Non-linear Memristive Synaptic Dynamics for Efficient Unsupervised Learning in Spiking Neural Networks

• DOI: 10.3389/fnins.2021.580909
• Publication Date: 2021-02-01T05:44:44Z
• Authors: Stefano Brivio; Denys R. B. Ly; Elisa Vianello; Sabina Spiga
• Keywords: Neuromorphic engineering; Memristor; Phase-change memory

Spiking neural networks (SNNs) are a computational tool in which the information is coded into spikes, as some parts of brain, differently from conventional (NNs) that compute over real-numbers. Therefore, SNNs can implement intelligent extraction real-time at edge data acquisition and correspond to complementary solution NNs working for cloud-computing. Both NN classes face hardware constraints due limited computing parallelism separation logic memory. Emerging memory devices, like resistive switching memories, phase change or memristive devices general strong candidates remove these hurdles applications. The well-established training procedures helped defining desiderata device dynamics implementing synaptic units. generally agreed requirements linear evolution conductance upon stimulation with train identical pulses symmetric increase decrease. Conversely, little work has been done understand main properties supporting efficient SNN operation. reason lies lack background theory their training. As consequence, have taken reference develop SNNs. In present work, we show that, CMOS/memristive SNNs, very different needs NN. System-level simulations trained classify hand-written digit images through spike timing dependent plasticity protocol performed considering various non-linear plausible dynamics. We consider bounded by artificial hard values natural toward asymptotic (soft-boundaries). quantitatively analyze impact resolution non-linearity synapses on network classification performance. Finally, demonstrate boundary enable higher performance realize best trade-off between accuracy required time. With obtained results, discuss how constitute technologically convenient development on-line