And researchers have employed machine learning techniques to train an artificial intelligence to figure out-- for itself-- how to draw human faces. Of course, it’s artwork, but is it art?
SALLY WARD-FOXTON: The most interesting thing really happened when they limited the amount of brush strokes that the AI was allowed. They reduced it from a thousand down to 20 strokes, and then the faces started to look like abstract art, which was incredible. But you could see that AI had clearly identified the features that make up the face: the eyes, the nose, the mouth. These abstractions that previously thought you could teach that to an AI using supervised learning with labeled data, but it turns out you can do it with reinforcement learning.
SALLY WARD-FOXTON: 最有趣的事发生在研究人员限制了AI被允许使用的笔触数量时。他们将笔触数从1000个减少到20个，然后这些画像开始看起来像抽象艺术，这简直不可思议。 但是你会看到AI已清楚地识别出构成面部的特征：眼睛、鼻子和嘴巴。在这之前人们认为可以使用带标签数据的监督学习，教会AI识别这些抽象事物，但事实证明，AI可以通过强化学习来做到这一点。
Researchers in artificial intelligence – AI – are experimenting with new techniques to teach machines. A research operation named DeepMind recently got our attention for teaching an AI to draw. The point wasn’t that an AI could learn to draw. That's been demonstrated a long time ago. The point was to find out how effective new machine learning techniques can be. Can they be used to get a computer learn to draw a human face without the absolute minimum of guidance?
The short answer is yes. But as always, we find endless fascination in the engineering details. Here’s Junko again, talking with EE Times editor Sally Ward-Foxton, who wrote the story for us.
答案是肯定的。 但是与往常一样，我们在工程细节上发现了真正引人关注的地方。 再次有请Junko与 EE Times编辑Sally Ward-Foxton来谈谈这个话题，Sally为我们撰写了这个故事。
JUNKO YOSHIDA: Hi, Sally. How are you?
JUNKO YOSHIDA: 嗨，莎莉，你好吗？
SALLY WARD-FOXTON: Good, thanks. Hi, Junko.
SALLY WARD-FOXTON: 嗨，Junko。我很好，谢谢。
JUNKO YOSHIDA: I understand that you are inspired by DeepMind Research's recent project. So I want you to describe what this research project entails.
JUNKO YOSHIDA: 我了解到，DeepMind Research最近的项目对你很有启发。我想请你描述一下这个项目研究的具体内容。
SALLY WARD-FOXTON: Sure, yes. This research is out of DeepMind, a company that Google bought a few years back. The company's overall aim is to build artificial general intelligence, but with this particular project, they've taught an AI agent how to paint. The project was presented at the Deep Learning Summit here in London by DeepMind Research scientist Ali Elami.
SALLY WARD-FOXTON: 好的。这项研究来自DeepMind，这是Google几年前收购的一家公司。该公司的目标是构建通用人工智能，但是通过这个特殊项目，他们已经教会AI代理如何绘画。DeepMind Research科学家Ali Elami最近在伦敦举行的深度学习峰会上介绍了该项目。
The researchers started by teaching AI to understand handwriting, and by extension, how to write. But this new research is much more complex. They basically trained the AI on photos on faces, and they gave access to a drawing program. They used My Paint, which is kind of similar to Photoshop, and they let it control the brush sizes, the colors, the weight of the brush strokes to try to create a realistic looking face by drawing.
研究人员从教AI了解手写开始，然后扩展到如何写作。 但是这项新研究要复杂得多。 他们基本是在面部照片上训练AI，并使用绘图软件。 他们使用的是My Paint，一款类似于Photoshop的软件，并让AI控制画笔的大小、颜色和画笔笔触的粗细，以尝试通过绘画创建逼真容貌。
The results were rather amazing, actually. On eetimes.com I've got some pictures that it drew. They're not photo-realistic, but they do look like drawings of faces. It's not trying to reproduce a target image, you know a specific photograph, rather it's trying to create an image that's different from the data set but looks like it belongs in that data set, if you see what I mean. And the results got even more interesting when the researchers restricted the amount of brush strokes the AI was allowed to use, because the pictures that came out of it started to look like abstract art, which was very cool.
JUNKO YOSHIDA: Let's go down one step deeper here. What sort of specific AI technologies did DeepMind guys use here?
JUNKO YOSHIDA: 让我们在这方面深入聊一聊。 DeepMind的研究人员使用了哪些特定的AI技术？
SALLY WARD-FOXTON: They used reinforcement learning, which is where you use two AI agents working together. An AI agent is just a neural network that takes some kind of action. So these two agents, one's doing the drawing. It's trained on the photos of faces and it tries to recreate them. And the other agent looks at the drawings and produces a feedback score, which is based on how close to the training data we felt that the drawing was. This feedback score is fed back to the agent who did the drawing so that it can improve.
SALLY WARD-FOXTON: 他们利用强化学习，使两个AI代理一起协作。 AI代理只是采取某种行动的神经网络，这两个AI代理中的一个负责绘图。它受过面部照片训练，并尝试重新创作这些照片。另一个AI代理审查图画并产生一个反馈分数，以表明图画与训练数据的接近程度。这一反馈分数将反馈给进行绘图的AI代理，以便可以改进。
The important thing really about reinforcement learning is, it uses unlabelled data, so you train it on these photographs of faces, but you don't say, this is the eyebrow, this is the ear, these are the teeth. It has to identify those features on its own. It has to learn to do that.
JUNKO YOSHIDA: So it's sort of like AI teaching another AI, right?
JUNKO YOSHIDA: 所以有点像一个AI在教另一个AI，对不对？
SALLY WARD-FOXTON: Exactly. Right, yeah.
SALLY WARD-FOXTON: 对，确实是这样。
JUNKO YOSHIDA: You said that the results were quite remarkable. Tell me, what did we learn from this project?
JUNKO YOSHIDA: 你说这项研究的成果非常值得关注。跟我讲讲，我们从这个项目中学到了什么？
SALLY WARD-FOXTON: It produced some very realistic-looking faces. Not specific faces; it just drew something which looks like it could be a face. Of course it hasn't seen a person drawing, or hasn't watched how people draw. But the steps it took in creating the major features like the eyes and nose first in the heavier strokes, and then it added these minor features, maybe like shadow or contour towards the end. All of this was learned by reinforcement learning.
SALLY WARD-FOXTON: 它画出了一些非常逼真的面孔。 没有特定的面孔； 它只是画了一些看起来像一张脸的图。 当然，它还没有看过一个人在绘画，也没有看过人们绘画的方式。 但是，在创建较重的笔触时首要步骤是画出主要特征（如眼睛和鼻子），然后添加这些次要特征，例如在最后添加阴影或轮廓。 所有这些都是通过强化学习来完成的。
The most interesting thing really happened when they limited the amount of brush strokes that the AI was allowed. They reduced it from a thousand down to 20 strokes, and then the faces started to look like abstract art, which was incredible. But you could see the AI had clearly identified the features that make up the face: the eyes, the nose, the mouth. These abstractions that previously thought they could teach that to an AI using supervised learning with labeled data, but it turns out you can do it with reinforcement learning.
最有趣的事发生在研究人员限制了AI被允许使用的笔触数量时。他们将笔触数从1000个减少到20个，然后这些画像开始看起来像抽象艺术，这简直不可思议。 但是你会看到AI已清楚地识别出构成面部的特征：眼睛，鼻子，嘴巴。 在这之前人们认为可以使用带标签数据的监督学习，教会AI这些抽象事物，但事实证明，AI可以通过强化学习来做到这一点。
JUNKO YOSHIDA: The way I see it is like asking an AI to draw is something similar to other projects that I've seen on the web in which I think it was IBM Watson, but scientists for example asked an AI to compose music. And it turns out it actually created a pretty cool music, right? So whether AI's drawing or music is better than what's developed by human beings is really a subjective call. I mean, you can't really say which is better.
JUNKO YOSHIDA: 要求AI绘画，给我的感觉是和我在网上看到的其他一些项目（类似于IBM Watson）相似，比方说科学家要求AI创作音乐。 事实证明，AI确实创造了很酷的音乐，对吗？ 所以，AI无论是绘画，还是音乐，是否都比人类做得好这个问题，实际上是一个主观的判断。 我是说，你不能真正断言哪个更好。
But let me tell you this: I actually draw. I do sketches, I do watercolors in my spare time. For somebody who loves art, I take a little offense when I hear about AI learning to draw. Should I feel threatened by this? Or should be expect to see AI would eventually find a way to provide somebody like myself with some useful AI tools? How do I think about this?
不过，我可以告诉你：实际上我是画画的， 我画素描，业余时间画些水彩画。 对于热爱艺术的人，当我听说AI学习绘画时，我会有点感觉被冒犯了。 我应该感到受威胁吗？ 还是应该期望看到AI最终会找到一种方法，为像我这样画画的人提供一些有用的AI工具？ 我应该怎么看待这事呢？
SALLY WARD-FOXTON: You shouldn't feel threatened by this, that's all. What the AI has produced looks like a drawing. It looks like art, but is it really art? Are the pictures creative? Or are they just random? I mean, what is art? As an electronic engineer, I don't even qualify to answer that question, right? But I think it's generally accepted that art requires thought and the intention to create art, which the AI obviously doesn't have. It's just trying to draw a realistic face. But then it doesn't know what a face is. So it's not trying to suggest any particular expression or mood with its drawing, or it's not trying to elicit an emotional response in the viewer like an artist would. The AI ultimately doesn't understand the subject of the drawing; it doesn't know what that face represents.
SALLY WARD-FOXTON: 你不应该对此感到威胁。 AI创造的东西看起来像一幅图画。它看起来像艺术品，但真的是艺术吗？ 图片有创意吗？ 还是只是随机的？ 我是说，什么是艺术？ 作为电子工程师，我甚至没有资格回答这个问题，对吗？ 但我认为的艺术，也是大众普遍认知的，艺术是需要思想的，需要创造意图的，这显然是AI不具备的。它只是想画出逼真的画像。 但是它并不知道“脸”是什么。 因此，它并不会在绘画中暗示任何特定的表情或心情，也不会像艺术家那样在观众中引起情感上的回应。 AI最终是不了解绘图主题的。 它不知道那张“脸”代表什么。
So no. I don't think this is "art." I think you could argue more successfully that the research team are a group of artists using AI as a tool to create art. And who knows? In the future you may find yourself using AI as a tool, helping you make drawings or helping generate ideas perhaps. Who knows?
所以，我不认为这能称为“艺术”。 我觉得你可以更有力的辩驳，研究团队是一群使用AI作为创作艺术工具的艺术家。 谁知道呢？ 将来，你可能会发现自己将AI用作工具，可能会帮助你绘画或创作。 谁知道呢？
JUNKO YOSHIDA: All right. Very well put. It was good to talk to you, Sally!
JUNKO YOSHIDA: 好的，聊得很愉快。很高兴和你交谈，Sally！
SALLY WARD-FOXTON: You too, Junko. Thank you.
SALLY WARD-FOXTON: 我也很高兴和你聊天，Junko。谢谢。