Neuromorphic Computing and Engineering (Jan 2024)

Time–frequency analysis using spiking neural network

  • Moshe Bensimon,
  • Yakir Hadad,
  • Yehuda Ben-Shimol,
  • Shlomo Greenberg

DOI
https://doi.org/10.1088/2634-4386/ad80bc
Journal volume & issue
Vol. 4, no. 4
p. 044001

Abstract

Read online

Time–frequency analysis plays a crucial role in various fields, including signal processing and feature extraction. In this article, we propose an alternative and innovative method for time–frequency analysis using a biologically inspired spiking neural network (SNN), encompassing both a specific spike-continuous-time-neuron-based neural architecture and an adaptive learning rule. We aim to efficiently detect frequencies embedded in a given signal for the purpose of feature extraction. To achieve this, we suggest using an SN-based network functioning as a resonator for the detection of specific frequencies. We developed a modified supervised spike timing-dependent plasticity learning rule to effectively adjust the network parameters. Unlike traditional methods for time–frequency analysis, our approach obviates the need to segment the signal into several frames, resulting in a streamlined and more effective frequency analysis process. Simulation results demonstrate the efficiency of the proposed method, showcasing its ability to detect frequencies and generate a Spikegram akin to the fast Fourier transform (FFT) based spectrogram. The proposed approach is applied to analyzing EEG signals, demonstrating an accurate correlation to the equivalent FFT transform. Results show a success rate of 94.3% in classifying EEG signals.

Keywords