Modulating STDP with Vectorized Backpropagation: A New Paradigm for Real-time Audio Prediction in Spiking Neural Networks

Abstract

Spiking Neural Networks (SNNs) are inspired by biological neural systems in the human brain. Theoretically, they were proven to consume less power than those non-spiking conventional neural networks due to their sparse spike activities in information transfer and processing. These make the SNNs crucial in state-of-the-art low-power applications widely present in both edge computing and mobile devices. In this thesis, a novel modulating Spike-Timing-Dependent-Plasticity (STDP) learning rule is proposed for multilayer SNNs. The global error feedback strategy has been used to optimize the performance of the proposed learning algorithm. Then, this learning rule is modified with vectorized backpropagation, therefore enabling continuous online learning, hence allowing real-time processing of data. This is extended to a Spiking Long-Short-Term-Memory (S-LSTM) network performing audio prediction on speech data. This enables the extension of SNNs to real-world applications where energy efficiency and real-time processing are imperative. Demonstrating competitive performance from this SNN framework in tasks such as S-LSTM audio predictions points toward outstanding computational efficiency improvements over conventional deep learning models. These strategies are significant pointers to the fact that SNNs be a promising approach for efficient, on-device learning across a wide spectrum of applications in Internet of Things (IoT) deployments.