Abstract
This paper presents a highly efficient CNN-LSTM hybrid classifier optimized for low-power wearable biomedical devices. By employing ternary weight quantization and a resource-sharing hardware architecture, the proposed FPGA implementation achieves a 1.3x improvement in efficiency (GOPs/Slice) over state-of-the-art accelerators. The system maintains high accuracy across various heart and muscle signal datasets, demonstrating its suitability for on-node physiological monitoring with minimal hardware footprint.
Introduction
We are currently in the era of "smart" wearables, yet most commercial devices remain limited to tracking simple metrics like step counts or heart rates. Analyzing complex physiological signals—such as multi-channel electrocardiograms (ECG) or surface electromyography (sEMG)—requires deep learning, which is notoriously power-hungry and memory-intensive.
The primary challenge lies in "on-node" computation. While sending raw data to the cloud consumes significant bandwidth and battery life, running deep neural networks (DNNs) on a tiny wearable device is constrained by limited silicon area and power. Most existing hardware accelerators focus on maximizing raw throughput for high-speed applications like video processing, which is inefficient for the relatively slow-moving signals (0-500 Hz) found in human biology.
Proposed Architecture and Methodology
The authors propose a specialized hardware-oriented time-series classifier that combines two powerful architectures to balance feature extraction and temporal modeling:
- Convolutional Neural Networks (CNN): Utilized for automated feature extraction from raw 1D signals.
- Long Short-Term Memory (LSTM): A type of Recurrent Neural Network (RNN) that excels at learning temporal dependencies in sequences.
To meet the strict power and area budgets of wearable devices, the design employs two key strategies:
- Ternary Weight Quantization: Instead of standard 32-bit floating-point numbers, weights are quantized to ternary values (-1, 0, and +1). This allows weights to be stored using only 2 bits, drastically reducing memory bandwidth and storage requirements.
- Resource Sharing via Finite State Machine (FSM): Rather than building a massive parallel processor, the system iterates through eight specific states. This approach trades off unnecessary high throughput for a massive reduction in hardware complexity and silicon area.
The design was implemented on a Xilinx Kintex-7 FPGA, focusing on maximizing GOPs/Slice (Giga-Operations Per Second per hardware unit), a critical metric for embedded efficiency.
Experimental Results and Discussion
The system was validated across diverse datasets, including synthetic chaotic dynamical systems (Logistic map and Lorenz attractor) and real-world biomedical benchmarks like CapgMyo (sEMG hand gestures) and PhysioNet (ECG heart sounds).
Key Insights
- Quantization as Regularization: The authors observed that binarizing or ternarizing weights acts as a form of regularization, helping the network generalize better to new data. Ternary weights, in particular, provide a superior accuracy-to-complexity ratio compared to binary weights, as the "zero" state allows the network to effectively prune unimportant connections.
- Efficiency Gains: The design achieved 124.0E-4 GOPs/Slice, which is 1.3x higher than previous state-of-the-art FPGA accelerators. The entire classifier requires only 508 FPGA slices, a fraction of the resources used by traditional high-throughput designs.
- Performance: Despite heavy quantization, the hardware maintained impressive performance, achieving ~97.7% accuracy on sEMG gesture recognition and 93-96% on various ECG benchmarks. The response time was less than 0.2 ms per input window, which is negligible compared to the 10 ms sampling interval of the biological signals.
Conclusion
This work demonstrates that a serialized, state-machine-based approach is far more appropriate for the "slow" nature of human physiology than high-parallelism GPUs. By prioritizing hardware efficiency (GOPs/Slice) over raw throughput, the proposed CNN-LSTM architecture enables sophisticated deep learning on the extreme edge. Future work will explore extending the convolution layers to handle multivariate time series and optimizing training techniques for highly quantized networks.
Summary updated