Towards a Scalable Neuromorphic Hardware for Classification and Prediction with Stochastic No-Prop Algorithms

Dan Christiani, Cory Merkel, Dhireesha Kudithipudi
Rochester Institute of Technology


Stochastic logic offers significant area efficiency when applied to high-density redundant neural architectures; while noisy chaotic fluctuations associated with stochastic learning systems have been proven to reduce overfitting and escape local minima. Feed forward networks with random weights are uniquely qualified for stochastic logic due to their static abstraction layer. In this research, we explore neuromorphic architectures for the emerging No Propagation (No-Prop) class of algorithms. Specifically, a scalable three layer No-Prop artificial neural network is designed using single-line bipolar stochastic digital logic. The network is defined in VHDL and remodeled in C# for behavioral testing. Preliminary results show that the network has the ability to generalize well, achieving 100% test accuracy for binary signal classification, and 64% test accuracy for five class hyper-spectral classification. Further, a discussion that highlights the need for more sophisticated behavioral models which efficiently simulate higher precision stochastic logic systems is also presented.