Brain-Inspired Framework for Image Classification with a New Unsupervised Matching Pursuit Encoding

Shiming Song, Chenxiang Ma, Qiang Yu

Abstract: The remarkable object recognition ability of biological systems allows individuals to have prompt and reliable responses to different stimuli. Despite many implementations, an efficient and effective one is still under exploring. Spiking neural networks (SNNs), following brain-like processing, provide a potential solution for efficient object recognition. The existing SNNs can benefit an efficient feature extraction from a temporal code, but they are vulnerable to noise, less adaptive and vitally poor in recognition accuracy. How could one make full use of the biological plausibility to improve their performance? In this paper, we propose a new temporal-based encoding method with unsupervised matching pursuit. Additionally, a unified SNN framework for image recognition is designed by integrating our encoding with recently advanced synaptic learning. We evaluate our approach on MNIST, with systematic insights into encoding capabilities, robustness to noise, learning efficiency and classification performance. The results highlight the effectiveness and efficiency of our spike-based approach. To date and the best of our knowledge, our approach achieves the best temporal-based accuracy performance. Moreover, our approach requires and consumes fewer number of neurons and spikes, making it significantly advantageous to fast and efficient computation. Our work also contributes to motivating new brain-inspired developments on image classification.