Abstract: The juxtaposition of pattern representations is reconfigured at each layer of a multi-layered perceptron. As the activity propagates through the network, the representations are transformed through a systematic progression such that the representation at the penultimate layer is computable at the final stage (e.g., linearly separability for a classification task). Simple networks trained on simple tasks show a tendency to collapse the representational space to a low dimensional manifold in the intermediate layers.
The similarity structure among the representations can be directly visualized at each layer if there are just 2 or 3 units.

Bio: Dr. Munro has been conducting research in the area of neural plasticity and network learning rules. He received his PhD in Physics from Brown University (1983) working on early neural models of the mammalian visual system. He then went on for postdoc positions at University of Trondheim, Norway, in a neurophysiology lab and then at UC San Diego, at the Institute for Cognitive Science, as part of the Parallel Distributed Processing research group. He joined the Pitt faculty in 1986 with a primary appointment in the School of Computing and Information, and has a secondary appointment in the Center for Neuroscience. Dr. Munro is also a member of the Center for the Neural Basis of Cognition, which spans relevant faculty across many departments, both at Pitt and CMU .

Learn more at https://www.sci.pitt.edu/people/paul-munro