Maximizing Efficiency of SNN-Based Reservoir Computing via NoC-Assisted Dimensionality Reduction

Spiking Recurrent Neural Networks (SRNNs) with online learning capabilities are sought after because they offer the ability to retrain and adapt to changing input patterns. However, achieving online learning in SRNN s presents challenges due to the intricate evolution of reservoir states and the interdependencies among errors in recurrent connections. Reservoir Computing (RC) streamlines this process by focusing solely on training the feedforward output layer, avoiding complexities associated with training of recurrent connections. Despite recent advancements, implementing online learning using Reservoir Computing (RC) still remains resource-intensive. To address this, we propose employing a circuit-switched NoC for routing spikes which can be leveraged in conjunction with dimensionality reduction technique to achieve notable resource savings. The RC architecture is designed for nonlinear channel equalization application and based on post-layout simulation results, reductions of approximately 30 % in area and 50 % in power consumption are achieved by reducing dimensions by 0.6 x, with no statistically significant impact on accuracy. Additionally, a Symbol Error Rate (SER) of 0.0027, comparable to state-of-the-art, was observed for a small reservoir size of 32. Post-layout simulations validate the proposed architecture's efficiency and potential for hardware implementations with online learning capabilities.