Deep Network Quantization via Error Compensation

On-Device Deep Learning: Survey on Techniques Improving Energy Efficiency of DNNs

Providing high-quality predictions is no longer the sole goal for neural networks. As we live in an increasingly interconnected world, these models need to match the constraints of resource-limited devices powering the Internet of Things (IoT) and embedded systems. Moreover, in the era of climate change, reducing the carbon footprint of neural networks is a critical step for green artificial intelligence, which is no longer an aspiration but a major need. Enhancing the energy efficiency of neural networks, in both training and inference phases, became a predominant research topic in the field. Training optimization has grown in interest recently but remains challenging, as it involves changes in the learning procedure that can impact the prediction quality significantly. This article presents a study on the most popular techniques aiming to reduce the energy consumption of neural networks' training. We first propose a classification of the methods before discussing and comparing the different categories. In addition, we outline some energy measurement techniques. We discuss the limitations identified during our study as well as some interesting directions, such as neuromorphic and reservoir computing (RC).

FAT: Frequency-Aware Transformation for Bridging Full-Precision and Low-Precision Deep Representations

Learning low-bitwidth convolutional neural networks (CNNs) is challenging because performance may drop significantly after quantization. Prior arts often quantize the network weights by carefully tuning hyperparameters such as nonuniform stepsize and layerwise bitwidths, which are complicated since the full- and low-precision representations have large discrepancies. This work presents a novel quantization pipeline, named frequency-aware transformation (FAT), that features important benefits: 1) instead of designing complicated quantizers, FAT learns to transform network weights in the frequency domain to remove redundant information before quantization, making them amenable to training in low bitwidth with simple quantizers; 2) FAT readily embeds CNNs in low bitwidths using standard quantizers without tedious hyperparameter tuning and theoretical analyses show that FAT minimizes the quantization errors in both uniform and nonuniform quantizations; and 3) FAT can be easily plugged into various CNN architectures. Using FAT with a simple uniform/logarithmic quantizer can achieve the state-of-the-art performance in different bitwidths on various model architectures. Consequently, FAT serves to provide a novel frequency-based perspective for model quantization.

Efficient Harmonic Neural Networks With Compound Discrete Cosine Transform Filters and Shared Reconstruction Filters

The harmonic neural network (HNN) learns a combination of discrete cosine transform (DCT) filters to obtain an integrated feature from all spectra in the frequency domain. HNN, however, faces two challenges in learning and inference processes. First, the spectrum feature learned by HNN is insufficient and limited because the number of DCT filters is much smaller than that of feature maps. In addition, the number of parameters and the computation costs of HNN are significantly high because the intermediate spectrum layers are expanded multiple times. These two challenges will severely harm the performance and efficiency of HNN. To solve these problems, we first propose the compound DCT (C-DCT) filters integrating the nearest DCT filters to retrieve rich spectrum features to improve the performance. To significantly reduce the model size and computation complexity for improving the efficiency, the shared reconstruction filter is then proposed to share and dynamically drop the meta-filters in every frequency branch. Integrating the C-DCT filters with the shared reconstruction filters, the efficient harmonic network (EH-Net) is introduced. Extensive experiments on different datasets demonstrate that the proposed EH-Nets can effectively reduce the model size and computation complexity while maintaining the model performance. The code has been released at https://github.com/zhangle408/EH-Nets.