Computing with Spikes: Legendre-based SNNs and Deep Local Learning toward Neuromorphic Intelligence

Abstract

The pursuit of Green AI has brought Neuromorphic Computing with Spiking Neural Networks (SNNs) to the forefront as an energy-efficient alternative to conventional Artificial Neural Networks (ANNs). This dissertation advances Neuromorphic Computing along four interconnected directions: the algorithmic design of reservoir-based SNNs, their deployment on Intel's Loihi-2 neuromorphic chip, a biologically-inspired local-learning method to train deep SNNs, and real-world neuromorphic applications spanning wireless domain and event-based vision. We begin by proposing a family of Legendre Delay Network (LDN) based SNNs for Time Series Classification (TSC): the Spiking Reservoir Computing (SRC) model, the Legendre-SNN (LSNN), and the Deep Legendre-SNN (DeepLSNN). These models use the LDN as a static reservoir to extract temporal features, followed by a trainable spiking network. Herein, we present the first spiking-TSC benchmark by evaluating the DeepLSNN on 102 TSC datasets. We then address the non-trivial challenge of deploying these heterogeneous models entirely on Loihi-2 - by implementing the quantized LDN on the chip's embedded Lakemont (LMT) cores and the spiking network on its Neuro-Cores; thus, this work contributes scarce technical documentation to program the LMT cores. We next propose "DALTON" – a Three-Factor-Rule-based local-learning method that effectively leverages the depth of Convolutional SNNs. Finally, we present two real-world neuromorphic applications on Loihi-2: 5G jamming detection and drone-recognition using our newly introduced related datasets. Across our works, we find that our neuromorphic methods are not only highly energy-efficient and low-latency, but also help close the performance gap between SNNs and ANNs, thereby paving the future of neuromorphic intelligence.