Extremely Low Bit Neural Network: Squeeze the Last Bit Out with ADMM

Although deep learning models are highly effective for various tasks, such as detection and classification, the high computational cost prohibits the deployment in scenarios where either memory or computational resources are limited. In this paper, we focus on model compression and acceleration of deep models. We model a low bit quantized neural network as a constrained optimization problem. Then, with the use of ADMM, we decouple the discrete constraint and parameters of network. We also show how the resulting subproblems can be efficiently solved with extragradient and iterative quantization. The effectiveness of the proposed method has been demonstrated in extensive experiments on convolutional neural network for image recognition, object detection, and recurrent neural network for language model.