On the heels of being awarded the Nobel Prize for computes science in March, Montreal’s AI icon Yoshua Bengio has been honoured with yet another award. The Killam Prizes are among Canada’s most prestigious awards for careers in research.
Montreal in Technology sat down with Dr. Bengio on May 2 to discuss these latest honours, bringing research and corporations together, managing expectations and the future of AI in general.

Steve La Barbera for Montreal In Technology:
Thank you very much doctor Bengio and congratulations on all your recent awards. This time it is the Killam award. Did you ever think that you’d be awarded a prize for natural sciences?

Yoshua Bengio:
That’s a good one! I guess, a lot of what I’ve been doing is inspired by the brain and collaboration with neuroscientists and people who do cognitive science. So, there’s pretty fundamental connections between artificial intelligence and a natural intelligence and especially in the work that we did in deep learning. I’m happy that I’m being awarded this is under a natural science.

SL - MiT:
You’ve been doing this what now 25, 30 years?

Yoshua Bengio:
Yes.

SL - MiT:
At what point did you realize that you were on to something?

Yoshua Bengio:
You know, I got really passionate about this field when I started doing research as a master’s student. So that’s in 1986. And as I developed maturity in the field, I became more and more convinced that we were doing the right thing and going into the right direction. But I did not expect that we would get this exponential growth and economic impact as it has happened in the last five years. This came much faster than I would have expected for sure. But, but in a way, we had a strong conviction that we were on the right direction and it would have a big impact on society.

SL - MiT:
In Canada, there are just so many university research projects which could be promising, but it’s just kind of really challenging for people to commercialize these ideas and get them out of the lab and into companies. What do you tell your research colleagues in terms of getting the ideas out of the lab and turning them into actual products and services?

Yoshua Bengio:
It’s true. It’s an art. Doing this right is very different from what we do as researchers doing the transfer, creating new products, creating companies. So, what we are trying to do is educate our research students, like PhD students, so they would learn about entrepreneurship. They would learn about the investing world. They could get connected to people who have that experience because the best thing at the end of the day, if you are coming from the research environment, is to collaborate with somebody who has experience in starting companies or in getting funding from VCs and so on.
So, it’s a matter of experience and knowledge. We can teach them things, we can connect them with the right people. I think a big mistake many people make in this and in many other cases, is that they think they can do everything by themselves and that they know enough. We all need others to help us in our projects, so we shouldn’t hesitate to reach out and a team up with other people.

SL - MiT:
Speaking of which, I saw there was an announcement, a Samsung press release this morning and can’t help but notice you’re wearing the Samsung badge now.
Yoshua Bengio:
This follows up the inauguration that happened yesterday for their new lab, which is going to be one of the corporate labs within Mila. So, a number of companies are putting researchers on the same site in the same building so they could interact more easily with the academic scientists at MILA.

SL - MiT:
The announcement I saw was an expansion of their presence here [at MILA]. Is that right?

Yoshua Bengio:
That’s right. What do you have a small office. Yeah. Um, and now it’s going to be a lot bigger. They recruited one of the professors to lead that effort and because it’s right embedded in the same place, the same building, I think it’s going to be very, it’s going to be more productive for them.

SL - MiT:
Sure. Do you a sense of how many people Samsung will have onsite after when this is fully up and running?

Yoshua Bengio:
So they have room and budget for expanding their research group here to 20 people. Which, you know, on a research scale is actually pretty good.

SL - MiT:
How do you manage that? Where you work with a company or several companies in your case, to develop technologies that are actually going to be in demand from large corporate customers? What’s the key to identifying that industry pull and building these relationships?

Yoshua Bengio:
It’s communication, right? If the people who understand the problems talk with the people who will understand potential solutions, then, good things can come out. So, it’s all about how we interact. It’s all about creating a space where that communication is going to be fluid. That’s why we have a tech transfer team, for example, at MILA. it’s a lot of their job to build this bridge between the academic knowledge and the needs of industry. We talk to startups, smaller medium enterprises, large companies.

SL - MiT:
And do you find you get more traction with any one or other of those groups?

Yoshua Bengio:
Yeah. There’s a difference between the Canadian culture of tech company and the Canadian culture of big corporations versus what you’d find in the US, particularly in tech. So unfortunately in Canada we have currently a very conservative culture. A corporate culture that isn’t yet willing and ready to take these sort of risks and long-term very ambitious bets, which are needed in tech and especially in AI.
Currently these kind of bets are being made by companies from the US or China or here we’re seeing South Korea. I wish things would change faster in Canada to allow companies to be more agile and be not lagging but leading the transformations that AI is going to bring to society. I see things changing in the right direction, and where it’s most obvious, basically, is in the smaller the companies, they’re more able to do this kind of cultural change right now.

SL - MiT:
Do you have a sense of why the culture is different here than in the US?

Yoshua Bengio:
I mean, if you go to the US and you look at non tech companies, they’re also very conservative. But, what happened is that in places like Silicon Valley, they’ve developed a separate corporate culture that is agile and is willing to invest on the 10 year horizon. Increase the ideas. And is willing to consider changing the process of, some industry. We are seeing that same type of drive currently in China. I’m not sure what are the ingredients for this. And I see some of it on a small scale here in Montreal in the AI ecosystem, but I think if we want to have a big impact, it needs to expand to a much larger set of companies.

SL - MiT:
When you work with these corporate partners, whether they be foreign or whoever, those who are buying in… Are there expectations on their part that something is going to be delivered within a certain timeframe? How do you manage that?

Yoshua Bengio:
Right. So that’s a very good question. And I think there are different deals that are being made with different professors or different research groups. Here at MILA we have a very high profile, in terms of international recognition and we are training so many very strong students with a very rare expertise.
Companies are actually willing to come and collaborate with us without very definite commitments in terms of deliverables because that’s the only way you can do basic research. That’s the only way you can do very innovative research. If you want to do that kind of research we’re doing here at MILA, you have to be in that frame of mind. If you develop a product that’s already well defined and understood, it’s a different story.
But in terms of the research collaborations, we’re able to work in a set up that’s fairly free of IP constraints, for example. Where there’s the understanding that we are going to be communicating what we do openly in publications and scientific meetings and so on. And that we are open to collaborations involving potentially multiple companies without very rigid walls separating the work done by one company, for one company or the work done by another company. So we’re lucky right now to be in a strong position to negotiate with companies these sorts of arrangements. That’s for large companies.
For startups and small companies, which we want to encourage, we understand that it’s a different kind of relationship where either their companies are actually started by our students, or they come to us because they need expertise. And then it’s more like a service that we’re giving them that will help them avoid big mistakes. And we have a group of tech transfer engineers and scientists who are doing this sort of consulting if you like, providing help in their projects.

SL - MiT:
For 2019 & 2020, what is Yoshua Bengio working on? What’s the next thing? What’s the new tech?

Yoshua Bengio:
Well, at this meeting today I talked about a research direction which I find really exciting. It has to do with the ability of learning systems to figure out causal relationships and how world works in terms of, if I do this, what will be the effect? This is something that current machine learning doesn’t do very well, but we’d have many, many applications if we can make important progress.
Current machine learning systems, they are really stupid. They don’t understand what they’re doing. They learn some tricks to solve the problem, but they don’t have an understanding of how some aspects of the world work. Like humans are. We are able to imagine how things could be if we did X or Y, or if someone else were to do something. It’s not something that current systems can do. We think that it’s a very important cognitive ability that we need to put in machines.

SL - MiT:
Just as an example, what would that be used for?

Yoshua Bengio:
Well the main advantage from a practical point of view is that it might address some of the lack of robustness. When you build an AI product these days, it’s trained in a particular way, with a particular dataset, particular data collected in some environment, and then you use it in a somewhat different way… Like let’s say you take pictures in Canada, but then you want to use your system in China, it might not work that great because of the changes between the setting where you collected the data and the setting where you want to use the system. Humans are much better than that at adapting to these changes in the environment.
So, we believe that the ability to better capture relationships in the world actually comes with the advantage of being able to generalize about changes in the data and the distribution that are otherwise difficult to achieve. And so, it would lead to more robust systems that are going to be more portable, that can be adapted from very little new data. This is an issue in medicine. This is an issue in all kinds of products. This could have a potentially very important impact. We call this transfer of learning. And by the way, this is one of the core missions of the research group at Element AI, which I co founded a few years ago.

SL - MiT:
Some of your students have worked on things like a voice replication technology and others are now doing the same thing for videos. Technology like this and other technologies related to artificial intelligence have huge potential but also kind of pose some ethical dangers. We’ve just discussed some of the biggest opportunities are in AI. Can you tell me what, what are some of the dangers and the challenges at this point?

Yoshua Bengio:
Absolutely. I think it’s really important that more people understand those and social aspects of AI and impact. Because, as these technologies become more and more powerful, they could both be used for good or they could be misused. And the only way we’re going to be able to minimize the misuse is by collective wisdom. By collective decision making by laws, regulations, social norms.
The kind of misuse I’m worried about includes things like military use. Killer drones that can target particular people, say all the journalists or all the members of parliament from a particular party. You could imagine really terrible things using these techniques. Another example is in surveillance, how that could be used to not only breach privacy but control people, and have authoritarian regimes use these tools to stay in place and be able to keep their power in a way that wasn’t possible before.
Other concerns include the potential damage due to bias and discrimination that can arise if those systems are not trained properly with the right data. And finally, maybe the thing that people worry most about when I talk to ordinary citizens is the potential for job loss due to automation.

SL - MiT:
Someone yelled at me about that yesterday.

Yoshua Bengio:
Yeah. And there’s good reasons to be concerned. Of course, I don’t have a crystal ball and nobody does, but the trend and the projections that economists are proposing suggest that within the next decade we could see fairly major impact on the job market in some areas, in many areas. It’s hard to predict exactly which areas, although some people have tried, but we need to prepare better for that. Whether it’s in the education system or the social safety net. I think changes in these very complex machines, government machines, take time and it’s better to get at it soon.

SL - MiT:
I’m glad you brought that up. Is it not better then to be developing the technology that will make some of these jobs obsolete but create jobs in and of itself? Is it not better to do that here in Montreal rather than have it done somewhere else?

Yoshua Bengio:
We really ought to create an AI industry here, not just because it will create jobs, but more importantly because if we create wealth and we would like that wealth to be returning to the people here and for that we need to be producers of AI rather than just passive consumers of AI that’s being developed in other countries.
But at the end of the day, I don’t think that the number of jobs that will be created in technology will match the number of jobs that could be lost. And it won’t be the same people. So, if you’re a factory worker that doesn’t have a lot of specialized skill, I don’t think you’re going to transform, even in a few years, into an AI engineer. I mean, we should try to help people do that transition, but many won’t be able to. And so we need to think how we organize society. Maybe we train them potentially for other things that will become more valuable, like skills that have to do with a human to human interactions. Because we need humans. We want humans in front of us in many cases. There’s, there’s a lot of thinking that needs to be done to prepare for that and avoid misery that would otherwise very likely happen.