Database Reference
In-Depth Information
In physics, a lot of the new results in astrophysics and high-energy physics actu-
ally rely very heavily on large data and complex statistical models. Things like
the discovery of dark energy, for example, co-discovered by Saul Perlmutter,
Nobel Prize winner, who is my counterpart of the Moore-Sloan Data Science
Initiative at UC Berkeley, was made using massive statistical analysis. Also, a
thing like the discovery of the Higgs boson was the result of massive statisti-
cal data analysis and results. Part of the system for this work was actually
designed by my NYU colleague, Kyle Cranmer, who designed the integration
for all the statistical models.
Data Science is also on its way to revolutionize social science. There is actu-
ally a big push from social scientists who would love to put their hands on
Facebook's data. Facebook does not give its data away because it's private for
users, and so that data is inaccessible for science. It's sad perhaps, but that's
the way that Facebook operates.
Gutierrez:
What is something that a smallish number of people know about
now that you think 5 or 10 years from now will be huge?
LeCun:
I think we'll have systems that do a much better job than they cur-
rently do at things like language translation. The difference will be that they
will actually understand what it is that they are translating. Current systems
that do translation don't actually understand what they are translating. It's just
glorified statistical pattern matching. Eventually, however, these systems will
understand more and more about the text that's being translated. The tech-
niques for language understanding, in general, will have wide applications, not
just for translation, but also for search indexing and intelligent agents. One of
the things that people have been asking me about in the last few months is the
movie
Her
and my thoughts on interaction with the intelligent agents. We're
not going to have this within 5 or 10 years. This is way beyond what we can
do, but intelligent systems that have some hint of common sense may appear
in the next decade.
Gutierrez:
What advice would you give to someone starting out?
LeCun:
I always give the same advice, as I get asked this question often.
My take on it is that if you're an undergrad, study a specialty where you can
take as many math and physics courses as you can. And it has to be the right
courses, unfortunately. What I'm going to say is going to sound paradoxical,
but majors in engineering or physics are probably more appropriate than say
math, computer science, or economics. Of course, you need to learn to pro-
gram, so you need to take a large number of classes in computer science to
learn the mechanics of how to program. Then, later, do a graduate program in
data science. Take undergrad machine learning, AI, or computer vision courses,
because you need to get exposed to those techniques. Then, after that, take all
the math and physics courses you can take. Especially the continuous applied
mathematics courses like optimization, because they prepare you for what's
really challenging.
