选自inverse机器之心编译 参与：杜夏德、曹瑞 自从人类有了意识（有人会告诉你这是在 2000 年前，有人会说在 20 万年前）人类学者一直在试图理解和定义意识。对意识最简单纯粹的概念描述或许是世界情境中的自我意识。但是如果不能理解意识的深层机制，那么所谓的理解不过是停留在表面。这就是为什么神经科学家通过指向大脑内的物理现象成功地介入了关于意识的持续对话。但是把形而上学与物理联系起来，仍然会产生某种准科学、准哲学的结果，反而会弄巧成拙，让外人笑话。 所以我们努力理解人工智能是否具有意识也就不足为奇，也难怪图灵测试越来越无法提供我们需要的答案。 因为我们无法证明人工意识，很多工程师和专家对这种观点都是都条件反射地不屑一顾。但是 Jürgen Schmidhuber 不是这样，他很喜欢这个古怪的游戏。他今年 53 岁，自上世纪八十年代以来一直致力于人工智能研究，在谷歌、苹果、微软和 IBM 的产品中都能找到他研究成果的蛛丝马迹。作为瑞士卢加诺大学的人工智能教授、瑞士人工智能实验室 IDSIA 的科学主任，及致力于打造第一个实用通用人工智能的创业公司 NNAISENSE 的主席，Schmidhuber 认为，目前的人工智能系统已经具有意识。他相信他已经做到了，这就是他为什么会让人产生自我膨胀的印象，同时他也是这个领域最具争议的人物。 Schmidhuber 所处的位置具有讽刺意味，如果他是正确的，他的信念将可能会被一些未来高级意识（hyperconscious AI）验证。Schmidhuber 是全球强调人工智能为数不多的人之一。所以 Inverse 问他在编程中是否看到了别人漏掉的什么东西，如果有，它与人类意识有什么关系。 Jürgen Schmidhuber 在国际健康论坛，2015 在你看来，未来人工智能的角色是什么？ 所谓智能—人类或人工—就是关于解决问题的能力。长期以来，我们已经尝试打造一个能解决所有问题的东西（general problem solver）。 这种 problem solver 能基于之前学到的技能去学习新的技能，能无限制地添加新技能，变得越来越通用。当然，如果我们能成功，它将会改变一切，因为每个计算问题、每个职业都会受到影响。 你说一些人工智能已经具有意识，你能解释一下为什么吗？ 我认为我们在 25 年前就已经有了一个基本的有意识的学习系统。那个时候，我就提出通用的学习系统要包括两个模块。 一个是一种循环网络控制器，学习将不断接收的数据——比如视频和疼痛传感器的疼痛信号，饥饿传感器上的饥饿信号——转换为行动。例如，当电池电量低时，传感器就会传来一些负数信息。这个网络能学习将所有这些输入及时转换成能成功的行动序列。例如，电池电量低时及时到达充电站，但如果路上没有碰到障碍，如椅子或桌子，就不需要唤醒疼痛传感器。 智能体（agent）一生的目的是最大限度地享受快乐，同时将痛苦最小化。这个目标简单明确，但是很难实现，因为你需要学很多东西。你看一个小婴儿它需要花上几年的时间去学习这个世界的运作，学习怎么与世界互动以实现目标。 自 1990 年以来，我们的智能体一直在尝试做同一件事情，使用一个额外的循环网络—一个无监督模块，去预测将要发生什么。它看起来在执行所有的行动，所有的观察都会有，使用这种经验来学习预测将要发生什么。因为它是循环网络，它能在一定程度上用所谓的预测编码以规律的形式预测未来。 例如，如果你有一条关于 100 个落下的苹果的视频，所有的苹果总是以同一个方式落下，你就能学习预测这些苹果是怎么落下的，你也无需分别存储这些预测，也就是说你能把这个视频压缩非常小的几个字节。 1963 年的 Jürgen 和他的父亲 Johann Schmidhuber 玩象棋 由于数据是从与环境的互动中得来，这种无监督模型网络——1990 年以来我一直叫它世界模型——能随着时间学习发现新的规则或者对称性、或各种重复。它能学习去用少量计算资源编码数据，这里的计算资源指的是更少的存储或者更少的计算时间。过去学习中有意识的东西会随着时间的推移变得自动和具有潜意识。 随着网络的进步，学习一个新的规律后，它可以通过查看无监督的世界模型在学会前后需要编码多少计算资源来测量其新洞察力的深度，前后的差异就是网络的「乐趣」。其洞察力的深度是一个数字，直奔第一个网络，也就是控制器，控制器的任务是将所有的反馈信号最大化，包括从各种内在的愉悦时刻（joy moment）和该网络之前没有的洞察力而来的反馈信号。愉悦时刻类似于一个科学家发现一个以前未知的新物理定律。 你能帮我在我的意识和经验的范围内理解所有的处理吗？ 当你行走在这个世界中时，你会遇到很多张人脸，这意味着你做的一些处理工作，在大脑中构建某种循环子网络来压缩观测历史（称之为 compactify）确实有效。「面部编码器」（face encoder）会与原型人脸相对应。所以当一个新面孔出现的时候，你需要做的就是要将这张面孔与原型的差异进行编码。 即便如此，我仍然不确定为什么我们能说这就是有意识。 关于意识有一件很重要的事情是智能体会注意到，在与世界的互动中有一件事总是存在，即智能体自己。 出于数据压缩的目的，用循环世界模型网络（recurrent world-model network）挑出一些神经元对智能体本身进行编码是非常有效率的。它能够通过创造一个符号将整个行为和感知的历史以及属于智能体的其他符号都进行压缩：可能是手、脚等。在解决新问题的过程中，不管你什么时候激活这些与自我符号相关的神经元，智能体都在思考着自身。 所以，在 1991 年的时候，我们就已经达到了这一点。当然，那只是意识的一种基本形式——不像你我的意识这样令人印象深刻，因为我们的大脑要比这些小人工智能体的大脑大得多。我们的大脑皮层中的连接可能多达 10 万亿种，但是目前最大的长短期记忆人工神经网络（Long-Short Term Memory, LSTM）也可能只拥有 10 亿种连接。相比较，你的大脑皮层要大 10 万倍，所以携带的意识当然也要比小小的人造大脑更加可观。但是几乎每 5 年，计算的成本就会便宜 10 倍。所以或许我们还需要 25 年的时间，才能第一次获得和大脑皮层连接一样多的长短期记忆人工神经网络。 人造大脑的电子连接比大脑皮层的连接要快。 Schmidhuber 和仿真机器人 用技术行话来说，真正的挑战是「知觉难题」（hard problem of consciousness）。这就像是一种经验一样，哲学家们称之为感质（qualia）。当你拥有了一种经验之后——看日落、听你最喜欢的乐队唱你喜欢的歌、闻卡布奇诺的味道等等——就可以从一些存在的事物当中获取一些经验。只是目前原因我们还不清楚。哲学家 David Chalmers 认为这个问题是一个「难题」，换句话说：「身体上的感受为什么会让我们的内心生活更加丰富呢？」你是否相信这个模型还会让这些人工智能产生感质？ 我认同这个观点。从行为上来讲，这些人工智能都非常的相似。当我们让这些人工智能和其他一些能够伤害它们的人工智能面对面的时候——例如在捕食与被捕食（predator-prey）的场景中——它们不想要受伤。当一个人工智能袭击另一个人工智能时，它对疼痛的感知就会上升——而第二个人工智能就会据此进行预测并且避免这种痛苦，比如说藏在窗帘或者是模拟相同的场景。所以说，从这些智能体的行为来看，它们并不喜欢这样。 这是我们很长时间以来所看到的。我们的人工智能正在试图避免痛苦，并将愉悦最大化——包括开心或是内心的喜悦，从洞察力到模式——因为它们已经内置了可以最大化的效用函数（utility function）或是奖励函数（reward function）。人类也生来具有一种奖励函数。这些人工生命的行为至少从性质上来看和高等级的动物或者说是人类非常相似。所以无论如何，我们都应该相信这是可复制的。 这就像是为了讨论思维又发明出了一门新的语言一样。当你在反省的时候，你会不会觉得自己就像是一台计算机一样？你是不是在想：「我大脑中的高级活动区正在处理这个问题，较低级的区域正在进行自动处理？」你是这样反省的吗？ 是的。我经常想这些洞察力是否都是来源于「第一原理」（First Principles），我是否能够通过的思考重新发现它们。我相信我可以，虽然我知道很多人都被反思所欺骗。但是对我来说非常清晰的一点是：这应该就是我正在做的事情。对我来说，我们不需要其他的一些东西来解释意识。我深信，我们已经具备了所有理解意识的基本成分，并且这一进程 25 年前就开始了。只是神经科学领域的人对人工神经网络研究的进展知之甚少，他们对一些简单的原理也不是太关注。但是我相信他们一定会对此了解更多。至少，这是我希望能够发生的。 最近在纽约，我在一场大会上谈到伦理和人工智能。我回答了现场提出的一个有些挑衅的问题，多多少少地重复了我从 20 世纪 80 年代以来持有的一个观点：「我必须要坦白：我的公司正在研发仿真机器人。我就是一个原型。虽然我可能没有意识，但是我很擅长伪装成具有意识的样子。」 原文链接：https://www./article/25521-juergen-schmidhuber-ai-consciousness ©本文由机器之心编译，转载请联系本公众号获得授权。 ✄------------------------------------------------ 加入机器之心（全职记者/实习生）：hr@almosthuman.cn 英文原文：For as long as humans have had consciousness — two or two hundred millennia depending on who you ask — human scholars have made great efforts to understand and define what that means. The most facile and purely conceptual description of consciousness might be that it is an awareness of the self within the context of the world. But without an understanding of the underlying mechanism, consciousness keeps chasing its tail. This is, in part, why neuroscientists have successfully interjected themselves in the ongoing conversation about consciousness by pointing to physical phenomena within the brain. But linking the metaphysical to the physical still results in the sort of quasi-scientific, quasi-philosophical overreach that gets academics laughed out of faculty lounges and labeled eccentric. No wonder we’re struggling to understand whether the artificial intelligences we build are conscious. No wonder the Turing Test increasingly falls short of providing the sort of answers we require. Because we cannot prove artificial consciousness, many engineers and experts are reflexively dismissive of the idea. But Jürgen Schmidhuber is not many engineers or experts and he’s fine playing the eccentric. He’s 53 years old and has been working on A.I. since the 1980s, which is why traces of his pioneering work in the field are surface in Google, Apple, Microsoft, and IBM products. A professor of A.I. at Switzerland’s University of Lugano, scientific director of the Swiss A.I. Lab IDSIA, and president of NNAISENSE (pronounced “nascence”), a startup that aims to build the first practical general purpose A.I., Schmidhuber believes that some current A.I. systems are already conscious. He believes he helped engineer it, which is both why he can come across as self-aggrandizing and why he’s one of the most interesting figures in a growing field. The irony of Schmidhuber’s positions is that if he turns out to be right, his belief will likely be retroactively validated by some future hyperconscious A.I.. Schmidhuber may be one of the few people on the planet ready to actively empathize with A.I. so Inverse asked him what he sees in the programming that other people miss and what — if anything — that has to do with human consciousness. What, in your mind, is the role of artificial intelligence in our future? All of intelligence — human or artificial — is about problem solving. For a long time, we have been trying to build general problem solvers that not only can solve one little problem here and another over there, but many different problems. [Problem solvers that can] learn new skills on top of previously learned skills, always adding new skills to the repertoire in an unlimited way, becoming more and more general problem solvers. Of course, to the extent that we succeed, this is going to change everything, because every computational problem, every profession, is going to be affected by this. You claim that some A.I.s are already conscious. Could you explain why? I would like to claim we had little, rudimentary, conscious learning systems for at least 25 years. Back then, already, I proposed rather general learning systems consisting of two modules. One of them, a recurrent network controller, learns to translate incoming data — such as video and pain signals from the pain sensors, and hunger information from the hunger sensors — into actions. For example, whenever the battery’s low, there’s negative numbers coming from the hunger sensors. The network learns to translate all these incoming inputs into action sequences that lead to success. For example, reach the charging station in time whenever the battery is low, but without bumping into obstacles such as chairs or tables, such that you don’t wake up these pain sensors. The agent’s goal is to maximize pleasure and minimize pain until the end of its lifetime. This goal is very simple to specify, but it’s hard to achieve because you have to learn a lot. Consider a little baby, which has to learn for many years how the world works, and how to interact with it to achieve goals. Since 1990, our agents have tried to do the same thing, using an additional recurrent network — an unsupervised module, which essentially tries to predict what is going to happen. It looks at all the actions ever executed, and all the observations coming in, and uses that experience to learn to predict the next thing given the history so far. Because it’s a recurrent network, it can learn to predict the future — to a certain extent — in the form of regularities, with something called predictive coding. For example, if you have a video of 100 falling apples, and all of these apples always fall down in the same way, you can learn to predict how they fall down, and what you can predict, you don’t have to store separately, which means that you can compress the entire video to a much smaller number of bits. As the data’s coming in through the interaction with the environment, this unsupervised model network — this world model, as I have called it since 1990 — learns to discover new regularities, or symmetries, or repetitions, over time. It can learn to encode the data with fewer computational resources — fewer storage cells, or less time to compute the whole thing. What used to be conscious during learning becomes automated and subconscious over time. As the network makes progress, and learns a new regularity, it can measure the depth of its new insight by looking at how many computational resources the unsupervised world model needs to encode the data before it learns that and afterwards. The difference between before and after: That is the “fun” that the network has. The depth of its insight, which is a number, goes straight to the first net, the controller, which has the task to maximize all the reward signals — including reward signals coming from such internal joy moments, from insights the network didn’t have before. A joy moment, like that of a scientist who discovers a new, previously unknown physical law. Can you help me understand all that processing within the context of my own consciousness and experience? As you are walking through the world, you are encountering lots of faces of humans, which means that it’s really efficient for you to compactify, to compress your observation history by constructing, in your brain, some sort of recurrent sub-network. A “face encoder,” which corresponds to something like a prototype face. When a new face comes along, all you have to do is encode the deviations from the prototype. I’m still not sure why we can say that it’s conscious, though. One important thing about consciousness is that the agent, as it is interacting with the world, will notice that there is one thing that is always present as it is interacting with the world — which is the agent itself. For data compression reasons, it’s really efficient for the recurrent world-model network to set a couple of neurons aside to encode this agent itself. It will be able to better compress the entire history of actions and perceptions by creating a symbol of itself, and additional symbols for things that belong to the agent: Maybe the hands, and the feet, and whatever. During the search for a solution to a new problem, whenever you wake up these neurons that are responsible for that self-symbol, then the guy, the agent, is basically thinking of itself. So we have had that since 1991. Sure, it’s just a rudimentary form of consciousness — not as impressive as your own, because your brain is much bigger than the brains of our little artificial agents. You have maybe 100,000 billion connections in your cortex, while today’s largest LSTM networks — long short-term memory networks — have maybe only a billion connections. So your cortex is still 100,000 times larger, and the consciousness that can be carried by it is more impressive than what we can fit into our little artificial brains. However, we just take notice that every five years, computing gets 10 times cheaper. So maybe we’ll need only another 25 years until — for the first time — we will have rather cheap LSTM networks which have as many connections as your entire cortex. And these cortex connections are much slower than the electronic connections of our artificial brains. In the technical jargon, the real challenge is known as the “hard problem of consciousness.” It’s the what it’s like of experience, which philosophers tend to call “qualia.” When you have an experience — when you watch a sunset, listen to your favorite band play your favorite song, smell a cappuccino, and so on — there’s something it’s like for you to have that experience. It’s just not clear why. Philosopher David Chalmers, who identified this as the “hard problem,” put it this way: “Why should physical processing give rise to a rich inner life at all?” Are you confident that this model also reproduces the qualia — the what it’s like — for these A.I.s? I do think so, yes. Behaviorally, our A.I.s are quite similar. When we confront them with other A.I.s who can hurt them — in predator-prey scenarios, for example — they don’t like to get hurt. Whenever one A.I. hits another A.I., whose pain sensors go up — this is something that the second A.I. can learn to predict and avoid, say, by hiding behind the curtain, or the equivalent in the simulation. So, of course, from the behavior of the agent, you see that it doesn’t like that. That’s what we have seen for a long time. Our A.I.s try to avoid pain, and they try to maximize pleasure — including fun, or internal joy, from insights into patterns — because they have a built-in utility function or reward function that they want to maximize. Humans also have such a reward function, already built-in as babies. And the behavior of these artificial beings is at least qualitatively similar to what we see in higher level animals, or in humans, and so on. So there is no reason whatsoever to believe that this is not replicable. It’s almost like you’ve invented a new language for discussing our own minds. When you introspect, do you think as if you were a computer? Are you thinking, “My higher level brain is dealing with this problem, while my lower level brain is running through these automatic processes?” Do you introspect in that way? Yeah. I often think about whether these insights, derived from first principles, whether I can rediscover them in my own thinking, and I believe I can, although I am aware that many people have been fooled by introspection. But it seems pretty obvious to me: That’s more or less what I’m doing. To me, it’s not obvious that there is a need for something else to explain consciousness. I’m pretty convinced that all the basic ingredients to understand consciousness are there, and have been there for a quarter-century. It’s just that people in neuroscience who maybe don’t know so much about what is going on in artificial neural network research, they are not yet so aware of these simple basic principles. But I’m sure they will learn more about that. At least that’s what I’m hoping is going to happen. Recently in New York, I spoke about this at a conference about ethics and A.I. I replied to a provocative question from the audience by repeating more or less what I have told smaller audiences since the 1980s: “I have to make a confession: My company is making androids. I am a prototype. I may not have consciousness, but I am good at faking it.” This interview has been edited for brevity and clarity. |
|
|
