SaleNet: A low-power end-to-end CNN accelerator for sustained attention level evaluation using EEG

Zhang, Chao; Tang, Zijian; Guo, Taoming; Lei, Jiaxin; Xiao, Jiaxin; Wang, Anhe; Bai, Shuo; Zhang, Milin · 2022 · Crossref

DOI: 10.1109/iscas48785.2022.9937323

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper introduces SaleNet, an end-to-end convolutional neural network (CNN) accelerator designed for low-power, real-time evaluation of sustained attention levels using prefrontal electroencephalogram (EEG) signals. The research addresses the need for portable brain-machine interfaces (BMIs) capable of monitoring attention for applications such as ADHD diagnosis and classroom interaction efficiency. While previous systems relied on handcrafted features or separate offline feature extraction modules, SaleNet employs a unified CNN architecture to directly map EEG inputs to attention states, aiming to overcome the high power consumption and complexity of existing solutions. The SaleNet architecture consists of four 1-D convolutional blocks, each containing a convolutional layer, batch normalization (BN), and ReLU activation, followed by a global average pooling (GAP) layer and a linear classifier. To enable deployment on resource-constrained hardware, the authors implemented a comprehensive model compression pipeline. This includes group convolution to reduce computational load, GAP to minimize linear layer parameters, and two pruning strategies: near-zero pruning to remove insignificant weights and a novel bias-driven pruning method that eliminates channels based on BN bias thresholds. Additionally, weight clustering and quantization were applied to reduce parameter precision, achieving a total compression ratio of 183.11x. The compressed model was implemented on a Xilinx Artix-7 FPGA, utilizing a process engine array and on-chip block RAM with a memory reuse scheme to optimize latency and power. Experimental results were obtained using a custom dataset from six subjects performing a number-clicking task, with five prefrontal EEG channels recorded at 250 Hz. The SaleNet achieved a subject-independent classification accuracy of 84.2% for binary attention level evaluation, which aligns with state-of-the-art performance. The hardware implementation demonstrated a competitive power consumption of 0.11 W and an energy efficiency of 8.19 GOps/W. Compared to prior FPGA-based biomedical signal processors, SaleNet offers a significant reduction in power while maintaining high accuracy, despite using a more complex end-to-end CNN structure than some lightweight alternatives. The significance of this work lies in demonstrating that end-to-end deep learning models can be efficiently deployed on low-power FPGA hardware for continuous physiological monitoring. By integrating aggressive model compression techniques with hardware-aware design, SaleNet provides a viable solution for portable, real-time BMI applications. The study highlights the effectiveness of bias-driven pruning and group convolution in reducing model size without substantial accuracy loss, offering a template for future low-power neural network accelerators in healthcare and human-computer interaction domains.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• neuro workload indices