走出实验室

这里是DeepMind播客，我是汉娜·弗莱。过去一年里，我深入了解了人工智能（AI）领域的最新研究进展。我们采访了众多科学家、研究员和工程师，探讨AI的发展现状与未来方向，追踪这项当今科学界最重大挑战之一的快速演进历程。如果你想在自己的AI探索之旅中获得启发，那么你来对地方了。在本系列节目至今的内容中，我们主要关注了AI在实验室和游戏领域的能力表现。

This is DeepMind, the podcast, and I'm Hannah Fry. Over the last year, I have been getting an inside look at current research into artificial intelligence or AI. And we've been talking to scientists, researchers, engineers about how things stand and where we're headed, tracing the fast moving story of one of the biggest challenges in science today. So if you want to be inspired on your own AI journey, then you've come to the right place. Up until now in this series, we have largely looked at what AI is capable of in the lab or in the world of games.

但AI工作者的终极目标是用技术解决现实世界的问题。当前，人们正尝试运用AI应对各种挑战——从交通拥堵预测到濒危物种监测。DeepMind团队也在此领域取得了多项突破。虽然我们可以称之为案例研究，但那样未免太无趣了。不如我们来段精彩的预告片式介绍如何？

But the ambition of people working in AI is to help solve problems in the real world. Right now, are trying to use AI to help with everything from predicting traffic jams to monitoring endangered species. And here at DeepMind, they've also been working on a few things. Now we could call them case studies, but where would the fun be in that? So I'll tell you what, let's have one of those big trailer things, shall we?

本期节目将揭示AI如何有望帮助数千名眼疾患者重见光明。

In this episode, we learn how AI might potentially help save the sights of thousands of eye disease sufferers.

我们获取第一个神经网络的输出结果后，可以像医生为患者做诊断那样对其进行深度解析。

We get the output from the first neural network and we can interrogate that as doctors making decisions for our patients.

助力破解蛋白质折叠之谜。

Help break down the enigma of protein folding.

我们能理解整个过程或其运作原理，从而根据氨基酸序列预测蛋白质结构。

We can learn what the process is or a description of the process so that we can take a sequence and then predict its structure.

对我们日益增长的能源需求产生重大影响。

An impact on our growing energy demands.

如果这些废热不设法处理掉，真的会熔化你的笔记本电脑。

If that waste heat isn't taken care of somehow, it will literally melt your laptop.

首先，我想向你介绍西姆斯·威瑟斯彭。西姆斯来自南卡罗来纳州，现任DeepMind的项目经理，她坚信人工智能能带来的潜在益处。

First, I want you to meet Sims Witherspoon. Sims is originally from South Carolina. She's now the program manager at DeepMind, and she is a big believer in the potential benefits that AI can bring.

当人们想到DeepMind时，如果联想到游戏，他们主要考虑的是研究层面。但试图解决智能问题，本质上就是利用这种智能让世界变得更美好。世界上不乏需要我们解决的问题，有些非常重大，有些相对较小。但对于那些极其复杂的领域——涉及海量数据、无数变量或庞大可能性组合空间的问题——有时对人类大脑而言确实令人生畏。

When people think about DeepMind, if they think about games, they're largely thinking about the research side. But trying to solve for intelligence is literally to use that intelligence to make the world a better place. The world has no shortage of problems we need to solve. We've got some really big ones, some smaller ones. But for the areas of immense complexity, things where there's lots of data and lots of permutations or a huge combinatorial space of possibilities, sometimes that's really daunting for human brains to try to figure out.

现实世界的事物就是如此复杂。

Stuff in the real world is just really complicated.

确实非常复杂。现实世界的数据杂乱无章。但如果我们能借助人工智能在这片混沌中找到路径，就能比仅靠人力更快地解决问题。

Speaker Yeah, it's really complicated. Real world data is really messy. But if we can use AI to try to find a path through that complexity, we can solve our problems faster than we could if we were trying to do it on our own.

这一点在医疗领域体现得最为明显。2016年，DeepMind与伦敦穆尔菲尔德眼科医院NHS基金会信托合作，尝试将深度学习应用于眼部扫描。随着英国视力受损患者数量预计到2050年将翻倍，这项试验可能产生巨大影响。皮尔斯·基恩是穆尔菲尔德眼科医院的顾问眼科医师，同时就职于国家健康研究所，他亲身体验过医生面临的压力。

Nowhere is that more true than in medicine. In 2016, DeepMind partnered up with Moorfields Eye Hospital NHS Foundation Trust in London to try and apply deep learning to eye scans. With the number of people suffering from sight loss in The UK set to double by two thousand and fifty, the trial had the potential to be hugely impactful. Piers Keene is a consultant ophthalmologist at Moorfields Eye Hospital, a National Institute for Health Research, and knows firsthand about the demands placed on doctors.

眼科领域面临的重大难题——不仅在英国，全球皆然——就是我们需要应对的患者数量极其庞大。特别是在英国国家医疗服务体系内，我们每年要处理近1000万次门诊预约，这占整个NHS门诊量的10%，而且过去五年这个数字增长了超过三分之一。

One of the huge problems that we have in ophthalmology, not just in The UK but all around the world, is the huge number of patients that we have to deal with. So in particular, in the National Health Service, we get nearly 10,000,000 clinic appointments in The UK every year. That's actually 10% of all clinic appointments across the whole NHS, and it's a number that's increased by more than a third in the past five years.

为什么增加了？

Why has it increased?

我认为增加的原因是人口老龄化。某些疾病如糖尿病的发病率上升，导致我们需要处理大量与之相关的眼部问题。举例来说，年龄相关性黄斑变性（AMD）是英国、欧洲及北美最常见的致盲原因。仅这一种疾病，在英国每天就有近两百人发展为致盲型AMD。

I think it's increased because of the aging population. Think it's increased because certain diseases like diabetes are on the rise, and so we just have to deal with a lot of eye problems related to that. So for example, I can tell you about just one condition, which is age related macular degeneration, or AMD. And AMD is the commonest cause of blindness in The UK, the commonest cause of blindness in Europe and in North America as well. And for just that one condition, nearly two hundred people develop the blinding forms of AMD every single day just in The UK.

我们面临的挑战是这些患者需要得到及时诊治。以2016年为例，我工作的伦敦摩尔菲尔德眼科医院收到了七千例社区转诊的疑似湿性AMD（该病的致盲类型）紧急病例。但在这七千例转诊中，只有八百名患者实际患有该疾病的严重类型。

And so the challenge that we have is that those people have to be seen and treated in an urgent fashion. The problem then is that in, for example, in 2016, Morfields Eye Hospital in London, where I work, received seven thousand urgent referrals from the community as possible wet AMD, the blinding form of this condition. But of those seven thousand referrals, only eight hundred patients actually had the severe form of the disease.

面对如此大量的紧急转诊，患者等待专科医生数周已成必然。在此期间，成千上万的人误以为自己患有威胁视力的眼疾而忧心忡忡（实际无需担心），同时数百名本可治愈的患者视力却在等待中持续恶化。皮尔斯向我讲述了他的患者伊莱恩·马诺尔的案例。

With so many people getting urgent referrals, it's inevitable that patients are going to have to wait weeks to be seen by a specialist. And during that time, you have thousands of people who think they've got an eye disease that will threaten their sight, who don't actually need to worry, and hundreds of people with a curable condition whose sight could be saved but is slipping away while they wait. Peers told me about one of his patients, Elaine Manor.

十多年前，在有效疗法出现之前，伊莱恩的左眼就因黄斑变性完全失明。2013年，她开始出现健眼视力模糊。社区验光师检查后告知：‘您的健眼可能正在发展AMD，现在已有有效疗法，需立即就诊。’但她获得的NHS医院预约却在六周之后。

Elaine lost her sight from macular degeneration in her left eye completely more than ten years ago before there was good treatment. And in 2013, she started to develop blurring of vision in her good eye. She went to her high street optometrist who looked into her eye and said, I think you're developing AMD in your good eye. You need to be seen and treated urgently because now we have good treatments for this. She got an appointment to another hospital in the NHS and it was six weeks later.

您能想象吗？当您在家眼睁睁看着健眼视力衰退，明明有现成疗法却被告知要等待六周？如果这是我的家人，我会希望他们在六天内——而非六周后——得到治疗。

So can you imagine if you're at home and you're losing your sight in your good eye and you're told you have to wait six weeks when there is a treatment that's available? And if that was my family member, I would want them treated in six days, not in six weeks.

这时人工智能就派上用场了，原理很简单：转诊至摩尔菲尔德的患者已由社区医生或验光师拍摄了眼底图像。这些被称为OCT扫描的二维/三维图像能显示50种不同眼病。若用人工智能先筛选图像并分诊患者，就能标记重症患者，让他们更快就诊、更早确诊治疗，从而可能挽救视力。

Here is where the AI comes in, and the idea is simple. Patients who are referred to Morfields will already have had pictures taken of the back of their eye by their doctor or the optometrist in their high street opticians. These are two and three-dimensional images known as OCT scans that can show up any one of 50 different diseases. If you can use artificial intelligence to filter through those images first and triage the patients, you can flag the people with a serious disease, get them in front of a doctor sooner, give them an earlier diagnosis, earlier treatment, and potentially save their sight.

在人工智能和深度学习取得近期成功之前，传统编程方法若要实现例如识别猫照片的功能，需要编写大量代码来描述猫的特征：猫有胡须、猫有尾巴。接着你还要考虑有些猫没有尾巴，有些猫没有毛发等等。你需要尝试编写成千上万行代码来描述这些特征。而采用深度学习时，就无需如此繁琐。

Before artificial intelligence and before the recent successes of deep learning, the traditional approach to programming an algorithm to recognize a photo of a cat, for example, would be you would write all the code to describe the features of a cat. Cat has whiskers, cat has a tail. Then you'd say some cats don't have a tail, some cats don't have fur, etc, etc. And you try and write thousands or hundreds of thousands of lines of code to describe that. With deep learning, don't do that.

通过深度学习，我们只需向神经网络展示大量样本——通常是成千上万张猫的图片，它就能自行提取关键特征并学会识别猫。我们现在用同样的方法处理眼部疾病。

With deep learning, we show many examples, often thousands or hundreds of thousands of pictures of cats to a neural network and it will extract the features of interest itself and learn how to recognize a cat. We simply do the same thing but with eye diseases.

对机器而言，识别猫和识别眼底图像本质上没有区别。

The machine doesn't care whether it's looking at cats or the backs of eyes, it's the same thing.

确实，很大程度上是这样。我认为这次合作之所以成功，是因为Moorfields眼科医院作为全球历史最悠久、规模最大的眼科医院之一，我们拥有海量的OCT扫描数据来训练这些神经网络。

Yeah, to a large extent, yes. And I think the reason why this collaboration has been successful so far is because Moorfields Eye Hospital is one of the oldest, one of the largest eye hospitals in the world, and we have huge numbers of OCT scans to train these neural networks.

那这套系统有效吗？

Does it work then?

我认为成果令人惊叹，简直难以置信。我们开发的算法在OCT扫描分诊方面，已达到与Morpheus机构世界顶尖专家同等的水平。

I think the results are amazing. I think they're jaw dropping. I think that the algorithm we've created is on a par with world leading experts at Morpheus in triaging these OCT scans.

但识别猫图片和诊断眼疾存在显著差异。面对如此重要的工作，如何确保算法判断正确？当医生不认同算法结论时，如何让他们确信自己看到了算法所见，或有足够把握推翻算法判断？皮尔斯告诉我，关键在于构建不仅能告知发现、还能展示发现过程的人工智能。为此，这个AI系统需要两个神经网络而非一个。

But there is quite a big difference between spotting cats pictures and picking out eye disease. For something so important, how do you know the algorithm is getting it right? How can a consultant be sure to see what the algorithm saw or feel confident to overrule it if they don't agree. Well, the key, Piers told me, is about building an AI that doesn't just tell you what it's found, but also shows you. And to do that, the AI needs not one neural network, but two.

第一个神经网络经过训练，能够识别扫描图像中的所有疾病特征，而第二个神经网络则被训练利用这些特征对扫描结果做出诊断。

So the first neural network is trained to identify all the disease features on the scan, and the second neural network is trained to take those disease features and to use them to make a diagnosis on the scan.

首先它会遍历图像，标记出那些看起来不完全正常、显得可疑的区域，然后第二个网络介入，解释所有这些区域的情况，并据此做出最终判断。

So it's first going through and kind of highlighting areas that don't look totally normal, that looks suspicious, and the second one's coming in, explaining what's going on in all of those, and using that come to a final decision.

没错，正是如此。

Exactly, yeah.

而且你能看到第一个神经网络标记出的所有区域。

And you can see all those areas that that first neural network has highlighted.

是的，这正是这种方法的重大优势之一。我们不仅能获取第一个神经网络的输出结果，还能以医生为患者做决策的方式对其进行验证。比如视网膜出血或液体渗漏导致视网膜水肿时，它会高亮显示所有这些特征。因此当诊断患者患有糖尿病眼病时，你能看到促成该诊断的典型特征，这给使用该系统的医疗专业人员带来了极大的信心保障。

Yes, so that's one of the great advantages of this approach. We get the output from the first neural network, and we can interrogate that as doctors making decisions for our patients. So if you've got bleeding in the retina, or if you've got leakage of fluid and water logging of the retina, it will highlight all of those features. So if you see that a patient has diabetic eye disease, then you can see the very typical features that have led it to make that decision, which gives a lot of, I think, reassurance for health care professionals who would be using this.

这种方法具有双重价值：不仅能增强临床医生的诊断信心，辅助他们完成现有诊断工作，更有望推动人类对眼睛本身认知的突破。2018年，另一组研究人员尝试通过视网膜图像的深度学习来预测患者性别。眼科医生最多只能做到50%的猜测准确率，但令人震惊的是，该算法的正确率高达97%。全球眼科专家至今无法理解算法从照片中识别出了什么特征，也没有任何理论能解释男女眼球结构可能存在差异。

There's a double whammy with this approach. Not only can it reassure the consultants, helping them with the diagnosis they're already doing, but there is hope that the AI might one day also be able to advance our understanding of the eye itself. In 2018, another group of researchers decided to see if they could use deep learning on images of the retina to predict the sex of the patient. Now the best an eye doctor could manage would be a fifty fifty guess, But to their astonishment, the algorithm got 97% right. No ophthalmologist in the world has any idea what it is that this algorithm is picking up on in the photograph or any theory as to why the male and female eye might be structurally different.

但AI确实发现了人类尚未理解的规律。在摩尔菲尔德的这项研究中，即便算法诊断错误，它可能仍捕捉到了专业医生未曾注意到的细节。

But the AI has found something that they're now trying to understand. And in the Moorfield study, even when the algorithm gets the diagnosis wrong, it might still be picking up on something that the professionals hadn't spotted.

有趣的是，当我们查看算法出错的案例时，实际上不得不退一步思考，因为其中一些案例似乎非常模棱两可、具有挑战性——可能算法给出了正确答案，而我们的黄金标准本身都值得商榷。真的吗？我们得到的结果简直令人震惊。

What was interesting was that when we looked at the cases that the algorithm got wrong, we actually had to take a step back because it seemed like some of those cases were very ambiguous, challenging cases where maybe the algorithm had made the right answer and our gold standard was at least open to debate. Really? So really kind of like jaw dropping the results that we were getting.

确实令人震惊。这就是人工智能在医学领域的初步尝试。但科学中最基础的问题之一呢？

Jaw dropping indeed. So that's AI dipping its toe into the world of medicine. But how about one of the most fundamental problems in science?

当我询问一位资深研究员他认为生物学中最重要的问题时，他列出的首要问题是理解大脑及其运作机制。他认为第二重要的问题是理解蛋白质如何折叠。

When I spoke to a very senior, researcher about what he thought were the most significant problems in biology, he his top problem was understanding the brain and how that works. His second problem that he thought was the most important was understanding how proteins fold.

这是DeepMind科学项目的产品经理桑迪·尼尔森。正如桑迪所言，蛋白质的重要性怎么强调都不为过——它是5000万篇科学论文中最常被提及的主题。

This is Sandy Nelson, a product manager for DeepMind's science program. And as Sandy told me, it's hard to overstate the importance of proteins. It is the most cited topic in 50,000,000 scientific papers.

我们在讨论疾病时熟悉的许多术语都与蛋白质有关。免疫系统的运作依赖蛋白质，激素调节人体众多功能也离不开蛋白质。当然还包括分子与蛋白质的相互作用，但蛋白质对疾病研究的重要性远不止于此。比如阿尔茨海默症等神经退行性疾病就与蛋白质直接相关。

So many of the terms we're used to when we're thinking about medical conditions are underlying proteins. We think about the immune system and how that works, well that's proteins. We think about hormones, we know that regulates so many functions in our body. Course, have a lot to do about molecules interacting with proteins, but there are many other ways in which proteins are important for thinking about disease. So we know for example Alzheimer's and some of those neurodegenerative diseases are to do with proteins, or proteins are implied.

蛋白质是所有生命系统的基石。如果完全展开，它们只是由氨基酸组成的长链，像丝带一样。但蛋白质会自我折叠，通过肽键形成巨大的三维结构。蛋白质可能的折叠方式数量极其庞大——就像折纸艺术，但这是令人难以置信复杂的人类微生物折纸，存在10的300次方种可能性。

Proteins are the building blocks of all living systems. Stretched out straight, they're just big, long chains of amino acids, a bit like a ribbon. But they fold in on themselves and make these giant three-dimensional structures stuck together with peptide bonds. Now the number of different ways a protein could fold is vast. Think origami here, except mind bogglingly complicated human microbiological origami with 10 to the power of 300 possibilities.

科学家们极其关注这些最终折叠成型的蛋白质的确切形状。

And scientists care a great deal about exactly what shape those final folded proteins end up as.

蛋白质之所以能参与众多生化过程，部分原因在于它们具有高度特异性。它们能精准作用于特定过程的关键节点，这种特异性源于其独特的空间构型。当我们把蛋白质视为活体生物中执行各种功能的分子工具时，若想理解某些蛋白质为何出错或设计干预方案，掌握从氨基酸序列到空间结构的形成机制，就是设计蛋白质或解析错误原因的第一步。

Part of the reason proteins are so useful for taking part in so many biochemical processes is because they're specific. They can target very, very specific points in some process. That specificity comes from the uniqueness of their shape. So when we think about proteins are the go to molecule for anything you need to do in a living animal, if you want to try and understand why some of the proteins have gone wrong, or to create some kind of intervention, understanding that process of creating structure from sequence is a first step on maybe designing proteins or understanding why it might go wrong.

无论是感光蛋白、抗体还是酶，蛋白质的功能都由其独特的三维结构决定。核心问题是：蛋白质如何从线性链状转变为最终折叠构型？那么研究目标是什么？是否意味着最终要实现——当我给出氨基酸序列时，你就能预测其空间结构？

The function of the protein, whether it's to detect light in the eye or fight disease or speed up reaction rates, is determined by its unique three-dimensional structure. The question is, how does the protein go from one state to the other, from the ribbon to the final folded structure? What is the objective here then? Is it that in the end, you want to create something where I tell you a sequence of amino acids and you tell me what the structure will look like?

从最基础层面来说确实如此。如果能达到实验室测定级别的准确度，将节省大量人力物力。

So at its simplest level, yes. If you could do it as accurately as it can be done in a lab, that saves a huge amount of effort.

理论上只需观测最终折叠结构即可。最常用的方法是让X射线穿透蛋白质晶体，通过衍射图谱反推结构。但这极其困难——每个蛋白质结构解析需耗费数十万美元，耗时数月甚至数年。事实上，马克斯·佩鲁茨仅因解出血红蛋白结构就获得了1962年诺贝尔奖。

In theory, you can just observe the shape of the final folded structure. The most common way of doing this is by bombarding crystals of the protein with x rays and inferring its shape from the way that these beams are scattered. But that is hard to do. It can cost hundreds of thousands of dollars for each protein structure and take months or even years of work. It's so hard, in fact, that Max Perrets won a Nobel Prize in 1962 just for figuring it out for one single protein, hemoglobin.

不过还有另一条路径。蛋白质的最终结构实际上由氨基酸链的组分特性决定，包括每个氨基酸所受的作用力和电荷分布。理论上可以用物理定律预测蛋白质折叠过程，但这需要海量计算。假设有超级计算机级别的算力，输入氨基酸序列就能输出结构预测。但问题在于现有计算能力远远不足。

There is an alternative, though. The final structure of the protein is actually determined by the chain of component parts, the forces and charges that are acting on each of those individual amino acids. So in theory, you could use the physics to predict how the protein ribbon is going to fold, but it's gonna take a lot of number crunching. So if you had a gigantic enough computer, you know, super computer level, I could give you a string of amino acids and you could tell me what shape it would end up as. But the problem is that we just don't have the computing power to crunch through it.

还达不到全面模拟所有作用力的水平。虽然能用化学物理原理解释折叠机制，但由于分子尺度和复杂度，作用力数量庞大到无法完整建模。

Not at the level of modeling all the forces. So we can explain why the protein folds the way it does using our understanding of chemistry and physics, But because of the size and the complexity of the molecules, there are so many forces, we can't model everything.

这正是人工智能的用武之地。

And here's where AI comes into it.

我们认为存在另一个抽象层次，或许能找到对所有作用力的概括性描述。单纯通过分析难以触及这一层次，但借助庞大的数据集——比如已知某种序列会以特定方式折叠且这一规律可靠——我们或许能通过机器学习理解这些过程或其描述机制，从而根据序列预测其结构。结构。

We think that there's another level of abstraction where we we think we can maybe find a summary description of all those forces. And that's again too hard to come at through analysis, but maybe we can learn that because we've got a huge data set which says, well, this sequence folds this way, and we know that that's reliably the case. So using machine learning, maybe we can learn what the processes or description of the process, so that we can take a sequence and then predict its structure. Structure.

这里有一个目标非常明确的问题：准确预测氨基酸链将如何折叠。而达成这一目标的可能路径数量极其庞大。人工智能正是破解这种复杂性的绝佳工具。唯一的问题在于，即便有人工智能相助，这些过程如此复杂，仅凭物理学原理仍无法清晰预测蛋白质的折叠方式。不过幸运的是，有个技巧可以简化问题，为AI提供先发优势。

Here is a problem with a very clear objective: correctly predict how a chain of amino acids is going to fold. And a vast, vast number of possible ways to get there. AI is perfectly placed to cut through that complexity. The only problem is that even with AI on your side, these things are so enormously complicated that you still can't cut a clear path to predicting how a protein might fold based only on the physics. Thankfully though, there is a trick that you can use to simplify the problem and give your AI a head start.

蛋白质的多样性反而有助于限定问题范围。不过作为数学家，我必须坦言这些内容理解起来相当困难。我会尽量放慢节奏为大家讲解。蛋白质和生物体一样，有着漫长的进化历史。有时它们可能只是氨基酸链上微小的随机突变。

The fact that proteins are so diverse can help you constrain the problem. Although, I should warn you, as a mathematician, I found this stuff pretty hard to get my head around. So I'm gonna try and walk you through it nice and slowly. Proteins, like organisms, have a long evolutionary history. They can sometimes be small random mutations in the string of amino acids.

偶尔会出现这样的情况：突变蛋白质与正常版本仅在一个角落存在差异，但若将其展开为氨基酸链，会发现突变标记出现在多个位置。想象你手中握着折叠成复杂形状的缎带，用记号笔在某个角落点一下。当展开缎带时，会看到墨迹沿其长度分布在多个位置。因此逆向思考时，如果从展开的缎带出发，发现几处异常标记暗示着某种一致性突变，就能确定无论蛋白质最终如何折叠，这些墨迹在成品中必定彼此相邻——这就是重要线索。

Every now and then, a mutant protein will differ from its normal version on just one of its corners, where if you unraveled it back into the ribbon of amino acids, the markers of that mutation would show up in more than one spot. You can imagine this as though you've got your folded ribbon scrunched up in some complicated shape in your hand, and then you take a felt tip pen to one corner of it. Now if you unfolded your ribbon and flattened it out, you would see that the pen would have stained various spots along its length. So working backwards then, if you start with a flattened ribbon and notice something strange in a few different places, marks that hint a consistent mutation, you know that however the protein ends up being folded, you found a big clue. Those stains must have to be next to each other in the final protein.

收集所有这些线索后，问题就大大简化了。这就像面对浩瀚的可能性景观时，你正在筑墙来划定探索范围？

Collect up all of those clues, and you have greatly simplified your problem. Is it like you've got this this vast sort of landscape of options, and you're trying to build walls to pen yourself in?

没错。正因为这些蛋白质体积庞大，可能折叠成无数种形状。我们需要找到能排除大量形状的线索，从而聚焦于更少量的可能性。

Yes. That's exactly right. Because these proteins are so large, they could fold in so many different shapes. So we need to find clues that allow us to eliminate a whole mass of shapes so we can concentrate just on a much smaller number.

你正在把问题规模缩小。

You're making the problem smaller.

没错。

That's right.

您正在收听的是DeepMind团队制作的播客。每两年会举办一次名为CASP（蛋白质结构预测关键评估）的大型蛋白质折叠竞赛。在为期三个月的比赛中，来自全球的学者们通过算法预测氨基酸结构。这些特定氨基酸的结构已通过传统观测手段得到确认，因此可以评判谁的预测最接近真实情况。2018年，DeepMind携其人工智能程序AlphaFold参加了比赛。

You're listening to a podcast from the people at DeepMind. Now every two years, there is a big protein folding competition called CASP, Critical Assessment of Structure Prediction Competition. Over the course of three months, academics from around the world compete to predict the structures of amino acids using algorithms. The structures of these particular amino acids have already been confirmed through traditional observation, so it's possible to judge who comes closest. And in 2018, DeepMind entered its AI program, AlphaFold.

我们研究了其他人进行蛋白质折叠的方法，发现他们如何利用进化信息。传统做法是采用二元约束条件，即判断两个氨基酸应该接触还是不应接触。而DeepMind的创新在于，我们计算了这些氨基酸之间不同距离的概率分布。这本质上意味着我们试图保留更多信息，或者说学习到一个能更好描述氨基酸邻近关系的函数。

We had a look at how other people did protein folding and we saw how they used evolutionary information. And what other people had been doing was they'd been looking at a sort of binary constraint, which said these two amino acids should be in contact or shouldn't be in contact. Whereas what DeepMind did is we looked at the probability of different distances between those amino acids. So that's really like just saying, well, we try to retain some more information or learn a better function for describing that relationship between proximity of amino acids.

所以用你构建的围栏来比喻的话——我现在把这个类比延伸得有点远了。

So in terms of the fences that you're building on your landscape, and I'm going quite far with this analogy now.

没关系，这是个很好的思路。

That's fine. That's a very good plan.

你们当时是在确保不丢失任何信息。

You were making sure that you weren't throwing away any information.

正是如此。我们设置的'围栏'定义更为精妙，或者说边界更清晰——它们其实不太像围栏，反倒更像小山丘。天啊，'小山丘'这个词用得好。完全正确。

That's right. Our our fences were, more subtly defined or a bit more clearly delineated or they were less fence like and more like just a sort of Hillocks. Oh, my god. Hillock. That's right.

很好。但这实际上让预测变得更准确了。

Nice. But that actually ended up making the prediction more accurate.

是的，没错。这是一个非常非常复杂的函数，但我们成功掌握了它。一旦我们能够学会保留这些额外信息，这就成为了我们系统取得成功的关键因素之一。

Yes. That's right. So it's it's a very, very complicated function, but we were able to learn that. And so once we're able to learn to kind of essentially retain that extra information, That's one of the key things that made our system more successful.

问题简化后，人工智能就能发挥所长。在令人屏息的三个月里，DeepMind AlphaFold团队致力于将氨基酸序列转化为三维折叠结构的预测。

With the problem reduced, the AI could get to work, doing what it does best. For three nail biting months, the DeepMind AlphaFold team worked on the turning sequences of amino acids into predictions of three-dimensional folded shapes.

我们没有任何明确信号能表明我们的表现如何。可以看到在很多情况下我们并不完美，因此很难判断我们相对于其他人的表现。当时有太多杰出的研究者发表了优秀成果。直到我们真正通过系统化评估，才非常非常艰难地弄清楚了我们的实际水平。

We didn't have any strong signal which told us how well we were performing. We could see that we were not perfect in many cases, So it was very hard to find out how well we were doing compared to other people. There were so many fantastic researchers that were publishing great results. Until we actually went through that sort of organised assessment, very, very hard to really figure out how well we were doing.

终于，决定性时刻

And then finally, it was the moment

到

of

来了。在给定的43条氨基酸链中，该团队成功最精确地预测出其中25条的结构。排名第二的团队仅预测正确3条——这个结果在任何标准下都令人震惊。你有点轻描淡写了，因为当这个结果公布时我和几位学者交流过，在DeepMind所有成果中，这是最让科学界振奋的一项。

truth. Of the 43 strings of amino acids they were given, the team came closest to correctly predicting the structure for 25 of them. The team that came in second only managed three, a staggering result by anyone's standards. You're kind of downplaying this because I, I was talking to a few academics when this this result came out, of all of the results that have come out of DeepMind, this is the one that's got the scientific community most excited.

是的。这是因为这是一个经典的科学领域。这是许多科学家长期研究的重大科学挑战，因此众多科学家对此深表关切。他们能看到其潜在影响，而且众所周知这是个非常棘手的问题。

Yes. And and that's because this is a classical scientific domain. It's a grand challenge in science that many people have worked on. So it's something that many, many scientists care about deeply. They can see what the potential, impact is, and it's been known to be a very hard problem.

所以我们能够在研究超过五十年的难题上实现阶段性突破。

So we've been able to make a step change on a hard problem that's been worked on for over fifty years.

那么从干预措施来看，这只是生物学家和科学家会为之兴奋的蓝天研究吗？比如理解蛋白质？还是说它最终可能对现实生活产生实际影响？

In terms of interventions then, is this just something that biologists and scientists will get very excited about in terms of kind of blue sky research, understanding protein? Or is it something that could end up having an impact in real people's lives?

我认为这与所有生物医学研究类似。它是基础性的，因此具有巨大影响力，本质上会影响许多方面，但需要转化为具体应用才能对人们生活产生直接影响。例如药物研发过程，部分环节在实验室进行且非常抽象，完全涉及化学反应，但最终会生产出我们能购买或开具的药物，从而影响生活。而这个研究就处于该过程的起始阶段。

So I guess it's similar to all sort of biomedical research. It's fundamental, so it has huge leverage, essentially will affect many, many things, but it needs to be translated into something specific for it to have immediate impact on people's lives. So, for example, if we think about the drug discovery process, part of that process goes on in labs and is very abstract and all to do with chemistry, but ultimately that process does produce medicines that we can buy or prescribe which ultimately will affect our lives. So this is at the start, the early part of that process, for example.

虽然蛋白质折叠的长期影响可能改变我们所有人，但它并非大多数人日常直接面对的话题。而我们共同面临的另一个问题是气候变化。还记得之前提到的西姆斯吗？她的团队决定专注于一个具体气候挑战——数据存储的能源消耗。他们从寻找能源浪费严重的环节入手。

Although the long term implications of protein folding have the potential to impact all of us, it's not exactly a topic that most of us are coming face to face with on a daily basis. But one issue that we are all facing is that of climate change. You remember Sims who you met earlier? Well, her team decided to focus their efforts on one specific climate challenge, energy consumption in data storage. And they started by looking for a place we're burning far more energy than we need to.

因为事实证明，你的电子邮件正是导致地球变暖的因素之一。

Because it turns out your emails are one of the things that are warming the planet.

想想我们每天在网上的活动，无论是发邮件、谷歌搜索，还是在YouTube看狗狗视频。要知道最热门的视频是猫咪视频，但对我来说还是狗狗视频更吸引人。

If you think about the things that we all do online every day, whether that's sending an email, doing a Google search, looking at dog videos on YouTube. You know, the number one video is cat videos. But for me, it's dog videos.

还有我。还有我。不错。

And me. And me. Nice.

所有这些都需要计算能力。你知道，我们发送的信息、存储的数据、传播的信息，所有这些都要通过一个物理空间，也就是数据中心。而这些我们在互联网上依赖的所有操作都需要消耗大量能源。

All of that requires compute power. And, you know, the information that we, you know, we send, data that we store, when information is disseminated, all of that runs through a physical space, I e, data center. And it takes a lot of energy to do all of those actions that we rely on on the Internet.

因为那里确实有实际的物理仓库，存放着所有这些猫咪视频。

Because there I mean, there are actual sort of physical warehouses that are holding all of those cat videos.

哦，是的。它们确实是物理空间。如果你想想它们消耗的能源量，一个大型工业级的数据中心消耗的能源相当于一个小城镇。我的意思是，这些设施规模庞大，运行需要大量能源，冷却也同样需要大量能源。

Oh, as they're yes. They are physical spaces. And if you think about the amount of energy they consume, a data center, you know, in a large industrial kind of setting can consume the same amount of energy as a small town. I mean, these things are massive, and they require a lot of energy to run. They also require a lot of energy to cool.

你收件箱里躺着的所有邮件、你同时播放的四条狗狗视频、你为下载这个播客向服务器发送的请求——每一样都需要某个数据中心的计算能力。整体而言，数据中心现在消耗着全球3%的能源，相当于凭空冒出一个新国家的能耗量。

All those emails sitting in your inbox, the four dog videos you're streaming simultaneously, the request you sent to the server to download this very podcast, every one of those things requires computing power in a data center somewhere. Collectively, data centers now use 3% of the world's energy, the equivalent of a whole new country that just popped up on

地图就在几年前。所有这些计算都会产生热量，大量的热量。想象一下当你在线看奈飞或四个视频时笔记本电脑有多烫，再把这个热度乘以一百万倍。如果这些废热不以某种方式处理，它真的会熔化你的笔记本，对数据中心来说则会熔毁服务器。这就是为什么你的笔记本装有风扇。

the map a few years ago. And all that computing generates heat, lots and lots of heat. If you imagine how hot your laptop gets when you're streaming Netflix or the four videos online, imagine that but multiply it times a million. If that waste heat isn't taken care of somehow, it will literally melt your laptop, or in the case of the data center, it will melt your server. That's why your laptop has a fan.

这就是为什么数据中心需要冷却系统。我们必须将它们保持在特定温度以防熔化，这样你我才能从YouTube上看到狗狗视频。

That's why data centers have cooling that needs to happen there. We have to keep them at a temperature so they don't melt, and you and I can get our dog videos off YouTube.

我猜光是给这些数据中心降温就要消耗大量能源。

And I guess just cooling down those data centers takes up a vast amount of energy.

确实如此。要知道，我们讨论的冷却系统规模堪比公交车，才能维持它们的低温运行。

Yes. It does. We are you know, we're talking about chillers required that are the size of buses in order to to keep them cool.

而这正是人工智能的用武之地。

And this is where AI comes in.

想象一下，你正在试图控制数据中心的冷却系统，而通常由设施管理员或数据中心操作员负责的人类操作员，只有两个调节旋钮来控制整个中心。这显然是极度简化的场景——比如一个风扇和空调。

So imagine you were trying to control the cooling of a data center, and a human being who, you know, is usually a facility manager, data center operator, just has two kind of dials to control. And that was all you had to do to control the entire center. Now, that is vast oversimplification. Like a fan and air cool. Yeah.

没错。就这两个选项。你可以推算出最佳方案——是用空调？用风扇？还是两者都用？

Exactly. Just those two. You could figure out the the best, you know, is it just air con? Is it just fan? Is it both?

或者都不用？毕竟选项并不多，对吧？这很容易判断。但当系统变成无数设备，每个都有可调节的参数，这些参数互相影响时，突然间你就面临数十亿种可能的组合。

Is it neither? Like, that's not that many options. Right? You could figure that out. But when it turns into a huge number of pieces of equipment with set points on every single one, which are all things you can change by some degree that then interact, all of a sudden, you've got a vast number, literally a number of options that is in the billions.

这对设施管理员、数据中心操作员或任何人类来说都太复杂了。因此我们认为这正是AI的理想应用场景——它能处理远超人脑容量的海量信息，帮我们从无数排列组合中找出最优解。但AI如何穿透这种复杂性？我们可以要求模型在保持数据中心特定温度的同时降低能耗，并告诉它所有可调控的系统参数。

And that's just too much for a facility manager, a data center operator, a human being to try to control. So this is where we think AI is, it's the perfect space for AI because AI can ingest a vast amount of information more than the human brain can and can help us figure out which of those permutations, which of those combinations actually is the optimal path forward. How does AI cut through all of this complexity, though? We can ask a model to figure out, okay, we want to keep the data center at a certain temperature, but we want to use less energy to do that. Here are all of the ways that you can manipulate this system.

请弄清楚这一点。场景看起来可能大不相同

Please figure it out. The setting might look quite different to

虽然与国际象棋或围棋不同，但核心理念完全相同。你有一个非常明确的目标：在尽可能少消耗能量的情况下保持中心冷却，而AI需要从海量可能性中找到实现路径。一旦成功，AI会告诉你整个中心所有调节阀应该如何设置。

a game of chess or Go, but the principal ideas here are exactly the same. Again, you have a very clear objective, namely, keep the center cool while using as little energy as possible, and a vast, vast number of possibilities of how to get there that the AI has to find a path through. And once it does The AI tells you how all of the dials should be set across the across the center.

没错。具体要调整哪些设定点，调整幅度是多少。效果如何？确实有效。这才是最棒的部分。

Exactly. What set points to change and by how much to change them. And does it work? It does work. That's the best part about it.

我们通过直接AI控制验证了这一点——即获取这些建议并让AI直接反馈到数据中心的物理基础设施中，经过多重安全约束后，我们发现冷却谷歌数据中心所需的能源减少了30%。

We saw that with direct AI control, I. E. Getting those recommendations and having AI feed them directly back into the physical infrastructure of the data center, going through lots of safety constraints, we saw a 30% reduction in the amount of energy required to cool Google data centers.

这非常惊人。

Which is massive.

这非常惊人，确实是个令人振奋的数字。

Which is massive and a really exciting number.

现在我得承认，我看过那张图表。是的，就是当AI直接控制所有调节阀时的效果图。简直令人震惊——原本起伏不定的能耗曲线变得平稳。

Now I have to confess, I've seen the graph. Yeah. Of what happened when you put the AI in in direct control of all of the dials. And it is staggering. I mean, you've got this this sort of bumpy line that goes along about how much energy is being used.

然后这一切看起来就像那些英镑暴跌的图表一样。

And then it's all it looks like those graphs where the pound crashes.

哦，那是

Oh. That's

真是个糟糕的消息。它就像从悬崖上跌落。然后你会看到它在图表底部徘徊，这时你把AI的控制权收回，切换回人工控制。它就会立刻反弹回原来的位置。这太神奇了。

terrible news. It just drops off a cliff. And then you kind of have it bumping along the bottom of the graph, at which point you you take the AI away from being in control and it's a switch back over to human control. And it just jumps straight back up to where it was before. It's amazing.

AI现在正在管理数据中心的冷却系统吗？

Is the AI running the cooling system in the data centers right now?

是的，这太棒了。既然我们已经证明这套系统运行良好，希望未来能推广到更多领域。随着数据量和实践经验的增加，换句话说，AI会随时间推移变得更出色。30%已经是个惊人的数字，但它还在持续提升。规则和启发式方法不会随时间进步，但AI可以。

Yes, is, which is fantastic. And hoping to roll them out to even more in the future now that we've proved that this works and works well. With more data, with more practice, in other words, the AI gets better over time. So 30% is a fantastic number, but it's increasing. Rules and heuristics don't get better over time, but AI does.

这些系统最棒的部分就在于此。

That's the best part about these systems.

这就是为什么DeepMind乃至全球AI实验室对AI潜力如此兴奋的原因。如果你想了解更多关于AI在能源、医疗和科学问题中的应用，或探索DeepMind之外的AI研究世界，每期节目的注释中都有大量实用链接。如果你有认为对其他听众有帮助的故事或资源，请告诉我们。你可以通过Twitter留言或发送邮件至team@podcastatdeepmind.com联系我们。你也可以用这个地址发送对本系列节目的问题或反馈。

And that is why there is so much excitement about AI's potential here at DeepMind and at AI labs around the world. If you would like to find out more about applying AI to energy, healthcare, and scientific problems, or explore the world of AI research beyond DeepMind, you'll find plenty of useful links in the show notes for each episode. And if there are stories or resources that you think other listeners would find helpful, then let us know. You can message us on Twitter or email the team@podcastatdeepmind.com. You can also use that address to send us your questions or feedback on the series.

现在，我们要不要再休息一下？

Now, shall we have another break?