DeepCABAC: Context-adaptive binary arithmetic coding for deep neural\n network compression

Abstract

We present DeepCABAC, a novel context-adaptive binary arithmetic coder for\ncompressing deep neural networks. It quantizes each weight parameter by\nminimizing a weighted rate-distortion function, which implicitly takes the\nimpact of quantization on to the accuracy of the network into account.\nSubsequently, it compresses the quantized values into a bitstream\nrepresentation with minimal redundancies. We show that DeepCABAC is able to\nreach very high compression ratios across a wide set of different network\narchitectures and datasets. For instance, we are able to compress by x63.6 the\nVGG16 ImageNet model with no loss of accuracy, thus being able to represent the\nentire network with merely 8.7MB.\n