Database Reference
In-Depth Information
a London company called Deep Mind. All of Deep Mind's code is essentially
built around Torch. Ronan Collobert who is one of the originators and main
developers of Torch, has joined Facebook. The other two main developers are
both former students of mine. Koray Kavukcuoglu is at Deep Mind. Clément
Farabet is at Twitter here in NewYork. There are other developers as well, but
these are the main three characters. Many other companies are using Torch7
for deep learning besides Facebook, Google, and Twitter.
Gutierrez:
What papers have been published on the work you're doing at
Facebook?
LeCun:
We had two papers at the CVPR 2014 conference. One of the papers
was on a system called PANDA, which basically does human detection in
images. The other paper was on a system called DeepFace that does face
recognition with very good accuracy. Both of those systems use convolu-
tional nets. Both of those projects, by the way, were started before I joined
Facebook. We also had two papers at the EMNLP 2014 conference on topics
such as text embedding—for things like hashtag prediction—and question
answering. There are other projects that were started more recently that
we have submitted papers on just in the last few weeks on essentially natural
language-related tasks.
A natural language model is a system that can predict, upon seeing a piece of
text, what word is going to come next in the text. It can't predict the exact
word, but it can predict the distribution of the words and can then measure
the accuracy of that prediction. So we've been working on basic techniques to
be able to do this—not necessarily because we're interested in language mod-
els, but because we're interested in the underlying techniques that can be used
for all kinds of different domains.Variations of recurrent neural nets, for exam-
ple, is one of these techniques that we are exploring. A particularly interesting
development is what we call “Memory Networks”, which are recurrent nets
augmented by a separate “short-term memory” that the neural net can use to
store temporary states. Using a memory network, we have been able to build
models of simple worlds, similar to the world of a text adventure game, from
which the model can answer complex questions after being told a number of
events. An impressive example is a paragraph that described the sequence of
major events in Lord of the Rings. Then we can ask questions like “where is
the ring?” (answer: Mount Doom), “where is Frodo?”, etc.
Gutierrez:
What lessons have you learned from being at Facebook and
working with their data sets?
LeCun:
Well, the data sets are truly gigantic. There are some areas where
there's more data than we can currently process intelligently. Content analysis
and understanding users are very important problems. And then there's the
big question of matching users' interest with content. There is an interesting
interplay in areas like image recognition and natural language understanding
