[22p-E102-1~14] FS.1 Focused Session "AI Electronics"

[22p-E102-5] Multimodal Spike Coding for Memcapacitive Neuromorphic System

Reon Oshio¹, Atsushi Sawada¹, 〇Mutsumi Kimura^1,2, Renyuan Zhang¹, Yasuhiko Nakashima¹ (1.NAIST, 2.Ryukoku Univ.)

Keywords:Multimodal Spike Coding, Memcapacitive Neuromorphic System

Neuromorphic systems have been investigated to accelerate machine learning calculations and reduce power dissipation by hardware architecture of SNN. Conventional rate coding is highly compatible with analog operation in ANN, and the technology developed recently can be employed. However, the rate coding needs many spikes and has problems of long latency and large power dissipation. Here, we introduce multimodal spike coding, which utilizes not only the spike frequency but the voltage amplitude and time width. The amount of information transmitted per spike increases than the conventional rate coding. Moreover, we adopt memcapacitors and charge pump circuits for synapse elements, which realize a voltage-domain synaptic operation consuming less power during switching. We have performed elementary analysis for the proposed ideas by the models and simulators supported by VDEC. The proposed ideas can successfully demonstrate SNN working.