Error-Tolerant Quantized Neural Network Based on Non-Weighted Arithmetic

This paper describes a non-weighted number representation called unitary-weight representation and its application to neural network hardware. Compared with well-used number representations such as fixed-point ones, in which each digit has a different weight, the unitary-weight representation, which is classified as one of the non-weighted number representations, has the property that it can suppress the effect of errors in the buffer memory of the neural network on the classification accuracy of the entire network. Through simulation results using a neural network for CIFAR-10 dataset, we show that the unitary-weight representation has higher error tolerance than conventional ones with equivalent bit lengths.