Interview with Yoshua Bengio, Pioneer of AI – 01
On June 7th, 2019 at the MILA (Montreal Institute for Learning Algorithms) in Montreal, Canada, I conducted my interview with Professor Yoshua Bengio, who is one of the pioneers of AI. He is well-known as the “father of AI” for his great contribution to developing so-called deep learning. He has received the 2018 ACM A.M. Turing Award with Geoffrey Hinton and Yann LeCun for major breakthroughs in AI.
In my interview, I asked him about the possibilities of AGI, biased data, people’s concerns about GAFA and China, the opportunities and risks of AI and the future of AI. All these questions are based on my previous experiences in the University of Cambridge as well as many international summits and conferences on AI I have been invited to recently.
Bengio is also noteworthy because he chooses to remain as an academic, staying at the University of Montreal as head of the MILA, while other AI leaders such as Geoffrey Hinton have left academia and now work for Google. Bengio continues to contribute to teaching students as well as engaging with local communities. He believes the education of future generations and people’s engagement with AI is crucial for the creation of a better society including AI. This is because he is aware of not only the opportunities but also the risks of AI. As he owns his startup, so-called Element AI, he is instrumental in buildinga bridge between academia and the business world.
This is my interview with Yoshua.
The Road to AGI
Yoshua Bengio : Did you have some questions for me?
Toshie Takahashi : Yes, of course. Thank you for taking the time. I’d like to ask you about AGI.
YB : Okay.
TT : I watched some of your videos and I understand you are very positive about AGI.
YB : No.
TT : No? I thought you showed a…
YB : I’m positive that we can build machines as intelligent as humans, but completely general intelligence is a different story. I’m not positive as to how humans might use it because we’re not very wise.
TT : Okay. So can you show a road map of how you could create AGI?
YB : Yes, the one I have chosen to explore.
TT : I spent some time in Cambridge, and some scholars, for example Professor John Daugman, the head of Artificial Intelligence Group at the University of Cambridge, said that AGI is an illusion created by science fiction because he said that we don’t even understand a single neuron so how can we create AGI?
YB : Yes, I disagree with him.
TT : Okay, so could you tell me about that?
YB : Sure. Having worked for decades on AI and machine learning I feel strongly that we have made very substantial progress, and in particular we have uncovered some principles, which today allow us to build very powerful systems. I also recognise that there’s a long way towards human level AI, and I don’t know how long it’s going to take. So I didn’t say we’ll find human level AI in five years, or ten years or 50 years. I don’t know how much time it’s going to take, but the human brain is a machine. It’s a very complex one and we don’t fully understand it, but there’s no reason to believe that we won’t be able to figure out those principles.
TT : I see. Mr. Tom Everitt at DeepMind said that he could create AGI in a couple of decades, maybe 20 or 30 years. Not too far in the future.
YB : How does he know?
TT : I don’t know. I asked him but he didn’t answer it.
YB : Nobody knows.
TT : Nobody knows. Yes, of course. When I met Professor Sheldon Lee Glashow, a Nobel Prize winning American theoretical physicist, he told us that we won’t have AGI. Or even if we have it, it’d be very far away.
YB : Possibly, so we don’t know. It could be ten years, it could be 100 years.
TT : Oh really?
YB : Yes.
TT : Okay.
YB : It’s impossible to know these things. There’s a beautiful analogy that I heard my friend Yann LeCun mention first. As a researcher, our progress is like climbing a mountain and as we approach the peak of that mountain we realise there’s some other mountains behind.
TT : Yes, exactly.
YB : And we don’t know what other higher peak is hidden from our view right now.
TT : I see.
YB : So it might be that the obstacles that we’re currently working on are going to be the last ones to reach human level AI or maybe there will be ten more big challenges that we don’t even perceive right now, so I don’t think it’s plausible that we could really know when, how many years, how many decades, it will take to reach human level AI.
TT : I see. But some people also say that we need a different breakthrough to create AGI. We need a kind of paradigm shift from our current approach. Do you think that you can see the road to reach if you keep on with deep learning? So this is a right road?
YB : We have understood as I said some very important principles through our work on deep learning and I believe those principles are here to stay, but we need additional advances that are going to be combined with things we have already figured out. I think deep learning is here to stay, but as is, it’s obviously not sufficient to do, for example, higher-level cognition that humans are doing. We’ve made a lot of progress on what psychologists call System 1 cognition, which is everything to do with intuitive tasks. Here is an example of what we’ve discovered, in fact one of the central ideas in deep learning: the notion of distributed representation. I’m very, very sure that this notion will stay because it’s so powerful.
The Future of AI
TT : You are really ahead of this…So what can you see of the future?
YB : I don’t see the future.
TT : Really? I thought that you could see a future we cannot see yet.
YB : No, I have research goals and I have chosen to explore particular directions because I believe in them, so the vision I have is that a big missing piece in our current machine learning systems is what people call common sense. So think about the intelligence of a two-year-old or the intelligence of a cat, we don’t even have that in machines right now. And that’s not even starting to think about things like language where we’re seriously lacking, especially when you look at the mistakes made by current state-of-the-art systems. That sort of common sense understanding includes things that current machine learning doesn’t do, like understanding cause and effect, and discovering cause and effect. It includes a broad understanding of the world, not just one specialised task. It includes the ability to discover this model of the world through unsupervised exploration. We rely today heavily on supervised learning where all of the high-level concepts have been defined by a human teacher or human labels. There are lots of aspects of intelligence that are currently at the frontier, and I’m not the only one exploring these things, which could make a big difference in a few years, but it’s hard to be sure.
TT : Is your goal to create a human or a superhuman?
YB : No, my goal is to understand general principles of intelligence, how an agent can become intelligent. I and many others would like to discover the equivalent of the laws of physics, but for intelligence.
TT : Yes.
YB : And presumably those principles would apply to humans, to animals, to aliens, who might be intelligent, to machines that we can build, so these would be very general principles and machine learning has already established some of those principles, but we’re still missing some important ones, I believe.
TT : But you’ve found a simple principle?
YB : Yes, several. But behind this, there’s a meta-principle, a scientific hypothesis, that intelligence could be explained by a few simple principles. We don’t know if that hypothesis is true, but the success of deep learning in the last few years is a good validation of that hypothesis. It’s consistent with that hypothesis because deep learning is built on a few very simple principles. Most of the complexity of the systems that are trained with deep learning is not in the learning mechanismsp; it’s in the data. The data contains the overwhelming share of the information in a current trained AI system, while only a little bit of information, relatively speaking, is in those principles, which are like the learning procedures.
Biased Data and AI for Humanity
TT : You said that the intelligence is from knowledge and knowledge is acquired from data?
YB : That’s right.
TT : But if data is biased, what happens? Some social scientists criticize most data for being based on male, Caucasian middle aged…
YB : That’s right.
TT : So if a young woman of colour applies, for instance, for an insurance policy, AI might say no because they don’t have enough data for those applicants.
YB : Yes, absolutely. I think there are technical solutions and social solutions to this problem. We have to change our social norms, for example, so that companies building products use technological solutions and logistical solutions, for example, in the way that the data is collected, in the way that it’s described and managed, and in the particular learning algorithms that are used because we know techniques that can mitigate the bias and discrimination. So we can probably include those techniques, but more importantly we need to make sure that companies and governments use them.
TT : Is that why you think it’s important that both social scientists and natural scientists work for AI together?
YB : Yes.
TT : I love the idea of “AI for humanity” as you have in the Mila here.
YB : Right, because the AI researcher might not realise some of the social issues that could be involved in the deployment. I think it’s particularly important for people who are doing research or development of products that is close to something that people will use, in large-scale deployment for example.