非同小可。

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

例如，我们团队开发的机器感知算法

会根据图片的内容 让你在谷歌相册的图片

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

based on what's in them.

出现在搜索结果中。

The flip side of perception is creativity:

感知的另一方面是创意：

turning a concept into something out there into the world.

将概念变成现实。

So over the past year, our work on machine perception

因此，这些年我们 在机器感知能力方面的工作

has also unexpectedly connected with the world of machine creativity

也意外地跟机器创意以及机器艺术

and machine art.

联系在了一起。

I think Michelangelo had a penetrating insight

我觉得米开朗基罗对感知和创意

into to this dual relationship between perception and creativity.

之间的双重关系有着深刻的见解。

This is a famous quote of his:

他有一句名言：

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

“每一块石头里都藏着一尊雕像，

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

而雕塑家的工作就是去发现它。”

So I think that what Michelangelo was getting at

我想米开朗基罗意思是

is that we create by perceiving,

我们通过感知来创造，

and that perception itself is an act of imagination

而感知本身是想象力的表现，

and is the stuff of creativity.

以及创意的来源。

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

而进行思考、感知和想象的器官，

of course, is the brain.

毫无疑问，就是大脑。

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

我想先简单地谈一谈

about what we know about brains.

我们对大脑的了解。

Because unlike, say, the heart or the intestines,

因为不像心脏或其它内脏，

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

你无法仅仅通过观察 就能看出点什么来，

at least with the naked eye.

至少仅凭肉眼看不出来。

The early anatomists who looked at brains

早期的解剖学家看着大脑，

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

给它的表面结构 取了各种充满想象力的名字。

like hippocampus, meaning "little shrimp."

比如说海马体，意思是“小虾子”。

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

但这些并不能告诉我们

about what's actually going on inside.

大脑里面究竟是怎样工作的。

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

我认为第一个真正对大脑的工作方式

into what was going on in the brain

有所洞悉的人，

was the great Spanish neuroanatomist, Santiago Ramón y Cajal,

是19世纪西班牙 伟大的神经解剖学家

in the 19th century,

圣地亚哥 · 拉蒙 · 卡哈尔 （Santiago Ramón y Cajal），

who used microscopy and special stains

他使用了显微镜以及某种特殊染色剂，

that could selectively fill in or render in very high contrast

有选择性地将大脑中的 单个细胞填充或者渲染上

the individual cells in the brain,

高对比度的颜色，

in order to start to understand their morphologies.

以便了解它们的形态。

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

这些就是他在19世纪

in the 19th century.

完成的的神经元手绘图。

This is from a bird brain.

这是一只鸟的大脑。

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

能看到这些形态各异的细胞，

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

甚至在当时对细胞学说 本身还是新鲜事物。

And these structures,

而这些结构，

these cells that have these arborizations,

像树枝一样分岔，

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

能够延伸到很长的距离——

this was very novel at the time.

这些在当时都是闻所未闻。

They're reminiscent, of course, of wires.

他们让人联想到的，当然是电线。

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

这对于很多19世纪的人 来说是显而易见的，

the revolutions of wiring and electricity were just getting underway.

因为那时电线和电力革命刚刚兴起。

But in many ways,

但是在许多方面

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

拉蒙 · 卡哈尔的神经解剖学 绘画，比如这一张，

they're still in some ways unsurpassed.

从某些方面来说是很卓越的。

We're still more than a century later,

一个多世纪后的我们，仍然在继续

trying to finish the job that Ramón y Cajal started.

尝试完成拉蒙 · 卡哈尔开启的事业。

These are raw data from our collaborators

提供这些原始数据的，是我们来自

at the Max Planck Institute of Neuroscience.

马克斯 · 普朗克 神经科学研究所的合作者。

And what our collaborators have done

他们的工作

is to image little pieces of brain tissue.

是对那些小块的脑组织进行成像。

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

这一整个样品的大小 是1立方毫米左右，

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

而我展示的只是它上面 很小很小的一块区域。

That bar on the left is about one micron.

左边那段比例尺的长度是1微米。

The structures you see are mitochondria

你看到的这个结构

that are the size of bacteria.

是一个细菌大小的线粒体。

And these are consecutive slices

这些是利用这个非常微小的组织

through this very, very tiny block of tissue.

所制作成的连续的切片。

Just for comparison's sake,

我们来做个对比。

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

通常一根头发的直径是 100微米左右。

So we're looking at something much, much smaller

所以我们看到的东西

than a single strand of hair.

比一根头发丝还要细很多。

And from these kinds of serial electron microscopy slices,

通过这些连续的电子显微镜切片，

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

人们可以重构出类似这样的 神经元三维图像。

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

某种程度上，这跟拉蒙 · 卡哈尔 所用的方式是一样的。

Only a few neurons lit up,

我们只对少量的神经元进行了突出显示，

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

否则我们不可能看到任何东西，

It would be so crowded,

因为那样一来画面会很拥挤，

so full of structure,

充满了组织结构，

of wiring all connecting one neuron to another.

充满了各个神经元间 纵横交错的通路。

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

显然，拉蒙 · 卡哈尔 有一点超前于他的时代，

and progress on understanding the brain

接下来的几十年间

proceeded slowly over the next few decades.

人们对大脑的理解进展非常缓慢。

But we knew that neurons used electricity,

但是我们已经知道， 神经元通过电流传导信息，

and by World War II, our technology was advanced enough

而到二战时，我们的技术 已取得了长足的进步，

to start doing real electrical experiments on live neurons

可以开始在活的 神经元细胞上做电流实验，

to better understand how they worked.

以便更好地理解它们的工作原理。

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

而电脑也正是在 这个时候被发明了出来，

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

它的发明是基于对大脑的模拟——

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

也就是阿兰 · 图灵 所称的“智能机器”理念，

one of the fathers of computer science.

图灵是计算机科学的开创者之一。

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

沃伦 · 麦卡洛克（Warren McCulloch）和 沃尔特 · 皮兹（Walter Pitts）看到了

of visual cortex,

拉蒙 · 卡哈尔所画的 大脑视觉皮层，

which I'm showing here.

就是我给你们看的这个。

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

这是负责处理我们视觉信息的大脑皮层。

And for them, this looked like a circuit diagram.

对他们来说，这看起来像一个电路图。

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

在麦卡洛克和皮兹的电路图上，

that are not quite right.

有许多细节并不是那么正确。

But this basic idea

但基本概念是对的，

that visual cortex works like a series of computational elements

他们认为视觉皮层工作起来 就像一系列计算机元件

that pass information one to the next in a cascade,

在同一个层级中传递信息，

is essentially correct.

这一点是对的。

Let's talk for a moment

我们再聊一聊

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

视觉信息处理模型需要做些什么。

The basic task of perception

感知的基本任务就是

is to take an image like this one and say,

抓取这样的图像并且告诉我们

"That's a bird,"

“这是一只鸟”，

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

这对我们的大脑来说非常简单。

But you should all understand that for a computer,

但对一台电脑来说，

this was pretty much impossible just a few years ago.

在几年前，这还是完全不可能的事。

The classical computing paradigm

传统的计算模式

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

很难完成这个任务。

So what's going on between the pixels,

像素、鸟的图像以及“鸟”这个词，

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

这三者之间所产生的联系，

is essentially a set of neurons connected to each other

本质上是在一个神经网络中各神经元

in a neural network,

相互连接的结果，

正如这张图所示。

as I'm diagramming here.

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

这种神经网络可能是生物学上的， 存在于我们大脑视觉皮层里，

or, nowadays, we start to have the capability

或者，现如今我们开始有能力

to model such neural networks on the computer.

在电脑上模拟这种神经网络。

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

我们来看一下它的工作原理。

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

可以将像素想像成第一层的神经元，

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

这实际上就是在 眼睛内部的工作原理——

that's the neurons in the retina.

是视网膜上的神经元。

And those feed forward

然后这些前馈信息

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

通过一层层神经元往下传递，

all connected by synapses of different weights.

这些神经元通过突触彼此连接。

The behavior of this network

这个神经网络的行为

is characterized by the strengths of all of those synapses.

是通过所有这些突触的强度来表达的，

Those characterize the computational properties of this network.

也塑造了这个网络的计算性能。

And at the end of the day,

最终，

you have a neuron or a small group of neurons

一个或者一小群神经元

that light up, saying, "bird."

会亮起来，说，“鸟”。

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

接下来我会将这三部分——

the input pixels and the synapses in the neural network,

输入的像素，神经网络中的突触，

and bird, the output --

以及“鸟”，这个输出结果——

by three variables: x, w and y.

用三个变量来表示：x、w和y。

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

在那张图片上可能会有一百万个x——

a million pixels in that image.

代表一百万个像素点。

There are billions or trillions of w's,

然后有几十亿或几万亿的w，

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

代表着神经网络中所有突触的权重。

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

只有很少数量的y，

of outputs that that network has.

代表整个网络的输出结果。

"Bird" is only four letters, right?

“Bird（鸟）"这个单词 只有四个字母，对吧？

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

我们假定这只是一个很简单的公式

x "x" w = y.

x 乘以 w 等于 y。

I'm putting the times in scare quotes

我把乘号打上了引号，

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

因为实际的过程要复杂得多。

is a very complicated series of mathematical operations.

牵涉到一系列非常复杂的数学运算。

That's one equation.

这是一个方程式，

There are three variables.

有三个变量。

And we all know that if you have one equation,

而我们知道在一个方程式中

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

通过两个已知数 你就能算出另一个未知数。

So the problem of inference,

所以这道推论题，

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

即判断出图中是一只鸟，

is this one:

可以这样来描述：

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

y是未知数，w跟x都是已知数。

You know the neural network, you know the pixels.

也就是神经网络和像素是已知的。

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

实际上这是一个相当简单的问题。

你只需要用2乘以3，就完事儿了。

You multiply two times three and you're done.

I'll show you an artificial neural network

我会给你们展示我们最近 完成的人工神经网络，

that we've built recently, doing exactly that.

它的工作原理正是如此。

This is running in real time on a mobile phone,

这是在一台在手机上 实时运行的神经网络，

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

当然，令人惊叹的是它自身的运算能力，

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

每秒钟可以进行 几十亿甚至几万亿次的

运算。

per second.

What you're looking at is a phone

你所看到的是一台手机的

looking at one after another picture of a bird,

相机对准了一张张含有鸟的图片，

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

并且它不只能判断出， “是的，这是一只鸟”，

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

而且还能用这种网络 来判断这些鸟的种类。

So in that picture,

因此在这张图片中，

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

x和w是已知的，y是未知的。

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

当然，我省略了非常复杂的那一部分，

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

也就是我们如何判断出w？

the brain that can do such a thing?

为什么大脑能做出这样的判断？

How would we ever learn such a model?

我们是如何学会这种模式的？

So this process of learning, of solving for w,

在学习以及解出w的过程中，

if we were doing this with the simple equation

如果我们使用简单的等式

in which we think about these as numbers,

将这些都想象成数字，

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

那这道题就简单了： 6 = 2 x W，

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

那么，用6除以2就可以得出答案。

The problem is with this operator.

现在的问题就是这个运算符号。

So, division --

除法——

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

我们用除法是因为它是乘法的逆运算。

但就像我刚才说的，

but as I've just said,

the multiplication is a bit of a lie here.

乘法表述在这里其实不太准确。

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

这是一个非常非常 复杂的非线性运算，

it has no inverse.

它没有逆运算。

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

所以我们要找出一个不使用除号

without a division operator.

就能解出这个方程式的方法。

And the way to do that is fairly straightforward.

其实非常简单。

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

只需要使用一点代数上的小技巧，

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

将6移到等式的右边。

Now, we're still using multiplication.

现在我们仍然使用乘法。

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

而这个0——我们就当它是一个误差。

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

换句话说，如果我们 能用正确的方法解出w，

then the error will be zero.

那么这个误差就为0。

And if we haven't gotten it quite right,

如果我们没有找到正确的答案，

the error will be greater than zero.

那么这个误差就会大于0。

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

所以现在我们可以通过 假设去缩小这个误差，

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

而这正是电脑所擅长的。

比如你最开始假设：

So you've taken an initial guess:

what if w = 0?

如果w = 0呢？

那么误差就为6。

Well, then the error is 6.

如果w = 1呢？误差就变成了4。

What if w = 1? The error is 4.

And then the computer can sort of play Marco Polo,

然后电脑就像玩游戏一样不断测试，

and drive down the error close to zero.

将误差降低到接近于0。

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

这样就逐步逼近了w的值。

通常来说，它不可能获得完全精确的值， 但是经过很多步运算以后，

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

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

我们得到了 w = 2.999， 已经足够精确了。

And this is the learning process.

以上就是这个学习过程。

So remember that what's been going on here

大家回想一下刚刚我们所做的，

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

我们用了很多已知的x和y的值，

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

通过迭代法去解出中间的w，

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

这也正是我们自己 在学习时所使用的方法。

We have many, many images as babies

在我们很小的时候， 会看到很多很多图像，

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

然后有人告诉我们： “这个是鸟，这个不是鸟。”

And over time, through iteration,

经过一段时间的重复，

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

我们解出了w，建立起了 神经元之间的连接。

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

那么现在，我们有了确定的 x和w。再要去解出Y

that's everyday, fast perception.

就会非常快了。

We figure out how we can solve for w,

我们找到解出w的方法，

that's learning, which is a lot harder,

这是一种学习，要困难得多，

because we need to do error minimization,

因为我们要用很多的训练样本，

去将误差最小化。

using a lot of training examples.

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

一年前，我们团队的 亚历克斯 · 莫尔德温采夫

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

决定做一个实验， 看如果给定已知的w和y，

given a known w and a known y.

去解出x，会发生什么。

In other words,

换句话说，

you know that it's a bird,

你已经知道那是一只鸟

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

并且也有一个接受过 鸟类识别训练的神经网络，

那么一只鸟的图像是怎样的呢？

but what is the picture of a bird?

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

我们发现，通过运用相同的 将误差最小化的步骤，

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

加上一个受过鸟类识别 训练的神经网络，

and the result turns out to be ...

我们就可以得到

a picture of birds.

一张含有鸟的图片。

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

这是一张由一个进行过 鸟类识别训练的

that was trained to recognize birds,

神经网络所生成的鸟的图片，

just by solving for x rather than solving for y,

仅仅是通过解出x，而不是y，

and doing that iteratively.

并且重复不断的运行。

Here's another fun example.

这是另外一个有趣的例子

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

是我们团队的迈克 · 泰卡制作的 ，

which he calls "Animal Parade."

他称之为“动物大游行”。

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

这让我想起了威廉 ·肯特里奇的作品，

in which he makes sketches, rubs them out,

他先画一些素描，然后擦掉，

makes sketches, rubs them out,

再画一些素描，再擦掉，

and creates a movie this way.

用这种方法创作了一部影片。

In this case,

在我们这个案例中，

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

迈克在一个旨在识别和辨认

in a network designed to recognize and distinguish

不同种类动物的神经网络中

different animals from each other.

将y变换成各种不同的动物。

这样你就得到了这个奇特的 动物图像的埃舍尔式变换效果。

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

Here he and Alex together have tried reducing

他和亚历克斯还一起尝试了

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

将这些y降低到一个二维空间内，

thereby making a map out of the space of all things

从而将被该神经网络识别出来的

recognized by this network.

所有对象放到一张图上来。

Doing this kind of synthesis

通过这样的合成

or generation of imagery over that entire surface,

或者在整个表面上生成图像，

在表面上不断的变换y， 你就创造出了一种图像——

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

一个包含该神经网络能够 分辨出来的所有对象的视觉图像。

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

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

所有的动物都在这儿， 犰狳在那个点上。

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

你也可以用其它的神经网络 实现类似的目的。

This is a network designed to recognize faces,

这是一个为识别和分辨出不同面孔

to distinguish one face from another.

而设计的神经网络。

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

这里，我们输入一个y值，代表“我”，

my own face parameters.

我自己的面部参数。

And when this thing solves for x,

当它在解出x的时候，

it generates this rather crazy,

就生成了这张集不同视角 于一体，相当不可思议的，

kind of cubist, surreal, psychedelic picture of me

立体的、超现实的、迷幻版本的

from multiple points of view at once.

我的面部图像。

The reason it looks like multiple points of view at once

它之所以看起来像是集不同视角于一体，

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

是因为这个神经网络被设计成将一张脸

of a face being in one pose or another pose,

在不同姿势、不同光线之间产生的

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

模棱两可的地方抹掉了。

So when you do this sort of reconstruction,

因此当你开始这项复原工作时，

如果不利用某种影像引导，

if you don't use some sort of guide image

or guide statistics,

或者统计引导，

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

那么你就会得到一种 令人困惑的多视角的图像，

because it's ambiguous.

因为它是模棱两可的。

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

这就是亚历克斯在复原 我的面部的优化流程中，

用他自己的脸作为 影像引导时所得到的图像。

during that optimization process to reconstruct my own face.

So you can see it's not perfect.

你可以看到它还不是十分完美。

There's still quite a lot of work to do

我们在完善这个优化流程方面

on how we optimize that optimization process.

还有许多的工作要做。

但是通过将我自己的脸 作为渲染过程中的引导，

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

rendered using my own face as a guide.

你已经可以得到一个 更清晰的面孔了。

You don't have to start with a blank canvas

你不需要完全从一块空白的画布

or with white noise.

或白噪音开始。

When you're solving for x,

当你在解出x时，

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

你可以从一个本身已经是 别的图像的x开始。

That's what this little demonstration is.

正如这个小小的展示那样。

This is a network that is designed to categorize

这是一个设计为用来将所有物品——

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

人造结构、动物等进行分类的神经网络。

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

我们从一张云图开始，

and as we optimize,

在优化过程中，

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

这个神经网络正在不停地计算 它在云中看到了什么。

And the more time you spend looking at this,

你花越多的时间盯着这张图，

the more things you also will see in the clouds.

你就会在云中看到越多的东西。

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

你也可以使用面部识别 神经网络去产生迷幻效果，

and you get some pretty crazy stuff.

然后就可以得到这种不可思议的东西。

(Laughter)

（观众笑声）

Or, Mike has done some other experiments

或者可以像迈克做的另外一个实验那样，

in which he takes that cloud image,

他还是利用那张云图，

hallucinates, zooms, hallucinates, zooms hallucinates, zooms.

使它幻化、再放大， 幻化再放大，幻化再放大.

And in this way,

这样一来，

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

我想你就可以得到 这个网络的神游状态，

or a sort of free association,

或者某种自由联想，

in which the network is eating its own tail.

仿佛这个网络正在吞噬自己的尾巴。

So every image is now the basis for,

因此每一张图都是 下一张图的基础，决定了

"What do I think I see next?

“我觉得接下来会看到什么？

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

接下来又会看到什么？ 接下来还会看到什么？”

I showed this for the first time in public

我第一次公开展示这些是在西雅图，

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

为一个团队做的一次名为 “高等教育”的讲座上——

this was right after marijuana was legalized.

刚好就在大麻合法化之后。

(Laughter)

（观众笑声）

So I'd like to finish up quickly

在结束我的演讲前，

by just noting that this technology is not constrained.

我想再提醒各位， 这种技术是不受限的。

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

我给你们看了一些纯粹的视觉实例， 因为它们看起来真的很有趣。

It's not a purely visual technology.

它不是一种纯粹的视觉技术。

Our artist collaborator, Ross Goodwin,

我们的合作者，艺术家罗斯 · 古德温

has done experiments involving a camera that takes a picture,

做了一个实验，他用相机拍了一张照片，

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

然后他背包里的电脑 基于这张照片的内容，

based on the contents of the image.

用神经网络作了一首诗。

And that poetry neural network has been trained

这个作诗的神经网络已经接受过

on a large corpus of 20th-century poetry.

大量的20世纪诗歌的训练。

And the poetry is, you know,

其实我觉得

I think, kind of not bad, actually.

那首诗还不赖。

(Laughter)

（观众笑声）

In closing,

下面，

I think that per Michelangelo,

再回到米开朗基罗那句名言，

I think he was right;

我想他是对的，

perception and creativity are very intimately connected.

感知和创意是密不可分的。

What we've just seen are neural networks

我们刚刚所看到的是一些

that are entirely trained to discriminate,

完全被训练成去区分，

or to recognize different things in the world,

或辨别世上的不同物品，

able to be run in reverse, to generate.

能够逆向运行、成生图像的神经网络。

One of the things that suggests to me

我从中受到的启发之一就是，

is not only that Michelangelo really did see

不仅米开朗基罗真的看到了

the sculpture in the blocks of stone,

石头中的雕像，

but that any creature, any being, any alien

而且任何的生物、任何人、任何外星人，

that is able to do perceptual acts of that sort

只要能够有这样的感知，

is also able to create

也就能够创造，

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

因为它们都运用了截然相同的机制。

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

另外，我想感知和创意决不是

uniquely human.

人类所特有的。

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

我们开始有了可以 完成这些事的电脑模型。

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

这应当不足为奇，因为大脑会运算。

And finally,

最后，

computing began as an exercise in designing intelligent machinery.

电脑运算最开始是作为 设计智能机器的一种练习。

It was very much modeled after the idea

它在很大程度上仿照了我们如何

of how could we make machines intelligent.

让机器变得智能这一理念。

And we finally are starting to fulfill now

而我们也终于开始能够实现

some of the promises of those early pioneers,

图灵、冯 · 诺依曼、

of Turing and von Neumann

麦卡洛克和皮兹

and McCulloch and Pitts.

这些先驱的一些期望了。

And I think that computing is not just about accounting

我觉得电脑不仅仅是拿来计算，

or playing Candy Crush or something.

或者玩游戏的。

From the beginning, we modeled them after our minds.

从一开始，我们就是 仿照大脑来制造它们的。

而它们也赋予了我们能够 更好的理解我们的大脑，

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

and to extend them.

并且拓展其潜力的能力。

Thank you very much.

非常感谢。

(Applause)

（观众掌声）