동적 시간 지연 인코딩 기반 멀티스케일 스파이크 신경망

Abstract

Spiking Neural Networks (SNNs), inspired by the structure and efficiency of biological neural systems, are emerging as a powerful alternative to traditional artificial neural networks, especially in low-power and real-time processing applications. SNNs process information through discrete spikes, mimicking the temporal dynamics of bio- logical neurons, which results in improved energy efficiency and computational effectiveness. This study proposes a novel dynamic temporal latency encoding method that adjusts spike timing based on input intensity through a tun- able hyperparameter, allowing for enhanced accuracy and adaptability. The method was evaluated on MNIST and TIDIGITS datasets, achieving remarkable accuracy rates of 99.12% and 100%, respectively, surpassing traditional encoding approaches. Additionally, experiments conducted on the Caltech101 dataset, which includes images of vary- ing resolutions, demonstrated the robustness and generalizability of the proposed encoding strategy. By optimizing the temporal characteristics of spike generation, this method effectively balances computational efficiency and accu- racy, making it particularly suitable for neuromorphic applications that demand real-time, high-accuracy performance. These findings underscore the potential of dynamic temporal latency encoding in advancing the capabilities of SNNs for practical, energy-efficient tasks in the rapidly evolving field of neuromorphic computing.