Figure 6.

Unsupervised STDP-based pattern learning. The neuron becomes selective to P = 2 patterns. (Top) Initial state. On the left, we plotted the neuron's potential as a function of time. Colored rectangles indicate pattern presentations. Next, we plotted the two spike patterns, coloring the spikes as a function of the corresponding synaptic weights: blue for low weight (0), purple for intermediate weight, and red for high weight (1). Initial weights were uniform (here at 0.7, so the initial color is close to red). (Middle) During learning. Selectivity progressively emerges. (Bottom) After convergence. STDP has concentrated the weights on the afferents which fire at least once in at least one of the pattern subsections, located at the beginning of each pattern, and whose duration roughly matches the optimal Δt (shown in green). This results in one postsynaptic spike each time either one of the two pattern is presented. Elsewhere both ${\overline{V}}_{\text{noise}}$ and σ_noise are low, so the SNR is high. In addition ${\overline{V}}_{\text{noise}}$ roughly matches the theoretical value ${\overline{V}}_{\text{noise}}^{\text{opt}}$ (shown in green), corresponding to the optimal SNR. We also show in green $V_{\text{max}}^{\text{opt}}$, the theoretical optimal value for V_max. However, the potential never reaches it, because the adaptive threshold is reached before.