Technical Interests
My primary technical interest is in systems that adapt: how to
analyze them, how to understand them, how to build them. Because
the most flexible and competent adaptive systems available to us
is the nervous system, I'm interested in artificial neural networks
and computational neuroscience. I'm most fascinated by the
construction of novel architectures and algorithms that enable us
to understand and attack previously unassailable problems, and to
understand previously mysterious aspects of nervous system
function.
For specific projects, see also the Brain and
Computation Lab Research Projects page. Here is a broad
overview of my personal interests.
- Blind source separation: We are working on better and
more modular and incremental methods to solve the ``cocktail party
problem,'' both in the classic (linear square mixing matrix) and
in the more difficult (fewer microphones than sources) cases. We
have been applying BSS to magnetoencephalographic data with great
success, and are looking for people at
all levels to work on that project: from undergraduates to staff,
grad students, and postdocs.
- Reinforcement learning in a weakly adversarial
domain: In the real world, one's actions modify the world,
typically to the detriment of similar actions in the future. I'd
like to understand how to perform as well as possible, under the
circumstances.
- Neural information and coding: How is information
represented and transformed in the nervous system? How are
these representations acquired and adapted?
- Egomotion: The process of estimating a camera's
motion efficiently, reliably, robustly, and using beautiful
mathematics.
- Neural networks: Learning algorithms,
generalization, relations to other techniques, handling time and
domain drift in a principled fashion, unsupervised learning,
information theory.
A secondary interest of mine is in programming systems, especially
advanced programming language design and implementation. We have a
nascent effort to build a new efficient advanced programming
language with generalized robust performant automatic
differentiation operators. The hope is to allow many numeric
algorithms and scientific computations to be expressed more
clearly and very succinctly.
Barak Pearlmutter <barak@pearlmutter.net>