人工智能的真相（2023）（2）

you get a billion years of PhD time it potentially saved.
阿尔法折叠对里夫卡这类人的影响是什么
What's the impact been of AlphaFold on people like Rivka?
影响是令人惊奇的
It's been amazing, the impact, and
仅仅在过去的两三年中
even over the last two or three years,
全球超过一百万科学家
over a million researchers around the world
都使用了阿尔法折叠和它预测的结构
have used AlphaFold in its structures.
几乎是全世界所有的科学家都在用
We think that's almost every biologist in the world.
谢谢你 德米斯
Thank you, Demis.
人工智能已经在解答
So AI is already answering really important problems
科学和医学界十分重要的问题了
in the world of science and medicine.
但在创造和艺术领域呢
But what about creativity and the arts?
很长一段时间 艺术家都认为
For a long time, artists assumed that
他们是不会受到人工智能的影响的
they were going to be immune from the impact of AI.
那么艺术家是否会受到人工智能发展的威胁呢
So are artists under threat from developments in AI?
人工智能一定就无法做到人类艺术家所做的事吗
Surely it can't do what a human artist does?
为了帮助我们理解 请欢迎艺术家埃里克·德拉斯
To help us understand, please welcome artist Eric Drass.
欢迎你 诶里克
Welcome, Eric.
诶里克 说一下你的职业
So, Eric, tell us what you do?
我是个视觉艺术家
Well, I'm a visual artist and I've been working with
我和人工智能合作已经五年了
and collaborating with AI for about five years now.
我们将要用人工智能在观众面前
OK. And we're going to use AI to create some pictures live
现场创作一些图片 对吗
in front of this audience, is that right?
我们会尝试一下
We're going to try.
你会用电脑控制这个过程 对吗
And you're going to steer the process through your laptop, right?
我们会向你们征询一些
So, what I need, we're going to ask you for some suggestions
要创造什么图片的建议
of pictures to create.
如果你想给人工智能的创作提建议
So put your hands up if you'd like to suggest something
就把手举起来
that you want the AI to create. OK.
告诉我们 你想看到什么图片 -苹果
Tell us, what picture would you like to see? - Apples.
就一个苹果吗 诶里克你觉得你能做到吗
An apple? Just an apple? OK, Eric, do we think we can do that?
我觉得我们应该可以
I think we can probably do that,
比如这样
yeah.
我都还没走回来
OK. I didn't even manage to get back to you
你就已经画出一个苹果的图片了
and you'd already come up with a picture of an apple!
已经好了 我们已经画了一个苹果了
It's already there! We've already got an apple.
这看起来确实像一个苹果
OK. That certainly looks like an apple.
你觉得我们能把它变得更有趣一些吗
Do you think we can make it a bit more interesting?
也许我们可以看看是否能用乐高积木制♥作♥它
Well, let's see if we can make it out of Lego, perhaps.
我不确定我能猜出来这是个苹果
I'm not sure I would have guessed that that was an apple!
但这确实是你能用乐高积木搭出来的
But, yeah, that's what you can do with Lego.
让我们再用一个例子 让我们看看别人
Let's have another, let's have a look at somebody else.
我们还没去过后面 让我走过来
We haven't been at the back here, so let's just come to the back.
好的
Ok.
你有什么建议 -一只狗
And you? - A dog.
一只狗 你想让他干什么吗 比如要吃的
A dog. Do you want it doing anything? Begging?
追赶一只球 -正在追赶一只球的狗
Chasing after a ball. A dog chasing after a ball.
很好的题目 诶里克轮到你了
Great prompt. Eric, off you go.
看你能不能再画一次
See if you can do the same again.
看你能不能在我走回来之前画出来
See if you can do it before I manage to get back down.
追赶一只球 很逼真
Chasing a ball, photorealistic!
尾巴好像太多了
Ok. There's a lot of tails there!
这只狗确实是有两条尾巴
Yeah, the dog does appear to have two tails! It does!
它有几条腿 诶里克
How many legs has it got, Eric?
很多条腿
Many, many legs!
所以人工智能只是从网上
So is it just going out on the internet,
找了一张图片
searching for that picture
来展示给我们看吗
and it's just pulling that in and showing it to us?
并不是
Well, no.
每一张图片都是由神经网络即时创造的
Each of these images is created dynamically on the fly
此前从未存在的
by the neural network and has never existed before.
这些图片以前都不存在吗
So that image, these images never existed before?
从来没有过 这就是所谓的文本生成图像
Not at all. This is what's known as text to image.
要将文本转换为图像 你需要两个部分
So to turn text into an image, you need two parts.
你需要一个可以理解文本的东西
You need something that can understand text,
以及能够创造图片的东西
and you need something that can make pictures.
如果我们看理解文本的这个第一部分
So if we look at the first part about understanding text,
神经网络的一部分能够告诉你
part of the neural network is able to tell you
它试图创造的图片
how well the picture
和你输入的文字有多匹配
it's trying to create matches the words that you've put in.
在这个例子里 这张图片 配文是
So in this case, this photograph, with the prompt
"篮子里的虎斑猫" 网络给出的分数是93%
"a tabby cat in a basket" gives us like a 93% score.
它对图片与文字的匹配度非常有信心
It's pretty confident that that's what it looks like.
如果我们给这张图片配一个不同的文本
If we give it the same picture with a different prompt
比如"伦敦巴士" 它只会打5%的分数
a London bus, for example, it only says 5%.
所以这里展示的
So what we've got here
就像我们之前听到过的强化学习
is like a reinforcement learning we heard about earlier.
这个信♥号♥♥会回到系统里
This is the signal that goes back into
告诉机器
the system to tell the machine
它试图创造的图片有多接近描述
how well it's doing when it's trying to make a picture.
创造图片的第二部分
Now, the second part is making the picture,
是使用扩散训练
which is using diffusion training.
这会有一些难以理解
Now, this is a bit of a head fry,
但我们会试着解释给你听
but we're going to try and explain it to you.
你拿一张图片
What you do is you take an image
扩散训练
并加入一些视觉噪点
and you add some visual noise to it,
加入一些随机的像素
adding a few random pixels to it.
用来教神经网络理解这些图片之间的关系
And we're teaching the network the relationship between these pictures.
我知道如果我加入一点噪点 图片会变成这样
So I know that if I add a bit more noise, it turns into this.
再加入一些会变成这样 然后你不停加入噪点
And a bit more and a bit more, and you keep adding noise.
图片正在消失
And the picture is disappearing.
图片会一直消失 直到只剩一大♥片♥噪点
And the picture disappears until you've only got a field of noise.
如果你用很多图片训练这个网络
And if you train that on lots and lots of images,
蹲着的虎斑猫
在篮子里的虎斑猫
虎斑猫的近身照
阳光下的虎斑猫
比如很多只猫配上不同的文本
for example, lots and lots of cats with lots of different text,
机器就能够通过每次移除一些噪点
the machine is then able to reconstruct images from the noise
从而从噪点中重建图像
by taking a bit of noise away each time.
所以如果我要尝试"一只虎斑猫"这个文本
So if I'm going to try the prompt "a tabby cat"
并且从纯粹的噪点开始
and I start with just pure noise,
在每一步 网络都会试着移除一些噪点
at each step, the network tries to take away the dots
一只虎斑猫
让图片变得更像一只虎斑猫
that will make it look a bit more like a tabby cat.
然后你不断重复这一过程
And you do that over and over and over again,
直到它给出图像
until the image emerges.
显然这套模型不止用过猫来训练
This model has not been trained only on cats, obviously.
它已经用互联网上数以百万计的图片
It's been trained on millions and millions and millions
连同与之相关的文本训练过了
of images from the internet, along with their associated text.
伦敦巴士 一只鸡的水彩画 切达干酪 下国际象棋
我们训练过巴士 水彩画
So we've got a bus or a watercolour painting
奶酪 人们下棋
or a block of cheese. People playing chess.
一旦它掌握了这些概念
And once it has all these concepts inside,
你就能把它们混合到一起了
you can start blending them together.
举个例子 我可以要求画虎豹猫下棋
So I can ask for a tabby cat playing chess, for example,
因为它明白棋是什么
because it understands what chess is,
明白猫是什么
it understands what cats are.
或是"伦敦巴士上的虎豹猫"
Or a tabby cat on a London bus.
这个随机化然后再去随机化的想法
Now the idea of randomising and then de-randomising,
引入噪点直到图片消失
just introducing noise until the image disappears
然后不知怎么的就能再从噪点恢复
and then somehow being able to kind of un-randomise,
说实话 这是过去几年里
this, to be honest, is one of the weirdest
人工智能最怪异也最绝妙的发展了
and most wonderful developments in AI over the last few years.
这些在后台都有着
And there's quite a lot of mathematics
十分巨大的运算量
going on behind the scenes.
所以尽管人工智能创造的图像是全新的
So even though the image that the AI creates is new,
它又是如何知道要创造什么呢 这也是值得思考的
it's worth thinking about how it knows what to create.
它需要训练数据
And it needs training data.
记住 人工智能需要训练数据
Remember, AI needs training data.
训练数据越多 它就会变得越好
The more training data it has, the better it's going to be.
因此许多艺术家担心他们的画
And so lots of artists are concerned that their images,
他们的创意作品被用于训练人工智能
their creative output, is being used to train AI,
从而能使其创造出如我们刚刚所见的图片
which then is used to create images like the one that we've seen.