Yann LeCun - Data Scientists at Work - page 45

LeCun: I'll start with the long-term goal. It's going to sound incredibly arro-
gant, but it's basically to solve AI. Of course, a lot of people have tried to do
this and failed, so we're not claiming we're going to succeed, at least we're not
giving a timeline. The long-term goal is to really understand intelligence, be it
natural or artificial, and try to make machines more intelligent so that they can
help people in a better way all around—help people communicate and help
people deal with the digital world. So that's the long-term goal.
The shorter-term goal is to understand content. People upload hundreds of
millions of images onto Facebook every day and also make billions of posts
and comments. One of the main things that Facebook has to do is essentially
decide what to show to users. Every single day users could be shown roughly
between 1500 to 2000 items. Users don't have time for that, so Facebook's
algorithms basically selects a couple hundred of those that the user does have
time to see every day. To do a good job at picking what to show users, we
have to essentially match the content to users' interest. Being able to figure
out what's in an image, what a post is about, what topic it is, and whether it's
a topic that a particular user is likely to be interested in allows us to better
match content to users' interests. So understanding content is really essential
for Facebook.
Gutierrez: How much deep learning is involved in this process?
LeCun: A lot of things we do are based on deep learning, but not everything.
Deep learning is very useful for us, particularly for things like images and some
text representation and some topic extraction, because deep learning shines
in situations where we have lots of data. Of course, Facebook certainly does
have lots of data, so that's why we use it. In a lot of situations, it beats the
approach of handcrafted features followed by classifiers just because of the
amount of data we have.
Gutierrez: Looking back at how your career has progressed, were you
following a path?
LeCun: Though there is no clear path in terms of institutions I've worked in,
there has certainly been a path in terms of the technical problems I've been
interested in. In fact, what I'm interested in has been pretty constant except
for short time period. I've always been really fascinated by AI and related
subjects since I was a kid. While I was undergrad in the late 1970s through
early 1980s, I studied electrical engineering. During this time, I did a bunch of
projects trying to figure out if we could make machines learn. I was always
convinced that the only way to make intelligent machines was to get into
learning, because every animal is capable of learning. Anything with a brain
basically learns.
I approached the problem by searching the literature for machines that could
learn and realized that, at least in the early 1980s, nobody was working on
these types of problems. The only literature I could find was from the 1960s