Database Reference
In-Depth Information
I've started working with clinical schizophrenia data partly because there are
people out there suffering from this disease and, in some sense, that's absurd.
The fact that disease is a “thing” is absurd. We're mechanistic, we should be
able to fix ourselves, and the world I want to exist in 20 years will solve that.
So going back to the goals, it's important to ask if we are closer to this world
or not?
Startup culture teaches you to be like Steve Jobs, in that you're right, everyone
else is wrong, and your vision will power through. Academic culture teaches
you that you're dumb and that you're probably wrong because most things
never work, nature is very hard, and the best you can hope for is working on
interesting problems and making a tiny bit of progress. Just doing that is seen
as an amazing career. So the question is: How do you reconcile these kinds
of things? I don't know, I struggle a lot with reconciling these two cultures in
myself.
Some of the best scientists out there are the ones who are extremely oppor-
tunistic—when they see novel ideas and how things suddenly fit together,
they drop everything else and work on that for a while. Others are consumed
by a single, all-encompassing vision and aggressively pursue that forever. The
downside, in many ways, is that the academic funding system really rewards
the former, in that if you have three
Nature
papers with no clear coherent tie
to them, it doesn't matter. “You have three
Nature
papers—congratulations,
Professor!”Whereas if you've been working on the same problem for 10 years,
but only making incremental progress—“Well, sorry, you're not getting tenure
at a place like MIT. I really hope you enjoyed working on the problem.”
That's one of the reasons why I'm not necessarily excited to go back into
academia, because the incentive structures are so confused around this issue.
Gutierrez:
What kind of tools have you used and do you use now?
Jonas:
From a technical point of view, I'm almost entirely a Python and C++
person. I do C++ for the heavy numerics and Python for basically everything
else. It's an extremely productive environment. It's nice because, as someone
with computer science training, I can do complicated things in Python. It's not
like MATLAB, where you have to jump through a million different hoops. And
I can drop down in C++ when I need it for speed.
Mathematically, a lot of what I work on is Bayesian models using Markov chain
Monte Carlo to try and do inference. I really like that universe because the
world is so simple when you think about it probabilistically. You can think
of a stochastic process that you can condition on your data and do the
inference. That's great! It means the set of math I have to know is actually
shockingly small, especially because often the problems that I'm working on
don't have data from a billion neurons yet—we have data from 100. And so
I'd much rather spend my time building complicated correct models and then,
