Presentation Information
[20a-A303-7]Improvement in Inference Accuracy in Non-ideal Situations of Event-based Vision Sensor and Computation-in-Memory by Modified Training Algorithms
〇(M2)Yinghao Sun1, Kazuhide Higuchi1, Chihiro Matsui1, Ken Takeuchi1 (1.Univ. Tokyo)
Keywords:
Event-based Sensor,Computation in Memory,ConvLSTM
EVS has been catching attention in the field of computer vision for its merits like high dynamic range, high temporal resolution, and low power consumption. Such characteristics make it suitable for in-sensor computing applications such as neural network CiM. However, both EVS and CiM suffers from non-idealities that greatly reduces the inference accuracy of the neural network. Such non-idealities include background noise events for EVS and device non-ideality that leads to bit-inversion error in CiM. Although there exist post-processing units that could mitigate these non-idealities, like noise filters for EVS, they often require extra hardware units. This work proposes a modified training method to mitigate the reduction in inference accuracy caused by the two non-idealities in ReRAM-based Tiny ConvLSTM CiM. The proposed method involves adding artificial noise-events that follow the distribution of real noise events in EVS to the training dataset. The method has been proved to enhance the inference accuracy of the CiM model under non-ideal situations.