Vector Quantization of Deep Convolutional Neural Networks With Learned Codebook

In this paper, we present a novel vector quantization technique, VQLC, for the compression of convolutional neural networks. Motivated by the hypothesis that better quantization codebooks exist beyond the neighbourhood of a pretrained model, we depart from the conventional strategy in which a codebook is constructed from a pretrained model and then fine tuned. In VQLC, the quantization codebook is randomly initialized and learned together with the training of the neural network. On several popular image classification benchmarks, we demonstrate the state-of-the-art performance of VQLC and show that VQLC is particularly effective on wide networks.