Speaker: Adam Charles, Georgia Tech

Title: Remember the Inputs—How a small group of nodes banded together to extend their short-term memory

Abstract:
Random neural networks have been observed to have astounding computational benefits when used as pre-processing for prediction and classification tasks. These networks are typically constructed as randomly connected systems which are driven by a streaming input. The network then accumulates this incoming information, and the network state can be read out and used in classification, prediction and estimation. In this presentation, I'll discuss how we can begin to quantify the computational power of these random networks by studying a proxy for computational power: the length of the short-term memory for such systems. Specifically I will show how tools from the compressive sensing literature can yield strong bounds on the short-term memory both for single input systems as well as multi-input systems.

Speaker Bio:
Adam Charles is currently a graduate student at CSIP at the Georgia Institute of Technology. Adam currently works with Dr. Christopher Rozell and is particularly fond of research in statistical signal processing, applications to remote sensing and neuroscience. Adam is a veteran of the Cooper Union ECE program and enjoys random walks on the beach.