SPEAKER: Doru Balcan TIME: Wednesday 12-1pm, April 18, 2007 PLACE: NSH 1507 TITLE: Characterization of Robust Linear Coding Solutions ABSTRACT: Many linear encoding approaches focus on representing information by approximating the underlying statistical density of the data, like PCA or ICA, or by developing encoding/decoding algorithms with desirable computational and representational properties, such as Fourier and wavelet-based codes. However, if the coefficients' precision is limited, optimality of the representation cannot be guaranteed. The issue of optimality under limited precision is a common practical concern, and it is also relevant to biological neural representations, where the coding precision of individual neurons has been reported to be as low as a few bits per spike. In this talk, I will present a new coding scheme called Robust (Linear) Coding, that makes use of arbitrarily many coding units to minimize reconstruction error. One characteristic of Robust Coding is that it can introduce redundancy in the code to compensate for channel noise, unlike PCA or ICA, which aim to reduce redundancy. We can completely analyze the under- and overcomplete cases, and identify necessary and sufficient conditions of the optimal encoder/decoder pair in each case. In the process, we find an exact formula for the lower bound of the error function, as well as an efficient procedure for computing the optima. Other Information: Joint work with Eizaburo Doi and Mike Lewicki.