University of Montreal professor Yoshua Bengio is well known for his groundbreaking work in artificial intelligence, most specifically for his discoveries in deep learning. Bengio has shared his research in more than 200 published journals and reports and most recently began imparting his AI knowledge to entrepreneurs in the start-up factory he co-founded, Element AI.

One is deep unsupervised learning, which is using deep learning principles to learn in an unsupervised way, meaning without human guidance or labels on the data that is fed into AI systems to train them.

Our main industrial systems derived from deep learning—such as speech recognition, machine translation, image search, self-driving cars, vision systems for blind people, etc.—take advantage of progress in deep supervised learning, yet humans are very good at unsupervised learning, and we need to make substantial progress in that direction to approach human-level AI.

And then understanding and generating natural language are also exciting areas, which are probably going to have a huge impact on applications such as dialogue, understanding documents, interfacing with computers, etc.

However, many hard problems remain open, like how to allow AI systems to automatically learn high-level semantics—in other words, how to represent more abstract concepts and the meaning of not just words but more complex thoughts, such as those we express in a phrase, a sentence or a paragraph.

We’ll probably see systems that can understand and do a good job at generating natural language ready within the next five years, while deep unsupervised learning is likely to be farther out. That will take many years of patient, fundamental research.

Reasoning is about combining elements of knowledge, so in order to do a good job at it you need to acquire that knowledge in the first place. Traditional reasoning methods are very powerful, but they miss the knowledge on which to reason. That is why reasoning has to be coupled with machine learning, which extracts the knowledge from data. Current systems that learn to reason are still in their infancy, but the very fact that we are able to learn to reason using attention and memory extensions of neural nets is extremely promising.

Difficult to say. Clearly one obstacle is simply to scale the current approaches. Another is to revisit the question of knowledge representation. Knowledge graphs were designed to be curated by humans, but we may need to learn other forms of representation which are more amenable to being extracted from data automatically—from documents, for example— by deep neural nets.

Absolutely. It could have a huge impact on everyone’s health, for example on tackling cancer, which is the main killer in our societies. AI will allow for much more personalized medicine and bring a revolution in the use of large medical datasets. We’ll see patient-specific treatments—for example, ones created using your genomic and expression data, which are much more likely to work. Currently we are using very blunt instruments to treat patients. This has the potential to change a lot. And we’ll also see a much more efficient use of doctors’ time.