Towards Optimization-Friendly Binary Neural Network

Nianhui Guo, Joseph Bethge, Hong Guo, Christoph Meinel, Haojin Yang

Abstract:

Binary neural networks (BNNs) are a promising approach for compressing and accelerating deep learning models, especially in resource-constrained environments. However, the optimization gap between BNNs and their full-precision counterparts has long been an open problem limiting their performance. In this work, we propose a novel optimization pipeline to enhance the performance of BNNs. The main approach includes three key components: (1) BNext, a strong binary baseline based on an optimization-friendly basic block design, (2) knowledge complexity, a simple yet effective teacher-selection metric taking the capacity gap between teachers and binary students under consideration, (3) consecutive knowledge distillation (CKD), a novel multi-round optimization technique to transfer high-confidence knowledge from strong teachers to low-capacity BNNs.
We empirically validate the superiority of the method on several vision classification tasks CIFAR-10/100 & ImageNet. For instance, the BNext family outperforms previous BNNs under different capacity levels and contributes the first binary neural network to reach the state-of-the-art 80.57\% Top-1 accuracy on ImageNet with 0.82 GOPS, which verifies the potential of BNNs and already contributes a strong baseline for future research on high-accuracy BNNs. The code will be publicly available at (blind URL, see supplementary material).

Submission Length: Regular submission (no more than 12 pages of main content)

Code: https://github.com/hpi-xnor/BNext

Assigned Action Editor: ~Sanghyuk_Chun1

License: Creative Commons Attribution 4.0 International (CC BY 4.0)

Submission Number: 1603