Anthony J. Bell Ph.D.
Redwood Center for Theoretical Neuroscience
132 Barker, MC #3190
Berkeley, CA 94720-3190
phone (415) 699 6502
fax (510) 643-4952
Here's my Salk web-page from way back.
Here's me giving a 30 minute talk Levels, Time and Models about Levels in Biology.
Here's me giving a 85 minute talk Emergence and Submergence in the Nervous System.
(The production on the latter is not so good, so here are the slides.
Also, if you wait a few minutes in, the audio drastically improves.)
What am I doing? If you watch either of these you will see, at least, where I am starting from.
There are 3 steps to uniting physics, biology and machine learning:
(1) Solve the time series density estimation problem (this will be very useful).
(2) Solve the sensory-motor density estimation problem
(3) Solve the levels density estimation problem
The clues to (2) and (3) lie in (1). The clues to (1) lie in non-equilibrium statistical mechanics. We are working on (1) and I think we've got it.
A paper on this Learning out of equilibrium will be ready shortly. Email me if you want it when it's ready.
The reinforcement learning literature, while elegant, has nothing relevant to say about these deep problems.
Nor, in my view, do generative models, which estimate fictional hidden variables.
And unfortunately there's no way around it: to really crack this we are going to have to get real.
That means we use real biology and real physics to guide us. Computational fantasizing has demonstrated its limits.
We will have to absorb and augment the emerging non-equilibrium statistical mechanics and, in the end, also (cough) quantum theory.
I know these are radical views, but still, after a lot of thought, I believe them to be correct. Fortunately, both these branches of physics,
in their core mathematical structure, are relatively simple.
Here's an unsatisfactory paper Towards a cross-level theory of neural learning that explains what I was thinking up till about 2008.
Here's my CV. (Ignore the stuff about the cyclist journalist - that's not me :)