Convolutional Spiking Neural Networks for Spatio-Temporal Feature Extraction
Artikel i vetenskaplig tidskrift, 2023

Spiking neural networks (SNNs) can be used in low-power and embedded systems e.g. neuromorphic chips due to their event-based nature. They preserve conventional artificial neural networks (ANNs) properties with lower computation and memory costs. The temporal coding in layers of convolutional SNNs has not yet been studied. In this paper, we exploit the spatio-temporal feature extraction property of convolutional SNNs. Based on our analysis, we have shown that the shallow convolutional SNN outperforms spatio-temporal feature extractor methods such as C3D, ConvLstm, and cascaded Conv and LSTM. Furthermore, we present a new deep spiking architecture to tackle real-world classification and activity recognition tasks. This model is trained with our proposed hybrid training method. The proposed architecture achieved superior performance compared to other SNN methods on NMNIST (99.6%), DVS-CIFAR10 (69.2%), and DVS-Gesture (96.7%). Also, it achieves comparable results compared to ANN methods on UCF-101 (42.1%) and HMDB-51 (21.5%) datasets.



Feature extractor



Ali Samadzadeh

Amirkabir University of Technology

Fatemeh Sadat Tabatabaei Far

Amirkabir University of Technology

Ali Javadi

Amirkabir University of Technology

Ahmad Nickabadi

Amirkabir University of Technology

Morteza Haghir Chehreghani

Chalmers, Data- och informationsteknik, Data Science och AI

Neural Processing Letters

1370-4621 (ISSN) 1573-773X (eISSN)

Vol. 55 6 6979-6995


Sannolikhetsteori och statistik

Datavetenskap (datalogi)

Datorseende och robotik (autonoma system)



Mer information

Senast uppdaterat