Network Architecture Search for Spiking Neural Networks using Zero Cost Proxies

Context

Spiking Neural Networks (SNNs) offer a promising alternative to traditional deep learning models and enable event-driven and energy-efficient computations. However, their training can be inefficient as their sparse and temporal nature makes optimization difficult. This increases the computational cost of architecture search and makes it a resource-intensive process.

In this work, we investigate how zero-cost proxies (ZCPs) can enable efficient evaluation of SNN architectures without the need for full training. By using ZCPs, we can estimate the potential performance of a model at minimal cost, making the search for neural architectures (NAS) much more efficient. The goal is to develop a NAS framework that integrates ZCPs to identify promising SNN architectures while reducing computational overhead.

Objectives

• To conduct research on the state of the art of zero-cost proxies for SNNs

• Develop a NAS framework for zero-cost based architecture Search for promising architectures

• Benchmarking the framework using relevant SNN datasets and existing models

Requirements

• Experience with ML frameworks such as Pytorch or TensorFlow

• Very good knowledge of Python

• Experience with SNN frameworks or JAX is an advantage