LIF-based Legendre Memory Unit: neuromorphic redesign of a recurrent architecture and its application to human activity recognition

Nowadays traditional artificial neural networks (ANNs) present a significant challenge in edge devices and embedded systems due to their high power consumption. This inefficiency in standard machine learning (ML) and deep learning (DL) solutions requires the exploration of more efficient alternatives. Spiking neural networks (SNNs), inspired by the interconnections of human brain neurons, constitute a valid alternative to address this challenge. However, the advancement of these brain-inspired models is constrained by the limited accessibility to dedicated neuromorphic hardware capable of integrating brain-inspired computational principles. Despite this low hardware availability, SNNs represent a step forward in achieving more efficient, brain-inspired computation for edge devices and embedded systems. The goal of neuromorphic computing is to emulate the intricating complexities that have evolved over millions of years to fine-tune neural processing. Driving inspiration from the behavior of biological neurons, neuromorphic models trace a unique path to computational efficiency and adaptability, making a significant break from the rigid, power-hungry architectures of conventional computing. Instead, they adopt a more flexible, energy-efficient, and inherently parallel paradigm. Unlike the traditional neural network, these brain-inspired models activate neurons only in response to certain events. This event-based approach simulates brain information processing. As a result, this leads to a more sparse neuron activation, decreasing the amount of energy consumption. The core part of this work is transforming a recursive cell architecture, the Legendre Memory Unit (LMU), into its neuromorphic version, named the LIF-based Legendre Memory Unit (L2MU). This architecture redesign reconfigures the internal states of the original architecture into populations of neurons that are interconnected through synapses, allowing components to communicate via synaptic currents and facilitating information flow through neuronal spikes in response to changes in current and voltage. As a benchmarking stage for the L2MU architecture, a comprehensive comparative analysis is presented focusing on the task of human activity recognition (HAR); including various alternatives built from the L2MU cell and other network architectures both in the neuromorphic field and in the non-spiking domain. This also includes an analysis of model compression techniques to keep spiking neural networks lightweight and efficient for low-power environments. The Thesis hence offers a comparative study with classical artificial neural networks, highlighting the advantages and trade-offs in terms of computational efficiency, energy consumption, and processing speed. Real-world scenarios are also considered, demonstrating the potential of L2MU for various edge computing applications such as real-time data processing.