On the Pareto Efficiency of Quantized CNN

Ting-Wu Chin, Pierce I-Jen Chuang, Vikas Chandra, Diana Marculescu

Abstract: Weight Quantization for deep convolutional neural networks (CNNs) has shown promising results in compressing and accelerating CNN-powered applications such as semantic segmentation, gesture recognition, and scene understanding. Prior art has shown that different datasets, tasks, and network architectures admit different iso-accurate precision values, which increase the complexity of efficient quantized neural network implementations from both hardware and software perspectives. In this work, we show that when the number of channels is allowed to vary in an iso-model size scenario, lower precision values Pareto dominate higher precision ones (in accuracy vs. model size) for networks with standard convolutions. Relying on comprehensive empirical analyses, we find that the Pareto optimal precision value of a convolution layer depends on the number of input channels per output filters and provide theoretical insights for it. To this end, we develop a simple algorithm to select the precision values for CNNs that outperforms corresponding 8-bit quantized networks by 0.9% and 2.2% in top-1 accuracy on ImageNet for ResNet50 and MobileNetV2, respectively.

Keywords: convolutional neural networks quantization, model compression, efficient neural network