Supervised Learning Algorithm for Deep Spiking Neural Networks Based on Instantaneous Error

In the development of neural networks, BPTT updates the weights by decomposing the time series into consecutive time steps and using a standard backpropagation strategy. Although this approach works well in some cases, it does not fully take into account the importance of temporal localization. To surpass this constraint, this paper innovatively introduces Multi_INSER, a multilayer spiking neural network supervised learning algorithm based on instantaneous error computation. The algorithm breaks through the constraints of traditional algorithms by focusing on the prediction error at the current moment only, and by accurately measuring the discrepancy between the expected and the real result. The gradient descent method is used for immediate weight adjustment, which realizes an efficient optimization strategy based on the instantaneous error gradient. Through a series of spike sequence learning experiments, this paper not only verifies the effectiveness of the Multi_INSER algorithm, but also deeply discusses the key factors affecting its learning performance. The results demonstrate that the algorithm is not only equally matched in terms of performance, but also shows significant advantages in time efficiency.