Database Reference
In-Depth Information
So these days I spend my time doing two things. First is working with
neuroscientists and answering neuroscience questions about how learning
and memory work. How does your brain take in new information, build
models internally, and then do things such as saving that to the neocortex?
Then second thing I spend my time doing is building new types of machine
learning models to find patterns in exactly this kind of data. I think that's really
where the field is headed, especially as we see a tremendous set of interests
in neuroscience data now.
I have been interested in the brain for most of my life. Even before college,
I was already fascinated with it. Then when I arrived at MIT, I double-majored
in electrical engineering and computer science [EECS] and brain and cognitive
sciences [BCS] as an undergraduate. I then did a master's in EECS, and then
finished it off with a PhD in BCS. So my background is computer science and
brain and cognitive sciences, and I have been studying it for many, many years.
This area is something I've been thinking about even when I was working in
the industry.
Going from being a startup CEO to working in a large company is a jarring
transition, because suddenly you have extra time after work. In this extra time,
I finished my PhD, and I continued doing some of the neuroscience research
I did in my PhD. I was always very interested with the idea of building machine-
learning signal processing technology to find patterns in neuroscience data
that we just couldn't find natively. As undergraduates we're taught things like
the Fourier transform or principal component analysis, but those are all like
fifty years old. Clearly the machine learning technology that's out there has
become a lot better at finding those sorts of patterns in data. So lately I've
been working with researchers at UCSF, Northwestern, MIT, Harvard, and
other academic institutions.
The Obama administration has spearheaded the new BRAIN initiative and
there's going to be all this high-throughput neural data around soon, but
frankly, not a lot of technology to even look at it—especially technology that's
accessible to regular scientists. And I say “regular scientists” partly because
the people running the experiments and formulating the questions are often
actually quite different from the people with the ten years of computational
or mathematical training necessary to build and understand the models. I am
trying to carve out a space in between them to be at the impedance matching
layer, serving as that buffer because I believe that you can be really productive
there.
Gutierrez:
Why is this important work now and ten years from now?
Jonas:
One of the reasons that I left Salesforce was in some sense that my
team was there, everything was in good hands, and I was looking for the next
big challenge. Aaron Levie, the Box CEO, at one point made the comment,
“One of the questions you have to ask yourself when doing a startup is why is
now the right time to do the startup?” I think Peter Thiel told him that.
