Mapping-Aware Spiking Neural Architecture Search

Description

Spiking Neural Networks (SNNs) are bio-inspired neural networks that promise higher energy efficiency compared to traditional Artificial Neural Networks (ANNs), though often at the cost of lower task performance While Neural Architecture Search (NAS) typically optimizes SNN topologies solely for accuracy, deploying them on resource-constrained hardware requires joint consideration of hardware mapping strategies. Current Hardware-Aware NAS approaches often neglect critical compiler and execution metrics (e.g., data reuse, memory tiling).

Targets

This thesis aims to expand a Mapping-Aware Spiking NAS framework that directly integrates hardware mapping strategies and compiler metrics into the architecture search loop.
Depending on the scope, focus can be placed on:

• Search Space: Adapting the SNN search space toward hardware- and compiler-friendly mapping patterns.
• Benchmarking: Evaluating across diverse datasets (e.g., event-based data) and integration of SpikeNAS-Bench
• Mapping Strategies: Systematically comparing mapping approaches.

Requirements

• Good Knowledge of Python and ML-Frameworks (PyTorch, Optuna)
• Good Knowledge of Neural Networks