Otters++: A Time-to-first-spike Based Energy Efficient Optical Spiking Transformer

Otters++ introduces a novel spiking neural network (SNN) architecture that leverages time-to-first-spike (TTFS) coding for energy-efficient inference, utilizing the natural signal decay in optoelectronic devices to compute the TTFS temporal term. The model establishes a layer-wise equivalence with quantized neural networks (QNNs) and employs a hybrid training method that combines SNN computation for the forward pass and QNN gradients for the backward pass, effectively addressing challenges like over-sparsity in training. Achieving an average score of 84.17% on the GLUE dataset, Otters++ demonstrates significant energy efficiency compared to previous spiking Transformer models, highlighting its potential for practical applications in energy-constrained environments.