自监督学习：智能的暗物质

你和 Ishan Mizrah 合著了文章《自监督学习：智能的暗物质》。请问什么是自监督学习，为什么它是智能的暗物质？

You co-wrote the article 'Self-Supervised Learning: The Dark Matter of Intelligence' with Ishan Mizrah. So let me ask: what is self-supervised learning and why is it the dark matter of intelligence?

我先说暗物质部分。显然，人类和动物在进行一种学习，而我们目前无法用机器或 AI 很好地复现。当今最流行的机器学习方法是监督学习和强化学习，它们极其低效。监督学习需要大量样本来学习任何东西，强化学习则需要数量惊人的试错才能让系统学会任何东西。这就是为什么我们没有自动驾驶汽车。要用监督学习解决难题，你需要大量人工标注。用强化学习，你需要某种方式模拟问题，以便进行大规模学习。为什么大多数青少年能在约 20 小时的练习中学会开车，而自动驾驶汽车即使经过数百万小时的模拟练习，也无法真正学会安全驾驶？显然我们遗漏了什么。很多人直觉认为人类利用背景知识学得更快。但背景知识是如何获得的？这才是大问题。婴儿在生命最初几个月如何通过观察（因为他们几乎无法在世界上行动）学习世界如何运作？他们学到了大量关于世界的背景知识，这可能是我们所谓常识的基础。这种学习不是学习任务，不是为了任何东西而受到强化；它只是观察世界并弄清楚世界如何运作，构建世界模型。我们如何在机器中复现这种学习？自监督学习就是试图复现这种学习的一种尝试。

I'll start by the dark matter part. There is obviously a kind of learning that humans and animals are doing that we currently are not reproducing properly with machines with AI. The most popular approaches to machine learning today are supervised learning and reinforcement learning, and they are extremely inefficient. Supervised learning requires many samples for learning anything, and reinforcement learning requires a ridiculously large number of trials and errors for a system to learn anything. That's why we don't have self-driving cars. To solve difficult problems with supervised learning, you need a lot of human annotation. With reinforcement learning, you need some way to simulate the problem so you can do large-scale learning. How is it that most teenagers can learn to drive a car in about 20 hours of practice, whereas even with millions of hours of simulated practice, a self-driving car can't actually learn to drive itself properly? Obviously we're missing something. It's quite obvious for a lot of people that humans use their background knowledge to learn faster. But how was that background knowledge acquired? That's the big question. How do babies in the first few months of life learn how the world works, mostly by observation, because they can hardly act in the world? They learn an enormous amount of background knowledge about the world that may be the basis of what we call common sense. This type of learning is not learning a task, it's not being reinforced for anything; it's just observing the world and figuring out how it works, building world models. How do we reproduce this in machines? Self-supervised learning is one attempt at trying to reproduce this kind of learning.

所以你关注的是纯粹的观察，甚至不是孩子互动的部分。只是坐在那里看爸爸妈妈走来走去、拿起东西。这就是你说的背景知识。也许甚至不是看爸爸妈妈，只是看着世界运转，睁眼闭眼，世界出现消失的基本信息。你是说，人类之所以能快速学会开车，是因为他们之前多年观察汽车在世界上如何运行，以及基本物体物理等背景知识。

So you're looking at just observation, not even the interacting part of a child. It's just sitting there watching mom and dad walk around, pick up stuff. That's what you mean by background knowledge. Perhaps not even watching mom and dad, just watching the world go by, having eyes open or closed, the very act of opening and closing eyes that the world appears and disappears. And you're saying that in order to learn to drive, the reason humans are able to learn to drive quickly is because of the background knowledge they were able to watch cars operate in the world in the many years leading up to it, the physics of basic objects, all that kind of stuff.

没错。物体的基本物理，你甚至不需要知道汽车如何工作，因为那可以很快学会。我常用的例子是：你开车在悬崖边，由于你对直觉物理的理解，你事先知道如果向右打方向盘，车会向右偏，冲下悬崖，掉下去，不会有好结果。但如果你是一个没有世界模型的表格化强化学习系统，你必须重复掉下悬崖数千次才能明白这是个坏主意，再重复数千次才能学会如何避免，再重复数百万次才能学会在所有情况下避免。

That's right. The basic physics of objects you don't even need to know how a car works, because that you can learn fairly quickly. The example I use very often is: you're driving next to a cliff, and you know in advance because of your understanding of intuitive physics that if you turn the wheel to the right, the car will veer to the right, run off the cliff, fall off the cliff, and nothing good will come out of this. But if you are a tabularized reinforcement learning system that doesn't have a model of the world, you have to repeat falling off this cliff thousands of times before you figure out it's a bad idea, and then a few more thousand times before you figure out how to not do it, and then a few more million times before you figure out how to not do it in every situation you ever encounter.

所以自监督学习仍然需要某种真相来源告诉它。你必须找到一种无需人类帮助或无需大量人类帮助就能从世界获取真相的方法。谜题在于世界提供了多少信号、多少真相，无论是人类世界（如 YouTube）还是更自然的世界。那么信号有多少？

So self-supervised learning still has to have some source of truth being told to it by somebody. You have to figure out a way without human assistance or without significant amount of human assistance to get that truth from the world. So the mystery there is how much signal is there, how much truth is there that the world gives you, whether it's the human world like YouTube or the more natural world. So how much signal is there?

诀窍在于：自监督设置中的信号比监督或强化设置中的信号多得多。这就像我那个蛋糕的比喻，有人称之为‘低蛋糕’。当你考虑让机器预测多少信息以及给机器多少反馈时：在强化学习中，每次试验你只给机器一个标量，告诉它‘你做得好’或‘做得差’，而且只是偶尔告诉它。我说‘你’，可能是宇宙告诉机器，但只是一个标量。因此，没有大量此类反馈的多次试验，你不可能学到非常复杂的东西。监督学习：每个样本你给机器几个比特。比如你在 ImageNet 上训练图像识别系统，有 1000 个类别，每个样本不到 10 比特信息。但自监督学习：理想情况下，你给机器展示一段视频，然后暂停，让机器预测接下来会发生什么。你让机器预测，然后让时间流逝，展示实际发生的情况，希望机器下次能预测得更好。你给机器的信息量巨大，因为那是一整段未来的视频片段。

Here's the trick: there is way more signal in a self-supervised setting than there is in either a supervised or reinforcement setting. This goes to my analogy of the cake, the 'low cake' as someone has called it. When you try to figure out how much information you ask the machine to predict and how much feedback you give the machine: at every trial in reinforcement learning, you give the machine a single scalar, you tell the machine 'you did good' or 'you did bad', and you only tell this to the machine once in a while. When I say 'you', it could be the universe telling the machine, but it's just one scalar. As a consequence, you cannot possibly learn something very complicated without many, many trials where you get many feedbacks of this type. Supervision: you give a few bits to the machine at every sample. Let's say you're training a system on recognizing images on ImageNet, there are 1000 categories, that's a little less than 10 bits of information per sample. But self-supervised learning: ideally, you would show a machine a segment of a video, then stop the video and ask the machine to predict what's going to happen next. You let the machine predict, then let time go by and show the machine what actually happened, and hope the machine will learn to do a better job at predicting next time. There's a huge amount of information you give the machine because it's an entire video clip of the future after the video clip you fed it in the first place.

所以对于语言和视觉，都有一个微妙看似简单的构造，但也许这代表了创造智能所需的东西，即填补空白。在空白中，听起来很蠢，但你能……有可能你可以用这种方式解决所有智能问题，仅靠语言？

So both for language and for vision, there's a subtle seemingly trivial construction, but maybe that's representative of what is required to create intelligence, which is filling the gap. So in the gaps, it sounds dumb, but can you... it is possible you can solve all of intelligence in this way, just for both language?

只是……

Just...

给一个句子然后续写，或者给一个带空白的句子，一些词被遮住，你来填词。对于视觉，给一系列图像并预测接下来会发生什么，或者填补中间发生的内容。你认为仅凭这种作为自监督学习信号的公式，能否解决视觉和语言的智能问题？

Give a sentence and continue it, or give a sentence with a gap, some words blanked out, and you fill in what words go there. For vision, you give a sequence of images and predict what's going to happen next, or you fill in what happened in between. Do you think it's possible that this formulation alone, as a signal for self-supervised learning, can solve intelligence for vision and language?

我认为这是我们目前最好的办法。至于这是否能一路通向人类级别的智能，或者只是猫级别的智能，还不清楚。但在人们提出的所有可能方法中，我认为这是最好的。所以我认为，智能系统填补空白——无论是预测未来、推断过去、还是填补缺失信息——这个想法是核心。比如，我现在就在填补你头后面是什么、你后脑勺长什么样这个空白，因为我对人体构造有基本了解。我不知道你接下来会说什么、什么时候说，也不知道你会不会朝这边或那边转头、往哪个方向看，但我知道你不会突然消失然后出现在三米外的走廊里，因为我知道根据物理规律什么是可能的、什么是不可能的。所以你有一个关于可能和不可能的模型，如果发生了出乎意料的事，你会很惊讶，然后必须重建你的模型。这就是世界模型：它告诉你如何填补空白。根据你的感知提供的关于世界状态的部分信息，你的世界模型会填补缺失的信息，包括预测未来、回溯过去、填补你未直接感知到的事物。这不必是纯通用的视觉或语言信息；你可以具体到像预测你在车道行驶时的控制决策。你有一系列来自车辆的图像，如果你录下了视频，知道车最终去了哪里，你就可以回溯时间，根据视觉信息预测车的行驶。这非常具体，是领域特定的。但问题在于，我们能否提出一种通用的方法来训练机器做这种预测或填补空白。目前，这种方法在自然语言处理领域取得了难以置信的成功。所有现代自然语言处理都是通过自监督方式预训练来填补空白：给出一串词，去掉其中 10%，然后训练一个巨大的神经网络来预测缺失的词。预训练完成后，你可以用学到的内部表示作为输入，进行监督训练或其他操作。这非常成功。在图像方面就没那么成功，尽管有进展，而且依赖于手动数据增强——我们稍后再谈。但视频训练尚未成功：让机器通过观看视频来学习表示视觉世界，还没有人真正成功做到。

I think that's our best shot at the moment. So whether this will take us all the way to human-level intelligence or something, or just cat-level intelligence, it's not clear. But among all the possible approaches that people have proposed, I think it's our best shot. So I think this idea of an intelligent system filling in the blanks—either predicting the future, inferring the past, filling in missing information—you know, I'm currently filling the blank of what is behind your head and what your head looks like from the back, because I have basic knowledge about how humans are made. And I don't know if you're going to say what at which point, whether you'll move your head this way or that way, which way you'll look, but I know you're not going to just dematerialize and reappear three meters down the hall, because I know what's possible and what's impossible according to physics. So you have a model of what's possible and what's impossible, and then you'd be very surprised if it happens, and then you'll have to reconstruct your model. So that's the model of the world: it tells you what fills in the blanks. Given your partial information about the state of the world from your perception, your model of the world fills in the missing information, and that includes predicting the future, retrodicting the past, filling in things you don't immediately perceive. And that doesn't have to be purely generic vision or visual information or generic language; you can go to specifics like predicting what control decision you make when you're driving in a lane. You have a sequence of images from a vehicle, and then you have information if you recorded on video where the car ended up going, so you can go back in time and predict what the car did based on the visual information. That's very specific, domain-specific. But the question is whether we can come up with a generic method for training machines to do this kind of prediction or filling in the blanks. So right now, this type of approach has been unbelievably successful in the context of natural language processing. Every modern natural language processing is pre-trained in a self-supervised manner to fill in the blanks: you show it a sequence of words, remove 10% of them, and then train some gigantic neural net to predict the words that are missing. Once you've pre-trained that network, you can use the internal representation learned by it as input to something you train supervised or whatever. That's been incredibly successful. Not so successful in images, although it's making progress, and it's based on manual data augmentation—we can go into this later. But what has not been successful yet is training for video: getting a machine to learn to represent the visual world, for example, by just watching video. Nobody has really succeeded in doing this.

好的，我们来做个高层次的概述。视觉和语言在种类和难度上有什么区别？你说人们还没能真正用自监督学习攻克视觉问题，但这不一定是因为它本质上更难。也许当我们谈到以图灵测试的完整精神通过图灵测试时，语言可能比视觉更难。这并不明显。那么在你看来，哪个更难，或者它们其实是同一个问题？我们越接近解决每一个，就越意识到它们都是一回事——都是同一块蛋糕。

Okay, well let's give a high-level overview. What's the difference in kind and in difficulty between vision and language? So you said people haven't been able to really crack the problem of vision open in terms of self-supervised learning, but that may not be necessarily because it's fundamentally more difficult. Maybe when we're talking about achieving passing the Turing test in the full spirit of the Turing test, language might be harder than vision. That's not obvious. So in your view, which is harder, or perhaps are they just the same problem? The farther we get to solving each, the more we realize it's all the same thing—it's all the same cake.

我认为我在寻找的是让它们看起来本质上是同一块蛋糕的方法，但目前它们还不是。学习世界模型或预测模型的主要问题是，预测从来不是单一结果，因为世界并非完全可预测。它可能是确定性的或随机的——我们可以深入哲学讨论——但即使是确定性的，也不是完全可预测的。所以如果我播放一个短视频片段，然后让你预测接下来会发生什么，这个视频片段有很多很多合理的延续，而且延续的数量随着你要求系统预测的时间间隔增长而增长。所以监督学习中的一个重要问题是如何表示这种不确定性，如何表示多个离散结果，如何表示连续的可能结果等等。如果你是一个经典的机器学习研究者，你会说：‘哦，你只要表示一个分布就行了，对吧？’我们知道如何在预测文本中缺失的词时做到这一点，因为你可以让神经网络为字典中的每个词给出一个分数——这是一个很大的数字列表，大概十万左右——然后你可以把它们转换成概率分布。这告诉你，当我说一个句子如‘猫在厨房里追空白’时，只有少数几个词有意义：可能是老鼠或激光点之类的。如果我说‘空白在草原上改变空白’，这两个词也有一堆合理的选项。因为你可以参考一个潜在的现实来填补这些空白。所以在草原上，你无法确定是狮子还是猎豹，你无法知道是斑马还是角马。但你可以通过一长串数字来表示不确定性。现在，如果我对视频做同样的事，让你预测一个视频片段，它就不是一组离散的潜在帧。你必须在高维连续空间中表示无限多个合理的多帧延续，而我们根本不知道如何正确做到这一点。

I think what I'm looking for are methods that make them look essentially like the same cake, but currently they're not. The main issue with learning world models or learning predictive models is that the prediction is never a single thing because the world is not entirely predictable. It may be deterministic or stochastic—we can get into the philosophical discussion about it—but even if it's deterministic, it's not entirely predictable. So if I play a short video clip and then I ask you to predict what's going to happen next, there are many, many plausible continuations for that video clip, and the number of continuations grows with the interval of time that you're asking the system to make a prediction for. So one big question with supervision is how you represent this uncertainty, how you represent multiple discrete outcomes, how you represent a continuum of possible outcomes, etc. And if you are a classical machine learning person, you say, 'Oh, you just represent a distribution, right?' And we know how to do that when we're predicting words missing in the text, because you can have a neural net give a score for every word in a dictionary—it's a big list of numbers, maybe a hundred thousand or so—and you can turn them into a probability distribution. That tells you, when I say a sentence like 'the cat is chasing the blank in the kitchen,' there are only a few words that make sense there: it could be a mouse or a laser spot or something like that. And if I say 'the blank is changing the blank in the savannah,' you also have a bunch of plausible options for those two words. Because you have an underlying reality that you can refer to to fill in those blanks. So you cannot say for sure in the savannah if it's a lion or cheetah or whatever; you cannot know if it's a zebra or gnu or wildebeest. But you can represent the uncertainty by just a long list of numbers. Now if I do the same thing with video and I ask you to predict a video clip, it's not a discrete set of potential frames. You have to represent a sort of infinite number of plausible continuations of multiple frames in a high-dimensional continuous space, and we just have no idea how to do this properly.

有限的高维，所以就像……你试图把它缩小到小于一百万的小有限集，类似这样？

Finite high-dimensional, so like... you try to get it down to a small finite set of under a million, something like that?

差不多吧。我的意思是，我们对语言中每一个可能的词都做分布，而且它居然有效，这有点荒谬。

Something like that. I mean, it's kind of ridiculous that we're doing a distribution over every single possible word for language and it works.

这真是个很蠢的做法。似乎应该有某种更压缩的表示来描述词的分布。你说得对，我同意。你有什么有趣的想法，关于如何以压缩的方式表示所有现实，从而可以形成分布？

Like that's a really dumb way to do it. There seems like there should be some more compressed representation of the distribution of the words. You're right about that, and so I agree. Do you have any interesting ideas about how to represent all the reality in a compressed way such you can form a distribution over it?

这是个大问题，你知道，怎么才能做好呢？我的意思是，当前 NLP 文本中自监督方法另一个真正愚蠢——我不该说愚蠢，但过于简单——的地方是，你不仅表示了一个巨大的词分布，而且对于缺失的多个词，这些分布基本上是相互独立的。而且你并没有为此付出太大代价。所以系统在我之前给出的句子中，如果它给狮子、猎豹一个概率，给瞪羚、角马、斑马另一个概率，这两个概率是相互独立的。但事实上它们并不独立；狮子实际上攻击比猎豹更大的动物。所以在这个过程中有一个巨大的独立性假设，这并不成立。原因是我们不知道如何正确表示符号组合序列上的分布，因为数量随符号长度呈指数增长，所以我们不得不使用技巧。但这些技巧可以绕过它，甚至不处理它。所以大问题是：是否存在某种文本的抽象潜在表示，能表明当我将狮子换成瞪羚、狮子换成猎豹时，我也必须将斑马换成瞪羚？

That's one of the big questions, you know, how do you do that right? I mean, what's kind of another thing that really is stupid about—I shouldn't say stupid, but simplistic—about current approaches to self-supervision in NLP in text is that not only do you represent a giant distribution over words, but for multiple words that are missing, those distributions are essentially independent of each other. And you don't pay too much of a price for this. So the system, in the sentence I gave earlier, if it gives a certain probability for a lion and a cheetah, and then a certain probability for a gazelle, wildebeest, and zebra, those two probabilities are independent of each other. And it's not the case that those things are independent; lions actually attack bigger animals than cheetahs do. So there's a huge independence hypothesis in this process which is not actually true. The reason for this is that we don't know how to properly represent distributions over combinatorial sequences of symbols essentially, because the number goes exponentially with the length of the symbols, and so we have to use tricks for this. But those techniques can get around it, don't even deal with it. So the big question is: would there be some sort of abstract latent representation of text that would say that when I switch lion for gazelle, lion for cheetah, I also have to switch zebra for gazelle?

是的，这个独立性假设……让我向你抛出一个我经常听到的批评，看看你怎么回应。这种填空只是统计；你没有学到任何深层概念。你只是在模仿过去的东西。你没有学到任何可以用来泛化世界的新东西。或者，好吧，让我说粗鲁的版本：这只是统计，不是智能。你怎么看？你通常怎么回应这种说法？

Yeah, so this independence assumption... Let me throw some criticism at you that I often hear and see how you respond. So this kind of filling in the blanks is just statistics; you're not learning anything like the deep underlying concepts. You're just mimicking stuff from the past. You're not learning anything new such that you can use it to generalize about the world. Or, okay, let me just say the crude version: it's just statistics, it's not intelligence. What do you have to say to that? What do you usually say to that if you kind of hear this kind of thing?

我不参与那些讨论，因为它们有点无意义。首先，很可能智能只是统计——只是特定类型的统计。是的，这里有个哲学问题：智能可能只是统计吗？但什么样的统计？所以如果你问：我们学到的世界模型是否具有某种因果概念？是的。所以如果批评来自那些说当前机器学习系统不关心因果的人——顺便说一句，这是错的——我同意他们。你的世界模型应该将你的行动作为输入之一，这将驱使你学习因果世界模型，知道对世界的干预会导致什么结果。或者你可以通过观察其他智能体在世界上行动并观察效果来做到这一点，例如其他人类。所以我认为在某种描述层次上，智能只是统计。但这并不意味着你不会拥有对事物有深层机械解释的模型。问题是如何学习它们。这是我感兴趣的问题。因为很多提出批评的人说那些机械模型必须来自其他地方——必须来自人类设计者，必须来自我不知道什么。显然我们学习它们，或者如果我们作为个体不学习它们，自然通过进化替我们学习。所以不管你怎么想，这些过程都是以某种方式被学习到的。

I don't get into those discussions because they are kind of pointless. So first of all, it's quite possible that intelligence is just statistics—it's just statistics of a particular kind. Yes, where this is the philosophical question: is it possible that intelligence is just statistics? But what kind of statistics? So if you are asking the question: are the models of the world that we learn, do they have some notion of causality? Yes. So if the criticism comes from people who say current machine learning systems don't care about causality, which by the way is wrong, I agree with them. You should have your model of the world with your actions as one of the inputs, and that will drive you to learn causal models of the world where you know what intervention in the world will cause what results. Or you can do this by observation of other agents acting in the world and observing the effect, other humans for example. So I think at some level of description, intelligence is just statistics. But that doesn't mean you won't have models that have deep mechanistic explanation for what goes on. The question is how do you learn them. That's the question I'm interested in. Because a lot of people who actually voice their criticism say that those mechanistic models have to come from someplace else—they have to come from human designers, they have to come from I don't know what. And obviously we learn them, or if we don't learn them as an individual, nature learns them for us using evolution. So regardless of what you think, those processes have been learned somehow.

所以如果你看人类大脑，就像我们人类内省大脑如何工作时，当我们思考什么是智能时，我们想到的是高层的东西，比如我们构建的模型，认知科学的概念，比如记忆和推理模块，几乎像这些高层模块。这是一个好的类比吗？我们是否忽略了暗物质，基本的低层机制，就像我们忽略操作系统的工作方式一样？我们只是在使用高层软件。我们忽略了在低层，神经网络可能在做类似统计的事情——抱歉可能用词不当且粗糙——但做这种填空式学习，不断更新模型，以便能够支持原始感官信息进行预测，然后在预测错误时调整。但当我们从高层看大脑时，感觉就像在下棋，我们在玩高层概念，把它们拼接起来，存入长期记忆。但实际上底层发生的是我们无法内省的东西，就是这种简单的、只是填空的大型神经网络，对吗？

So if you look at the human brain, just like when we humans introspect about how the brain works, it seems like when we think about what is intelligence, we think about the high-level stuff like the models we've constructed, concepts like cognitive science, like concepts of memory and reasoning module, almost like these high-level modules. Is this a good analogy? Are we ignoring the dark matter, the basic low-level mechanisms, just like we ignore the way the operating system works? We're just using the high-level software. We're ignoring that at the low level the neural network might be doing something like statistics—sorry to use this word probably incorrectly and crudely—but doing this kind of fill-in-the-gap kind of learning and just kind of updating the model constantly in order to be able to support the raw sensory information to predict it and then adjust to the prediction when it's wrong. But when we look at our brain at the high level, it feels like we're playing chess, like we're playing with high-level concepts and we're stitching them together and we're putting them into long-term memory. But really what's going underneath is something we're not able to introspect, which is this kind of simple large neural network that's just filling in the gaps, right?

嗯，好吧，这里有很多问题和答案。首先，神经科学中有一个完整的学派，特别是计算神经科学，喜欢预测编码的想法，这与我之前谈到的自监督学习中的想法密切相关。所以一切都是关于预测：智能的本质是预测能力，大脑所做的一切都是试图从一切预测一切。对吧？这实际上是基本原则。如果你愿意，自监督学习试图重现这种预测思想。这是任务无关学习的一个基本机制。下一步是：你对重现哪种智能感兴趣？当然，我们都想重现人类的高层认知过程。但像机器，我们甚至还没有达到重现猫大脑学习过程的水平。最智能的系统也没有家猫那么多的常识。那么猫是如何学习的？猫并不做很多推理；它们当然有因果模型，它们当然有——因为很多猫能弄清楚如何通过行动影响世界来获得……

Well, okay, so there's a lot of questions and answers there. So first of all, there's a whole school of thought in neuroscience, computational neuroscience in particular, that likes the idea of predictive coding, which is really related to the idea I was talking about in self-supervised learning. So everything is about prediction: the essence of intelligence is the ability to predict, and everything the brain does is trying to predict everything from everything else. Okay? And that's really sort of the underlying principle. If you want, that self-supervised learning is trying to kind of reproduce this idea of prediction. That's kind of an essential mechanism of task-independent learning, if you want. The next step is: what kind of intelligence are you interested in reproducing? And of course, we all think about trying to reproduce sort of high-level cognitive processes in humans. But like with machines, we're not even at the level of even reproducing the learning processes in a cat brain. The most intelligent systems don't have as much common sense as a house cat. So how is it that cats learn? And cats don't do a whole lot of reasoning; they certainly have causal models, they certainly have—because many cats can figure out how they can act on the world to get...

它们确实拥有一个绝佳的直观物理模型，当然包括自身身体的动力学，还有猎物等等。它们相当聪明，只用了大约 8 亿个神经元。我们远未达到能复现这种能力的水平。所以某种程度上，我可以说，在我们搞清楚能否复现猫的能力之前，先别担心人类能做的高级认知、长期规划和推理。话虽如此，这种学习世界模型的能力，是让机器也能推理的关键。每次演讲我都会说，机器学习有三大挑战。第一个是让机器学会表征世界，并提出自监督学习。第二个是让机器以与基于梯度的学习兼容的方式进行推理，因为这就是深度学习的核心。第三个是我们不知道如何解决的问题，至少我不知道：能否让机器学习行动计划的层次化表征？我们知道如何用卷积网络和 Transformer 训练它们学习感知的层次化表征，但行动计划呢？能否让它们自发地学习好的层次化行动表征，也是基于梯度的？所有这些都需要一定程度上的可微，这样你才能应用基于梯度的学习，而这正是深度学习的本质。所以，背景知识、以可微方式推理并与背景知识深度整合的能力，然后基于背景知识在世界上制定层次化计划。

They certainly have a fantastic model of intuitive physics, certainly of the dynamics of their own bodies, but also of prey and things like that. They're pretty smart; they only do this with about 800 million neurons. We are not anywhere close to reproducing this kind of thing. So to some extent, I could say let's not even worry about the high-level cognition and long-term planning and reasoning that humans can do until we figure out whether we can even reproduce what cats are doing. That said, this ability to learn world models is the key to the possibility of learning machines that can also reason. Whenever I give a talk, I say there are three main challenges in machine learning. The first is getting machines to learn to represent the world, and proposing self-supervised learning. The second is getting machines to reason in ways that are compatible with gradient-based learning, because that's what deep learning is all about. The third is something we have no idea how to solve, at least I have no idea: can we get machines to learn hierarchical representations of action plans? We know how to train them to learn hierarchical representations of perception with convolutional nets and transformers, but what about action plans? Can we get them to spontaneously learn good hierarchical representations of actions, also gradient-based? All of that needs to be somewhat differentiable so that you can apply gradient-based learning, which is really what deep learning is about. So it's background knowledge, the ability to reason in a way that is differentiable and deeply integrated with that background knowledge, and then given that background knowledge, be able to make hierarchical plans in the world.

在经典最优控制中，有一种叫做模型预测控制的方法，自 20 世纪 60 年代初就存在了。NASA 用它来计算火箭的轨迹。基本思想是，你有一个火箭的预测模型，给定 t 时刻的状态和一个动作，它会给出 t 加Δt 时刻的状态。如果你有这个模型，形式是神经网络或一组可以反向传播梯度的公式，你就可以做基于梯度的模型预测控制。你在时间上展开模型，输入一个假设的动作序列，然后有一个目标函数来衡量系统成功与否。通过时间反向传播，你可以找出最优的动作序列，使系统达到最佳最终状态。这是一种推理形式，基本上是规划。机器人学中的许多规划系统都基于此。例如，一个开车的青少年有一个很好的汽车动力学模型。如果你向右打方向盘，而那里是悬崖，你就会掉下悬崖。你不需要非常精确的模型就能预测到这一点，你可以在脑海中模拟并决定不这样做。你可以想象不同的场景，然后采取最有利场景的第一步，并重复规划过程。这叫做滚动时域模型预测控制。在经典最优控制中，世界模型通常不是学习的；它大多是确定性的，并且是手工构建的。未来十年人工智能的重大挑战是如何让机器学习预测世界模型，处理不确定性和真实世界的复杂性。不仅仅是火箭或机械臂的轨迹，而是其他一切：人的行为、像水或树木这样的物理系统，复杂的事物——人类可以毫不费力地为其开发抽象表征和预测模型，但我们仍然不知道如何用机器做到这一点。

In classical optimal control, there is something called model predictive control, which has been around since the early 1960s. NASA uses it to compute trajectories of rockets. The basic idea is that you have a predictive model of the rocket, which given the state at time t and an action, gives you the state at time t plus delta t. If you have this model in the form of a neural net or some set of formulas that you can backpropagate gradients through, you can do gradient-based model predictive control. You unroll the model in time, feed it a hypothesized sequence of actions, and have an objective function that measures how well the system succeeded. By backpropagation through time, you can figure out the optimal sequence of actions to get the system to the best final state. That's a form of reasoning, basically planning. Many planning systems in robotics are based on this. For example, a teenager driving a car has a good dynamical model of the car. If you turn the wheel to the right and there is a cliff, you'll run off the cliff. You don't need a very accurate model to predict that, and you can run this in your mind and decide not to do it. You can imagine different scenarios and take the first step in the most favorable scenario, then repeat the planning process. That's called receding horizon model predictive control. In classical optimal control, the model of the world is not generally learned; it's mostly deterministic and built by hand. The big challenge of AI for the next decade is how to get machines to learn predictive models of the world that deal with uncertainty and the real world in all its complexity. Not just the trajectory of a rocket or a robot arm, but everything else: people's behavior, physical systems like water or trees, complex things that humans have no trouble developing abstract representations and predictive models for, but we still don't know how to do with machines.

你把世界的博弈论性质放在哪里？你的行动不仅对环境动态做出反应，还会影响它，尤其是在有其他人类参与的情况下。这是第四点，还是以某种方式整合到了你的行动层次化表征中？

Where do you put the game-theoretic nature of this world, where your actions not only respond to the dynamic nature of the environment but also affect it, especially if there are other humans involved? Is this point number four, or is it somehow integrated into the hierarchical representation of action in your view?

我认为它是整合的。只是让事情变得更复杂。人类复杂且不易预测，这使你的世界模型更加复杂。

I think it's integrated. It just makes it more complicated. The fact that humans are complicated and not easily predictable makes your model of the world much more complicated.

就像在纽约这里，想下车时，你减速会传递某种信号，加速也会传递某种信号，这是一种舞蹈，一场异步的棋局。

To get out like here in New York, trying to get out of the car, you slowing down is going to signal something, you speeding up is going to signal something, and that's a dance. It's an asynchronous chess game.

我不知道。所以感觉这不仅仅是……我的意思是，我想你可以把所有这些小互动整合到一个巨大的模型中，因为它不像国际象棋那么复杂。这只是我们做的一个小舞蹈，一起跳个小舞，然后我们就搞定了。

I don't know. So it feels like it's not just... I mean, I guess you can integrate all of them into one giant model, like the entirety of these little interactions, because it's not as complicated as chess. It's just like a little dance we do, a little dance together, and then we figure it out.

嗯，在某些方面它比国际象棋复杂得多，因为它是连续的，以连续的方式不确定。

Well, in some ways it's way more complicated than chess because it's continuous, it's uncertain in a continuous manner.

它并不觉得更复杂，之所以不觉得更复杂，是因为我们就是这样。我们进化来解决这个问题。这是我们进化来解决的那种问题，所以我们很擅长，因为大自然让我们擅长。大自然并没有让我们擅长国际象棋；我们完全不懂国际象棋。

It doesn't feel more complicated, but it doesn't feel more complicated because that's what we are. We've evolved to solve this. This is the kind of problem we've evolved to solve, and so we're good at it because nature has made us good at it. Nature has not made us good at chess; we completely suck at chess.

是的，事实上，这就是为什么我们把它设计成游戏，让它有挑战性。如果说国际象棋和围棋的最新进展让我们意识到什么，那就是人类在这些事情上真的很糟糕，非常糟糕。你知道，在 AlphaGo 之前有个故事，最顶尖的围棋选手认为他们可能只比他们称为“神”的理想选手差两三个子。实际上，不是，他们差九到十子。我的意思是，我们就是不行。所以我们不擅长，这是因为我们的工作记忆有限。我们不擅长做这种树搜索，而计算机比我们擅长得多。但我们更擅长学习世界的可微分模型。我的意思是，我说可微分，在某种意义上……我不应该说可微分是指我们回溯到它，而是指我们的大脑有某种机制来估计某种梯度。

Yeah, in fact, that's why we designed it as a game, to be challenging. And if there is something that recent progress in chess and Go has made us realize, it's that humans are really terrible at those things, like really bad. You know, there was a story right before AlphaGo that the best Go players thought there were maybe two or three stones behind an ideal player that they would call God. In fact, no, they are like nine or ten stones behind. I mean, we're just bad. So we're not good at it, and it's because we have limited working memory. We're not very good at doing this tree exploration that computers are much better at doing than we are. But we are much better at learning differentiable models of the world. I mean, I said differentiable in the kind of... I should say not differentiable in the sense that you know we went back for up to it, but in the sense that our brain has some mechanism for estimating gradients of some kind.

是的，这就是让我们高效的原因。所以如果你有一个智能体，它由一个世界模型组成——在人脑中基本上是整个大脑的前半部分——和一个目标函数，在人类中是两件事的组合：一个是你的内在动机模块，位于大脑底部的基底神经节，它测量疼痛、饥饿等即时感受和情绪；另一个是相当于强化学习中所谓的“评论家”，它是一个预测未来情况结果的模块。所以它不是一个成本函数，也不是一个目标函数，而是一个经过训练的终极目标函数预测器，它也是可微分的。所以如果所有这些——你的成本函数、你的评论家、你的世界模型——都是可微分的，那么你就可以使用基于梯度的方法进行规划、推理、学习，做智能体该做的一切。

Yeah, and that's what makes us efficient. So if you have an agent that consists of a model of the world, which in the human brain is basically the entire front half of your brain, an objective function, which in humans is a combination of two things: there is your sort of intrinsic motivation module, which is in the basal ganglia at the base of your brain, that's the thing that measures pain and hunger and things like that, immediate feelings and emotions; and then there is the equivalent of what people in reinforcement learning call a critic, which is a sort of module that predicts ahead what the outcome of a situation will be. So it's not a cost function, but it's sort of not an objective function, but it's sort of a trained predictor of the ultimate objective function, and that also is differentiable. So if all of this is differentiable—your cost function, your critic, your world model—then you can use gradient-based methods to do planning, to do reasoning, to do learning, to do all the things that an intelligent agent would do.

而基于梯度的学习，你的直觉是什么，这可能是解决智能的核心。所以在你看来，不需要基于逻辑的推理？

And the gradient-based learning, like what's your intuition, that's probably at the core of what can solve intelligence. So you don't need logic-based reasoning in your view?

我不知道如何让基于逻辑的推理与高效学习兼容。是的，好吧，我的意思是有一个大问题，也许是一个哲学问题，我们可以问：是不是我们从工程和计算机科学中知道的所有学习算法都是通过优化某个目标函数来进行的？对吧。所以我们可以问一个问题：大脑中的学习是否最小化了一个目标函数？它可能是多个目标函数的组合，但它仍然是一个目标函数。第二，如果它确实优化了一个目标函数，它是否通过某种梯度估计来实现？它不需要是反向传播，但需要某种高效估计梯度的方法，其复杂度与运行推理本身相当，因为你不能通过扰动大脑中的权重来观察效果。你可以通过扰动来估计梯度，但对我来说，大脑使用某种零阶黑箱无梯度优化似乎非常不可信，因为它比梯度优化低效得多。所以它必须有某种估计梯度的方法。

I don't know how to make logic-based reasoning compatible with efficient learning. Yeah, and okay, I mean there is a big question, perhaps a philosophical question, that we can ask: is it that all the learning algorithms we know from engineering and computer science proceed by optimizing some objective function? Right. So one question we may ask is: does learning in the brain minimize an objective function? It could be a composite of multiple objective functions, but it's still an objective function. Second, if it does optimize an objective function, does it do it by some sort of gradient estimation? It doesn't need to be backprop, but some way of estimating the gradient in an efficient manner whose complexity is on the same order of magnitude as actually running the inference, because you can't afford to do things like perturbing a weight in your brain to figure out what the effect is. You can do sort of estimating gradient by perturbation, but to me it seems very implausible that the brain uses some sort of zeroth-order black box gradient-free optimization because it's so much less efficient than gradient optimization. So it has to have a way of estimating gradients.

有没有可能某种基于逻辑的推理在某些方面作为创建目标函数的有用机制而出现？也许它是一种创建知识库的机制，然后可以查询，就像一种基于梯度方式学习的高效知识表示？

Is it possible that some kind of logic-based reasoning emerges in pockets as a useful mechanism for creating objective functions? Maybe it's a mechanism for creating knowledge bases that can then be queried, like an efficient representation of knowledge that's learned in a gradient-based way?

嗯，所以我认为有很多不同类型的智能。首先，我认为我们想到的那种逻辑推理，可能源于 20 世纪 70 年代和 80 年代的经典人工智能，我认为人类很少使用它，而且并不特别擅长。但我们根据解决这些罕见问题的能力来评判彼此；这叫做智商测试。我想是的。比如我不太擅长国际象棋。

Well, so I think there are a lot of different types of intelligence. First of all, I think the type of logical reasoning that we think about, maybe stemming from classical AI of the 1970s and 80s, I think humans use that relatively rarely and are not particularly good at it. But we judge each other based on our ability to solve those rare problems; it's called an IQ test. I think so. Like I'm not very good at chess.

我一直在评判你，因为实际上，以你的血统，我肯定你擅长国际象棋。没有刻板印象，不是所有刻板印象都是真的。好吧，我国际象棋很烂。

I'm judging you this whole time because well, we actually with your heritage, I'm sure you're good at chess. No stereotypes, not all stereotypes are true. Well, I'm terrible at chess.

所以，你知道，但我认为也许我拥有的另一种智能是这种从推理中构建世界模型的能力，当然也从数据中构建。这些模型通常更偏向类比，对吧？所以它是通过模拟和类比进行推理，你用一个模型来应用于新情况，即使你从未见过那种情况。你可以把它与你之前遇到的情况联系起来，你的推理更类似于某种内部模拟。所以你在模拟当你建造一个木箱之类的东西时会发生什么。对吧？你可以提前想象以特定方式切割木头的结果，你会用螺丝还是钉子等等。当你与某人互动时，你也有那个人的模型，并在心中带着这个模型与那个人互动，告诉对方你认为对他们有用的东西。所以我认为这种构建世界模型的能力基本上是智能的本质，以及利用它来规划满足特定标准的行动的能力。

So, you know, but I think perhaps another type of intelligence that I have is this ability of building models of the world from reasoning obviously, but also from data. And those models generally are more kind of analogical, right? So it's reasoning by simulation and by analogy, where you use one model to apply to a new situation even though you've never seen that situation. You can sort of connect it to a situation you've encountered before, and your reasoning is more akin to some sort of internal simulation. So you're kind of simulating what's happening when you're building, I don't know, a box out of wood or something. Right? You can imagine in advance what would be the result of cutting the wood in this particular way, are you going to use screws or nails, or whatever. When you are interacting with someone, you also have a model of that person and sort of interact with that person having this model in mind to kind of tell the person what you think is useful to them. So I think this ability to construct models of the world is basically the essence of intelligence, and the ability to use it then to plan actions that will fulfill a particular criterion.

所以，如果那是智能的基本暗物质——形成背景模型的能力——你直觉认为需要多少知识？

So if that's the fundamental sort of dark matter of intelligence, this ability to form a background model, what's your intuition about how much knowledge is required?

我甚至不知道如何衡量这个问题的答案。但不管怎样，它都装在大约 8 亿个神经元里。猫的神经元不到 10 亿，狗有 20 亿。乘以 1000 就得到突触数量。我认为几乎所有知识都是通过自监督学习获得的，由饥饿等根深蒂固的目标函数驱动。自然告诉我们饿了，但没告诉我们怎么找食物；那是大脑其他部分要搞定的。

I don't even know how to measure an answer to that question. But whatever it is, it fits in about 800 million neurons. A cat has less than a billion neurons, a dog has two billion. Multiply that by a thousand and you get the number of synapses. I think almost all of it is learned through self-supervised learning, driven by ingrained objective functions like hunger. Nature tells us we're hungry, but not how to feed ourselves; that's something the rest of the brain has to figure out.

有趣的是，可能还有更深层的目标函数在背后。饥饿可能只是大脑试图维持稳态的一个症状。

It's interesting because there might be deeper objective functions underlying the whole thing. Hunger may be just one symptom of the brain trying to maintain homeostasis.

核心可能只是一个非常愚蠢的目标函数，但行为就是这样驱动的。基底神经节驱使我们做不同于猩猩或猫的事情。比如，我们的基底神经节驱使我们寻求陪伴，因为自然发现我们需要社交。猩猩是独居的；它们避开同类。所有这些行为都不是智能的一部分。人们说你永远不会有智能机器，因为人类智能是社交性的，但看看猩猩或章鱼：它们几乎不互动，却仍然非常聪明。我们过分高估了语言作为智能基础的重要性。

It could be just one really dumb objective function at the core, but that's how behavior is driven. The basal ganglia drive us to do things that differ from orangutans or cats. For example, our basal ganglia drive us to seek company because nature figured out we need to be social. Orangutans are solitary; they avoid others. All those behaviors are not part of intelligence. People say you'll never have intelligent machines because human intelligence is social, but look at orangutans or octopuses: they barely interact and still get really smart. We give way too much importance to language as a substrate of intelligence.

所以，要解决家猫级别的智能问题，你觉得可以在荒岛上完成？就一只猫坐在那里看海浪，有正确的驱动力？

So to solve the house cat intelligence problem, you think you could do it on a desert island? Just a cat sitting there looking at the waves, with the right set of drives?

是的，它需要正确的驱动力来学习适当的东西。比如，人类婴儿天生就有学习站立和行走的欲望。欲望是硬连线的；具体怎么做是学来的。硬连线这类东西非常简单。

Yes, it needs the right set of drives to learn the appropriate things. For example, baby humans are driven to learn to stand up and walk. The desire is hard-wired; how to do it precisely is learned. It's very simple to hardwire this kind of stuff.

但走路肯定有更深层的需求。一开始走得很烂，看起来很可怕；爬行安全得多。

But there's got to be a deeper need for walking. It seems scary because you suck at it at first; crawling is much safer.

嗯，你有某种东西驱使你去做，这是人类发展的一部分。这被理解了吗？不完全是。为什么要用两条腿站起来？这真的很难。大多数动物不用两条腿；它们用四条。那稳定得多。这是个非常难的问题。有多少鸟类用两条腿搞定了？严格来说，它们有两只脚。鸡、恐龙——很多都是。我刚知道霸王龙吃草，不吃其他动物。霸王龙可能是个友好的宠物。

Well, you have this thing that drives you to do it, which is part of human development. Is that understood? Not entirely. What's the reason to get on two feet? It's really hard. Most animals don't get on two feet; they get on four. That's much more stable. It's a really hard problem. How many birds have figured it out with two feet? Technically, they have two feet. Chickens, dinosaurs—many of them. I'm just now learning that T-Rex was eating grass, not other animals. T-Rex might have been a friendly pet.

你怎么看 François Chollet 提出的通用智能测试？你对如何解决智商测试有什么直觉？

What do you think about the test for general intelligence that François Chollet put together? What's your intuition about how to solve an IQ-type test?

我觉得这完全不在我的视野范围内，不太相关。短期内，更接近你工作的问法是：如何用极少的样本数据解决 MNIST？答案可能是自监督学习：先学习表示图像，然后在此基础上学习识别手写数字只需要少量样本。我们在人类身上观察到这一点：给幼儿看一本有几张例子的图画书，他们就学会了。

I think it's so outside of my radar screen that it's not really relevant. In the short term, one way to ask closer to your work is: how do you solve MNIST with very little example data? The answer is probably self-supervised learning: learn to represent images, and then learning to recognize handwritten digits on top of that will only require a few samples. We observe this in humans: you show a young child a picture book with a few examples, and they learn.

给孩子看几张大象的照片，他就知道大象是什么了。我们今天在实际系统中也看到了这一点：我们用海量图像训练图像识别系统，要么是完全自监督，要么是非常弱监督。例如，你可以训练一个神经网络来预测人们在 Instagram 上输入的标签。你可以用数十亿张图像来做这件事，因为每天都有数十亿张图像出现，所以训练数据量基本上是无限的。然后你从系统学到的输出中取几层以下的输出表示，将其作为输入提供给任何你想要的目标的分类器，效果相当好。这就是迁移学习，或者说弱监督迁移学习。人们也在用自监督学习快速推进这类场景。我猜这将是自监督学习的未来。

A couple pictures of an elephant and that's it, the child knows what an elephant is. And we see this today with practical systems: we train image recognition systems with enormous amounts of images, either completely self-supervised or very weakly supervised. For example, you can train a neural net to predict whatever hashtag people type on Instagram. You can do this with billions of images because billions appear per day, so the amount of training data is essentially unlimited. Then you take the output representation, a couple layers down from the output of what the system learned, and feed this as input to a classifier for any object in the world that you want, and it works pretty well. So that's transfer learning, or weakly supervised transfer learning. People are making very fast progress using self-supervised learning for this kind of scenario as well. My guess is that that's going to be the future for self-supervised learning.

你认为需要多少清理来过滤恶意信号？比如很多人用 Instagram 标签来做 SEO，但这些标签并不完全代表图像内容。他们会放一张猫的照片，然后打上“科学”、“棒极了”、“有趣”之类的标签……

How much cleaning do you think is needed for filtering malicious signal? Like a lot of people use hashtags on Instagram to get good SEO, but that doesn't fully represent the contents of the image. They'll put a picture of a cat and hashtag it with 'science', 'awesome', 'fun'...

几年前，我在 Facebook（现 Meta）的同事处理这个问题的方法是，他们只选择了大约 17,000 个标签，这些标签对应的是具有视觉内容的物理事物或场景。所以你不会看到“tbt”之类的标签。

The way my colleagues who worked on this project at Facebook, now Meta, a few years ago dealt with this is that they only selected something like 17,000 tags that correspond to physical things or situations that have some visual content. So you wouldn't have 'tbt' or anything like that.

所以他们只保留了一组精选的标签？

So they keep a very select set of hashtags?

是的，但数量仍然在 10,000 到 20,000 之间，所以相当大。

Yes, but it's still on the order of 10,000 to 20,000, so it's fairly large.

你能讲讲数据增强吗？数据增强到底是什么，怎么用？也许还有视频的对比学习？这里有什么酷想法？

Can you tell me about data augmentation? What the heck is data augmentation and how is it used? Maybe contrastive learning for video? What are some cool ideas here?

数据增强是通过以不改变图像本质的方式扭曲图像来人为增加训练集大小的想法。你拿一个数字，稍微移动它，改变大小，旋转，倾斜，添加噪声等等。用增强数据训练监督分类器效果更好。过去几年，这变得非常有趣，因为许多用于预训练视觉系统的自监督学习技术都基于数据增强。这个基本技术最初源于我在 90 年代初的工作，杰夫·辛顿也在差不多同一时间研究过。我当时称之为孪生网络。基本上，你取两个相同的网络副本，它们共享相同的权重，然后展示同一物体的两个不同视图。这两个视图可能通过数据增强获得，也可能是同一场景从不同角度、不同时间拍摄的两张照片，或者同一个人的两张照片。然后你训练神经网络产生一个输出表示，使得这两张图像的表示尽可能接近。你希望系统学习一个不变函数，其输出在特定变换下不会改变。这很容易做到。复杂的是如何确保当你展示两张不同的图像时，系统会产生不同的结果。如果你没有为此设置特定机制，系统会忽略输入，为每个输入产生一个常数向量。这叫做坍缩。

Data augmentation is the idea of artificially increasing the size of your training set by distorting the images in ways that don't change the nature of the image. You take a digit and shift it a little bit, change the size, rotate it, skew it, add noise, etc. It works better if you train a supervised classifier with augmented data. Over the last couple years, it's become really interesting because a lot of self-supervised learning techniques to pre-train vision systems are based on data augmentation. The basic technique was originally inspired by work I did in the early 90s and Jeff Hinton also worked on around the same time. I used to call this a Siamese network. Basically, you take two identical copies of the same network, they share the same weights, and you show two different views of the same object. Those two views may have been obtained by data augmentation, or maybe it's two different views of the same scene from a camera that you moved, or at different times, or two pictures of the same person. Then you train the neural net to produce an output representation such that the representations for those two images are as close to each other as possible. You want the system to learn a function that will be invariant, whose output will not change when you transform the inputs in those particular ways. That's easy to do. What's complicated is how to make sure that when you show two images that are different, the system will produce different things. If you don't have a specific provision for this, the system will just ignore the input and produce a constant vector that is the same for every input. That's called collapse.

如何避免坍缩？

How do you avoid collapse?

有两个想法。一个是我在 90 年代初与贝尔实验室的同事简·布罗姆利等人提出的，我们现在称之为对比学习，就是使用负样本。你取已知不同的图像对，展示给网络，然后将两个输出向量推离彼此。这最终会保证语义相似的事物产生相似的表示，不同的事物产生不同的表示。我们实际上是在一个签名验证项目中提出这个想法的。我们收集同一个人的多个签名，训练神经网络产生相同的表示，并强制系统为不同的签名产生不同的表示。这个问题是由当时 AT&T 的子公司 NCR 的人提出的。他们有兴趣将签名的表示存储在信用卡磁条的 80 个字节上。所以我们想出了这个主意：用一个有 80 个输出的神经网络，量化成字节，这样我们就可以编码签名。然后这个编码用于比较签名是否匹配。它确实有效，但他们最终没有使用，因为没人在乎。美国的金融支付系统在这方面比欧洲落后得多。签名到底有什么用？没人看，没人在乎。

There are two ideas. One idea I proposed in the early 90s with my colleagues at Bell Labs, Jane Bromley and a couple others, which we now call contrastive learning, is to have negative examples. You have pairs of images that you know are different, and you show them to the network, and then you push the two output vectors away from each other. This will eventually guarantee that things that are semantically similar produce similar representations, and things that are different produce different representations. We actually came up with this idea for a project of doing signature verification. We would collect multiple signatures from the same person, train a neural net to produce the same representation, and force the system to produce different representations for different signatures. The problem was proposed by people from NCR, a subsidiary of AT&T at the time. They were interested in storing a representation of the signature on the 80 bytes of the magnetic strip of a credit card. So we came up with this idea of having a neural net with 80 outputs that we would quantize to bytes, so that we could encode the signature. That encoding was then used to compare whether the signature matches. It actually worked, but they ended up not using it because nobody cares. The American financial payment system is incredibly lagging in that respect compared to Europe. What's the purpose of signatures anyway? Nobody looks at them, nobody cares.

所以这就是对比学习。你需要正样本对和负样本对。

So that's contrastive learning. You need positive and negative pairs.

是的，但问题是，尽管我写了关于这个的原始论文，我实际上并不非常看好它，因为它在高维空间中不起作用。如果你的表示是高维的，两个事物有太多不同的方式，所以你需要大量的负样本对。有一个相对较新的实现来自谷歌多伦多小组，杰夫·辛顿是资深成员。它叫做 SimCLR，是实现这个想法的一种特定方式。

Yes, but the problem is that even though I wrote the original paper on this, I'm actually not very positive about it because it doesn't work in high dimension. If your representation is high dimensional, there are just too many ways for two things to be different, so you would need lots and lots of negative pairs. There is a particular implementation of this that is relatively recent from the Google Toronto group, where Jeff Hinton is the senior member. It's called SimCLR, a particular way of implementing this idea.

运行特定的目标函数。现在，我更热衷的是非对比方法。也就是其他方式来确保不同输入的表示是不同的。这实际上基于杰弗里·辛顿在 90 年代初与当时的学生苏·贝克尔提出的一个想法。它基于最大化两个系统输出之间互信息的思想。你只展示正样本对，只展示你知道有些相似的图像对，然后训练两个网络，使它们具有信息量，同时尽可能相互提供信息。基本上，一个表示必须能从另一个预测出来。他提出了这个想法，在 90 年代初发表了几篇论文，然后几十年都没有进展。我和我在 FAIR 的博士后们，特别是现在在芬兰阿尔托大学担任初级教授的斯特凡诺尼，复兴了这个想法。我们提出了一个叫做 Barlow Twins 的方法，这是一种利用某些假设最大化向量信息内容的特定方式。我们还有一个更新的版本，叫做 VicReg，代表方差-不变性-协方差正则化。这是过去 15 年里我在机器学习中最兴奋的事情。我真的非常兴奋。

Contracting running the particular objective function. Now what I'm much more enthusiastic about these days is non-contrastive methods. So other ways to guarantee that the representations would be different for different inputs. And it's actually based on an idea that Jeff Hinton proposed in the early 90s with a student at the time, Sue Becker. And it's based on the idea of maximizing the mutual information between the outputs of the two systems. You only show positive pairs. You only show pairs of images that you know are somewhat similar, and you train the two networks to be informative but also to be as informative of each other as possible. So basically, one representation has to be predictable from the other essentially. And he proposed that idea, had a couple of papers in the early 90s, and then nothing was done about it for decades. And I kind of revived this idea together with my postdocs at FAIR, particularly a postdoc called Stefanoni, who's now a junior professor in Finland at University of Aalto. We came up with something called Barlow Twins, and it's a particular way of maximizing the information content of a vector using some hypotheses. And we have another version of it that's more recent now called VicReg, which stands for Variance Invariance Covariance Regularization. And it's the thing I'm the most excited about in machine learning in the last 15 years. I'm really, really excited about this.

对于那种非对比学习方法，什么样的数据增强是有用的？我们是在说这个吗？这没那么重要吗？还是说这似乎是步骤中非常重要的一部分？

What kind of data augmentation is useful for that non-contrastive learning method? Are we talking about? Does that not matter that much? Or it seems like a very important part of the step.

是的，如何生成相似但足够不同的图像。没错，这是一个重要的步骤，也是一个烦人的步骤，因为你需要知道哪些增强操作不会改变物体的本质。所以标准做法是，很多在这个领域工作的人都在使用这种类型的失真。基本上，你进行几何失真。一种只是稍微移动图像，叫做裁剪。另一种稍微改变尺度。另一种旋转它。另一种改变颜色。你可以调整色彩平衡或饱和度之类的。另一种是模糊它。另一种是添加噪声。所以你有一个标准操作的目录。人们尝试对不同的算法使用相同的操作，以便进行比较。但有些算法，有些自监督算法，实际上可以处理更大、更激进的数据增强，而有些则不能。这使整个事情变得困难。但这就是我们讨论的那种失真。所以你用这些失真进行训练，然后去掉网络的最后几层，将表示作为分类器的输入。你在 ImageNet 上训练分类器，然后测量性能。有趣的是，那些非常擅长消除无关信息（即图像之间的失真）的方法，确实很好地消除了它。结果，你不能在这些系统中使用这些表示来进行物体检测和定位，因为那些信息已经丢失了。所以你需要做的数据增强类型取决于你最终希望系统解决的任务。我们今天使用的标准数据增强只适用于物体识别或图像分类，不适用于定位之类的事情。

Yeah, how you generate the images that are similar but sufficiently different. That's right, it's an important step and it's also an annoying step because you need to have that knowledge of what augmentations you can do that do not change the nature of the object. So the standard scenario, which a lot of people working in this area are using, is you use the type of distortion. So basically, you do geometric distortion. One basically just shifts the image a little bit, it's called cropping. Another one kind of changes the scale a little bit. Another one kind of rotates it. Another one changes the colors. You can do a shift in color balance or something like that, saturation. Another one sort of blurs it. Another one adds noise. So you have like a catalog of standard things. And people try to use the same ones for different algorithms so that they can compare. But some algorithms, some self-supervised algorithms, actually can deal with much bigger, more aggressive data augmentation, and some don't. So that kind of makes the whole thing difficult. But that's the kind of distortions we're talking about. And so you train with those distortions, and then you chop off the last layer, a couple layers of the network, and you use the representation as input to a classifier. You train the classifier on ImageNet, let's say, or whatever, and measure the performance. And interestingly enough, the methods that are really good at eliminating the information that is irrelevant, which is the distortions between those images, do a good job at eliminating it. And as a consequence, you cannot use the representations in those systems for things like object detection and localization because that information is gone. So the type of data augmentation you need to do depends on the task you want eventually the system to solve. And the type of standard data augmentation that we use today are only appropriate for object recognition or image classification. They're not appropriate for things like localization.

你能帮我理解为什么定位会这样？你是说它不擅长分类负样本，所以不能用于定位？

Can you help me understand why localization is... So you're saying it's just not good at classifying the negative, so that's why it can't be used for localization?

不，只是你训练系统的方式：你给它一张图像，然后给它同一张图像经过平移和缩放后的版本，并告诉它这是同一张图像。所以系统基本上被训练成消除关于位置和大小的信息。现在你想用它来找出物体在哪里以及它的大小，比如一个边界框。它仍然可以在图像中找到物体，只是不擅长找到物体的精确边界。

No, it's just that you train the system: you give it an image, and then you give it the same image shifted and scaled, and you tell it that's the same image. So the system is basically trained to eliminate the information about position and size. So now you want to use it to figure out where an object is and what size it is, like a bounding box. It can still find the object in the image, it's just not very good at finding the exact boundaries of that object.

有趣。这是一个有趣的哲学问题：物体定位到底有多重要？我们痴迷于测量图像分割，痴迷于完美地知道物体的边界，而可以说这对于理解场景内容并不那么重要。另一方面，我认为在进化上，动物的第一个视觉系统基本上都是关于定位的，很少关于识别。在人类大脑中，有两条独立的通路，一条用于识别场景或物体的性质，另一条用于定位物体。所以你使用第一条通路，称为腹侧通路，来判断你在看什么。另一条通路，背侧通路，用于导航、抓取和其他一切。基本上，生存所需的大部分事情都是定位和检测。

Interesting. Which is an interesting sort of philosophical question: how important is object localization anyway? We're obsessed with measuring image segmentation, obsessed with perfectly knowing the boundaries of objects, when arguably that's not that essential to understanding the contents of the scene. On the other hand, I think evolutionarily, the first vision systems in animals were basically all about localization, very little about recognition. And in the human brain, you have two separate pathways for recognizing the nature of a scene or object and localizing objects. So you use the first pathway, called the ventral pathway, for telling what you're looking at. The other pathway, the dorsal pathway, is used for navigation, for grasping, for everything else. And basically, a lot of the things you need for survival are localization and detection.

相似性学习或对比学习，这些非对比方法等同于理解某物吗？仅仅因为你知道一只变形的猫和一只未变形的猫是同一只猫，这是否意味着你在某种程度上理解了猫是什么？我的意思是，这显然是一种肤浅的理解，但你认为这种方法的极限在哪里？这只是自监督学习道路上的一个技巧吗？我们能走得很远很远吗？

Is similarity learning or contrastive learning, are these non-contrastive methods the same as understanding something? Just because you know a distorted cat is the same as a non-distorted cat, does that mean you understand what it means to be a cat to some extent? I mean, it's a superficial understanding obviously, but what is the ceiling of this method do you think? Is this just one trick on the path to doing self-supervised learning? Can we go really, really far?

我认为我们可以走得很远。所以如果我们弄清楚如何使用这类技术，也许非常不同，但关键是从视频中训练系统进行视频预测，我认为我们将有一条通往机器某种程度的物理常识的道路，我不会说是无限的，但确实是一条道路。而且我还认为，从像视觉这样的高吞吐量通道学习世界如何运作的能力，是迈向真正人工智能的必要步骤。换句话说，我相信具身智能。我不认为我们可以纯粹从文本中训练出智能机器，因为我认为文本中包含的关于世界的信息量与我们所需知道的相比微不足道。例如，人们已经尝试这样做 30 年了，对吧？Cyc 项目之类的，基本上写下所有已知的事实，并希望某种常识会出现。

I think we can go really far. So if we figure out how to use techniques of that type, perhaps very different, but the signature to train a system from video to do video prediction essentially, I think we'll have a path towards, I wouldn't say unlimited, but a path towards some level of physical common sense in machines. And I also think that ability to learn how the world works from a sort of high-throughput channel like vision is a necessary step towards real artificial intelligence. In other words, I believe in grounded intelligence. I don't think we can train a machine to be intelligent purely from text because I think the amount of information about the world that's contained in text is tiny compared to what we need to know. So for example, people have attempted to do this for 30 years, right? The Cyc project and things like that, basically writing down all the facts that are known and hoping that some sort of common sense will emerge.

我认为这基本上是无望的。但让我举个例子：你拿一个物体，我向你描述这个情况。我拿一个物体，把它放在桌子上，然后推桌子。你完全明白物体会随着桌子一起被推动，对吧？因为它放在上面。我相信世界上没有任何文本能解释这一点。所以如果你训练一个尽可能强大的机器，比如你的 GPT-5000 或别的什么，它永远学不到这个。这些信息根本不存在于任何文本中。

I think it's basically hopeless. But let me take an example: you take an object, I describe the situation to you. I take an object, I put it on the table, and I push the table. It's completely obvious to you that the object will be pushed with the table, right? Because it's sitting on it. There's no text in the world, I believe, that explains this. And so if you train a machine as powerful as it could be, you know, your GPT-5000 or whatever it is, it's never going to learn about this. That information is just not present in any text.

嗯，像 PSYC 项目这样的问题，我认为梦想是拥有大约一千万个这样的事实，给你一个先发优势，就像父母引导你一样。现在我们人类不需要父母告诉我们桌子会移动——呃，抱歉，智能手机也会随着桌子移动——但我们通过其他方式得到了很多指导。所以有可能我们可以给它一个快速的捷径。

Well, the question like with the PSYC project, the dream I think is to have like 10 million facts like that that give you a head start, like a parent guiding you. Now we humans don't need a parent to tell us that the table will move—uh, sorry, the smartphone will move with the table—but we get a lot of guidance in other ways. So it's possible that we can give it a quick shortcut.

那猫呢？它们知道这个，但它们是进化来的，所以不，它们像我们一样学习物理。

What about cats? The guy knows that, but they evolved, so no, they learn like us the physics of stuff.

嗯，是的，所以你是说你把很多智能归因于后天培养，而不是先天。是的，我们似乎有一个非常低效的、可以说是进化的过程，把我们从细菌变成了今天的我们。从底层开始，现在我们在这里。确实如此。问题是这有多根本？整个硬件的本质，然后有没有办法走捷径？如果它是根本性的，如果不是，如果大多数智能，我们一直在谈论的大多数酷东西，主要是后天培养，主要是训练出来的，我们通过观察世界来弄清楚，我们可以通过坐在那里形成你所说的那个又大又漂亮又性感的背景模型，那么好吧，那么你需要像也许它全是监督学习一路到底。所以，无论是什么让人类智能与其他动物不同，很多人认为是语言和逻辑推理之类的东西，它不可能那么复杂，因为它只在过去一百万年里才出现。而且它只涉及我们基因组中不到百分之一的部分，这是人类基因组和黑猩猩基因组之间的差异。所以它不可能那么复杂，它不可能那么根本。我的意思是，大多数复杂的东西已经存在于猫和狗中，当然还有灵长类动物，非人类灵长类动物。人类的那个小东西可能只是关于社会互动和在集体中维持想法的能力。这听起来非常戏剧化，非常令人印象深刻，但从机制上讲，可能并非如此。它是，但我们还没到那一步。比如，你知道，我们有，我的意思是，这是要解决的问题列表中的第 634 个。所以世界的基本物理是第一位的。

Well, yeah, so you're saying you're putting a lot of intelligence onto the nurture side, not the nature side. Yes, we seem to have, you know, a very inefficient, arguably, process of evolution that got us from bacteria to who we are today. Started at the bottom, now we're here. So true. The question is how fundamental is that? The nature of the whole hardware, and then is there any way to shortcut it? If it's fundamental, if it's not, if most of intelligence, most of the cool stuff we've been talking about, is mostly nurture, mostly trained, we figure it out by observing the world, we can form that big beautiful sexy background model that you're talking about just by sitting there, then okay, then you need to then like maybe it is all supervised learning all the way down. So whatever it is that makes human intelligence different from other animals, which a lot of people think is language and logical reasoning and this kind of stuff, it cannot be that complicated because it only popped up in the last million years. And it only involves less than one percent of our genome, which is the difference between human genome and chimp genome or whatever. So it can't be that complicated, it can't be that fundamental. I mean, most of the complicated stuff already exists in cats and dogs and certainly primates, non-human primates. That little thing with humans might be just something about social interaction and ability to maintain ideas across a collective of people. It sounds very dramatic and very impressive, but it probably isn't, mechanistically speaking. It is, but we're not there yet. Like, you know, we have, I mean, this is number 634 in the list of problems to solve. So basic physics of the world is number one.

快速岔开一下，关于数据增强。很多是硬编码的，而不是学习的。你有没有直觉，也许会有一些奇怪的数据增强，比如生成式的数据增强，比如对图像做一些奇怪的事情，从而改进相似性学习过程？所以不仅仅是那种愚蠢的简单扭曲。

Just a quick tangent on data augmentation. A lot of it is hard-coded versus learned. Do you have any intuition that maybe there could be some weird data augmentation, like generative type of data augmentation, like doing something weird to images which then improves the similarity learning process? So not just kind of dumb simple distortions.

你摇头，只是说即使是简单的扭曲也足够了。我认为不，我认为数据增强是一种暂时的必要之恶。所以人们现在正在研究两件事：一是自监督的类型，比如试图将人们在语言中使用的自监督类型迁移到图像上，这基本上是一种去噪自编码器方法，对吧？所以你拿一张图像，你遮挡，你遮住它的一部分，然后你训练一个巨大的神经网络来重建缺失的部分。直到最近，还没有有效的方法。所有用于图像的自编码器方法都没有产生很好的表示。但 Menlo Park 的 FAIR 小组现在出了一篇论文，效果很好。所以这不需要文档，只需要遮罩。好的，只对图像进行遮罩。好的，对，所以你遮住图像的一部分，然后训练一个系统，在这种情况下是 Transformer，因为 Transformer 将图像表示为不重叠的块，所以很容易遮住块之类的东西。

By you shaking your head, just saying that even simple distortions are enough. I think no, I think that augmentation is a temporary necessary evil. So what people are working on now is two things: one is the type of self-supervision, like trying to translate the type of self-supervision people use in language to images, which is basically a denoising autoencoder method, right? So you take an image, you block, you mask some parts of it, and then you train some giant neural net to reconstruct the parts that are missing. And until very recently, there was no working method for that. All the autoencoder-type methods for images weren't producing very good representations. But there's a paper now coming out of the FAIR group in Menlo Park that actually works very well. So that doesn't require documentation, that requires only masking. Okay, only masking for images. Okay, right, so you mask part of the image and you train a system, which in this case is a transformer, because the transformer represents the image as non-overlapping patches, so it's easy to mask patches and things like that.

好的，那么我的问题就转移到了那个问题上。遮罩：为什么遮罩应该是正方形或长方形？

Okay, then my question transfers to that problem. The masking: why should the mask be a square or rectangle?

这没关系。就像，你知道，我认为我们将来可能会想出一些遮罩的方法，基本上是随机的。

It doesn't matter. Like, you know, I think we're gonna come up probably in the future with sort of, you know, ways to mask that are, you know, kind of random essentially.

嗯，我的意思是它们已经是随机的了，但不，不，但像一些有挑战性的东西，比如最优挑战性的。所以，我的意思是，也许这是一个不适用的比喻，但似乎有一个互动元素，就像你几乎在玩一张图像，就像我们在脑海中玩图像的方式。

Well, I mean they are random already, but no, no, but like something that's challenging, like optimally challenging. So like, I mean, maybe it's a metaphor that doesn't apply, but it seems like there's an interactive element with it, like you're almost like playing with an image, and like it's like the way we play with an image in our minds.

不，但这就像 dropout，就像玻尔兹曼机训练。你知道，每次你看到一个感知，你也可以以某种方式扰动它，然后训练过程的原则是最小化干净版本和损坏版本之间表示输出的差异，本质上。你可以实时做到这一点，对吧？所以你知道，机器是怎么工作的？你展示一个感知，告诉机器那是你的输入神经元的一个好的活动组合，然后你要么让它们自由活动而不将它们钳制到值，要么你只对子集这样做。你在做的是训练系统，使得整个网络的稳定状态无论看到完整输入还是只看到部分输入都是一样的。你知道，去噪自编码器方法基本上是同一回事，对吧？你训练一个系统来重现输入，完整的输入，并填补空白，无论哪些部分缺失。这真的是基本原则。你可以想象，甚至在脑中，某种神经原则，其中神经元振荡，对吧？所以它们获取活动，然后暂时关闭，迫使系统的其余部分在没有它们帮助的情况下重建输入。我的意思是，你可以想象或多或少生物学上合理的过程。

No, but it's like dropout, it's like Boltzmann machine training. You know, every time you see a percept, you also can perturb it in some way, and then the principle of the training procedure is to minimize the difference of the output of the representation between the clean version and the corrupted version, essentially. And you can do this in real time, right? So you know, what's the machine work like this? You show a percept and you tell the machine that's a good combination of activities of your input neurons, and then you either let them go their merry way without clamping them to values, or you only do this with a subset. And what you're doing is you're training the system so that the stable state of the entire network is the same regardless of whether it sees the entire input or whether it is only part of it. You know, the denoising autoencoder method is basically the same thing, right? You're training a system to reproduce the input, the complete input, and fill in the blanks regardless of which parts are missing. And that's really the underlying principle. And you could imagine, sort of, even in the brain, some sort of neural principle where neurons oscillate, right? So they take their activity and then temporarily they kind of shut off to force the rest of the system to basically reconstruct the input without their help. And I mean, you could imagine more or less biologically plausible processes.

而且我猜有了这种去噪自编码器、遮罩和数据增强，你不用担心效率问题，你可以想怎么做就怎么做，随着时间的推移变得更好。因为我在想，你可能想在这些过程的方式上变得聪明。

And I guess with this denoising autoencoder and masking and data augmentation, you don't have to worry about being super efficient, you can just do as much as you want and get better over time. Because I was thinking like you might want to be clever about the way you do all these procedures.

但只有当每次迭代成本高昂时才会这样，但实际上并非如此。还有一种无需显式数据增强的数据增强方式：通过等待来增强。在视频预测中，你观察一个视频片段，观察它的后续，然后尝试使用联合嵌入架构学习一种表示，使得未来片段的表示能够轻松地从观察到的片段的表示中预测出来。你认为 YouTube 有足够的原始数据来学习如何成为一只猫吗？

But that's only if it's somehow costly to do every iteration, but it's not really. And then there is data augmentation without explicit data augmentation: data augmentation by waiting. In video prediction, you observe a video clip, observe the continuation, and try to learn a representation using joint embedding architectures such that the representation of the future clip is easily predictable from the representation of the observed clip. Do you think YouTube has enough raw data from which to learn how to be a cat?

我认为是的。数据量不是限制。它需要一些筛选，也许是合适类型的数据。如果只看猫视频，你可能还需要看一些讲座之类的。

I think so. The amount of data is not the constraint. It would require some selection, maybe the right type of data. Down the rabbit hole of just cat videos, you might need to watch some lectures or something.

不，你不会。如果它观看关于智能的讲座并从中学习，那该有多元啊？观看你在纽约大学的讲座，学习如何变得智能。你对多模态学习怎么看？我们一直在讨论视觉和语言，将它们结合起来，也许还有音频。短期来看有很多有趣的东西，但它们并没有解决重要的问题，那些大的挑战。

No, you wouldn't. How meta would that be if it watches lectures about intelligence and learns from that? Watches your lectures at NYU and learns how to be intelligent. What's your take on multimodal learning? We've been talking about vision and language, combining those together, maybe audio. There's a lot of things I find interesting in the short term, but they're not addressing the important problem, the big challenges.

我认为像多任务学习、持续学习、对抗性问题——它们在相对短期内具有很大的实际意义，但我认为它们不是根本性的。主动学习，甚至某种程度上的强化学习——我认为一旦我们弄清楚如何进行自监督表示学习或学习可预测模型，这些东西就会变得过时、无用或容易解决。我认为这才是整个社区应该关注的，至少是那些对根本性问题感兴趣或推动 AI 进入下一阶段的人。当然，在具有短期影响的实际问题上也有大量有趣的工作。

I think things like multitask learning, continual learning, adversarial issues—they have great practical interest in the relatively short term, but I don't think they're fundamental. Active learning, even reinforcement learning to some extent—I think those things will become either obsolete or useless or easy once we figure out how to do self-supervised representation learning or learning predictable models. I think that's what the entire community should be focusing on, at least people interested in fundamental questions or pushing the envelope of AI towards the next stage. But of course, there is a huge amount of interesting work in practical questions with short-term impact.

谈论时间尺度很困难，因为当太阳熄灭时，整个人类文明最终都会被摧毁。即使埃隆·马斯克成功实现了多行星殖民，最终一切都会变成巨大的黑洞。不过那还需要一段时间。但那种逻辑可以用来论证一切都是无意义的。我的意思是，多任务学习可能是在我们试图解决自监督学习和背景知识这个更普遍问题的同时，实现非常接近智能的东西。特斯拉自动驾驶团队所做的工作给我留下了深刻印象。他们研究驾驶问题，询问需要感知的所有事物。他们制定了一百多个不同的任务，部署系统，从遇到麻烦的人那里获取数据，然后决定是否添加任务。他们分别关注每个单独的任务。

It's difficult to talk about temporal scale because all of human civilization will eventually be destroyed when the sun dies out. Even if Elon Musk is successful with multi-planetary colonization, eventually everything becomes giant black holes. That's going to take a while though. But that logic can be used to say it's all meaningless. I'm saying that multitask learning might be the thing that achieves something very akin to intelligence while we're trying to solve the more general problem of self-supervised learning and background knowledge. I've been very impressed by what the Tesla Autopilot team is doing. They're taking the problem of driving and asking what all the things you have to perceive. They formulate a bunch of different tasks, over a hundred tasks, deploy it, get data back from people who run into trouble, and figure out whether to add tasks. They focus on each individual task separately.

我会把安德烈·卡帕西的演讲分为两部分：一部分是关于门，另一部分是关于 ImageNet 有多糟糕。他在这两个话题之间来回切换。ImageNet 糟糕的意思是，你不能只用一个基准测试；你需要一整套基准测试来了解你的系统实际表现如何。我同意他的观点。他是一个非常明智的人。很明显，如果你面临一个需要在相对短时间内解决的工程问题，尤其是当问题迫在眉睫时，你就得走捷径。你可能认为正确的长期解决方案涉及花哨的自监督学习，但这里关系到人的生命。你必须进行系统性的工程、微调、改进、试错。这没什么错；这就是工程，是把技术推向世界。在此之前你必须确保它万无一失。宏伟的想法和原则就先放一放吧。但我把自己定位在这个过程的上游，相当上游。我思考柏拉图式的形式。最终我希望那些东西被使用，但如果社区需要五到十年才意识到这是正确的事情，那也没关系。我以前就经历过。如果你回顾 2000 年代中期，你想识别汽车或人脸。你可以使用卷积网络，或者更传统的计算机视觉技术，比如兴趣点检测器、SIFT 特征，再加上 SVM。当时数据集太小，手工工程的方法比卷积网络效果更好。卷积网络在当时可用的硬件上有点慢。当数据集变大、GPU 可用时，情况发生了变化。这两个因素让人们改变了想法。你可以看看 AI 或模式识别的所有分支的历史；都有类似的轨迹。人们一开始都拼命做工程——无论是光学字符识别、语音识别、计算机视觉、图像识别、自然语言理解、翻译。你一开始都拼命做工程。

I would classify Andrej Karpathy's talk in two ways: one was about doors, and the other about how much ImageNet sucks. He kept going back and forth. ImageNet sucks meaning you can't just use a single benchmark; you need a giant suite of benchmarks to understand how well your system actually works. I agree with him. He's a very sensible guy. It's very clear that if you're faced with an engineering problem that needs to be solved in a relatively short time, particularly if you have it breathing down your neck, you're going to have to take shortcuts. You might think the right long-term solution involves fancy self-supervision, but you have human lives at stake. You have to do systematic engineering, fine-tuning, refinements, trial and error. There's nothing wrong with that; it's called engineering, putting technology out in the world. You have to ironclad it before you do this. So much for grand ideas and principles. But I'm placing myself upstream of this, quite a bit upstream. I think about Platonic forms. Eventually I want that stuff to get used, but it's okay if it takes five or ten years for the community to realize this is the right thing to do. I've done this before. If you look back in the mid-2000s, you wanted to recognize cars or faces. You could use convolutional nets or more conventional computer vision techniques with interest point detectors, SIFT features, and an SVM on top. At that time, datasets were so small that hand-engineered methods worked better than ConvNets. ConvNets were a little slow with the hardware available. There was a change when datasets became bigger and GPUs became available. Those two factors made people change their minds. You can look at the history of all sub-branches of AI or pattern recognition; there's a similar trajectory. People start by engineering the hell out of it—be it optical character recognition, speech recognition, computer vision, image recognition, natural language understanding, translation. You start by engineering the hell out of it.

你开始获取所有知识：关于图像形成的先验知识、字符形状、形态学操作、特征提取、傅里叶变换、矩等等。人们想出了成千上万种表示图像的方法，以便之后能轻松分类。语音识别也一样：花了二十年才找到好的前端来预处理语音信号，保留所说的内容信息，但去掉说话者的身份信息，比如倒谱系数。文本也是如此：你需要做实体识别、句法分析、词性标注、从句的树状表示等等，然后才能做其他事情。所以一开始就是这样：拼命手工设计。然后你有了数据，有了更强大的计算机，也许你懂一些统计学习，于是开始使用机器学习，通常只是手工系统上的一小层，你手动提取特征。如今，标准做法是用深度学习系统端到端地训练整个模型，让它自己学习特征。语音识别和 OCR 系统已经完全端到端：一个巨大的神经网络接收原始波形，输出字符序列。除了内嵌在神经语言模型中的东西，没有显式的语言模型。翻译和其他任务也一样。所以你看到这种从手工设计到更多学习的持续演进。我认为生物学也是如此。

You start to acquire all knowledge: the prior knowledge about image formation, shape of characters, morphological operations, feature extraction, Fourier transforms, moments, whatever. People have come up with thousands of ways of representing images so they could be easily classified afterwards. Same for speech recognition: two decades to figure out a good front end to pre-process speech signals so that the information about what is being said is preserved but most of the information about the identity of the speaker is gone, like cepstral coefficients. Same for text: you do entity recognition, parsing, part-of-speech tagging, tree representation of clauses, all that stuff before you can do anything. So that's how it starts: engineer the hell out of it. Then you start having data and more powerful computers, maybe you know something about statistical learning, so you start using machine learning, usually a small sliver on top of your handcrafted system where you extract features by hand. Nowadays, the standard way is to train the entire thing end-to-end with a deep learning system, and it learns its own features. Speech recognition and OCR systems are completely end-to-end: a giant neural net that takes raw waveforms and produces a sequence of characters. There's no explicit language model other than what's ingrained in the neural language model. Same for translation and all kinds of stuff. So you see this continuous evolution from less handcrafting and more learning. I think it's true in biology as well.

我们可能对此有分歧，也许没有。在最后这一点上，你提到了主动学习。我觉得主动学习，即数据选择和交互性，需要成为这个巨大神经网络的一部分。你不能仅仅作为一个观察者来进行自监督学习。你必须有一个系统，无论是物理机器人还是数字机器人，它与世界互动，以有缺陷的方式行事，并随时间改进，从而形成自监督学习。你不能只给它一大片数据。

We might disagree about this, maybe not. In this one little piece at the end you mentioned active learning. It feels like active learning, which is the selection of data and interactivity, needs to be part of this giant neural network. You cannot just be an observer to do self-supervised learning. You have to have a system, whether a physical robot or a digital robot, that's interacting with the world and doing so in a flawed way and improving over time to form the self-supervised learning. You can't just give it a giant sea of data.

好吧，我既同意又不同意。我同意有两个方面。第一，如果你想要一个能预测行动后果的因果世界模型，训练这个模型需要你采取行动。你需要在世界中行动并观察效果，才能学习因果模型。但这并不明显，因为你可以观察他人并推断他们与你相似，然后从中学习。但那样你就得硬编码那部分，比如镜像神经元。目前不清楚如何在机器中做到这一点，所以行动对于因果模型可能是必要的。第二，主动学习会去探索你不知道的东西，即关于世界的不确定性区域。你可以通过系统性的探索来解决这种不确定性。如果你知道自己不知道，就会产生好奇心。在动物世界中，不同物种有不同水平的好奇心。猫和老鼠非常好奇，狗则不那么好奇。所以好奇心可能有用，但它只是让过程更快，而不是让过程存在。

Okay, I agree and I disagree. I agree in two ways. First, if you want a causal model of the world that allows you to predict the consequences of your actions, to train that model you need to take actions. You need to be able to act in the world and see the effect to learn causal models. But that's not obvious because you can observe others and infer they are similar to you, then learn from that. But then you have to hardwire that part, like mirror neurons. It's not clear how to do this in a machine, so action may be necessary for causal models. Second, active learning goes for the jiggler of what you don't know, areas of uncertainty about your world. You can resolve this uncertainty by systematic exploration. If you know you don't know, it makes you curious. Across the animal world, different species have different levels of curiosity. Cats and rats are incredibly curious, dogs not so much. So curiosity could be useful, but it just makes the process faster, not make the process exist.

那么主动学习让哪个学习过程更高效？我们还没有回答这个问题。这才是更根本的问题。如果主动学习或交互将效率提高了几个数量级，那确实如此，但本质上还是同一件事。建立关于如何以自监督方式构建背景模型的直觉，无论高效还是低效，都是核心问题。

So what learning process is it that active learning makes more efficient? We haven't answered that question yet. That's the more fundamental question. If active learning or interaction increases efficiency by several orders of magnitude, that's true but fundamentally still the same thing. Building up the intuition about how to construct background models in a self-supervised way, efficient or inefficient, is the core problem.

你怎么看约书亚·本吉奥谈论意识以及这些概念？我不知道意识是什么。这是个好开场白。在某种程度上，关于意识的很多说法让我想起人们在 17 或 18 世纪发现眼睛工作原理时提出的问题：眼睛后方的图像是倒立的，你怎么看到正立的？以我们今天所知，我们意识到这个问题没有意义或很荒谬。我认为关于意识的很多说法都是这种性质。不过，有很多我非常尊敬的聪明人在谈论这个话题，比如我在纽约大学的同事大卫·查默斯。我对意识有一个非正统的民间推测假说。

What do you think about Yoshua Bengio talking about consciousness and all these concepts? I don't know what consciousness is. It's a good opener. To some extent, a lot of what is said about consciousness reminds me of questions people asked in the 17th or 18th century when they discovered how the eye works and that the image at the back of the eye is upside down. How is it that you see right side up? With what we know today, we realize this question doesn't make sense or is ridiculous. I think a lot of what is said about consciousness is of that nature. That said, there are a lot of really smart people I respect, like David Chalmers, my colleague at NYU, talking about this topic. I have a kind of unorthodox folk speculative hypothesis about consciousness.

我们在讨论这个音频世界模型。我认为我们的整个前额叶皮层就是我们世界模型的引擎。但当我们专注于某个特定情境时，我们全神贯注于那个情境，无法关注其他事情。这表明我们的前额叶皮层只有一个世界模型引擎。这个引擎可以根据当前情境进行配置。所以无论是用木头做盒子、在高速公路上开车还是下棋，我们都有一个单一的世界模型，我们将其配置到当前情境中。

We're talking about this audio world model. I think our entire prefrontal cortex is the engine for our world model. But when we are attending to a particular situation, we are focused on that situation and cannot attend to anything else. That suggests we have only one world model engine in our prefrontal cortex. That engine is configurable to the situation at hand. So we are building a box out of wood, or driving down the highway, or playing chess. We have a single model of the world that we configure into the situation.

这就是为什么我们一次只能专注于一项任务。如果一个任务我们反复做，它就会从那种使用世界模型和预测的刻意推理，也许还有我之前提到的模型预测控制，转变为更潜意识的东西，变得自动化。我不知道你是否和象棋大师下过棋。我十步之内就被干掉了，对吧？而我需要花 15 分钟思考我的走法，但对面的大师只需几秒钟就能反应。他不需要思考；这已经成为潜意识的一部分，因为基本上就是模式识别了。开车也一样：最初几小时你开车时非常专注，不能做其他事。然后经过 20-30 小时的练习，50 小时后，它变成潜意识；你可以和旁边的人聊天，诸如此类。除非情况变得不可预测，那时你就得停止说话。所以这表明你脑子里只有一个模型，这可能暗示意识本质上就是配置你这个世界模型的模块。你需要某种执行监督者来为当前情况配置你的世界模型，这引出了一个非常有趣的概念：意识不是我们大脑能力的产物，而是大脑局限性的产物。因为我们只有一个世界模型，所以我们必须有意识。如果我们有和遇到的情况一样多的世界模型，那么我们可以同时处理所有情况，就不需要这种我们称之为意识的执行控制了。

At hand, which is why we can only attend to one task at a time. Now, if there is a task that we do repeatedly, it goes from the sort of deliberate reasoning using a model of the world and prediction, and perhaps something like model predictive control which I was talking about earlier, to something that is more subconscious, that becomes automatic. So I don't know if you've ever played against a chess grandmaster. I get wiped out in ten moves, right? And I have to think about my move for like 15 minutes, and the person in front of me, the grandmaster, would just react within seconds. He doesn't need to think about it; that's become part of the subconscious because it's basically just pattern recognition at this point. Same with driving: the first few hours you drive a car, you're really attentive, you can't do anything else. Then after 20-30 hours of practice, 50 hours, it becomes subconscious; you can talk to the person next to you, things like that. Unless the situation becomes unpredictable, and then you have to stop talking. So that suggests you only have one model in your head, and it might suggest the idea that consciousness basically is the module that configures this world model of yours. You need to have some sort of executive overseer that configures your world model for the situation at hand, and that leads to the really curious concept that consciousness is not a consequence of the power of our minds, but of the limitation of our brains. Because we have only one world model, we have to be conscious. If we had as many world models as there are situations we encounter, then we could do all of them simultaneously and we wouldn't need this sort of executive control that we call consciousness.

是的，有意思。不知何故，那个执行控制器，我是说意识难题，生物学中有某种化学物质创造了一种感觉，就像体验这些事情的感觉。这就是那个难题：那到底是什么，为什么有用？也许更实际的问题是：为什么感觉这是真正的你在体验，而不仅仅是信息在处理，这样有用？

Yeah, interesting. And somehow that executive controller, I mean the hard problem of consciousness, there's some kind of chemicals in biology that's creating a feeling like it feels to experience some of these things. That's the hard question: what the heck is that and why is that useful? Maybe the more pragmatic question: why is it useful to feel like this is really you experiencing this versus just like information being processed?

这可能只是我们进化方式的一个很好的副作用，感觉对自己做出的决定、感知、试图维持的模型有一种所有权感非常有用。就像你拥有这个东西，它是你唯一拥有的，如果失去它，那会很糟糕，所以你真的应该给大脑发送一些关于它的信号。

It could be just a very nice side effect of the way we evolved that's just very useful to feel a sense of ownership to the decisions you make, to the perceptions you make, to the model you're trying to maintain. Like you own this thing and it's the only one you got, and if you lose it it's going to really suck, so you should really send the brain some signals about it.

你认为哪些想法可能是真的，但大多数人，或者至少很多人不同意你，比如说在机器学习领域？

What ideas do you believe might be true that most, or at least many people, disagree with you with, let's say in the space of machine learning?

嗯，这取决于你和谁说话，但我认为肯定有一群人是先天论者，他们认为世界上很多基本的东西都是我们大脑中硬连线的。比如世界是三维的：那是硬连线的吗？比如物体恒存性：那是我们在三个月大之前学会的，还是与生俱来的？认知科学家之间存在分歧。我认为这些东西实际上很容易学习。V1 中的方向边缘检测器是学习的还是硬连线的？我认为它们是学习的。它们可能在出生前就学会了，因为从视网膜产生信号来训练边缘检测器非常容易。再说一次，这些东西可以在睁开眼睛几分钟内学会。自 20 世纪 90 年代以来，我们有算法可以在完全无监督的情况下，用相当于几分钟实时时间学习方向检测器。所以这些东西必须是学习的。还有那些麻省理工学院的实验，你把视神经连接到雪貂幼崽的听觉皮层，那个听觉皮层基本上变成了视觉皮层。所以显然那里发生了重新布线。所以我认为很多人们认为非常基本以至于需要硬连线的东西，实际上很多都是学习的，因为它们很容易学习。

Well, it depends who you talk to, but I think certainly there are a bunch of people who are nativists, right, who think that a lot of the basic things about the world are kind of hardwired in our minds. Things like the world is three-dimensional, for example: is that hardwired? Things like object permanence: is that something we learn before the age of three months or so, or are we born with it? There are disagreements among cognitive scientists. I think those things are actually very simple to learn. Is it the case that the oriented edge detectors in V1 are learned or are they hardwired? I think they are learned. They might be learned before birth because it's really easy to generate signals from the retina that will train edge detectors. Again, those are things that can be learned within minutes of opening your eyes. Since the 1990s, we have algorithms that can learn oriented detectors completely unsupervised with the equivalent of a few minutes of real time. So those things have to be learned. There's also those MIT experiments where you plug the optical nerve onto the auditory cortex of a baby ferret, and that auditory cortex becomes a visual cortex essentially. So clearly there's rewiring taking place there. So I think a lot of what people think are so basic that they need to be hardwired, I think a lot of those things are learned because they are easy to learn.

所以你非常重视学习的力量。你怀疑哪些东西可能不是学习的？有什么东西是不能学习的吗？

So you put a lot of value in the power of learning. What kind of things do you suspect might not be learned? Is there something that could not be learned?

你的内在驱动力不是学习的。它们是使人类成为人类，或使猫不同于狗的东西。基本驱动力在我们的基底神经节中是硬连线的。有人在研究这类东西，在强化学习背景下称为内在动机。所以这些是目标函数，其中奖励不是来自外部世界，而是由你自己的大脑计算。你自己的大脑计算你是否快乐，它衡量你的舒适或不舒适程度。因为这是你的大脑在计算，大概它也知道如何估计这个的梯度。所以当你的目标是内在的时候，学习更容易。所以那必须是硬连线的。对结果做出长期预测的评判者，也就是最终结果，那是学习的。感知是学习的，你的世界模型也是学习的。但让我举个例子说明为什么评判者可能是学习的。如果我走向你，我伸手过桌子捏你的胳膊，你完全出乎意料。你不会预料到这一点。我一直预料到，但好吧，对吧？为了故事起见，是的。好的，你的视觉神经节会亮起来，因为会疼。现在你的世界模型包括这样一个事实：如果我把手伸向你的胳膊，我可能会捏你。所以如果我再次尝试，你会退缩，那就是你的评判者，你对最终痛苦的预测器，预测当你退缩以避免时会发生坏事的系统。所以即使是那个也可以学习。这绝对允许你定义一些目标。你是一个学童，早上醒来去上学，不是因为你一定喜欢早起上学，而是你知道有一个长期目标你试图优化。

Your intrinsic drives are not learned. They are the things that make humans human, or make cats different from dogs. The basic drives are kind of hardwired in our basal ganglia. There are people working on this kind of stuff, called intrinsic motivation in the context of reinforcement learning. So these are objective functions where the reward doesn't come from the external world; it's computed by your own brain. Your own brain computes whether you're happy or not, it measures your degree of comfort or discomfort. And because it's your brain computing this, presumably it also knows how to estimate gradients of this. So it's easier to learn when your objective is intrinsic. So that has to be hardwired. The critic that makes long-term prediction of the outcome, which is the eventual result of this, that's learned. And perception is learned, and your model of the world is learned. But let me take an example of why the critic might be learned. If I come to you, I reach across the table and I pinch your arm, complete surprise for you. You would not have expected this. I was expecting that the whole time, but yes, right? Let's say for the sake of the story, yes. Okay, your visual ganglia is going to light up because it's going to hurt. And now your model of the world includes the fact that I may pinch you if I approach my hand to your arm. So if I try again, you're going to recoil, and that's your critic, your predictor of your ultimate pain, the system that predicts that something bad is going to happen when you recoil to avoid it. So even that can be learned. That is definitely what allows you to define some goals. The fact that you're a school child, you wake up in the morning and you go to school, it's not because you necessarily like waking up early and going to school, but you know that there is a long-term objective you're trying to optimize.

欧内斯特·贝克尔，我不确定你是否熟悉这位哲学家，他写了《拒斥死亡》这本书，他的观点是人类的核心动机之一是我们对死亡的恐惧，对死亡的害怕。这就是我们与猫不同的地方。猫只是在生存，它们……

Ernest Becker, I'm not sure if you're familiar with the philosopher, he wrote the book Denial of Death, and his idea is that one of the core motivations of human beings is our terror of death, our fear of death. That's what makes us unique from cats. Cats are just surviving, they do...

你认为人类什么时候意识到自己会死？是早期习得的吗？在什么时刻你真正理解死亡是什么？我认为大多数人并没有真正理解死亡。大多数人相信你会去天堂之类的。

When do you think humans understand that they're going to die? Is it learned early on? At what point do you realize what death really is? I think most people don't actually realize what death is. Most people believe you go to heaven or something.

对此我要反驳一下。欧内斯特·贝克尔和谢尔登·所罗门等人的观点我觉得很有说服力。在人生早期，大约七八岁时，会有一些时刻你深刻体验到这种认识带来的恐惧。你会想：‘天哪，这就是神秘，就是恐惧。’就像你是猎物，是坐在黑暗丛林里的小鹿，四周充满恐惧。于是你回到心灵的舒适区，那里有深刻的意义，假装自己是不朽的。宗教对此有帮助。你可以用各种方式自欺欺人：沉迷于忙碌，设定小目标，以为一切会永远持续。你隐约知道自己会死，但并没有真正理解。这似乎是人类本性的一个核心独特方面：我们能够真正理解生命是有限的。

To push back on that, Ernest Becker and Sheldon Solomon and those folks have ideas I find compelling. There are moments early in life, around seven or eight years old, when you deeply experience the terror of this realization. You think, 'Holy crap, this is the mystery, the terror.' It's like you're a little prey, a baby deer sitting in the darkness of the jungle, surrounded by terror. So you go back to the comfort of your mind, where there is deep meaning, and pretend you're immortal. Religion helps with that. You can delude yourself in all kinds of ways: lose yourself in busyness, have little goals, think it will go on forever. You kind of know you're going to die, but you don't really understand it. That seems a core unique aspect of human nature: that we can really understand that this life is finite.

这里面有很多不同点。首先，我不认为我们和猫之间有质的区别。区别在于我们有更好的长期预测能力，所以我们对世界如何运作有更好的理解。我们有比猫更好的规划引擎。但规划的动机是什么？

There are a bunch of different things there. First, I don't think there is a qualitative difference between us and cats. The difference is that we have a better long-term ability to predict, so we have a better understanding of how the world works. We have a better planning engine than cats. But what's the motivation for planning?

我认为这只是拥有更好规划引擎的副作用，因为它让我们更聪明。智能的本质是预测能力。因为我们更聪明，作为副作用，我们也能对自己的未来存在与否做出预测。

I think it's just a side effect of having a better planning engine because it makes us smarter. The essence of intelligence is the ability to predict. Because we are smarter, as a side effect we also have the ability to make predictions about our own future existence or lack thereof.

你说宗教对此有帮助。我认为宗教有害；它让人们担心死后会发生什么。如果你相信死后就不存在了，那就完全解决了问题。至少你不担心死后的事。我不知道你为什么担心这一生，因为这是你唯一拥有的。

You say religion helps with that. I think religion hurts; it makes people worry about what happens after death. If you believe you just don't exist after death, that solves the problem completely. At least you're not worrying about what happens after death. I don't know why you worry about this life because that's the only one you have.

我不知道我会不会这么说。欧内斯特·贝克尔说，而且我基本同意，即使你相信没有上帝，你仍然会深深担忧。仍然有一种对这一切奥秘的深深担忧。它就这样结束，这怎么说得通？我不认为我们能真正理解。我们的生命、意识和自我如此多地投入在这个存在中。科学不断把人类从神坛上拉下来，但对我们个体人类来说，我们不喜欢被拉下神坛。

I don't know if I would say that. Ernest Becker says, and I agree with him more than not, that you do deeply worry even if you believe there's no god. There's still a deep worry about the mystery of it all. How does it make sense that it just ends? I don't think we can truly understand that. So much of our life, consciousness, and ego is invested in this being. Science keeps bringing humanity down from its pedestal, but for us individual humans, we don't like to be brought down from a pedestal.

但你看，你对此很坦然。欧内斯特·贝克尔会说，你坦然是因为那对你来说是一种更平静的存在，但你并非真的坦然；你在逃避。事实上，一些在人生早期经历过深度创伤的人，在寻求广泛治疗之前，会说‘我很好’。就像当你和真正愤怒的人交谈时：‘你怎么样？’‘我很好。’问题在于发生了什么。我有过一次濒死体验：17 岁时发生了一次非常严重的摩托车事故。

But see, you're fine with it. What Ernest Becker would say is you're fine with it because that's just a more peaceful existence for you, but you're not really fine; you're hiding. In fact, some people who experience deep trauma earlier in life, before they seek extensive therapy, will say 'I'm fine.' It's like when you talk to people who are truly angry: 'How are you doing?' 'I'm fine.' The question is what's going on. I had a near-death experience: a very bad motorbike accident when I was 17.

但这并没有影响我对那个话题的反思。我基本上只是在扮演一点魔鬼代言人的角色，反驳并思考：真的有可能接受死亡吗？而另一面，我认为对 AI 和机器人来说更有趣的是，将这一点作为动机集合的一部分有多重要？不仅仅是避免从屋顶掉下来之类的事情，而是思考旅程的终点。如果你听斯多葛学派的话，这是一个很好的动力，它增加了紧迫感。所以也许真正恐惧死亡或意识到它，可能会给当下赋予更深的意义和紧迫感，让人充分活好。

So but that didn't have any impact on my reflection on that topic. So I'm basically just playing a bit of a devil's advocate, pushing back and wondering: is it truly possible to accept death? And the flip side, that's more interesting I think for AI and robotics, is how important is it to have this as one of the suite of motivations? To not just avoid falling off the roof or something like that, but ponder the end of the ride. If you listen to the Stoics, it's a great motivator, it adds a sense of urgency. So maybe to truly fear death or be cognizant of it might give a deeper meaning and urgency to the moment, to live fully well.

也许我并不同意这一点。我的意思是，我认为激励我的是更多地了解人性。我认为人性和人类智能是一个巨大的谜团，是一个科学谜团，此外还有哲学等等。但我是科学的忠实信徒。我确实有一种信念，对于像大脑和心智这样的复杂系统，理解它的方法是通过你构建的人工制品来重现它，因为当你试图构建它时，你会知道它的本质。就像我以前用过的类比：我相信，就像我们只有在开始建造飞机时才真正理解空气动力学一样，这帮助我们理解了鸟类如何飞行。所以我认为这里有一个类似的过程，我们还没有一个完整的智能理论，但构建智能人工制品将帮助我们发展一些基础理论，不仅涵盖人工制品，也涵盖人类和生物智能。

Maybe I don't disagree with that. I mean, I think what motivates me here is knowing more about human nature. I think human nature and human intelligence is a big mystery, it's a scientific mystery, in addition to philosophical and etc. But I'm a true believer in science. And I do have kind of a belief that for complex systems like the brain and the mind, the way to understand it is to try to reproduce it with artifacts that you build, because you know what's essential to it when you try to build it. The same way I've used this analogy before: I believe the same way we only started to understand aerodynamics when we started building airplanes, and that helped us understand how birds fly. So I think there's a similar process here, where we don't have a full theory of intelligence, but building intelligent artifacts will help us perhaps develop some underlying theory that encompasses not just artificial implements but also human and biological intelligence in general.

所以你是问这个关于各种不同智能实体或智能问题的有趣人选。你对图灵测试或中文房间问题有什么看法？如果我们创造了一个表现出许多智能和意识属性的 AI 系统，你有多大把握认为那个实体是智能或有意识的？你现在正在构建具有智能的系统，并且有关于它们性能的指标，但那个指标是外部的。那么你呢？你愿意称一个东西为智能吗？你会像大多数人类一样，再次不高兴被从意识或智能的神坛上拉下来吗？

So you're an interesting person to ask this question about all kinds of different other intelligent entities or intelligences. What are your thoughts about the Turing test or the Chinese room question? If we create an AI system that exhibits a lot of properties of intelligence and consciousness, how comfortable are you thinking of that entity as intelligent or conscious? So you're trying to build now systems that have intelligence, and there's metrics about their performance, but that metric is external. So how are you? Are you okay calling a thing intelligent? Are you going to be like most humans and be, once again, unhappy to be brought down from a pedestal of consciousness or intelligence?

不，我会非常高兴通过构建具有类似能力的机器来更多地了解人性、人类心智和人类智能。如果这导致人类从神坛上降下一级，我完全没问题。这就是生活的现实。所以我对此没问题。现在你问我一些我可能有很多人不同意的观点。我认为如果我们考虑一个自主智能系统的设计，假设我们在让机器学习世界模型、预测世界模型方面取得了一定成功，我们构建了内在动机目标函数来驱动该系统的行为。系统还有感知模块，使其能够估计世界状态，然后有某种方式找出优化特定目标的动作序列。如果它有一个之前描述的那种批评者，那个让你在我第二次试图捏你时缩回手臂的东西，那么一个智能自主机器将会有情感。我认为情感是自主智能的一个组成部分。如果你有一个由内在动机和目标驱动的智能系统，如果它有一个批评者，允许你提前预测一个情况的结果是好是坏，它就会有情感。当它预测结果不好且需要避免时，它会有恐惧。当它预测结果好时，它会有兴奋。如果它有与人类建立关系的驱动力，以人类的方式，它就会具有社交性，对吧？所以它会有关于依恋之类的情感。所以我认为科幻片中看到指挥官数据有一个可以关闭的情感芯片，我觉得那很荒谬。

No, I'll be very happy to understand more about human nature, human mind, and human intelligence through the construction of machines that have similar abilities. And if a consequence of this is to bring down humanity one notch from its pedestal, I'm just fine with it. That's just the reality of life. So I'm fine with that. Now you were asking me about things that opinions I have that a lot of people may disagree with. I think if we think about the design of an autonomous intelligence system, so assuming that we are somewhat successful at getting machines to learn models of the world, predicting models of the world, we build intrinsic motivation objective functions to drive the behavior of that system. The system also has perception modules that allow it to estimate the state of the world, and then have some way of figuring out the sequence of actions to optimize a particular objective. If it has a critic of the type that was describing before, the thing that makes you recall your arm the second time I tried to pinch you, an intelligent autonomous machine will have emotions. I think emotions are an integral part of autonomous intelligence. If you have an intelligent system that is driven by intrinsic motivation, by objectives, if it has a critic that allows you to predict in advance whether the outcome of a situation is going to be good or bad, it's going to have emotions. It's going to have fear when it predicts that the outcome is going to be bad and something to avoid. It's going to have elation when it predicts it's going to be good. If it has drives to relate with humans, in some ways the way humans have, it's going to be social, right? And so it's going to have emotions about attachment and things of that type. So I think the sci-fi thing where you see Commander Data having an emotion chip that you can turn off, I think that's ridiculous.

这里有一个困难的哲学社会问题：你认为会有像机器人民权运动那样的时刻吗？先不说运动，但像最高法院那样的讨论，认为特定种类的机器人、特定种类的系统应该享有与人类相同的权利，因为它们可以像人类一样受苦，诸如此类的事情？

Here's the difficult philosophical social question: do you think there will be a time like a civil rights movement for robots? Where, forget the movement, but a discussion like the Supreme Court that particular kinds of robots, particular kinds of systems, deserve the same rights as humans because they can suffer just as humans can, all those kinds of things?

嗯，也许吧。也许不会。想象一下，人类可以死去然后被恢复，比如你可以被 3D 打印出来，你的大脑可以被最精细地重建。在这种情况下，我们对权利的想法会改变。如果你总是有备份，你总是可以恢复，也许谋杀的重要性会降低一个档次。没错。但你做危险事情的欲望，比如跳伞、赛车或飞机特技，可能会增加。做很多这样的事情或探索危险区域会没问题。这会改变你的关系。所以现在机器人很可能会是这样，因为它们可能基于与当今技术有些相似的技术，而且你总是可以有备份。所以这是可能的。我不知道你是否喜欢电子游戏，但有一款游戏叫《暗黑破坏神》。我的儿子们是这款游戏的超级粉丝。事实上，他们制作了一款受其启发的游戏。我的三个儿子共同经营一个游戏设计工作室。他们大约一年前推出了一款游戏。但在《暗黑破坏神》中，有一种叫做“硬核模式”的东西，如果你死了，你就完了，就是这样。所以对于 AI 系统来说，它们有可能成功运行，并且我们以某种方式对待它们，因为它们必须融入人类社会。它们必须能够死亡，不允许复制。事实上，复制是非法的。对人类来说也是可能的，比如克隆将是违法的，即使可能。因为克隆不是复制，对吧？我的意思是，你不能复制一个人的心智和经历，对吧？它只是一个延迟的双胞胎。但这就是我们之前谈论计算机时说的。

Well, perhaps. Perhaps not. Imagine that humans were that you could die and be restored, like you could be 3D reprinted and your brain could be reconstructed in its finest details. Our ideas of rights will change in that case. If you can always have a backup, you could always restore, maybe the importance of murder will go down one notch. That's right. But also your desire to do dangerous things like skydiving or race car driving, or airplane aerobatics, would probably increase. It would be fine to do a lot of those things or explore dangerous areas. It would kind of change your relationship. So now it's very likely that robots would be like that, because they'll be based on perhaps technology that is somewhat similar to today's technology, and you can always have a backup. So it's possible. I don't know if you like video games, but there's a game called Diablo. My sons are huge fans of this. In fact, they made a game that's inspired by it. My three sons have a game design studio between them. They came out with a game about a year ago. But in Diablo, there's something called hardcore mode, which if you die, you're gone, that's it. And so it's possible with AI systems for them to be able to operate successfully and for us to treat them in a certain way because they have to be integrated in human society. They have to be able to die, no copies allowed. In fact, copying is illegal. It's possible with humans as well, like cloning will be illegal even what's possible. Because cloning is not copying, right? I mean, you don't reproduce the mind of the person and experience, right? It's just a delayed twin. But then it's what we were talking about with computers.

你可以复制自己，完美保存心智状态，但这可能是非法的，因为这会破坏系统的动机。假设你有一个家用机器人，它经过一些预训练，能做很多事情，但有个性，让它与众不同。你和它一起生活了五年，彼此产生了感情，它了解你很多。或者它可能是一个虚拟助手，住在你的 AR 眼镜里。那个系统中的智能有点像你的孩子或博士生，因为里面有很多你的东西。如果它是一个活物，你会无偿这么做。如果是你的孩子，他可以过自己的生活，你从他那里学到东西并不意味着你拥有他。但如果是一个你训练的机器人，你可能拥有知识产权。我指的是永久价值，你的一部分在里面。如果机器人被毁而没有备份，你会失去很多投资，就像一个人去世一样。但你也拥有知识产权，因为系统是根据你的存在微调的，所以它是原始模型的独特实例。还有隐私问题：如果机器人有自己的意志，决定为别人工作，或者觉得和你生活不下去，那么它从你那里学到的一切，你怎么删除个人信息？这是一个伦理问题：你能为了隐私而抹去一个智能机器人的心智吗？你不能对人类这样做，你可以让他们闭嘴，但不能完全控制他们，也不能抹去他们。这是关系的问题：分手后你无法抹去另一个人。对于机器人，我认为也必须一样，我们的互动需要风险才能深刻体验。所以你必须能够失去你的机器人朋友，它可能会去发推文说你坏话。但你能为了保护隐私而谋杀机器人吗？如果它决定离开，我认为对于某些机器人，如果宣布它有感知能力，专为人类互动设计，那么你不能谋杀它，就像谋杀人类一样。但如果你备份了机器人，保存在硬盘上，未来这可能非法，就像盗版一样。但那是你自己的机器人。你可以抹去它的记忆，让它不再知道你的信息，但你技术上还有备份。最高法院会有精彩演讲，说你可以抹去机器人的心智就像抹去人类的心智，我们都能受苦。会有像奥巴马那样的人物演讲，说机器人和人类一样，都能受苦、希望、养家。情感是人类互动和人机互动中迷人的强大方面。如果它们能展现情感，最终会让我们深刻思考人权，我们珍视人类和动物的什么。这就是为什么机器人和 AI 很好，它让我们提出真正的好问题，困难的问题。你问到了中文房间论证，如果看起来真实，它是否真实？我认为中文房间论证很荒谬。对于不了解的人，中文房间基本上是通过遵循一个巨大的算法代码书来模仿智能系统的行为，这本书告诉你每种情况下如何回应。但这真的是智能吗？它就像一个巨大的查找表：当这个人说这个，你回答那个。这是智能吗？因为智能似乎是一个更有趣、更复杂的机制，而不是查找表。我不这么认为。真正的问题是：你能以某种方式机械化智能吗？即使涉及学习，答案当然是肯定的。第二个问题是：假设你能在不同于生物硬件的硬件上重现智能，比如计算机，你能在所有人类智能的领域匹配人类智能吗？这是强 AI 的假设。我的回答是毫无疑问的“是”，机器最终会在所有人类智能领域变得比人类更智能。这不是明天的事，需要很长时间，不管埃隆和其他人声称什么。这比很多人想象的要难得多。那些五年前认为更简单的人，现在意识到需要更长时间，包括 DeepMind 的一些人。

that you'll be able to copy you right you'll be able to perfectly save pickle the the the mind state and it's possible that that would be illegal because that goes against um that will destroy the motivations of the system okay so let's say you you have a domestic robot okay sometime in the future yes and uh the domestic robot you know comes to you kind of somewhat pre-trained you know it can do a bunch of things yes but it has a particular personality that makes it slightly different from the other robots because that makes them more interesting and then because it's you know it's live with you for five years you've you've grown some attachment to it and vice versa and it's learned a lot about you or maybe it's not a household robot maybe it's uh maybe it's a virtual assistant that lives in your you know augmented reality glasses or whatever right uh you know the horror movie type thing right um and that system to some extent the the intelligence in that system is a bit like your child or maybe your phd student in a sense that there's a lot of you in that in that machine now right yeah and so if it were a living thing you would do this for free if you want right if it's your child your child can you know then live his or her own life and you know the fact that they learn stuff from you doesn't mean that you have any ownership of it right yeah but if it's a robot that you've trained perhaps you have some uh yeah intellectual property claim about intellectual property oh i thought you meant like uh permanent value in the sense this part of you is in well there is permanent value right so you would lose a lot if that robot were to be destroyed and you you had no backup you would lose a lot you know you a lot of investment you know kind of like a uh you know a person dying you know um that that a friend of a friend of you was dying or or a co-worker or something like that um but also uh you have like intellectual property rights in the sense that that that system is fine-tuned to your particular existence so that's now a very unique instantiation of that original background model whatever it was that arrived and then there are issues of privacy right because now imagine that that robot has its own kind of volition and decides to work from someone else yes or kind of you know thinks life with you is sort of untenable or whatever right um now all the things that that system learned from you uh you know how can you like you know delete all the personal information that that system knows about you yeah i mean that would be kind of an ethical question like you know can you erase the the mind of a of a intelligent robot uh to protect your your privacy yeah you can't do this with humans you can ask them to shut up but that you don't have complete power over them can't erase humans yeah it's the problem with relationships you know that you break up you can't you can't erase the other human with robots i think it'll have to be the same thing with robots that that risk that there has to be um some risk to our interactions to truly experience them deeply it feels like so you have to be able to lose your robot friend and that robot friend to go tweeting about how much of an you are but then are you allowed to you know murder the robot to protect your private information yeah probably decides to leave i have the situation that for robots with with certain like it's almost like uh regulation if you declare your robot to be let's call it sentient or something like that like this this robot is designed for human interaction then you're not allowed to murder these robots it's the same as murdering other humans well but what about you do a backup of the robot you do preserve on the on a hard drive or the equivalent in the future that might be illegal just like it's like priority uh piracy is illegal but it's your own it's your own robot right but you can't you don't but then but then you can wipe out his brain so this robot doesn't know anything about you anymore but you still have technically a certain existence because you backed it up and then there'll be these great speeches at the supreme court by saying oh sure you can erase the mind of the robot just like you can erase the mind of a human we both can suffer there'll be some epic like obama type character with a speech that we we like the robots and the humans are the same we can both suffer we can both hope we can both all those all those kinds of things raise families all that kind of stuff it's it's uh interesting for these just like you said emotion seems to be a fascinatingly powerful aspect of human human interaction human robot interaction and if they're able to exhibit emotions at the end of the day that's probably going to have us deeply consider human rights like what we value in humans what we value in other animals that's why robots and ai is great it makes us ask uh really good questions the hard questions yeah but you ask about you asked about the chinese room type argument you know is it real if it looks real yeah i think the chinese room argument is the ridiculous one so so so for people who don't know chinese room is uh you ca you can i don't even know how to formulate it well but basically you can mimic the behavior of an intelligent system by just following a giant algorithm code book that tells you exactly how to respond in exactly each case but is that really intelligent it's like a giant lookup table when this person says this you answer this when this person says this you answer this and if you understand how that works you have this giant nearly infinite lookup table is that really intelligence because intelligence seems to be a mechanism that's much more interesting and complex than this lookup table i don't think so so the i mean the real question comes down to do you think uh you know you can you can mechanize uh intelligence in some way even if that involves uh learning and the answer is of course yes there's no question there's a second question then which is uh assuming you can uh reproduce intelligence in sort of different hardware than biological hardware you know like computers uh can you you know match uh human intelligence in all the domains in which humans are intelligent is it possible right so that's the hypothesis of a strong ai the answer to this in my opinion is unqualified yes this will swell happen at some point there's no question that machines at some point will become more intelligent than humans in all domains where humans are intelligent this is not for tomorrow it's going to take a long time regardless of what you know elon and others have claimed or believed this is a lot a lot harder than many of many of those guys think it is and many of those guys who thought it was simpler than that years you know five years ago now i think it's hard because it's been five years and they realize it's it's gonna take a lot longer that includes a bunch of people deepmind for example but um

我认为中文房间论证很荒谬。对于不了解的人，中文房间基本上是通过遵循一个巨大的算法代码书来模仿智能系统的行为，这本书告诉你每种情况下如何回应。但这真的是智能吗？它就像一个巨大的查找表。这是智能吗？智能似乎是一个更有趣、更复杂的机制，而不是查找表。我不这么认为。真正的问题是：你能以某种方式机械化智能吗？即使涉及学习，答案当然是肯定的。第二个问题是：假设你能在不同于生物硬件的硬件上重现智能，比如计算机，你能在所有领域匹配人类智能吗？这是强 AI 的假设。我的回答是毫无疑问的“是”。这会在某个时刻发生。毫无疑问，机器会在所有人类智能领域变得比人类更智能。这不是明天的事，需要很长时间，不管埃隆和其他人声称什么。这比很多人想象的要难得多。许多五年前认为更简单的人现在意识到需要更长时间，包括 DeepMind 的一些人。

I think the Chinese room argument is ridiculous. For people who don't know, the Chinese room is basically you can mimic the behavior of an intelligent system by just following a giant algorithm code book that tells you exactly how to respond in each case. But is that really intelligent? It's like a giant lookup table. Is that really intelligence? Intelligence seems to be a mechanism that's much more interesting and complex than this lookup table. I don't think so. The real question is: can you mechanize intelligence in some way, even if that involves learning? The answer is of course yes. There's a second question: assuming you can reproduce intelligence in different hardware than biological hardware, like computers, can you match human intelligence in all domains? That's the hypothesis of strong AI. The answer in my opinion is unqualified yes. This will happen at some point. There's no question that machines will become more intelligent than humans in all domains where humans are intelligent. This is not for tomorrow; it's going to take a long time, regardless of what Elon and others have claimed. It's a lot harder than many of those guys think. Many who thought it was simpler five years ago now realize it's going to take a lot longer. That includes a bunch of people at DeepMind, for example.

哦，有意思，我其实还没和 DeepMind 的人联系过。但有些东西，比如埃隆或民主……我的意思是，有时候你的角色必须设定一些更近而不是更远的截止日期，来制造紧迫感，因为你知道，你必须相信不可能的事情是可能的，才能实现它。当然，这也有另一面，但很奇怪……如果你想做成什么事，就不能太愤世嫉俗。

Oh interesting, I haven't actually touched base with the DeepMind folks. But some of it, Elon or democracy... I mean, sometimes your role, you have to kind of create deadlines that are nearer than farther away, yeah, to kind of create an urgency, because you know, you have to believe the impossible as possible in order to accomplish it. And there's of course a flip side to that coin, but it's a weird... you can't be too cynical if you want to get something done.

完全同意。但我的意思是，你必须激励人们去从事某些雄心勃勃的事业。所以，这肯定比我们想象的要难得多，但我毫不怀疑这会发生。现在人们有点担心这对人类意味着什么。他们会从神坛上被拉下来，降低好几个档次。这是好是坏？我的意思是，这只会赋予更多力量，对吧？它实际上是人类智能的放大器。

Absolutely, I agree with that. But I mean, you have to inspire people right to work on certain ambitious things. So you know, it's certainly a lot harder than we believe, but there's no question in my mind that this will happen. And now people are kind of worried about what that means for humans. They are going to be brought down from their pedestal, you know, a bunch of notches with that. And is that going to be good or bad? I mean, it's just going to give more power, right? It's an amplifier for human intelligence, really.

说到做酷而雄心勃勃的事情，Facebook AI Research 最近庆祝了它的八岁生日，或者你可以纠正我。回顾过去，FAIR 这八年有哪些成功、失败和经验教训？也许你还可以介绍一下新成立的 Meta AI 处于什么位置，它与 FAIR 的关系如何？

So speaking of doing cool ambitious things, the Facebook AI Research group has recently celebrated its 8th birthday, or maybe you can correct me on that. Looking back, what has been the successes, the failures, the lessons learned from the eight years of FAIR? And maybe you can also give context of where does the newly minted Meta AI fit into, how does it relate to FAIR?

让我稍微介绍一下整个组织架构。FAIR 几乎是在八年前成立的，当时还不叫 FAIR，几个月后才有了这个名字。我加入 Facebook 时，有一个叫 AI 小组的团队，大约有 12 名工程师和几位科学家，比如十名工程师和两名科学家这样。我作为总监管理了三年半，招聘了最初几位科学家，建立了文化，组织了团队，向 Facebook 领导层解释了基础研究是什么，如何在工业界运作，以及它需要保持开放等等。我认为它取得了无可置疑的成功，因为 FAIR 同时产出了一流的研究，推动了科学和技术进步，提供了像 PyTorch 这样的开源工具，同时对当时的 Facebook 和现在的 Meta 产生了直接或间接的影响——Meta 现在许多系统都基于 FAIR 启动的研究项目。如果把深度学习从 Facebook 服务中拿掉，公司基本上会崩溃。如今它完全围绕 AI 构建，对运营至关重要。三年半后我换了角色，成为首席科学家，不再管理 FAIR 的日常事务，而是更多思考战略之类的事情，并开展自己的研究。我有了自己的研究小组，从事自监督学习等工作，这是当总监时没时间做的。现在 FAIR 由 Joelle Pineau 和 Antoine Bordes 共同管理，因为 FAIR 现在分为两部分：FAIR Labs 是自下而上的科学家驱动研究，FAIR Excel 则更有组织性，适合需要更多关注和工程支持的大项目。Joelle 领导 FAIR Lab，Antoine Bordes 领导 FAIR Excel。它们分布在全球各地。毫无疑问，公司领导层认为这是一项非常值得的投资，这意味着它会长期存在。如果你非要用我不喜欢的商业模型来说，FAIR 尽管是一个基础研究实验室，但通过其他团队间接为公司带来了巨大价值。我卸任三年半后，Facebook AI 也成立了，它是一个更大的组织，涵盖了 FAIR。FAIR 包含在其中，但也有其他专注于应用研究或 AI 技术高级开发的团队，更贴近公司产品。基础研究成分少一些，但仍然是研究——这些团队也发表了很多论文，人员非常出色。但它是一种将可能非常实验性的 AI 技术从实验室原型扩展到可用产品的方式。所以 FAIR 是 Meta AI 的一个子集。它变得像 KFC 一样，只保留 F，没人关心 F 代表什么。我们很快会知道，可能在 2021 年底。这不是一个巨大的变化。Meta FAIR？听起来不太好，但品牌团队正在决定，他们犹豫了一段时间。他们告诉我们，他们会给出答案，是 FAIR 改名，还是只改变 F 的含义。

So let me tell you a little bit about the organization of all this. FAIR was created almost exactly eight years ago. It wasn't called FAIR yet; it took that name a few months later. At the time I joined Facebook, there was a group called the AI group that had about 12 engineers and a few scientists, like ten engineers and two scientists or something like that. I ran it for three and a half years as a director, hired the first few scientists, set up the culture, organized it, explained to the Facebook leadership what fundamental research was about and how it can work within industry, and how it needs to be open and everything. I think it's been an unqualified success in the sense that FAIR has simultaneously produced top-level research, advanced the science and the technology, provided open source tools like PyTorch and many others, but at the same time has had a direct or mostly indirect impact on Facebook at the time, now Meta, in the sense that a lot of systems that Meta is built around now are based on research projects that started at FAIR. If you were to take out deep learning out of Facebook services now and Meta more generally, I mean the company would literally crumble. It's completely built around AI these days, and it's really essential to the operations. So what happened after three and a half years is that I changed role. I became chief scientist, so I'm not doing day-to-day management of FAIR anymore. I'm more of a kind of think about strategy and things like that, and I conduct my own research. I have my own research group working on self-supervision and things like this, which I didn't have time to do when I was director. So now FAIR is run by Joelle Pineau and Antoine Bordes together, because FAIR is kind of split in two now. There's something called FAIR Labs, which is sort of bottom-up scientist-driven research, and FAIR Excel, which is slightly more organized for bigger projects that require a little more focus and more engineering support and things like that. So Joelle leads FAIR Lab and Antoine Bordes leads FAIR Excel. Where are they located? It's always delocalized all over. So there's no question that the leadership of the company believes that this was a very worthwhile investment, and what that means is that it's there for the long run. If you want to talk in these terms, which I don't like, there's a business model where FAIR, despite being a very fundamental research lab, brings a lot of value to the company, either mostly indirectly through other groups. Now what happened three and a half years ago when I stepped down was also the creation of Facebook AI, which was basically a larger organization that covers FAIR. So FAIR is included in it, but also has other organizations that are focused on applied research or advanced development of AI technology that is more focused on the products of the company. So less emphasis on fundamental research, less fundamental, but it's still research. I mean, there's a lot of papers coming out of those organizations, and people are awesome and wonderful to interact with. But it serves as a way to scale up AI technology which may be very experimental and sort of lab prototypes into things that are usable. So FAIR is a subset of Meta AI. It's become like KFC; it'll just keep the F, nobody cares what the F stands for. We'll know soon enough, probably by the end of 2021. This is not a giant change. Meta FAIR? Well, Meta doesn't sound too good, but the brand people are kind of deciding on this, and they've been hesitating for a while now. They tell us they're going to come up with an answer as to whether FAIR is going to change name or whether we're going to change just the meaning of the F.

哦，好主意。我会保留 FAIR，改变 F 的含义。这是我的偏好。

Oh, that's a good call. I would keep FAIR and change the meaning of the F. That would be my preference.

你知道，我倾向于把 F 变成 Fundamental（基础）。哦，那正是我研究的。哦，那太好了。对，然后 Meta AI，这就是 FAIR。FAIR，对。但人们还是会叫它 FAIR，对吧？没错，我喜欢。现在 Meta AI 是 Reality Labs 的一部分。Meta，新的 Facebook 叫 Meta，它大致分为 Facebook、Instagram、WhatsApp 和 Reality Labs。Reality Labs 涉及 AR、VR、远程呈现通信、技术等。你可以把它看作 Meta 的新产品和技术部分的组合。

You know, I would tend to turn the F into Fundamental. Oh, that's what I researched. Oh, that's really good. Yeah, then Meta AI, so this would be FAIR. FAIR, yeah. But you know, people will call it FAIR right? Yeah, exactly. I like it. And now Meta AI is part of Reality Labs. So Meta, the new Facebook is called Meta, and it's kind of divided into Facebook, Instagram, WhatsApp, and Reality Labs. Reality Labs is about AR, VR, telepresence communication, technology, and stuff like that. You can think of it as a combination of new products and technology part of Meta.

机器人的触觉传感是在那里吗？我看到你发过相关帖子。那是我在机器人派对上提到的。实际上是 FAIR，就是它。哦，好的。是的，还有触觉手套，对吧？那更像是 Reality Labs 的研究。我必须有研究。不过，触觉传感器非常有趣，比如将这种模态整合到整个传感套件中非常有意思。

Is that where the touch sensing for robots, I saw that you were posting about that? That's what I touched on for robotics party. FAIR actually, that's it. Oh, okay. Yeah, this is also... but there is the haptic glove, right? Yes, that has like... that's more Reality Labs research. I have to have research. But by the way, the touch sensors are super interesting, like integrating that modality into the whole sensing suite is very interesting.

那么你对元宇宙怎么看？你对 Facebook 和 Meta 在世界中角色的这种扩展怎么看？

So what do you think about the metaverse? What do you think about this whole expansion of the view of the role of Facebook and Meta in the world?

嗯，元宇宙确实应该被视为下一个……

Well, the metaverse really should be thought of as the next...

互联网的下一步，是让连接他人或内容的体验更引人入胜。我们进化并习惯于在三维环境中活动，在那里我们可以看见他人、近距离交谈，远处的人也能听到我们。现实世界中存在许多社会惯例，我们可以尝试移植过来。那么，这种体验到底能有多吸引人？人们会愿意整天戴着一副大护目镜吗？也许不会。但如果体验足够吸引人，也许就会。或者如果设备只是一副眼镜，技术足够进步的话。AR 的概念更容易理解：你戴着增强现实眼镜，里面有个虚拟助手帮你处理日常事务。但 AR 需要与现实互动，而 VR 可以完全脱离现实，给你自由。设计 VR 世界可能更容易。但你可以想象元宇宙是混合的：元宇宙中的物体可以叠加在现实世界上，或者只存在于虚拟现实中。

Step in the internet, trying to make the experience more compelling of being connected either with other people or with content. We are evolved and trained to evolve in 3D environments where we can see other people, talk to them when we're near them, and when other people are far away can hear us. There are a lot of social conventions in the real world that we can try to transpose. Now, how compelling is it going to be? Will people be willing to do this if they have to wear a huge pair of goggles all day? Maybe not. But if the experience is sufficiently compelling, maybe so. Or if the device is just a pair of glasses, technology makes sufficient progress for that. AR is a much easier concept to grasp: you have augmented reality glasses that contain a virtual assistant to help you in daily life. But with AR you have to contend with reality; with VR you can completely detach yourself from reality, so it gives you freedom. It might be easier to design worlds in VR. But you can imagine the metaverse being a mix: objects that exist in the metaverse pop up on top of the real world or only exist in virtual reality.

让我问个尖锐的问题。Facebook，现在的 Meta，被媒体描绘成社会的净负面，有时甚至是破坏性和邪恶的。你曾反驳这一点，为 Facebook 辩护。你能解释一下你的辩护吗？

Let me ask the hard question. Facebook, now Meta, has been painted by the media as a net negative for society, even destructive and evil at times. You've pushed back against this, defending Facebook. Can you explain your defense?

一些媒体描述的公司并不是我们在内部工作时所了解的公司。可能会有人说员工不了解真实情况，但我是副总裁，我对公司运作有相当清晰的了解。我当然不是无所不知，但我有不错的视野。他们描述的邪恶，我根本看不到。有一个很容易接受的说法：世界上所有坏事，你朋友相信疯狂言论的原因，社交媒体、尤其是 Facebook 是方便的替罪羊。但你必须看数据。Facebook 是否在政治上让人们两极分化？有学术研究证明吗？青少年使用 Instagram 更多是否自我评价更低？人们越用 Facebook 是否越容易对对立观点激动？一项又一项研究表明这些都不成立。这些是独立学术研究，不是 Facebook 或 Meta 资助的。斯坦福大学的研究，以及我在纽约大学的一些同事的研究，我和他们没有联系。最近有一项研究，他们付钱给前南斯拉夫地区的人们，在斯雷布雷尼察大屠杀纪念日前一段时间不使用 Facebook。结果发现，那些人不使用 Facebook 后反而更加两极分化，而更多使用 Facebook 的人两极分化程度更低。还有一项斯坦福大学经济学家的研究，试图找出美国日益两极分化的原因，这种趋势在马克·扎克伯格出生前就已经持续了 40 年。如果有原因，那也不是 Facebook 或社交媒体。你可以说社交媒体加速了它，但这是一个持续的演变。与其他国家比较，比如西德、丹麦或其他国家，他们同样使用 Facebook，却没有变得更两极分化，反而更不两极分化。如果你想找因果关系，你可以找到替罪羊，但找不到原因。要解决问题，你必须找到真正的原因。让我恼火的是，人们指责 Facebook 做了别人做的坏事，而那些别人却没有任何行动。顺便说一句，那些别人包括《华尔街日报》的所有者，所有这些文章都是在那上面发表的。

The company being described in some media is not the company we know when we work inside. It could be claimed that employees are uninformed about what really goes on, but I'm a vice president; I have a pretty good vision of what goes on. I don't know everything, obviously, but I have decent vision. This evil being described, I just don't see it. There is an easy story to buy: all the bad things in the world, the reason your friend believes crazy stuff, there's an easy scapegoat in social media, Facebook in particular. But you have to look at the data. Does Facebook polarize people politically? Are there academic studies showing this? Do teenagers think less of themselves if they use Instagram more? Do people get more riled up against opposite sides if they are more on Facebook? Study after study shows none of this is true. These are independent studies by academics, not funded by Facebook or Meta. A study by Stanford, by some of my colleagues at NYU with whom I have no connection. There was a recent study where they paid people in the former Yugoslavia to not use Facebook for a while before the anniversary of the Srebrenica massacres. It turned out those people ended up more polarized than at the beginning, and the people who were more on Facebook were less polarized. There's a study from Stanford economists that tried to identify the causes of increasing polarization in the US, which has been going on for 40 years before Mark Zuckerberg was born, continuously. If there is a cause, it's not Facebook or social media. You could say social media accelerated it, but it's a continuous evolution. Compare with other countries like West Germany, Denmark, or other countries; they use Facebook just as much and they're not getting more polarized, they're getting less polarized. If you want to find a causal relationship, you can find a scapegoat but not the cause. To fix the problem, you have to find the right cause. What riles me up is that people are accusing Facebook of bad deeds done by others, and those others are not doing anything about them. By the way, those others include the owner of the Wall Street Journal, in which all those papers were published.

我应该提一下，我正在和 Yann LeCun 做这个播客，还有马克·扎克伯格。可能你可以和他们进行这些对话，因为我觉得很有意思。即使 Facebook 有一些可衡量的负面影响，你也不能孤立地看待它。你必须考虑它连接我们的所有积极方式。就像每一项技术，你不能只说分裂增加了。可能谷歌搜索引擎也造成了分裂增加。我们必须考虑它给世界带来了多少信息。我敢肯定，如果只看分裂，维基百科也造成了更多分裂。我们必须看世界的整体背景，以及它是否让世界变得更好。

I should mention that I'm talking to Yann LeCun for this podcast, and also Mark Zuckerberg. Probably these conversations you can have with them because it's very interesting to me. Even if Facebook has some measurable negative effect, you can't just consider that in isolation. You have to consider all the positive ways it connects us. Like every technology, you can't just say there's an increase in division. Probably Google search engine has created an increase in division. We have to consider how much information it brought to the world. I'm sure Wikipedia created more division if you just look at division. We have to look at the full context of the world and whether it made a better world.

印刷机造成了更多分歧，对吧？印刷机发明时，最早印刷的书是《圣经》之类的东西，这让人们可以自己阅读《圣经》，而不是只从欧洲的牧师那里获得信息。这引发了新教运动和 200 年的宗教迫害与战争。这是印刷机的负面副作用。社交网络远没有印刷机那么糟糕，但没有人会说印刷机是个坏主意。这很大程度上是认知问题，而且这里有很多不同的激励因素在起作用。

The printing press has created more differences, right? When the printing press was invented, the first books printed were things like the Bible, and that allowed people to read the Bible by themselves, not get the message uniquely from priests in Europe. They created the Protestant movement and 200 years of religious persecution and wars. That's a bad side effect of the printing press. Social networks aren't nearly as bad as the printing press, but nobody would say the printing press was a bad idea. A lot of it is perception, and there are a lot of different incentives operating here.

也许快速说一句：既然你是 Facebook 和 Meta 的高层领导之一，我相信 Facebook 涉及许多需要解决的惊人技术挑战。其中很多可能是在计算机基础设施方面。

Maybe a quick comment: since you're one of the top leaders at Facebook and Meta, I'm sure Facebook involves a lot of incredible technological challenges that need to be solved. A lot of it probably is on the computer infrastructure.

硬件方面，我是说那真是巨大的一笔投入。也许你能给我一些背景，关于 Shrek 的生活中有多少是 AI，多少是底层计算，多少是到处飞着处理商业事务？同样对于扎克伯格，马克·扎克伯格，他们真的很专注于 AI。

Hardware, I mean it's just a huge amount. Maybe can you give me context about how much of Shrek's life is AI and how much of it is low-level compute, how much of it is flying all around doing business stuff? And the same with Zuckerberg. Mark Zuckerberg, they really focus on AI.

我的意思是，当然，在 FAIR 创建前夕以及之后至少一年里，马克非常非常专注于 AI，投入了大量精力。这就是他的风格：当他对某件事感兴趣时，他会阅读所有相关的东西。比如，他在加入之前就读过我的一些论文。他学到了很多，做了笔记。Shrek 也非常投入。我是说，Shrek 有一种我尽管不再年轻但仍试图保持的东西，那就是对科学和技术的好奇心，他肯定有。他也是一个很棒的人，作为管理者，与人打交道等等。马克其实也是。所以我的意思是，他们都是非常人性化的人。就马克而言，考虑到他的经历，他惊人地有人情味。媒体描绘的他的个性完全是错的。但你必须知道如何应对媒体，所以我也把部分责任归咎于他。你必须像乐队指挥一样，以某种方式引导媒体和公众，向他们传达真实的自己。如果内心有深度和善良，这很难，他可能并不擅长。所以，他需要学习。看到这种情况很遗憾，我会和他谈谈。但 Shrek 正在慢慢退下来。看到长期任职的人慢慢离开总是令人难过，我想是时间到了。我认为他已经完成了自己设定的目标，而且他有家庭优先事项等等。我理解。13 年左右，这是一段不错的历程，在硅谷基本上就是一辈子。是的，因为那是狗年。

I mean certainly, in the run-up to the creation of FAIR and for at least a year after that, if not more, Mark was very, very much focused on AI and was spending quite a lot of effort on it. And that's his style: when he gets interested in something, he reads everything about it. You know, he read some of my papers, for example, before he joined. And so he learned a lot about it, like notes, right. And Shrek was really into it also. I mean, Shrek is really kind of, you know, has something I've tried to preserve despite my not-so-young age, which is a sense of wonder about science and technology, and he certainly has that. He's also a wonderful person, I mean, in terms of as a manager, dealing with people and everything. Mark also, actually. So I mean, they're very human people. In the case of Mark, shockingly human, given his trajectory. I mean, the personality of him that is painted in the press is just completely wrong. But you have to know how to play the press, so I put some of that responsibility on him too. You have to, it's like the director of an orchestra, you have to play the press and the public in a certain kind of way where you convey your true self to them. If there is depth and kindness, it's hard, and he's probably not the best at it. So yeah, you have to learn. And it's sad to see. I'll talk to him about it. But Shrek is slowly stepping down. It's always sad to see folks be there for a long time and slowly, I guess time. I think he's done the thing he set out to do, and he's got family priorities and stuff like that. I understand. After 13 years or something, it's been a good run, which in Silicon Valley is basically a lifetime. Yeah, you know, because it's dog years.

欧洲的会议刚刚结束。让我回到另一件事。你发帖说你合著的论文被欧洲会议拒绝了，正如你自豪地说的，被拒绝了。你能开个玩笑吗？是的，我知道。你能描述一下这篇论文，它的想法是什么吗？也许这也是一个很好的机会来问一下，审稿过程的优缺点是什么，什么有效，什么无效。

Europe's the conference just wrapped up. Let me just go back to something else. You posted the paper you co-authored was rejected from Europe as you said proudly in quotes rejected. Can you joke? Yeah, I know. Can you describe this paper and like what was the idea in it? And also maybe this is a good opportunity to ask what are the pros and cons, what works and what doesn't about the review process.

好的，我先谈谈论文，之后再谈审稿过程。这篇论文叫 VICReg。我之前提到过：方差不变性协方差正则化。这是一种技术，一种非对比学习技术，用于我所说的联合嵌入架构。孪生网络就是联合嵌入架构的一个例子。联合嵌入架构是……让我稍微回溯一下。如果你想做监督学习，你可以通过预测来实现。假设你想训练系统预测视频。你给它一个视频片段，训练系统预测该视频片段的后续。现在，因为你需要处理不确定性，因为有很多可能的后续，你需要以某种方式处理。你需要一种方法让系统能够产生多个预测。我知道的唯一方法是通过所谓的潜变量。所以你有一些隐藏的向量或变量，你可以在一个集合上变化它，或者从一个分布中抽取。当你在这个集合上变化这个向量时，输出预测会在一个可能的预测集合上变化。这被称为生成潜变量模型。现在有一种替代方法来处理不确定性。不是直接预测片段的下一帧，而是也将这些帧通过另一个神经网络。所以你现在有两个神经网络：一个看视频片段的初始部分，另一个在训练期间看后续部分。你要做的是学习这两个视频片段的表示，这些表示对视频片段本身信息量最大，但使得你可以轻松地从第一个片段的表示预测第二个片段的表示。你可以用最大化互信息之类的东西来形式化。但你要的是两个视频片段的信息性表示，它们可以相互预测。这意味着第二个视频片段中有很多细节是无关的。假设一个视频片段是摄像机平移一个场景。房间的一部分将被揭示，我可以某种程度上预测房间会是什么样子，但我可能无法预测地面纹理的细节、瓷砖的边界等等。这些是无关的细节，也许我的表示会消除它们。所以我需要训练第二个神经网络，使得当后续视频片段在所有可能的后续上变化时，表示不变。明白了吗？所以这和你做相似性学习是一样的，对吧？所以这是处理预测中多模态的两种方式。第一种方式，你用潜变量参数化预测，但本质上预测像素。第二种方式，你不预测像素；你预测像素的抽象表示，并保证这个抽象表示包含尽可能多的关于输入的信息，但丢弃了你实际上无法预测的所有东西。我以前是第一种方法的忠实粉丝，事实上，在这篇与 Chen Mishra 合作的论文和博客文章《智能的暗物质》中，我有点提倡这种方法。但在过去一年半里，我完全改变了想法。我现在是第二种方法的忠实粉丝，这是因为在过去一年半到两年里提出的一小批算法，包括 VICReg，它的前身我提到过的 Barlow Twins，来自 DeepMind 朋友的方法 BYOL，以及其他一些类似的方法。它们都基于联合嵌入的思想。其中一些有明确的准则，是互信息的近似。

Yeah, let me talk about the paper first. I'll talk about the review process afterwards. The paper is called VICReg. This is I mentioned that before: Variance Invariance Covariance Regularization. And it's a technique, a non-contrastive learning technique for what I call joint embedding architecture. So Siamese nets are an example of joint embedding architecture. So joint embedding architecture is... let me back up a little bit. If you want to do supervised learning, you can do it by prediction. So let's say you want to train your system to predict video. You show it a video clip and you train the system to predict the continuation of that video clip. Now, because you need to handle uncertainty, because there are many continuations that are plausible, you need to handle this in some way. You need a way for the system to be able to produce multiple predictions. And the only way I know to do this is through what's called a latent variable. So you have some sort of hidden vector or variable that you can vary over a set or draw from a distribution. As you vary this vector over a set, the output prediction varies over a set of plausible predictions. That's called a generative latent variable model. Now there is an alternative to this to handle uncertainty. Instead of directly predicting the next frames of the clip, you also run those through another neural net. So you now have two neural nets: one that looks at the initial segment of the video clip, and another one that looks at the continuation during training. What you're trying to do is learn a representation of those two video clips that is maximally informative about the video clips themselves, but such that you can predict the representation of the second video clip from the representation of the first one easily. You can sort of formalize this in terms of maximizing mutual information, something like that. But what you want is informative representations of the two video clips that are mutually predictable. That means there are a lot of details in the second video clip that are irrelevant. Let's say a video clip consists of a camera panning a scene. There's going to be a piece of that room that is going to be revealed, and I can somewhat predict what that room is going to look like, but I may not be able to predict the details of the texture of the ground and where the tiles are ending and stuff like that. Those are irrelevant details that perhaps my representation will eliminate. So what I need is to train this second neural net in such a way that whenever the continuation video clip varies over all the plausible continuations, the representation doesn't change. Got it? So it's the same kind of thing as you're doing with similarity learning, right? So these are two ways to handle multi-modality in a prediction. In the first way, you parameterize the prediction with a latent variable, but you predict pixels essentially. In the second one, you don't predict pixels; you predict an abstract representation of pixels, and you guarantee that this abstract representation has as much information as possible about the input but drops all the stuff that you really can't predict essentially. I used to be a big fan of the first approach, and in fact, in this paper with Chen Mishra, this blog post 'The Dark Matter of Intelligence', I was kind of advocating for this. But in the last year and a half, I've completely changed my mind. I'm now a big fan of the second one, and it's because of a small collection of algorithms that have been proposed over the last year and a half or two years to do this, including VICReg, its predecessor called Barlow Twins which I mentioned, a method from our friends at DeepMind called BYOL, and a bunch of others that work similarly. So they're all based on this idea of joint embedding. Some of them have an explicit criterion that is an approximation of mutual information.

有些方法会奏效，但我们并不真正知道原因。已经有很多理论论文，但为什么它有效？不，这并不糟糕，因为我们去掉它后它仍然有效。所以存在很大的争论。但重要的是，我们现在有了一系列非对比性联合嵌入方法，我认为这是自切片面包以来最棒的东西。我对此非常兴奋，因为我认为这是我们构建预测性世界模型的最佳途径，同时学习世界的层次化表示，保留世界中重要的部分，消除无关的部分。

Some others will work, but we don't really know why. There have been lots of theoretical papers, but why does it work? No, it's not bad because we take it out and it still works. So there's a big debate. But the important point is that we now have a collection of non-contrastive joint embedding methods, which I think is the best thing since sliced bread. I'm super excited about this because I think it's our best shot for techniques that would allow us to build predictive world models and at the same time learn hierarchical representations of the world, where what matters about the world is preserved and what is irrelevant is eliminated.

这种表示是针对图像序列中的前后空间，还是针对单张图像？

Is the representation for before and after in the space in a sequence of images, or is it for single images?

它可以用于单张图像，也可以用于序列。不一定是图像；这可以应用于文本、音频或几乎任何信号。我在寻找通用的方法，不局限于特定模态。

It would be either for a single image or for a sequence. It doesn't have to be images; this could be applied to text, audio, or just about any signal. I'm looking for methods that are generally applicable, not specific to one particular modality.

这篇论文背后有什么故事？

What's the story behind this paper?

这篇论文描述了一种这样的方法，即 VICReg 方法。第一作者是一位名叫 Adrien Bard 的学生，他是 FAIR 巴黎的驻站博士生，由我和 Jean Ponce 共同指导，Jean Ponce 是巴黎高等师范学院的教授，也是 INRIA 的研究主任。这是法国一个很棒的项目，博士生可以在工业界完成博士学业。这篇论文是我前博士后 Stefan Dune 与 Li Jing 和 Yurish Montar 等人合作的 Barlow Twins 论文的后续。审稿人的一个主要批评是 VICReg 与 Barlow Twins 没有足够大的区别，但我的印象是，它本质上是修复了几个 bug 的 Barlow Twins。最终，人们会使用它。我已经习惯了被拒稿，所以它可能会被拒，但我非常兴奋，因为人们在使用它。它已经被拒了好几次了。

This paper describes one such method, the VICReg method. The first author is a student called Adrien Bard, who is a resident PhD student at FAIR Paris, co-advised by me and Jean Ponce, a professor at École Normale Supérieure and research director at INRIA. This is a wonderful program in France where PhD students can do their PhD in industry. This paper is a follow-up on the Barlow Twins paper by my former postdoc Stefan Dune, with Li Jing and Yurish Montar and others from FAIR. One of the main criticisms from reviewers is that VICReg is not different enough from Barlow Twins, but my impression is that it's Barlow Twins with a few bugs fixed. In the end, this is what people will use. I'm used to stuff being rejected, so it might be rejected, but I'm exceptionally excited because people use it. It's already been rejected a bunch of times.

更深层次的问题是同行评审和会议。计算机科学独特之处在于会议备受推崇。同行评审过程与期刊类似，但速度更快。这是快速发表成果的好方法，但也有许多相同的缺陷。如果你想提出新想法，就会受到抵制。有些自私的人能推断出谁提交了论文，然后对此发牢骚。

The deeper question is about peer review and conferences. Computer science is unique in that conferences are highly prized. The peer review process is similar to journals but accelerated. It's a nice way to get stuff out quickly, but it has many of the same flaws. If you want to do new ideas, you get pushed back. There are self-interested people who can infer who submitted it and be cranky about it.

有很多社会现象。一个是因为该领域呈指数级增长，绝大多数人非常年轻。随着领域规模开始饱和，经验不足的审稿人问题会减少。后果是年轻审稿人试图通过找论文的缺陷来让自己轻松。大多数论文都有缺陷，即使是好论文。这样做很容易。但重要的是论文中是否有新想法，可能影响人们思考问题的方式，即使实验不完美。这些社会现象的结合造成了一种疾病，困扰过其他领域如语音识别，人们追逐基准上的数字。如果论文对主流方法带来增量改进，就更容易被接收。那些论文对我来说很无聊。它们并非无用，但不是我感兴趣的。试图取得新进展的论文通常不会被接收。幸好我们有 arXiv。我坚信评审应该由成千上万的人来做，而不是两个人。

There's a lot of social phenomena. One is that because the field has been growing exponentially, the vast majority of people in the field are extremely junior. As the field size starts saturating, you will have less of that problem of inexperienced reviewers. A consequence is that young reviewers try to make their life easy by finding flaws in papers. Most papers have flaws, even good ones. It's easy to do that. But what's important is whether there is a new idea in that paper that is likely to influence the way people think about the problem, even if the experiments are not great. That combination of social phenomena creates a disease that has plagued other fields like speech recognition, where people chase numbers on benchmarks. It's much easier to get a paper accepted if it brings an incremental improvement on a mainstream method. Those are boring papers to me. They're not useless, but they're not the ones I'm interested in. Papers that try to strike new advances generally don't make it. Thankfully, we have arXiv. I'm a huge believer that reviews should be done by thousands of people, not two.

你是否看到未来很多优秀论文只会出现在 arXiv 上，在一个名为 Twitter 和互联网的持续会议上展示？

Do you see a future where a lot of strong papers will just be on arXiv, presented on an ongoing continuous conference called Twitter and the internet?

这不是精英主义与否的问题。这是为那些认为自己没有能力自行判断的人提供推荐和认可的问题。如果你依赖他人的意见，并信任那些人或团体为你评估论文，这可以节省时间。这是一个集体推荐系统。实际上，我在 10-15 年前就考虑过很多，当时在 NeurIPS 上讨论与 Yoshua Bengio 创建 ICLR。我写了一份文档，描述了一个评审系统：你把论文发布在 arXiv 或 OpenReview 等仓库上，然后可以组成一个评审实体，比如期刊或程序委员会，列出成员。该团体可以选择评审这篇论文。

It's not a question of being elitist or not. It's a question of recommendation and seal of approval for people who don't see themselves as having the ability to do so by themselves. It saves time if you rely on other people's opinion and trust those people or groups to evaluate a paper for you. It's a collective recommender system. I actually thought about this a lot 10-15 years ago, when there were discussions at NeurIPS about creating ICLR with Yoshua Bengio. I wrote a document describing a reviewing system where you post your paper on a repository like arXiv or OpenReview, and then you can form a reviewing entity, like a journal or program committee, listing the members. That group can choose to review the paper.

无论是自发还是非自发地审阅一篇论文，论文与会议或评审机构之间不再有排他性关系。任何评审机构都可以审阅任何论文，或者选择不审阅，然后给出评价。这不是发表或不发表的问题；它只是一个评价和评论，公开署名由评审机构负责。如果它由评审机构署名，你知道它是评审机构成员之一。所以如果评审机构是 Lex Fridman 的优选论文，对吧？就像弗里德曼写评论一样。是的。所以对我来说，那是一个美丽的系统，我认为。但除此之外，感觉还应该有一个针对评审者、评审机构的声誉系统。不是针对个人评审者，而是评审机构。当然。但即使在那里面，也有评审者。因为还有另一件事：这不仅仅是声誉，而是激励个人做出卓越表现。现在，在学术环境中，激励是内在的，只是想做好工作。但老实说，这不足以激励人们真正出色地阅读论文，在错误和缺陷中发现美，诸如此类。对吧？如果你是第一个发现一篇有影响力的论文的人，并且你为这一发现感到自豪，那会给你巨大的激励。这实际上是我提案的重要部分。我将其描述为：如果你对论文的评价能预测未来的成功，是的，是的，好的，那么你的声誉作为评审机构应该上升。所以是的，没错。我的意思是，我甚至有一个硕士生，他实际上是图书馆学和计算机科学的硕士生，他精确地计算出了这应该如何运作，包括公式和一切。

Review a particular paper spontaneously or not, there is no exclusive relationship anymore between a paper and a venue or reviewing entity. Any reviewing entity can review any paper, or may choose not to, and then give an evaluation. It's not published or published; it's just an evaluation and a comment, which would be public, signed by the reviewing entity. And if it's signed by the reviewing entity, you know it's one of the members of reviewing entities. So if the reviewing entity is Lex Fridman's preferred papers, right? It's like Friedman writing a review. Yes. So for me, that's a beautiful system, I think. But what's in addition to that, it feels like there should be a reputation system for the reviewers, for the reviewing entities. Not the reviewers individually, the reviewing entities. Sure. But even within that, there are reviewers too. Because there's another thing here: it's not just the reputation, it's an incentive for an individual person to do great. Right now, in the academic setting, the incentive is kind of internal, just wanting to do a good job. But honestly, that's not a strong enough incentive to do a really good job in reading a paper and finding the beautiful amidst the mistakes and the flaws and all that kind of stuff. Right? Like if you're the person that first discovered a powerful paper and you get to be proud of that discovery, then that gives a huge incentive to you. That's a big part of my proposal, actually. I described that as: if your evaluation of papers is predictive of future success, yes, yes, okay, then your reputation should go up as a reviewing entity. So yeah, exactly. I mean, I even had a master's student who was a master student in library science and computer science, actually, kind of work out exactly how that should work, with formulas and everything.

那么在实施方面，你认为这是可行的吗？我的意思是，我一直在和不同的人谈论这个，比如创办 Open Review 的 Andrew McCallum。我们最初为 ICML 选择 Open Review 的原因，即使当时对他们来说还很早，是因为我希望 ICML 最终能开创这种系统。所以 ICML 保留了开放评审的理念，即评审与论文一起发表，我认为这非常有用。但在很多方面，它在其他方面又回归到了更传统的会议形式。我的意思是，我不运营 ICML，我只是基金会的主席，但运营它的人应该决定如何运营，我不会告诉他们该怎么做，因为他们是志愿者，我非常感谢他们所做的工作。所以我对我们不够创新感到遗憾。

So in terms of implementation, do you think that's something that's doable? I mean, I've been sort of talking about this to various people, like Andrew McCallum who started Open Review. And the reason why we picked Open Review for ICML initially, even though it was very early for them, is because my hope was that ICML was eventually going to kind of inaugurate this type of system. So ICML kept the idea of open reviews, where the reviews are published with the paper, which I think is very useful. But in many ways, that's kind of reverted to more of a conventional type conferences for everything else. And I mean, I don't run ICML, I'm just the president of the foundation, but the people who run it should make decisions about how to run it, and I'm not going to tell them because they're volunteers and I'm really thankful that they do that. So I'm saddened by the fact that we're not being innovative enough.

是的，我也是。我希望这能改变。是的，因为科学传播的广泛性，但计算机科学思想的传播，是让这些思想产生影响力的方式，我认为。我认为这在很大程度上是因为人们心中有一个目标，即对作者的公平性，以及基本上能够计算分数并准确给予学分。但这是以科学进步为代价的。所以在某种程度上，我们正在减缓科学的进步。我们真的实现了公平吗？我们没有实现公平。我们有偏见。我们在做双盲评审，但偏见仍然存在。有不同种类的偏见。

Yeah, me too. I hope that changes. Yeah, because the communication of science broadly, but communication of computer science ideas, is how you make those ideas have impact, I think. And I think a lot of this is because people have in their mind an objective which is fairness for authors and the ability to count points basically and give credits accurately. But that comes at the expense of the progress of science. So to some extent, we're slowing down the progress of science. And are we actually achieving fairness? We're not achieving fairness. We have biases. We're doing double-blind review, but the biases are still there. There are different kinds of biases.

你写道，涌现现象，即大量简单元素相互作用所表现出的集体行为，是当初让你进入神经网络的原因之一。我喜欢元胞自动机。我喜欢简单的相互作用元素以及它们涌现出的东西。你认为我们理解复杂系统如何从如此简单的相互作用组件中涌现出来吗？

You write that the phenomenon of emergence, collective behavior exhibited by a large collection of simple elements in interaction, is one of the things that got you into neural nets in the first place. I love cellular automata. I love simple interacting elements and the things that emerge from them. Do you think we understand how complex systems can emerge from such simple components that interact simply?

不，我们不知道。这是一个巨大的谜团。对物理学家来说也是一个谜，对生物学家来说也是一个谜。我们周围的宇宙似乎变得越来越复杂，而不是越来越简单？我的意思是，这是物理学的一个奇特性质，尽管有热力学第二定律，但我们似乎有进化、学习等，它们至少可以在局部增加复杂性，而不是减少它。所以也许宇宙的终极目的就是变得更复杂，拥有这些美丽复杂的小区域。

No, we don't. It's a big mystery. Also it's a mystery for physicists, a mystery for biologists. How is it that the universe around us seems to be increasing in complexity and not decreasing? I mean, that is a kind of curious property of physics that despite the second law of thermodynamics, we seem to have evolution and learning etc., which can at least locally increase complexity, not decrease it. So perhaps the ultimate purpose of the universe is to just get more complex, have these small pockets of beautiful complexity.

这种自动化的涌现和复杂系统是否给你一些直觉，或指导你对机器学习系统和神经网络等的理解？或者这些对你来说现在是独立的概念？

Does that sort of automated these kinds of emergence and complex systems give you some intuition or guide your understanding of machine learning systems and neural networks and so on? Or are these for you right now separate concepts?

嗯，它让我进入了这个领域。你知道，我上大学时通过一本非常好的书发现了感知机的存在。那本书是关于乔姆斯基和皮亚杰的辩论，麻省理工学院的西摩·帕佩特在那本书中赞扬了感知机。我第一次听说学习机器，对吧？所以我开始挖掘文献，找到了那些论文和书籍，它们基本上是 50 年代和 60 年代关于自组织系统的研讨会或会议的转录。所以有一系列关于自组织系统的会议。这些书中的一些，你实际上可以在互联网档案馆找到数字版本。里面有一些引人入胜的文章，作者是一个名字几乎被遗忘的人，海因茨·冯·福斯特。他是一位移民到美国的德国物理学家，在 50 年代和 60 年代研究自组织系统。他在伊利诺伊大学厄巴纳-香槟分校创建了一个生物计算机实验室（BCL），专门研究神经网络。不幸的是，那是在神经网络流行度接近尾声的时候，所以那个实验室从未真正蓬勃发展。但他写了很多关于自组织和自组织之谜的论文。他有一个例子：想象你在太空中，没有重力，你有一个大盒子，里面装有磁铁。你知道，矩形磁铁，一端是北极，另一端是南极。你轻轻摇晃盒子，磁铁会相互吸附，可能自发地形成一个复杂的结构。这可能是自组织的一个例子。但你有很多例子。神经网络在许多方面都是自组织的一个例子。这是一个谜，比如在物理系统、混沌系统等中，这种模式形成可能实现什么。你知道，生命的涌现，诸如此类。那么这是如何发生的呢？这是一个巨大的谜团。

Well, it got me into it. You know, I discovered the existence of the perceptron when I was a college student, by a really good book. And it was a debate between Chomsky and Piaget, and Seymour Papert from MIT was kind of singing the praise of the perceptron in that book. And I, the first time I heard about the learning machine, right? So I started digging the literature and I found those papers, those books which were basically transcriptions of workshops or conferences from the 50s and 60s about self-organizing systems. So there was a series of conferences on self-organizing systems. And these books on this, some of them are you can actually get them at the Internet Archive, the digital version. And there are fascinating articles in there by this guy whose name has been largely forgotten, Heinz von Foerster. He's a German physicist who immigrated to the US and worked on self-organizing systems in the 50s and 60s. He created at the University of Illinois Urbana-Champaign a Biological Computer Laboratory (BCL), which was all about neural nets. Unfortunately, that was kind of towards the end of the popularity of neural nets, so that lab never kind of thrived very much. But he wrote a bunch of papers about self-organization and the mystery of self-organization. An example he has is: you take, imagine you are in space, there's no gravity, you have a big box with magnets in it. You know, rectangular magnets with north pole on one end, south pole on the other end. You shake the box gently, and the magnets will kind of stick to themselves and probably form a complex structure spontaneously. That could be an example of self-organization. But you have lots of examples. Neural nets are an example of self-organization in many respects. And it's a bit of a mystery how, like what is possible with this pattern formation in physical systems, in chaotic systems, and things like that. You know, the emergence of life, things like that. So how does that happen? It's a big mystery.

感觉理解某些受限情境下涌现的数学原理，可能有助于我们创造智能，就像给系统加点料。因为在复杂系统中，涌现能让你用很少的东西获得很多，这似乎是性能大幅提升的捷径。但有一个缺失的概念：如何衡量复杂性？

It feels like understanding the mathematics of emergence in some constrained situations might help us create intelligence, like add a little spice to the systems. Because in complex systems with emergence, you seem to be able to get a lot from little, so that seems like a shortcut to big leaps in performance. But there's a missing concept: how do you measure complexity?

是的，从本科时代起我就对此着迷。我们没有衡量复杂性的好方法，至少无法很好地解读已有的度量。例如，柯尔莫哥洛夫复杂度——生成一个比特串的最短程序长度——定义时有一个常数，这个常数可能非常大。贝叶斯概率论中也有类似概念，复杂度是概率的负对数。但你需要一个分布模型和一个先验，这就像柯尔莫哥洛夫复杂度中选择计算机一样。每种复杂性度量都有一个任意的加法常数，可以任意大。如果没有好的度量，我们怎么能有关于事物如何变得更复杂的理论呢？

Yes, it's something I've been fascinated by since my undergrad days. We don't have good ways of measuring complexity, or at least interpreting the measures we have. For example, Kolmogorov complexity—the length of the shortest program that generates a bit string—is defined up to a constant, which can be very large. There are similar concepts from Bayesian probability theory, where complexity is the negative log of probability. But you need a model of the distribution and a prior, which plays the same role as the choice of computer in Kolmogorov complexity. Every measure of complexity has an arbitrary additive constant that can be arbitrarily large. How can we have a theory of how things become more complex without a good measure?

人们在生物学中研究这个问题，比如生命起源或识别其他行星上的生命。复杂性与生命相关，但它取决于观察者。

People study this in biology, like the origin of life or recognizing life on other planets. Complexity is associated with life, but it's in the eye of the beholder.

举个例子。如果我给你看数字图像，由于相邻像素相关，图像有结构。现在对所有像素施加一个随机排列——图像看起来杂乱无章，更复杂。但绝对意义上它们完全相同。如果你知道这个排列，你可以还原它。想象一副眼镜能还原这个排列——原本复杂的东西变得简单。所以如果人类和 aliens 戴着不同的排列眼镜看宇宙，我们觉得简单的，对他们可能很复杂，反之亦然。复杂性确实取决于你的感知系统运行什么算法。我认为在很好地理解复杂性之前，我们不会有关于智能、自组织或进化的理论，而我们知道复杂性取决于观察者。

Let me give an example. If I show you images of digits, there's structure because neighboring pixels are correlated. Now apply a random permutation to all pixels—the images look disorganized, more complex. But they are exactly the same in absolute terms. If you knew the permutation, you could undo it. Imagine glasses that undo that permutation—what looked complicated becomes simple. So if humans and aliens see the universe with different permutation glasses, what we perceive as simple might be complicated to them, and vice versa. Complexity truly depends on what algorithm you're running in your perception system. I don't think we'll have a theory of intelligence, self-organization, or evolution until we have a good handle on complexity, which we know is in the eye of the beholder.

我们可能因为戴着不同的眼镜而无法探测到外星物种，这很遗憾——他们对局部性的概念可能与我们不同。这与物理学中的问题有关，比如恢复黑洞中丢失的信息，这依赖于复杂性的概念。

It's sad that we might not detect alien species because we're wearing different glasses—their notion of locality might differ from ours. This connects to questions in physics, like recovering information lost in a black hole, which relies on notions of complexity.

你能描述一下你构建一个富有表现力的电子管乐器（EWI）的个人探索吗？它是什么？构建它需要什么？

Can you describe your personal quest to build an expressive electronic wind instrument (EWI)? What is it and what does it take to build it?

我是个喜欢捣鼓的人。我喜欢用电子和机械部件造东西。我的第一个爱好是做模型飞机，但我也在正式学习之前自学了电子学，因为音乐。我表哥是个有抱负的电子音乐人，有一台模拟合成器，我在高中时帮他改装并搭建了音序器。我喜欢 80 年代的前卫摇滚——比如马哈维什努管弦乐团、天气预报、Yes、创世纪、温柔巨人。所以电子学和音乐结合了。我在高中和大学初期演奏巴洛克和文艺复兴音乐，吹奏竖笛、克鲁姆管，还有一点双簧管。我是个管乐器演奏者，但我一直想演奏即兴音乐。除了学萨克斯，唯一的方法就是演奏电子乐器，指法类似萨克斯，通过合成器控制获得广泛音色。从 80 年代末开始，我有过几把雅马哈和赤井的乐器，但从未完全满意，因为表现力不足且价格昂贵。它们测量呼吸压力和嘴唇压力，但我想要更多。

I'm a tinkerer. I like building things with electronics and mechanics. My first hobby was building model airplanes, but I also taught myself electronics before studying it, because of music. My cousin was an aspiring electronic musician with an analog synthesizer, and I modified it and built sequencers for him in high school. I love progressive rock from the 80s—bands like Mahavishnu Orchestra, Weather Report, Yes, Genesis, Gentle Giant. So electronics and music combined. I played baroque and renaissance music in an orchestra, playing recorder, crumhorn, a bit of oboe. I'm a wind instrument player, but I always wanted to play improvised music. The only way short of learning saxophone was to play electronic instruments with sax-like fingering and a wide variety of sound via synthesizer control. I had several from Yamaha and Akai since the late 80s, but I was never completely satisfied due to lack of expressivity and high cost. They measure breath pressure and lip pressure, but I wanted more.

你可以用手指改变各种参数，但它们真的不如原声乐器那样富有表现力，对吧？你听到约翰·科尔特兰演奏两个音符，就能听出他独特的声音，或者迈尔斯·戴维斯吹小号，因为声音反映了他们的生理结构，声道形状塑造了声音。那么电子乐器如何做到这一点呢？

You have various parameters you can vary with fingers, but they're not really as expressive as an acoustic instrument, right? You hear John Coltrane play two notes and you hear his unique sound, or Miles Davis playing the trumpet because the sound reflects their physiology, the shape of the vocal tract shapes the sound. So how do you do this with an electronic instrument?

很多年前，我遇到一个叫大卫·韦塞尔的人。他是伯克利的教授，在那里创建了音乐技术中心，他对这个问题很感兴趣。所以我多年来一直在思考这个问题。最后，因为我在家，在我的工作室里——它也是我的 Zoom 房间和家庭办公室，在新泽西——我开始认真考虑打造自己的电子乐器。

Many years ago I met a guy called David Wessel. He was a professor at Berkeley and created the center for music technology there, and he was interested in that question. So I kept thinking about this for many years. Finally, because I was at home in my workshop—which also serves as my Zoom room and home office, in New Jersey—I started being serious about building my own electronic instrument.

那个新泽西的工作室里还有什么？有没有你做的疯狂东西留在地板上？

What else is going on in that New Jersey workshop? Is there some crazy stuff you built left on the workshop floor?

很多疯狂的东西。用各种微控制器做的电子设备，还有奇怪的飞行装置。我仍然热爱飞行；这是家族遗传。我小时候我爸带我入坑。他小时候就做模型飞机，他是机械工程师，自学了电子学。他在 60 年代末 70 年代初自己做了早期的无线电控制系统。这让我对工程、科学和技术产生了兴趣。

A lot of crazy stuff. Electronics built with microcontrollers of various kinds, and weird flying contraptions. I still love flying; it's a family disease. My dad got me into it when I was a kid. He was building model airplanes when he was a kid, and he was a mechanical engineer who taught himself electronics. He built his early radio control systems in the late 60s early 70s. That got me into engineering, science, and technology.

你对其他形式的飞行也有兴趣吗，比如无人机或四轴飞行器？

Do you also have an interest in flight in other forms, like drones or quadcopters?

在无人机成为消费品之前，我自己做了一个，用微控制器、陀螺仪和加速度计来稳定，还写了固件。当它变成可以买到的标准产品时，就无聊了。我就不做了，不再有趣了。我在它流行之前就做了。

Before drones were a consumer product, I built my own with a microcontroller, gyroscopes, and accelerometers for stabilization, writing the firmware. When it became a standard thing you could buy, it was boring. I stopped doing it; it was not fun anymore. I was doing it before it was cool.

你会给今天的高中生和大学生什么建议？他们梦想做大事，比如在智能领域解决基本问题，无论是为了职业还是生活。

What advice would you give to a young person today in high school and college that dreams of doing something big, like solving fundamental problems in the space of intelligence, both for their career and in life?

尝试对大问题产生兴趣：什么是智能，宇宙由什么构成，生命是怎么回事，甚至像时间是什么这样的疯狂大问题——没人知道时间是什么。然后学习有长久价值的基础知识：数学、物理、工程。如果你要在学习 iPhone 移动编程和量子力学之间选择，选量子力学。你会学到你从未想到存在的东西，比如路径积分，它无处不在——和贝叶斯积分是同一个公式。量子力学和其他成熟领域中的思想有更长的保质期，你会间接用到它们。学习经典力学，比如拉格朗日量，非常有用。学习统计物理，因为机器学习的数学都来自 19 世纪末 20 世纪初的统计物理学家。比如乔治·帕里西刚刚因复制方法获得诺贝尔奖，它用于变分推理。如果你学电子工程，学信号处理，学傅里叶变换，它是图神经网络的基础——一个很有前途的 AI 新子领域。

Try to get interested by big questions: what is intelligence, what is the universe made of, what's life all about, even crazy big questions like what is time—nobody knows what time is. Then learn basic things that have a long shelf life: math, physics, engineering. If you have a choice between learning mobile programming on iPhone or quantum mechanics, take quantum mechanics. You'll learn things you have no idea exist, like path integrals, which are used everywhere—the same formula for Bayesian integration. The ideas within quantum mechanics and other solidified fields have a longer shelf life and you'll use them indirectly. Learn classical mechanics, like Lagrangians, which are hugely useful. Learn statistical physics, because all the math for machine learning came from statistical physicists in the late 19th and early 20th centuries. For example, Giorgio Parisi just got a Nobel Prize for the replica method, used for variational inference. If you do electrical engineering, take signal processing and learn Fourier transforms, which are at the basis of graph neural networks—a promising new subarea of AI.

如果你对应用更感兴趣，一个很有前途的方向是将 AI、机器学习和深度学习应用于科学，解决大问题。在 Meta FAIR，我们启动了一个名为 Open Catalyst 的项目。这是一个开放的合作项目，利用深度学习设计新的化合物或材料，以促进氢和氧的分离。如果你能用电力高效地将氧从氢中分离出来，你就解决了气候变化。在沙漠上铺满太阳能电池板，全天生产氢气，然后运到需要的地方。你就有了可控的、可运输的能源。大规模高效储能，比如生产氢气，就能解决气候变化。另一个方法是让核聚变工作。问题是超高温等离子体不稳定。也许通过深度学习，你可以找到控制器来稳定等离子体，实现实用的聚变反应堆。这很推测，但值得尝试，因为回报巨大。谷歌有一个由约翰·普拉特领导的团队在研究这个。所以，把科学、物理、生物和化学中的许多问题转化为可学习的问题，看看机器能否学会。我们无法从第一性原理理解复杂材料的性质。如果我们能设计新材料，就能制造更高效的电池、更快的电子设备、更轻的汽车或飞机材料、更好的燃料电池。如果我们有好的氢燃料电池，就可以用于很多事情。

Something very promising if you're more interested in applications is applying AI, machine learning, and deep learning to science to solve big problems. At Meta FAIR, we started a project called Open Catalyst. It's an open collaborative project using deep learning to design new chemical compounds or materials that facilitate separating hydrogen from oxygen. If you can efficiently separate oxygen from hydrogen with electricity, you solve climate change. Cover a random desert with solar panels, produce hydrogen all day, and ship it wherever needed. You have controllable power that can be transported anywhere. Large-scale efficient energy storage like producing hydrogen solves climate change. Another way is figuring out how to make fusion work. The problem is super hot plasma is unstable. Maybe with deep learning you can find controllers to stabilize plasma for practical fusion reactors. That's speculative, but worth trying because the payoff is huge. There's a group at Google working on this led by John Platt. So convert as many problems in science, physics, biology, and chemistry into learnable problems and see if a machine can learn them. We don't understand properties of complex materials from first principles. If we could design new materials, we could make more efficient batteries, faster electronics, lighter materials for cars or airplanes, better fuel cells. If we had good hydrogen fuel cells, we could use them for many things.

给飞机和交通工具提供动力，汽车，我们就不会再面临航空运输的二氧化碳排放问题了。所以有很多地方可以用到 AI，这还没提医学、生物学等等。蛋白质折叠，弄清楚如何设计蛋白质使其在特定位置与另一个蛋白质结合，因为这就是设计药物的方式。所以深度学习可以用于所有这些，如果我们能将其用于这些领域，那将是巨大的进步。

Power airplanes and transportation, cars, we wouldn't have CO2 emission problems for air transportation anymore. So there's a lot of those things where AI can be used, and this is not even talking about medicine, biology, and everything like that. Protein folding, figuring out how to design proteins that stick to another protein at a particular site, because that's how you design drugs. So deep learning would be used for all of this, and that would be enormous progress if we could use it for that.

最近材料物理学的一个例子：取单原子层的石墨烯，就是六边形网格上的碳，单原子厚度。在上面再放一层，以某个神奇的角度扭转，比如三度，它就变成了超导体。没人知道为什么。我想知道这是怎么发现的，但这就是机器学习可以真正发现的东西。嗯，也许不是，但有迹象表明，通过机器学习，我们可以训练一个系统，使其成为某些复杂涌现现象的现象学模型，比如超导，这些现象用通常的还原论方法很难从第一性原理描述。但我们可以有深度学习系统，在训练了足够多的样本后，从系统的描述中预测其属性。

Here's an example from recent material physics: take a monoatomic layer of graphene, just carbon on a hexagonal mesh, single atom thick. Put another one on top, twist them by some magic number of degrees, three degrees or something, it becomes a superconductor. Nobody has any idea why. I want to know how that was discovered, but that's the kind of thing machine learning can actually discover. Well, maybe not, but there is a hint that with machine learning we could train a system to be a phenomenological model of some complex emerging phenomenon, like superconductivity, where the phenomenon is too difficult to describe from first principles with the usual reductionist method. But we could have deep learning systems that predict the properties of a system from a description of it after being trained with sufficiently many samples.

EPFL 的 Pascal Fua 有一家初创公司，他训练了一个卷积网络来预测固体的空气动力学特性。你可以通过运行计算流体动力学生成任意多的数据。所以给出一个机翼翼型或某种形状，运行计算流体动力学，得到阻力和升力。你可以生成大量数据，训练一个神经网络来做这些预测，现在你就有了一个关于阻力和升力作为形状函数的可微分模型。所以你可以做反向传播，优化形状以获得你想要的属性。

This guy Pascal Fua at EPFL has a startup where he basically trained a convolutional net to predict the aerodynamic properties of solids. You can generate as much data as you want by running computational fluid dynamics. So you give a wing airfoil or some shape, run computational fluid dynamics, get drag and lift as a result. You can generate lots of data, train a neural net to make those predictions, and now you have a differentiable model of drag and lift as a function of the shape. So you can do backward propagation and optimize the shape to get the properties you want.

太不可思议了。太不可思议了。除此之外，你可能还应该读一点文学和历史，以获得灵感和智慧，因为毕竟所有这些技术都必须在人类世界中运作，而人类世界是复杂的。

That's incredible. That's incredible. And on top of all that, probably you should read a little bit of literature and a little bit of history for inspiration and for wisdom, because after all, all these technologies will have to work in the human world, and the human world is complicated.

这是一次精彩的对话。我很荣幸今天能和你交谈。感谢你在 Meta 的 FAIR 所做的所有出色工作，感谢你这么多年对一切事物都充满热情。你是机器学习社区的希望灯塔。非常感谢你今天抽出宝贵时间与我交谈。太棒了。

This is an amazing conversation. I'm really honored that you talked with me today. Thank you for all the amazing work you're doing at FAIR at Meta, and thank you for being so passionate after all these years about everything that's going on. You're a beacon of hope for the machine learning community. Thank you so much for spending your valuable time with me today. That was awesome.

谢谢你邀请我。很愉快。

Thanks for having me on. It was a pleasure.

感谢收听与 Yann LeCun 的对话。要支持本播客，请查看描述中的赞助商。现在让我用艾萨克·阿西莫夫的话作为结束：'你的假设是你观察世界的窗户。偶尔擦洗它们，否则光线将无法照进来。'感谢收听，希望下次再见。

Thanks for listening to this conversation with Yann LeCun. To support this podcast, please check out our sponsors in the description. And now let me leave you with some words from Isaac Asimov: 'Your assumptions are your windows on the world. Scrub them off every once in a while, or the light won't come in.' Thank you for listening and hope to see you next time.