BoolNet: Towards Energy-Efficient Binary Neural Networks Design and Optimization

Nianhui Guo, Joseph Bethge, Haojin Yang, Kai Zhong, Xuefei Ning, Christoph Meinel, Yu Wang

Keywords: energy efficient network, low-bit compression

TL;DR: BoolNet innovatively optimizes Binary Neural Networks for enhanced energy efficiency and accuracy through advanced architecture design and optimization.

Abstract: Binary Neural Networks (BNNs) have been advancing towards bridging the accuracy gap with their 32-bit counterparts, often by integrating additional 32-bit components. However, such designs, including the use of 32-bit values for feature maps and residual shortcuts, pose challenges for hardware accelerators with constraints on memory, energy, and computing resources. Addressing the critical question of balancing accuracy with energy consumption in BNNs, we introduce \emph{BoolNet}, a novel BNN architecture minimizing the use of 32-bit components. \emph{BoolNet} employs 1-bit values for storing feature maps, achieving a significant balance between efficiency and performance. Our experiments on the ImageNet dataset show that \emph{BoolNet} attains 63.0\% Top-1 accuracy, along with a 2.95-fold reduction in energy consumption compared to recent state-of-the-art BNNs. The code and trained models will be made available at: \texttt{(URL in final version)}.

Submission Number: 5