Fig. 3: Results on the STORE-RECALL working memory task.
From: Efficient and robust temporal processing with neural oscillations modulated spiking neural networks
a The model architecture employed for solving the STORE-RECALL task. It consists of four groups of encoding neurons that convert input signals into spike trains, which are then processed by either a Rhythm-SRNN or a non-Rhythm-SRNN to produce the output. b Top: Input spike trains corresponding to the four groups of encoding neurons. Each input is encoded within a 100 ms encoding time window, following a Poisson distribution with an average firing rate of 50 Hz. Middle: Output spike raster of hidden neurons. Bottom: Temporal evolution of output predictions. c Comparison of recall errors between Rhythm-ALIF and Rhythm-DEXAT and their non-Rhythm counterparts across three runs with different random seeds. Error bars indicate standard deviations. d and e Loss curves and recall errors during the training process. Solid lines represent average performance, while shaded areas indicate standard deviation across three runs with different random seeds.
