Bioinspired networks with nanoscale memristive devices that combine the unsupervised and supervised learning approaches

This work proposes two learning architectures based on memristive nanodevices. First, we present an unsupervised architecture that is capable of discerning characteristic features in unlabeled inputs. The memristive nanodevices are used as synapses and learn thanks to simple voltage pulses which implement a simplified "Spike Timing Dependent Plasticity" rule. With system simulation, the efficiency of this scheme is evidenced in terms of recognition rate on the textbook case of character recognition. Simulations also show its extreme robustness to device variations. Second, we present a supervised architecture that can learn if the classification of every input is given. Simulations prove its efficiency. A good robustness to device variation is seen, but not to the level of the unsupervised approach. Finally, we show that both approaches can be combined, with variation robustness higher than in the supervised case. This opens important prospects, like the possibility to first train the system in an unsupervised way with unlabeled data, while still benefiting of the simplicity to program a supervised system.