Analyzing frameworks and strategies for converting neural networks for neuromorphic processors

Abstract

Neuromorphic computing employs innovative algorithms to mimic how the human brain interacts with the world, aiming to achieve capabilities that closely resemble human cognition. Neuromorphic processors are based on an entirely new computing paradigm and they come with new Machine Learning (ML) algorithms. Programming a neuromorphic processor often entails creating a Spiking Neural Network (SNN) that closely mimics the biological neural networks and can be deployed to the neuromorphic processor. Neuromorphic processors leverage these asynchronous, event-based SNNs to achieve substantial increase in power and performance over conventional architectures. However, training such networks is difficult due to the non differentiable nature of spike events. This thesis investigates the frameworks and strategies for converting standard neural networks, particularly Deep Learning (DL) models, to SNNs suitable for neuromorphic processors. Focusing on three neuromorphic platforms-Brainchip Akida, Intel Loihi, and SynSense-the thesis aims to develop a standardized conversion pipeline, address current limitations, and conduct metric-based analyses of the model developed using different frameworks. The thesis utilizes the PilotNet model, a Convolutional Neural Network (CNN) designed for autonomous driving, to evaluate the conversion processes and performance on the selected neuromorphic framework. Results demonstrate the varying degrees of efficiency and challenges associated with each neuromorphic processor, providing insights for optimizing the conversion process and further advancing neuromorphic computing for practical applications such as autonomous driving, robotics, and edge computing. The findings emphasize the need for continued development in conversion techniques and the optimization of neuromorphic hardware to fully harness the potential of AI-driven systems.