10:30 a.m. MEC, room 114. John Chiasson of Boise State University, will be presenting a seminar “Deep Learning via Spike Timing Dependent Plasticity”. John Chiasson received his B.S in mathematics from the University of Minnesota, M.S. in electrical engineering from Washington State University, and Ph.D. in control sciences from the University of Minnesota. He has published three textbooks, “Modeling and High-Performance Control of Electric Machines,” “An Introduction to Probability Theory and Stochastic Processes,” and “An Introduction to System Modeling and Control.”
Deep learning refers to having multiple layers in a neural network (NN). Currently, Deep Convolutional Neural Networks (CNN) are used due to their success in object recognition, speech recognition, etc. These networks are trained using the computationally intensive back propagation algorithm, which is an optimization method to (hopefully) find the global minimum based on gradient descent. In the 1990s the phenomena of spiking timing dependent plasticity (STDP) was discovered and studied by neuro-scientists. It is inherently a simple computational tool which has led to the consideration of STDP as an alternative method in training Convolutional Neural Networks. Specifically, STDP is an unsupervised learning approach where all the learning (weight updating) is done locally. That is, the weight (synapse) update between two neurons depends only the spike coming from out of the first neuron into the second neuron and any spike coming out of the second neuron. The literature reports that deep spiking networks can recognize the MNIST digits 0 through 9 within 1% of the accuracy of the best conventional CNN network. Further, these networks have the promise of being able to implement them in analog electronics with very low power requirements.
For more information, please visit coen.boisestate.edu/ece/seminars/.