[Paper Review] Towards Accurate Binary Convolutional Neural Network

Xiaofan Lin, Cong Zhao+1 more|arXiv (Cornell University)|Nov 30, 2017

Neural Networks and Applications392 citations

TL;DR

This paper introduces ABC-Net, a scheme to train binary CNNs using multiple binary weight bases and multiple binary activations, achieving near full-precision accuracy on ImageNet while enabling efficient bitwise computation.

ABSTRACT

We introduce a novel scheme to train binary convolutional neural networks (CNNs) -- CNNs with weights and activations constrained to {-1,+1} at run-time. It has been known that using binary weights and activations drastically reduce memory size and accesses, and can replace arithmetic operations with more efficient bitwise operations, leading to much faster test-time inference and lower power consumption. However, previous works on binarizing CNNs usually result in severe prediction accuracy degradation. In this paper, we address this issue with two major innovations: (1) approximating full-precision weights with the linear combination of multiple binary weight bases; (2) employing multiple binary activations to alleviate information loss. The implementation of the resulting binary CNN, denoted as ABC-Net, is shown to achieve much closer performance to its full-precision counterpart, and even reach the comparable prediction accuracy on ImageNet and forest trail datasets, given adequate binary weight bases and activations.

Motivation & Objective

• Motivate low-memory, low-power CNN deployment on mobile/hardware by binarizing weights and activations.
• Propose a method to approximate full-precision weights with linear combinations of multiple binary bases.
• Enhance information flow through networks by using multiple binary activations.
• Demonstrate that with adequate bases, binary CNNs can approach full-precision accuracy on ImageNet.

Proposed method

• Represent real-valued weights as a linear combination of M binary weight bases with B_m in {-1,+1}.
• Compute B_m from the shifted, normalized real weights to enable backpropagation with STE.
• Train using straight-through estimators to backpropagate through the binary bases and activations.
• Represent real-valued activations as a linear combination of N binary activations A_n with coefficients beta_n, enabling richer information passage.
• Perform convolution as a sum of M×N binary convolutions, enabling bitwise operations when activations are also binary.
• Use batch normalization and learned shift parameters to stabilize and optimize multiple binary activations.

Experimental results

Research questions

• RQ1 How close can binary CNNs come to full-precision accuracy on ImageNet using multiple binary weight bases and multiple binary activations?
• RQ2 What is the impact of varying the number of weight bases M and activations N on classification accuracy?
• RQ3 Can weight approximation as a whole (or channel-wise) with binary bases effectively replace full-precision weights in CNNs?
• RQ4 How does ABC-Net compare to existing quantization/binzarization methods on large-scale datasets?

Key findings

Model	W (weights bitwidth)	A (activations bitwidth)	Top-1	Top-5
Full-Precision Resnet-18	32	32	69.3%	89.2%
BWN	1	32	60.8%	83.0%
DoReFa-Net	1	4	59.2%	81.5%
XNOR-Net	1	1	51.2%	73.2%
BNN	1	1	42.2%	67.1%
ABC-Net [5 binary weight bases, 5 binary activations]	1	1	65.0%	85.9%
ABC-Net [5 binary weight bases, full-precision activations]	1	32	68.3%	87.9%

• ABC-Net with 5 binary weight bases and 5 binary activations achieves Top-1 65.0% and Top-5 85.9% on ImageNet with ResNet-18, nearing full-precision performance (69.3% Top-1, 89.2% Top-5).
• Increasing M and N consistently improves accuracy, with larger configurations yielding smaller gaps to full-precision models (e.g., M=5, N=5 reduces the Top-1 gap to about 4.3%).
• Compared to state-of-the-art binary/low-precision networks (BWN, XNOR-Net, BNN, DoReFa-Net), ABC-Net with multiple bases/activations delivers substantially higher Top-1/Top-5 accuracy while maintaining binary weights and activations.
• Channel-wise weight approximation offers potentially finer-grained modeling than whole-weight approximation, though more training cost is involved.
• Using full-precision activations with binary weights (or vice versa) also yields competitive results, illustrating flexibility in the ABC-Net design.