Visualizing Loss Landscapes for Neuromorphic Keyword Spotting

Context

Spiking Neural Networks (SNNs), often described as the third generation of artificial neural networks (ANNs), offer the promise of high energy efficiency and increased biological plausibility. However, despite their potential, SNNs often underperform compared to conventional ANNs, especially in terms of accuracy and training stability. In recent years, gradient-based training methods, such as surrogate gradients and Backpropagation Through Time (BPTT), have become more common for training SNNs. While these techniques have enabled progress, the underlying loss landscapes that govern optimization in SNNs remain poorly understood.

This thesis aims to investigate and visualize the loss landscape of SNNs to better understand their optimization challenges. By analyzing how specific properties shape the loss surface (e.g. the choice of surrogate loss), the thesis seeks to uncover factors that influence trainability and performance. The work will involve training SNNs under controlled conditions and visualizing their loss landscapes using methods adapted from conventional deep learning. A particular focus will be placed on the Spiking Heidelberg Digits (SHD) dataset as a benchmark for performance evaluation.

Targets

• Visualize the loss landscape of SNNs under different training and architectural conditions

• Analyze the impact of properties like connectivity and surrogate gradients on loss geometry

• Train SNNs on the Spiking Heidelberg Digits dataset and evaluate how loss landscape characteristics correlate with performance

Requirements

• Experience with ML frameworks like JAX or PyTorch

• High Python proficiency