Noise-assisted persistence and recovery of memory state in a memristive spiking neuromorphic network
- Authors: Surazhevsky I.A.; Demin V.A.; Ilyasov A.I.; Emelyanov A.V.; Nikiruy K.E.; Rylkov V.V.; Shchanikov S.A.; Bordanov I.A.; Gerasimova S.A.; Guseinov D.V.; Malekhonova N.V.; Pavlov D.A.; Belov A.I.; Mikhaylov A.N.; Kazantsev V.B.; Valenti D.; Spagnolo B.; Kovalchuk M.V.
- Publication year: 2021
- Type: Articolo in rivista
- OA Link: http://hdl.handle.net/10447/509843
Abstract
We investigate the constructive role of an external noise signal, in the form of a low-rate Poisson sequence of pulses supplied to all inputs of a spiking neural network, consisting in maintaining for a long time or even recovering a memory trace (engram) of the image without its direct renewal (or rewriting). In particular, this unique dynamic property is demonstrated in a single-layer spiking neural network consisting of simple integrate-and-fire neurons and memristive synaptic weights. This is carried out by preserving and even fine-tuning the conductance values of memristors in terms of dynamic plasticity, specifically spike-timing-dependent plasticity-type, driven by overlapping pre- and postsynaptic voltage spikes. It has been shown that the weights can be to a certain extent unreliable, due to such characteristics as the limited retention time of resistive state or the variation of switching voltages. Such a noise-assisted persistence of memory, on one hand, could be a prototypical mechanism in a biological nervous system and, on the other hand, brings one step closer to the possibility of building reliable spiking neural networks composed of unreliable analog elements.