Online Learning and Classification of EMG-Based Gestures on a Parallel Ultra-Low Power Platform Using Hyperdimensional Computing

Abstract

This work presents a wearable EMG gesture recognition system based on the hyperdimensional (HD) computing paradigm, running on a programmable Parallel Ultra-Low-Power (PULP) platform. The processing chain includes efficient on-chip training, which leads to a fully embedded implementation with no need to perform any offline training on a personal computer. The proposed solution has been tested on 10 subjects in a typical gesture recognition scenario achieving 85% average accuracy on 11 gestures recognition, which is aligned with the State-Of-the-Art (SoA), with the unique capability of performing online learning. Furthermore, by virtue of the Hardware (HW) friendly algorithm and of the efficient PULP System-on-Chip (SoC) (Mr. Wolf) used for prototyping and evaluation, the energy budget required to run the learning part with 11 gestures is 10.04mJ, and 83.2µJ per classification. The system works with a average power consumption of 10.4mW in classification, ensuring around 29h of autonomy with a 100mAh battery. Finally, the scalability of the system is explored by increasing the number of channels (up-to 256 electrodes), demonstrating the suitability of our approach as universal, energy-efficient biopotential wearable recognition framework.

Noninvasive detection of focal seizures in ambulatory patients

“…68 Such approaches would benefit from the next generation of ultra-low-power multicore platforms with embedded machine-learning accelerators, which can offer many advantages in terms of parallelization capabilities to execute complex algorithms and process multimodal data inputs in complex real-life wearable setups. 69 The ideal solution delineated above is not yet available and will require some years to be developed, tested, and validated. Intermediate suboptimal devices that could prove useful to a subgroup of patients (eg, hypermotor seizures) are welcome and shall bring important knowledge to the field.…”

“…Cutting edge self‐learning algorithms, such as generative adversarial networks, which proved highly effective for image processing, might also carry significant progress in FS detection and forecasting 68 . Such approaches would benefit from the next generation of ultra–low‐power multicore platforms with embedded machine‐learning accelerators, which can offer many advantages in terms of parallelization capabilities to execute complex algorithms and process multimodal data inputs in complex real‐life wearable setups 69 …”

A CNN-LSTM Hybrid Model for Wrist Kinematics Estimation Using Surface Electromyography

“…classifier-based pattern recognition (PR) and regression, have been extensively investigated in recent literature. Unlike PR-based methods which discriminate hand gestures in a discrete and sequential manner [3], regression models focus on continuous wrist kinematics estimation [4] and thus can promote simultaneous and proportional control in multiple degrees of freedoms (DoF). Several MLbased regression methods, including linear regression (LR), artificial neural network (ANN), kernel ridge regression, support vector regression (SVR) and random forest (RF), have been extensively exploited in both off-line simulations [5][6][7][8][9] and real-time prosthetic control [1].…”

Classification of Electromyographic Hand Gesture Signals Using Modified Fuzzy C-Means Clustering and Two-Step Machine Learning Approach

“…At first, EMG signals are prone to contaminations caused by real-world factors such as electrode shift and muscle fatigue. Small changes of the EMG traces may result in great impacts on classification results [17], [18]. Also, due to the limitation of data acquisition, the number of available data tends to be insufficient for deep learning's training [15].…”

“…Also, due to the limitation of data acquisition, the number of available data tends to be insufficient for deep learning's training [15]. To compensate the weak robustness of the established models and the insufficient training data, more data are employed for training in some literatures [17], [18], which further complicates the classification. In addition, the trained models tend to be subject-specific and are not easy to be generalized to other datasets, since different subjects have different data characteristics such as contractions' strength, skin thickness and muscular mass [18].…”