Mathematical structure of Information Distortion methods Tomas Gedeon, Department of Mathematics, Montana State University

In this talk we introduce Information Distortion function used in neural decoding problem and compare it to a closely related function used in Information Bottleneck. We discuss the annealing approach to finding an optimal solution for both functions. We show that there is a close relationship between the first phase transition in the annealing and the Approximate Normalized Cut for certain graph G. We also show that we can explicitely compute values of the annealing parameter where the phase transitions happen. As a consequence the annealing process for both functions can be characterized as Deterministic annealing.