PCBNN: An Algorithm–Hardware Co-Designed Binary Neural Network for Low-Power Edge Inference
Tsinghua Science and Technology · 2026-07
DOI: 10.26599/tst.2026.9010073 ↗Kurz gefasst
Fachartikel in der Zeitschrift Tsinghua Science and Technology, 2026-07. Verfasst von Haihui Liu und 5 weiteren Autor:innen.
Abstract (Originalsprache)
Deploying deep neural networks on resource-constrained edge computing systems requires high inference efficiency, low power consumption, and compact model representation, while maintaining competitive accuracy on large-scale tasks. To address these challenges, this paper pro-poses PCBNN, an algorithm-hardware co-designed binary neural network for low-power edge inference on FPGAs. PCBNN introduces a parallel binary convolutional architecture that exploits fine-grained hardware parallelism to accelerate inference. On the algorithmic side, a contrastive learning framework is designed to explicitly model the correlation between binary activations and their full-precision counterparts, thereby mitigating accuracy degradation from binarization. Furthermore, a channel-segmentation scheme enables a fully binarized input layer, preserving critical features while reducing computational complexity. Building on these components, an FPGA-based accelerator is implemented to support efficient and low-latency BNN inference. The experimental results demonstrate that PCBNN effectively utilizes FPGA parallelism and hardware resources to achieve substantial improvements in both accuracy and energy efficiency compared with state-of-the-art BNN accelerators.
Zitieren
APA
Liu, H., An, B., Mei, Y., Zhu, Y., Zhang, G., & Liang, Y. (2026). PCBNN: An Algorithm–Hardware Co-Designed Binary Neural Network for Low-Power Edge Inference. Tsinghua Science and Technology. https://doi.org/10.26599/tst.2026.9010073
Vancouver
Liu H, An B, Mei Y, Zhu Y, Zhang G, Liang Y. PCBNN: An Algorithm–Hardware Co-Designed Binary Neural Network for Low-Power Edge Inference. Tsinghua Science and Technology. 2026. doi:10.26599/tst.2026.9010073
BibTeX
@article{Liu2026,
author = {Liu, Haihui and An, Baoxu and Mei, Yaxin and Zhu, Yunxuan and Zhang, Guangxue and Liang, Yuzhu},
title = {PCBNN: An Algorithm–Hardware Co-Designed Binary Neural Network for Low-Power Edge Inference},
journal = {Tsinghua Science and Technology},
year = {2026},
doi = {10.26599/tst.2026.9010073}
}
EUVIMED