IBM's TrueNorth does the same, it is a common practice in neural network chips for inference. I think doing that kind of tricks for training is much more valuab

IBM's TrueNorth does the same, it is a common practice in neural network chips for inference. I think doing that kind of tricks for training is much more valuable. Doing training at hundreds of GPUs and inference with chips at mobile, can only be a temporary solution for deep learning market.