Robust Model Mapping Optimization For Non-Ideal Computing In-Memory

Summary form only given, as follows. A complete record of the panel discussion was not made available for publication as part of the conference proceedings. All binary computing-in-memory (CIM) designs for binary neural network (BNN) acceleration becomes popular due to lower cost and similar performance. However, it suffers from practical non-ideal peripheral circuit effect, which degrades the model performance significantly. To tackle above non-ideal effect, this paper proposes a robust model mapping method for BNN with the decomposed calibration for the convolution part and a gradient descent based optimization mapping for batch normalization (BN) part. The proposed approach can effectively restore the performance without extra hardware cost for computer vision tasks such as classification tasks on CIFAR-10, tiny ImageNet and object detection tasks.