- 3:00 pm on Thursday, January 31, 2013
- 4:00 pm on Thursday, January 31, 2013
- MCS 148
Title: High dimensional information processing with limited resources in
neural systems. Abstract: As is the case in many information processing systems, in the nervous system information processing is performed by multiple neuronal networks each having access to different types and amounts of information. To understand information processing in the brain, we should thus study learning approaches that are based on computationally efficient information sharing methods. This is the focus of this talk.
The first part of the talk will focus on a network of observers making local observations concerning an unknown vector. Each node faces a local identification problem, in the sense that it cannot consistently estimate
the parameter in isolation. Employing a novel local message passing algorithm, I will show that despite local identification problems, local estimates can be as efficient as any ideal global estimator.
In the second part of the talk, I will discuss optimal decoding, information rates and dimensionality reduction of high dimensional spatio-temporally correlated spiking activities. In particular, I will show that neural populations with strong history-dependent (non-Poisson) effects carry exactly the same information as do simpler equivalent populations of non-interacting Poisson neurons with matched firing rates.
Finally, I will present a scalable and robust method for extracting as much information as possible from the simultaneously recorded activity of
tens of thousands of neurons. The methodological innovation of our work is to use a regularizer that is robust to occasional discontinuities, and
nevertheless if there is enough evidence in the data, enforces similarity
between nearby neurons, all in an adaptive fashion.