Biologically plausible deep learning for recurrent spiking neural networks.
Seminar | December 19 | 12-1:30 p.m. | 560 Evans Hall
Despite widespread success in deep learning, backpropagation has been criticized for its biological implausibility. To address this issue, Hinton and Bengio have suggested that our brains are performing approximations of backpropagation, and some of their proposed models seem promising. In the same vein, we propose a different model for learning in recurrent neural networks (RNNs), known as McCulloch-Pitts processes. As opposed to traditional models for RNNs (such as LSTMs) which are based on continuous-valued neurons operating in discrete time, our model consists of discrete-valued (spiking) neurons operating in continuous time. Through our model, we are able to derive extremely simple and local learning rules, which directly explain experimental results in Spike-Timing-Dependent Plasticity (STDP).