And the Bit Goes Down: Revisiting the Quantization of Neural Networks
This addresses memory efficiency for deploying neural networks on resource-constrained devices like CPUs, representing a strong incremental improvement in quantization techniques.
The paper tackles the problem of reducing memory footprint in convolutional networks by introducing a vector quantization method that preserves output quality, achieving a 20x compression on ResNet-50 to 5MB with 76.1% top-1 accuracy on ImageNet and 26x compression on Mask R-CNN.
In this paper, we address the problem of reducing the memory footprint of convolutional network architectures. We introduce a vector quantization method that aims at preserving the quality of the reconstruction of the network outputs rather than its weights. The principle of our approach is that it minimizes the loss reconstruction error for in-domain inputs. Our method only requires a set of unlabelled data at quantization time and allows for efficient inference on CPU by using byte-aligned codebooks to store the compressed weights. We validate our approach by quantizing a high performing ResNet-50 model to a memory size of 5MB (20x compression factor) while preserving a top-1 accuracy of 76.1% on ImageNet object classification and by compressing a Mask R-CNN with a 26x factor.