Authors: Xiang He, Dongcheng Zhao, Yang Li, Guobin Shen, Qingqun Kong, Yi Zeng

Published on: March 23, 2023

Impact Score: 8.22

Arxiv code: Arxiv:2303.13077

Summary

• What is new: The introduction of a knowledge transfer loss comprising domain alignment loss and spatio-temporal regularization, alongside a sliding training strategy for improving SNN generalization on event-based datasets.
• Why this is important: Spiking neural networks suffer from overfitting and limited performance on event-based datasets due to smaller data scales and less annotation.
• What the research proposes: Using static images to aid SNN training on event data, incorporating domain alignment and spatio-temporal regularization to solve feature distribution inconsistencies.
• Results: Improved performance on neuromorphic datasets (N-Caltech101, CEP-DVS, N-Omniglot) over state-of-the-art methods.

Technical Details

Technological frameworks used: nan

Models used: Spiking Neural Networks (SNNs)

Data used: Static images and event-based neuromorphic datasets (N-Caltech101, CEP-DVS, N-Omniglot)

Potential Impact

Neuromorphic computing markets, companies in AI and machine learning, particularly those developing or utilizing spiking neural networks.

Want to implement this idea in a business?

We have generated a startup concept here: NeuroShift.