Structured Binary Neural Networks for Accurate Image Classification and Semantic Segmentation

Abstract

In this paper, we propose to train convolutional neural networks (CNNs) with both binarized weights and activations, leading to quantized models specifically for mobile devices with limited power capacity and computation resources. Previous works on quantizing CNNs seek to approximate the floating-point information using a set of discrete values, which we call value approximation, but typically assume the same architecture as the full-precision networks.In this paper, however, we take a novel "structure approximation" view for quantization-it is very likely that a different architecture may be better for best performance. In particular, we propose a "network decomposition" strategy, named Group-Net, in which we divide the network into groups. In this way, each full-precision group can be effectively reconstructed by aggregating a set of homogeneous binary branches. In addition, we learn effective connections among groups to improve the representational capability. Moreover, the proposed Group-Net shows strong generalization to other tasks. For instance, we extend Group-Net for highly accurate semantic segmentation by embedding rich context into the binary structure. Experiments on both classification and semantic segmentation tasks demonstrate the superior performance of the proposed methods over various popular architectures. In particular, we outperform the previous best binary neural networks in terms of accuracy and major computation savings.

Binary neural networks: A survey

“…But it also should be noted that the binary models still face a great challenge, especially when the activations are quantized to 1-bit. For semantic segmentation task, as shown in Table 7, the very recent method [122] achieved high accuracy using only 1-bit, which is almost the same as the full-precision model. But it is unknown how it works and the actual speed up of that method still needs to be verified.…”

“…Thus the binary neural networks are also applied to other common tasks such as object detection and semantic segmentation. in terms of accuracy and major computation savings [122]. In [123], SeerNet considers feature-map sparsity through low-bit quantization, which is applicable to general convolutional neural networks and tasks.…”

A Review of Recent Advances of Binary Neural Networks for Edge Computing

“…Group-Net [24] also improves the performance of 1-bit CNNs through structural design. Inspired by the fact that a floating point number in a computer is represented by a fixed-number of binary digits, Group-Net proposes to decompose a network into binary structures while preserving its representability, rather than directly doing the quantization via "value decomposition".…”

Larq: An Open-Source Library for Training Binarized Neural Networks

“…Since then, the field of BNNs and closely related Ternary Neural Networks has become a prime candidate to enable efficient inference for deep neural networks. Numerous papers have explored novel architectures (Liu et al, 2018;Rastegari et al, 2016;Zhu, Dong, & Su, 2019;Zhuang, Shen, Tan, Liu, & Reid, 2019) and optimization strategies (Alizadeh, Fernández-Marqués, Lane, & Gal, 2019;Helwegen et al, 2019), and the accuracy gap between efficient BNNs and regular DNNs is rapidly closing.…”