4:15 PM - 4:30 PM
[K-8-04] Noise Event Injection Training to Mitigate Inference Accuracy Degradation due to Non-Idealities of Event-based Vision Sensor and Computation-in-Memory
This paper proposes artificial noise injection training for Computation-in-Memory (CiM) to mitigate inference accuracy degradation due to non-idealities of event-based vision sensor (EVS) and CiM for ConvLSTM. EVS intrinsically outputs undesirable event data as noise events even under constant illumination. In addition, CiM has non-ideality that causes bit error. Those non-idealities degrade inference accuracy of neural networks. To mitigate the degradation of inference accuracy, the proposed method employs artificial noise events injected to training dataset. The simulation results reveal the inference accuracy is improved by 59.9 % compared to without noise event injection training at 0.5% bit-error-rate (BER) of CiM.