CLIF-Adjoint: Accelerating Training of CLIF with Adjoint Method

Surrogate gradients offer an effective approach for training spiking neural networks (SNNs), allowing them to achieve classification results comparable to those of artificial neural networks. Recent research has highlighted the enhanced performance of spiking neurons, particularly CLIF and similar types. However, the high cost of training these new neurons with surrogate gradient methods limits their efficiency. In this work, we introduce CLIF-adjoint neurons and utilize the adjoint method to speed up the training process. We measured the CLIF-adjoint neurons on multiple classification tasks, including static datasets and neuromorphic datasets. Our experiments demonstrated that using CLIF-adjoint neurons can significantly reduce the training time of SNNs while achieving classification accuracy that matches the original neurons. Additionally, with time steps 20, the training speed was improved by 2.05 times with CLIF-adjoint neurons.