Computer Vision News - November 2018

11 Computer Vision News Yann LeCun but they were a little too possessive about IP to really be open in the same way. There was no tradition of open source and things like that. Now, it’s different, and this is a new way of doing industry research. I think some of the other companies are being influenced a little bit by us. For example, in the last five years, Google has become much more open about what they do in research than they were in the past. They’re still a little secretive, but they’re definitely much more open than they used to be. I have a witness to what you said you are trying to implement at Facebook! Pauline Luc, whom I interviewed last month, told me : “ My lab at Facebook is the same as my lab at the school. ” That’s right. I know Pauline’s work very well of course, because I participate in her project and co-author the papers. I think the work she’s doing is amazing. Finally, what would you say is the biggest accomplishment you would like to achieve before you retire? Finding good ways to do self- supervised learning in a generic way. I’ve been working on computer vision, but I’m not a computer vision person. I don’t see myself as a computer vision person. At least not entirely. My interest is really in learning, so I like to find ways to get machines to learn how the world works, by observation. That means learning under the presence of uncertainty. If you give a machine a segment of video and you ask it to predict what’s going to happen next, there’s many things that can happen. Of all the possible things that can happen, one of them is going to be the thing that actually happens in the video, but there are many possible scenarios that may happen. When you train a machine to predict videos, if you’re not careful, it produces a blurry prediction which is sort of an average of all the possible scenarios that can occur. That’s a bad prediction. One of the technical problems we’re trying to solve is: how do you train a machine in a situation where what you’re asking it to predict is not a single thing, but a set of potential things? You can formulate this mathematically as predicting a probability distribution instead of a single point. But we don’t know how to represent probability distribution in high-dimensional continuous spaces. I think it’s going to be a combination of, again, figuring out what are the essential concepts there, and coming up with simple architectures that are easily understandable and can deal with that problem of representing uncertainty. GAN is a promising approach. But GANs are not very well understood. They converge sometimes, but when they work they work amazingly well. They don’t work every time, so we need to find ways to either understand how they work or find other techniques, then use those techniques to get machines to learn as much background knowledge as possible by just observing the world through video, images, etc. Then, once the machine has learned a good model of the world, it will only require a few samples or a few trials to learn any particular task. That’s my goal for the next few years... “GANs are not very well understood” “That’s my goal for the next few years!” Guest