Deep Spiking Neural Networks Driven by Adaptive Interval Membrane Potential for Temporal Credit Assignment Problem

One of the core challenges in biological and machine learning is how neural networks can discover the predictive features hidden in distracting streams of unrelated sensory activity via delayed feedback, commonly known as the "temporal credit-assignment (TCA) problem". By matching the output spikes of neurons to the number of features, the aggregated-label (AL) learning is a promising approach to target this problem. However, existing mainstream AL learning algorithms either only focus on a few information of features or consider the imprecise temporal loss. In this paper, we propose a novel adaptive interval membrane potential driven aggregated-label learning algorithm, named IMPD-AL. This algorithm not only captures the adaptive interval temporal information but also integrates the loss of all spike firing times in the precise interval most relevant to features. Experimental results on speech recognition (TIDIGITS), dynamic digit recognition (N-MNIST), dynamic image recognition (CIFAR10-DVS), and dynamic gesture recognition (DVS128 Gesture) show that the IMPD-AL algorithm outperforms the state-of-the-art (SOTA) AL algorithm in SNNs on both classification task and TCA task.