Figure 3. Learning in a WTA network with a mixture of software and memristor synapses.

(a) Diagram of the 2-neuron, WTA network used in this work. (b) Evolution of neuron specializations S_i to patterns 0110 and 1001 as weights change over successive events, illustrating the interplay between the two neurons. Inset: close-up of first 60 trials. (c) Computed membrane potentials of each neuron to both prototype patterns according to their weights at every trial illustrating the intrinsic pattern preferences of each neuron, that is independent of their interaction in the WTA network. (d) Evolution of hardware (synapses 0–3, enclosed in thick, black frame) and software (synapses 4–7) weights. (e,f) Responses of the WTA network to the initial (e) and final (f) 41 input samples. The fire count of both the hardware synapse neuron (orange) and the software synapse neuron (turquoise) is shown for patterns 0110 and 1001, and patterns that differ from these prototypes in one position (0110_δ and 1001_δ). The different pattern groups are perfectly segregated by the end of the run.