Figure 3

Memory maintenance characteristics of learning with sigmoidal stochastic S-STDP. (a) Memory maintenance curves obtained with sigmoidal stochastic S-STDP in the cases of \(k= 2\) and \(3\) (with \(p=0.13\) and \(q=0.03\)) and \(k=4\) (with \(p=0.2\) and \(q=0.08\)). For comparison, those obtained with conventional (\(k=1\)) and deterministic S-STDP are shown (taken from Fig. 1b). Furthermore, the memory maintenance curves in the case where the sigmoidal rule of \(k=3\) with \(p=0.13\) is applied only for potentiation (depression is performed with the conventional rule with \(q=0.008\)), and in the opposite case (with \(p=0.04\) and \(q=0.03\)) are also plotted. Colour intensity maps of synaptic weights afferent to \(5\times 5=25\) excitatory neurons out of 400 in the second layer are shown after the initial training and 50,000 and 100,000 additional trainings. Each pixel in the map correspond to a weight. Thus, each panel contains \(28\times 28\times 25=\mathrm{19,600}\) pixels. White and red indicate 0 and 1, respectively. The intermediate values for continuous weights are represented by yellowish colours. (b) Colour intensity maps showing the results of the benchmark test of memory maintenance (described in the text).