Online Transformers with Spiking Neurons for Fast Prosthetic Hand Control

2022 IEEE Biomedical Circuits and Systems Conference (BioCAS), Journal Year: 2023, Volume and Issue: 30, P. 1 - 6

Published: Oct. 19, 2023

Fast and accurate online processing is essential for smooth prosthetic hand control with Surface Electromyography signals (sEMG). Although transformers are state-of-the-art deep learning models in signal processing, the self-attention mechanism at core of their operations requires accumulating data large time-windows. They therefore not suited processing. In this paper, we use an attention sliding windows that allows a transformer to process sequences element-by-element. Moreover, increase sparsity network using spiking neurons. We test model on NinaproDB8 finger position regression dataset. Our sets its new terms accuracy NinaproDB8, while requiring only very short time 3.5 ms each inference step, reducing number synaptic up factor ×5.3 thanks results hold great promises wearable sEMG systems control.

Language: Английский

---

Evaluating Spiking Neural Network on Neuromorphic Platform For Human Activity Recognition

Published: Oct. 3, 2023

Energy efficiency and low latency are crucial requirements for designing wearable AI-empowered human activity recognition systems, due to the hard constraints of battery operations closed-loop feedback. While neural network models have been extensively compressed match stringent edge requirements, spiking networks event-based sensing recently emerging as promising solutions further improve performance their inherent energy capacity process spatiotemporal data in very latency. This work aims evaluate effectiveness on neuromorphic processors applications. The case workout with wrist-worn motion sensors is used a study. A multi-threshold delta modulation approach utilized encoding input sensor into spike trains move pipeline approach. spikes then fed direct-event training, trained model deployed research platform from Intel, Loihi, efficiency. Test results show that spike-based workouts system can achieve comparable accuracy (87.5\%) popular milliwatt RISC-V bases multi-core processor GAP8 traditional ( 88.1\%) while achieving two times better energy-delay product (0.66 \si{\micro\joule\second} vs. 1.32 \si{\micro\joule\second}).

Language: Английский

---

Feed-forward and recurrent inhibition for compressing and classifying high dynamic range biosignals in spiking neural network architectures

2022 IEEE Biomedical Circuits and Systems Conference (BioCAS), Journal Year: 2023, Volume and Issue: unknown, P. 1 - 5

Published: Oct. 19, 2023

Neuromorphic processors that implement Spiking Neural Networks (SNNs) using mixed-signal analog/digital circuits represent a promising technology for closed-loop real-time processing of biosignals. As in biology, to minimize power consumption, the silicon neurons' are configured fire with limited dynamic range and maximum firing rates restricted few tens or hundreds Herz. However, biosignals can have very large range, so encoding them into spikes without saturating neuron outputs represents an open challenge. In this work, we present biologically-inspired strategy compressing high-dynamic SNN architectures, three adaptation mechanisms ubiquitous brain: spike-frequency at single level, feed-forward inhibitory connections from neurons belonging input layer, Excitatory-Inhibitory (E-I) balance via recurrent inhibition among output layer. We apply encoded both asynchronous delta modulation method energy-based pulse-frequency method. validate approach silico, simulating simple network applied gesture classification task surface EMG recordings.

Language: Английский

---

FFCSLT: A Deep Learning Model for Traffic Police Hand Gesture Recognition Using Surface Electromyographic Signals

IEEE Sensors Journal, Journal Year: 2024, Volume and Issue: 24(8), P. 13640 - 13655

Published: March 6, 2024

Using surface electromyography (sEMG) signals for gesture recognition can significantly improve the effects of recognition. Therefore, this article proposed a CNN-SE-LSTM-TCN feature fusion network (FFCSLT) traffic police based on characteristics sEMG signals. First, an acquisition system with six-channel sensors was developed acquiring during human movement, and dataset hand gestures (TPG) constructed, which contains total 36000 sets data. Then, squeeze-and-excitation (SE) block adaptive channel weighting added top depthwise separable convolutional (DSCN) to enhance spatial features between each in FFCSLT network. Meanwhile, temporal (TCN) integrated into long short-term memory (LSTM) extract additional Finally, comparing experiments other methods were taken two datasets: self-collected TPG dataset, widely used Sensor Data Ninapro DB1. The experimental results show that our model has accuracy 98.89% 96.52% DB1 is 2.22% 0.75% higher than suboptimal methods, respectively. To further validate network, we also performed variety ablation studies.

Language: Английский

---

Hand gestures classification of sEMG signals based on BiLSTM- Metaheuristic Optimization and Hybrid U-Net-MobileNetV2 Encoder Architecture

Research Square (Research Square), Journal Year: 2024, Volume and Issue: unknown

Published: April 17, 2024

Surface electromyography (sEMG) data has been extensively utilized in deep learning algorithms for hand movement classification. This paper aims to introduce a novel method gesture classification using sEMG data, addressing accuracy challenges seen previous studies. We propose U-Net architecture incorporating MobileNetV2 encoder, enhanced by Bidirectional Long Short-Term Memory (BiLSTM) and metaheuristic optimization spatial feature extraction motion recognition. Bayesian is employed as the approach optimize BiLSTM model's architecture. To tackle non-stationarity of signals, we utilize windowing overlapping method, augmented with additional signals architectures. The encoder extract relevant features from spectrogram images. Edge computing integration leveraged further enhance innovation enabling real-time processing decision-making closer source. Six standard databases were utilized, achieving an average 90.23% our proposed model, showcasing 3-4% improvement 10% variance reduction. Notably, Mendeley Data, BioPatRec DB3, DB1 surpassed advanced models their respective domains accuracies 88.71%, 90.2%, 88.6%, respectively. Experimental results underscore significant enhancement generalizability recognition robustness. offers fresh perspective on prosthetic management human-machine interaction, emphasizing its efficacy improving reducing control interaction machines through edge integration.

Language: Английский

---