https://github.com/eve-ning/snn_voice

SNN Hybrid Model for Voice Recognition
https://github.com/eve-ning/snn_voice

Last synced: 13 days ago
JSON representation

SNN Hybrid Model for Voice Recognition

• Host: GitHub
• URL: https://github.com/eve-ning/snn_voice
• Owner: Eve-ning
• Created: 2022-08-28T14:04:56.000Z (over 2 years ago)
• Default Branch: master
• Last Pushed: 2023-04-12T08:29:26.000Z (over 1 year ago)
• Last Synced: 2024-10-27T18:24:49.661Z (2 months ago)
• Language: Jupyter Notebook
• Size: 9.58 MB
• Stars: 0
• Watchers: 3
• Forks: 0
• Open Issues: 1
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
[![Python package](https://github.com/Eve-ning/snn_voice/actions/workflows/python-package.yml/badge.svg)](https://github.com/Eve-ning/snn_voice/actions/workflows/python-package.yml)

# Retraining SNN Conversions: CNN to SNN for Audio Classification Tasks

This is the repository for the paper titled in the heading, submitted for my NTU FYP

## Abstract

Efficient yet powerful models are in high demand for its portability and affordability.

Amongst other methods such as model-pruning, is limiting neural network operations to

sparse event-driven spikes: Spiking Neural Networks (SNNs) aims to unravel a new di-

rection in machine learning research. A significant amount of SNN literature straddles

upon mature works of artificial neural networks (ANNs) by migrating its architecture

and parameters into SNNs, optimizing the migration to retain as much performance as

possible. We spearhead a novel approach: the architecture is migrated and retrained

from scratch. We hypothesize that this new direction will unravel concepts that cur-

rently bottlenecks improvements in the field of SNN conversions. Furthermore, alike

Transfer Learning, inspire future efforts of fine-tuning a well converted model through

training.

This paper presents our analysis of training converted Convolutional Neural Networks

(CNNs) to SNNs on audio classification models. Results show that (1) SNN conver-

sions consistently underperforms CNNs marginally during training, however we also

show that model complexity has a possible association with this margin. (2) SNN con-

verts doesn’t necessarily approach the performance of its CNN counterparts asymptot-

ically by increasing the number of time-steps. (3) SNN training from scratch is costly

and impractical with current hardware and dedicated SNN optimization techniques are

necessary. (4) Enabling the SNN membrane decay rate to be learned doesn’t signifi-

cantly affect performance. This paper provides valuable insights into the perspective of

retraining converted SNNs for audio classification, and serves as a reference for future

studies and hardware implementation benchmarks.