Database Reference
In-Depth Information
mentors. In neuroscience it's Sebastian Seung and Eero Simoncelli. These are
people my age or slightly younger or older who I know are interested in the
same questions and have slightly different takes on things and we always learn
from each other. I cherish all the time I can spend with them. There are also
lots of younger people whose work I find amazing. I've tried to hire them
all at Facebook! In terms of people whose work I find inspiring, they're not
necessarily in my field. They are people like Richard Feynman, Albert Einstein,
or people with similar breakthroughs. What I find inspiring about them is
their particular intellectual way of approaching problems, which I find fascinat-
ing, if not mysterious, and which I'd like to be able emulate.
Gutierrez:
What do you look for in other people's work?
LeCun:
The work that I tend to be interested in is innovative, creative work,
which occasionally doesn't make it to the big conferences because it's too
innovative and too creative. The review process for big conferences tends
to focus on incremental improvements on mainstream models, which I find
completely boring. I mean, it's useful, so don't get me wrong. I don't want to
say people should stop doing this type of work, but it's not like I learn much
from that. It's just that I am interested in the innovative creative work. I learn
much more from that type of work.
Another type of work that interests me is the type that is very careful experi-
mental work, which is very impressive. This is where the work that's been
done is relatively straightforward and obvious, but it's great that people have
done this work so exceedingly well that you now have a piece of data that you
can point to and say, “This actually works if you do it right.” In fact, to some
extent, convolutional nets were like that to many of us.
I remember that I'd been out of my postdoc with Geoff Hinton for about a
year and I gave a talk at NIPS in 1989 on convolutional nets. It was the first big
talk on convolutional nets that I gave, and Geoff Hinton was in the audience.
I had started thinking about things related to the topics I was speaking about
when I was a postdoc with him inToronto, so he knew a bit about my thoughts.
At the end of my talk he said, “This is a very good talk you gave. Basically, the
result is that if you do all the sensible things right, it actually works.”
It was at the same time a compliment and not a compliment in that he was
sort of saying, “This is not particularly innovative. We knew that this was the
right thing to do, and it's nice that you've done it and shown that it's the right
thing to do.” I think there is something to be said for that, and I find this inter-
esting in other people's work as well. Of course, these ideas weren't nearly
as obvious to other people as they were to us. But the best ideas are always
obvious a posteriori. Some people are more impressed by mathematical virtu-
osity—you know, really impressive technical work at a very high mathematical
level. Some of that work is occasionally useful, though not always. I'm very
