[F-7] In-Memory and Unconventional Computing

[F-7-02] Tiny and Error Torrent Convolutional LSTM for Event-based Vision Sensor with ReRAM Computation-in-Memory

〇Tomoki Kobayashi¹, Kazuhide Higuchi¹, Naoko Misawa¹, Chihiro Matsui¹, Ken Takeuchi¹ (1. Univ. of Tokyo (Japan))

Presentation style: Online

https://doi.org/10.7567/SSDM.2022.F-7-02

This paper proposes tiny and error tolerant Convolu-tional LSTM (ConvLSTM) for event-based vision sensor (EVS) data processing with ReRAM Computa-tion-in-Memory (CiM). To realize tiny ConvLSTM CiM, reduction of the bit precision of each ConvLSTM layer and weight is considered. On the other hand, reducing the bit precision degrades inference accuracy and error tolerance. ReRAM has low switching current and high scalability, but ReRAM has non-ideality such as conductance variation due to program/read disturb. Thus, the bit precision of the weight is optimized to accept bit error rate of 0.1%. Therefore, the proposed tiny ConvLSTM can reduce the size of memory by 91% compared with conventional ConvLSTM.