Computer Vision News - November 2018

science w as working on neural nets or even sim ple machine learning. There was a litt le bit of machine learning in the conte xt of AI, but a very, very small communi ty. I met some people who had tho ught about the question of emerging properties of networks compose d of lots of simple, connected elements, which is really what neural nets are. I got in touch with them and discovere d that there was an internatio nal community that was starting t o work on neural nets. I read the paper by John Hopfield on Hopfield nets, and discovered the existence of Geoff Hin ton and Terry Sejnowski. They had just published a paper on Boltzman n machines, which I thought was won derful and I really wanted to meet the m. Then I s tarted my PhD. My advisor didn’t kn ow anything about neural nets and said: “ I can sign the papers, you seem sm art enough, but I can’t really help you from a technical point of view. ” I had a scholarship through ESIEE, m y engineering alma mater in Paris. Eve ntually I discovered a version of backpr op on my own around 1984 and ende d up meeting Terry Sejnowski and Geof f Hinton at conferences in France i n 1985. At one of these conferenc es, I met Larry Jackel and John De nker from Bell Labs who eventuall y were going to hire me. I did a postdoc with Geoff Hinton in Toronto in 1987/8 8. Larry Jackel had started a group at B ell Labs on neural net hardware and hired me right after my postdoc. Let’s get to the talent part of how you got here. I interviewed Yoshua Bengio and he told me : “ I did not succeed because I am clev erer than the others, but because I kno w how to focus very well. ” Would yo u agree with that sentence? If that is his secret, what is yours? I surround mysel f by people who are smarter than me , so I certainly don’t see myself as pa rticularly talented in many areas. I’m m uch more impressed by other people. For example, I have a very long-term interaction and collaboration with Léon Bottou. He is a well-known figure in machine learning and is better than me in almost every way! [ he laughs ] O ne thing I like to do, and perhaps it’s something that I do quite well, is try to really get to the core of what is the central problem behind a questio n. How do we get machines to lear n? Things like that. Sort of genera l orientations and intuitions abou t what are the important pro blems. Simplifying questions down to their core. Sometimes an id ea or concept seems very complicated because of all the complicated math s you have to use, but at the core th ere is usually a very simple idea. I do n’t want to compare myself to Richard Feynman, but his way of thinking was very similar. In the sense that it’s trying to ask the elementary ques tions and to reduce everything to th e simplest questions you can imagine. I’m not at all putting myself in the same league. As for Yoshua, Yos hua is very disciplined and organised. I’m not. I’m very messy. Well, you’re Frenc h! Pauline Luc 5 Computer Vision News “I read the paper by John Hopfield on Hopfield nets, and discovered the existence of Geoff Hinton and Terry Sejnowski” Yann LeCun Guest