Richard Bright: Can we begin by you saying something about your background?
Yoshua Bengio: I was trained as a computer scientist. I got my PhD in 1991. I did post docs in the USA at MIT and Bell Labs, then I took up a position here at the University of Montreal in 1993. I’ve been building up a group here which has been focussing on neural networks – now called Deep Learning – that has become the largest deep learning academic group in the world. Many things have happened over the last decade as this technology has unfolded.
RB: Your work includes research around deep learning architectures. What is the relationship between Machine Learning and Deep Learning?
YB: Actually, to answer this I am going to draw three circles. There is AI, inside of which is Machine Learning, inside of which in Deep Learning. So, AI is the quest for algorithms to build intelligent machines and Machine Learning is a particular approach to do that which relies on machines learning from data and examples so that knowledge is acquired mostly by observing and interacting with the world. Deep Learning is a particular approach to Machine Learning which follows up on decades of work on neural nets, which was inspired by the things we know about brains. It gets its name because of a focus on representation and multiple levels on representation as a core ingredient in those learning systems.
RB: And common to both is experience?
YB: Common to both Deep Learning and Machine Learning is experience, observations or data that the machine is learning from. So, you have the word ‘learning’ because the machine gets born with very little knowledge and then it acquires more knowledge by observing, either passively or actively, as when you think about a robot which does things and observes the consequences.
RB: How can a machine ‘learn’ without human input?
YB: Well, just like a mouse learns without human input. Many animals learn in that way, with a culture which is very weak. We have a very rich culture as humans and so, of course, our learning is driven a lot by that. That said, a lot of the learning that we do in machine learning is also driven by human input. In fact, the most successful machine learning or deep learning to date is what is called ‘supervised learning’ where it is very much driven by human guidance. Imagine that you had to take your baby in hand and that for every decision or muscle movement that the baby has to take you would have a parent telling the baby what to do as to what the right action would be. So that’s supervised learning, this is the way we are solving problems these days. Of course, it’s far from the human ability, in terms of autonomy: humans, mammals and birds can learn in a much more autonomous way, they are very powerful learners.
RB: The February issue aims to explore what it means (and will mean) to be human in the age of Artificial Intelligence. It would be good to start with how we define intelligence, both in terms of human intelligence and our growing understanding of animal intelligence. It is obviously a complex, multi-faceted entity, (and maybe difficult to answer?) but how would you define intelligence?
YB: There’s a lot of confusion about the meaning of intelligence and thus the meaning of artificial intelligence. Intelligence, in the technical circles, has to do with the ability to make good decisions, even in an environment which changes or in an environment of which we know very little at the onset. For this, we have to learn and adapt, but it’s about taking the decisions, and in order to take good decisions an intelligent agent needs knowledge. It can get knowledge in many different ways but, as I said earlier, machine learning is about acquiring knowledge. So, with that definition what you see is that, even an ant is intelligent, even a bee is intelligent. It’s just that the set of things that it can do right is different and smaller than the set of things that a human can do well. And so, we already have intelligent machines, they’re just not as intelligent as us. They can outperform us on a few things and are totally ignorant of most of the things that humans are able to do.
When we talk about human-level intelligence, in terms of AI and machines, we mean a level of intelligence that is comparable with that of humans both in its strength but also in its scope. In other words, we can understand many different aspects of the world and then we can use that knowledge to do many different tasks across all those aspects of the world. Of course, we are far from that with AI but we actually made a lot of progress since the beginning of AI research.
RB: How does, or how can, machine intelligence enhance human intelligence?
YB: It’s already doing it. When you are using Google or other search engines this is extending your own intelligence. Everybody now who has access to the internet uses these tools multiple times a day, just to find information. Even just using a laptop to find information about all the exchanges you have had with other people by e-mail is extending your intelligence. There are many ways in which technology has been extending humans, what’s different with AI and computers is that it’s mostly extending our cognitive abilities, whereas previous industrial revolutions brought broad extensions of our speed, our muscle power, or our ability to fly when we couldn’t before. So, now it’s our ability to think, to solve problems at an intellectual level that computers and AI are extending.
RB: Do you think that, following the work being done with Evolutionary AI, our views on what is intelligence might change?
YB: You mean using evolutionary algorithms?
YB: So, that’s one approach, but it hasn’t been particularly successful. It falls under a more general pattern, which includes both evolution and learning, where there is a gradual improvement. You have a candidate solution (a learning agent or a population) and it gradually gets better and gets to know more things or do things better in some sense. That is handled with the mathematics of optimisation, whether you’re dealing with evolutionary learning or dealing with machine learning. That’s a very central in AI and in machine learning. The only problem with evolutionary methods is that they are very slow for the kind of computing resources we currently have. Of course, if you can have a billion individuals on the planet, each trying different configuration of genes, then that is very efficient, but if you consider a single computer, or even a single brain, then evolutionary methods don’t seem to be sufficiently efficient.
RB: So, do you think that if it was a single computer that it would take a long time, but if they were connected then that would speed up the process?
YB: Yes, if you had a million computers connected then I think evolutionary methods would be an interesting tool. The same way that we learn as individuals but also we also learn as a group. Actually, we learn in two ways as a group, we learn through the evolution of our genes, that’s very slow but that’s what our species is doing, so the group of humans with all their genes is evolving, and there is an optimisation leading to better genes. A much faster kind of evolution is happening through cultural evolution, where we are sharing the information about what we learn through culture rather than through our genes. That’s also the process of science, by the way. All of these things are interesting but unless you have access to a very large number of computing machines it might be a different kind of optimisation, it might be more of the kind you have in your individual brain.
RB: How would you characterize your view on consciousness?
YB: One problem with this word is that it means different things to different people. I like the definition in Wikipedia which talks about different aspects of consciousness and one of the aspects I find most interesting is related to attention (“being aware of an external object or something within oneself”). So, there are things that we’re conscious of, that come to our mind. This is something I am studying in my research, how do we build neural networks that have the equivalent of this attentive consciousness that brings pieces of knowledge and pieces of our recent experience to a special place, which is our consciousness, so we can use those pieces in a privileged way in order to decide the next thing we’re going to be doing. So that’s one aspect of consciousness, another aspect is self-knowledge, or self-consciousness. You have different degrees of self-knowledge, even a very simple robot which knows its position has a self-consciousness. It is not necessarily something magical or new, we can build that in machines. There are other things that people associate with consciousness, such as qualia for example, which are about subjective impressions we get from our perceptions.
RB: A super-intelligent AI could register information and solve problems and that would far exceed even the brightest human minds, but being made of a different substrate (eg. silicon), would it have conscious experience? Or would it only be able to attain certain aspects of consciousness, such as, for example, learning and memory?
YB: So, it’s going to be our design choice. If you think about a search engine as an example, you can imagine having a very intelligent search engine that knows all of the knowledge of the world and can give you answers to questions but doesn’t have any self-consciousness. It’s really a machine that mechanically gets information and can answer questions like an oracle. So you could have AI without consciousness. It’s totally conceivable for me that we could have AI without consciousness and we will build machines that don’t have any more consciousness than a toaster and yet are very smart.
That’s the example of the search engine. The other example would be if we build a robot that has to function in the world, then it’s going to have to have some kind of self-consciousness. It needs such a self-consciousness in order to act in the world in a way that takes into account its state and role in its environment. It needs to know that it exists and can act. Presumably we will have machines that are intelligent and don’t have consciousness, as well as machines that are intelligent and do have consciousness, depending on their intended use.
RB: Given what we know about our own minds, can we expect to intentionally create artificial consciousness? Or will AI become conscious by itself?
YB: We will create it. I’m not at all a believer in the way of thinking about the future in terms of singularity. At least, as far as I can see in the future we will be the ones building those machines. We design learning machines, we design their objectives and so we will choose to give some form of consciousness to those machines, as needed. I think some AIs will be conscious because we design them to become conscious.
There are some interesting questions to which I’m not yet sure what to say about them, which have to do with not just consciousness but the notion of a moral agent. I heard some interesting discussions about this and one interesting argument was the notion of ‘person’ in a society and the moral responsibility or the moral duties that we have to persons in society is something that has to do with the place of that agent in society, the fact that we need to set up some rules and some ways of functioning as a group for society to work. So we shouldn’t necessarily project these social attributes of humans automatically to AI and machines. Even though they might be intelligent and have some form of consciousness, that doesn’t mean that they should be ‘persons’ in the sense that we use for people in the their social role in terms of responsibility, of the duties of society towards persons and expected from persons. I think we should be careful not to necessarily aggregate the notions of personhood and of intelligence. It turns out that for humans we have intelligence and we have ‘personhood’ and moral agency, but it doesn’t have to be that way. For machines we could, and we probably should, separate those things and keep them distinct. There is a danger that companies might request that their robots be considered like persons, but I think that would just be a legal trick to avoid the responsibility, that these robots might do bad things and those companies can be sued. If they are ‘persons’ then it’s the robot that can be sued. On the other hand, if you think about the very hypothetical scenario of an artificial being which is a cognitive clone of a human (see one of the Black Mirror plots) most people would agree that we should avoid acting in a way that would make that being suffer. So it is not an easy question and we probably need to consider the duties of society towards such intelligent and conscious but non-human beings.
RB: So there are a lot of ethical issue to be considered which need to be addressed before we create an artificial consciousness?
YB: We already are creating artificial consciousness, even though it may be stupid and simple. There is no black and white thing that suddenly produces artificial consciousness, we do it all the time but at different degrees. It’s just short lived and it’s stupid, but science moves in small steps and we are gradually going to build smarter and smarter machines which have more and more self-consciousness. I think where it becomes tricky, and really I don’t have any answer, is that this is obviously connected to the debate on animal welfare and vegetarian issues. Where do we put the line between the entities that are intelligent and conscious (to some degree) that we wish to protect for ethical reasons? For example, we wish to protect weak people in our society or babies because they can’t fend for themselves, and we’ll even protect humans who are struck by mental health problems and have very little intelligence left, but we generally don’t protect cows or pigs, which are very smart and similar to us. So, where is the line, where should it be? If robots have something like our sophisticated emotions in the future, which is not totally crazy, should their wellbeing be protected as well? And in what sense? These are hard, ethical questions that need more thought.
RB: This goes back to my earlier question about the growing understanding of animal intelligence. For example, 20 or 30 years ago primates were not considered to have emotions, and many animals have emotions, not just primates.
YB: Exactly, and from research from the AI point of view emotions start looking not so mysterious anymore and it is something that we could put in. A very simple AI that uses what’s called ‘reinforcement learning’ already has primitive emotions such as fear and pleasure, in the sense that we design it so that it will try to avoid something dangerous and it will seek out other things which are rewarding. When it senses something that could be dangerous, in terms of what it needs to avoid, then in a sense it’s like experiencing fear in trying to avoid those things. So they have really primitive emotions already, in simple AI systems, in playing games, for example.
RB: Does consciousness require embodiment?
YB: Well, that’s an open question. It depends on what you mean by embodiment. You could have a machine embodied in a virtual world, for example. So, is that embodiment? From an information processing point of view, it doesn’t really matter if it’s the real world or if it’s a simulated world (it’s only a matter of complexity of the environment), it’s all information processing, learning and decision-making. There is a technical question as to whether we need some form of embodiment, or rather grounding in an environment, in order to build sufficiently intelligent machines. There are a lot of researchers who think, as I do, that some form of grounding in an environment (with perception as well as action) helps the intelligent learner to make sense of the world and connect its internal presentations with what it means in the world. There is now a lot of research in that direction, to make sure that words are associated with meaning relative to some environment. For example, we could easily have an AI that doesn’t have a body in a traditional sense, they don’t have to be living in robots, they could be just living in computers and they could still be self-conscious.
RB: Is consciousness the key to Artificial Intelligence, or we talking about various levels of AI?
YB: Artificial Intelligence, and intelligence in general, is orthogonal to consciousness, these are two different things. You could have an AI which has no consciousness at all, like a really smart search engine, and you could have an AI that has consciousness. An embodied system probably needs some form of consciousness, I don’t see how it could do things without it, without a self-consciousness at least. So, yes, the two things are really separate and one doesn’t necessarily imply the other. You could have consciousness and be really stupid. An insect has a form of self-consciousness, it knows where it is, it knows what it wants and goes and fetches food, and avoids predators etc. It has a very primitive form of consciousness and a very primitive form of intelligence.
RB: And they have a collective consciousness.
YB: Yes, absolutely, there is also a notion of collective consciousness. Through culture there is a ‘group think’ that develops, evolves and comes to conclusions, collectively we take decisions, especially with democracy, and so it’s not something that’s limited to a single individual.
RB: How far should we take AI?
YB: Oh, wow, I don’t know! From the point of view of an AI researcher my concern is that currently machines are too stupid and it may take hundreds of years before we get to human-level intelligence, maybe thousands of years, or maybe just decades. Frankly, it’s impossible to predict. If anyone tells you that in 2047 we will reach singularity, this is bullshit, because we don’t have enough information to make so precise predictions. So it’s more honest to abstain. My concern on the question of safety isn’t about AI becoming too smart, my concern is about human beings taking advantage of technology at the expense of other humans and of the rest of the planet. I’m more concerned about humans and their failures, which could make us self-destruct or simply hurt each other and there are many ways in which that could happen. As an example, with military applications such as ‘killer robots’ and so on, the ‘Big Brother’ applications which governments or companies could use to monitor everybody and control them. So, there are dangers, but they are mostly due to misuse of AI by either individuals or companies or governments.
RB: So, that fits well with the theme of this issue and we have to monitor the human rather than the artificial intelligence.
YB: That’s right. I think the progress in AI should hopefully come at a pace which is matching the progress in human wisdom and mostly human wisdom is about better organised societies, where people are most likely not to hurt each other. Unfortunately, right now, if we had human-level AI it could be used by many bad actors who simply do not care enough about the common good.
RB: What are some of the challenges you hope to address in the coming years?
YB: My research has been focussed, since the beginning of the Deep Learning quest, on the question of representation. What I mean by this is that computers are not right now just learning about how the world works, in the space of pixels and sounds, they are learning abstractions, they are learning representations of the world, representations of what their sensors give them, representations of images, representations of sounds etc. One of the big questions is how do we learn good representations and, up to now, we’ve been helping those computers in figuring out what matters in a representation, by telling them about the concepts behind the things we want them to recognize in images or in sounds, such as words or names of things. But humans are able to figure out a lot about the world without necessarily being told. As an example, a two year old child learns in the first two years about physics. Of course, she doesn’t take physics courses, her parents don’t need to give her Newton’s or Einstein’s equations, she just observes the world, plays with things around her and figures out what we call ‘intuitive physics’, like gravity, solid objects, pressure and so on. So that kind of autonomous learning and ability to understand the hidden abstractions and use them for predicting the future, this is not what we are doing well yet with computers but it is something that humans are able to do. We could call that ‘unsupervised learning’ or ‘autonomous learning’ and this is something that relies on learning abstractions without being told what the right abstractions are in the first place. This is where my research has been and continues to be aiming towards.
Get the Full Experience
Read the rest of this article, and view all articles in full from just £10 for 3 months.