Event-Driven Implementation of a Physical Reservoir Computing Framework for superficial EMG-based Gesture Recognition

TL;DR

Real-time gesture recognition from surface EMG on edge devices is challenged by power and latency constraints. This work fuses a physical reservoir computing framework, Rotating Neuron Reservoir (RNR), with a spiking neural network (SNN) and an event-based encoding to convert sEMG into spike trains, enabling near-sensor processing. The spiking RNR (sRNR) uses 48 input spike streams fed into 12 parallel 10-neuron reservoirs (480 outputs) and trains only a readout via either an SVM or a delta-learning rule with Softmax, achieving up to 80.3% test accuracy on Ninapro DB2 (50 gestures) and 74.6% with a linear SVM, demonstrating high performance with low training cost. This approach advances lightweight neuromorphic gesture recognition and paves the way for hardware implementations on neuromorphic chips with minimal training.

Abstract

Wearable health devices have a strong demand in real-time biomedical signal processing. However traditional methods often require data transmission to centralized processing unit with substantial computational resources after collecting it from edge devices. Neuromorphic computing is an emerging field that seeks to design specialized hardware for computing systems inspired by the structure, function, and dynamics of the human brain, offering significant advantages in latency and power consumption. This paper explores a novel neuromorphic implementation approach for gesture recognition by extracting spatiotemporal spiking information from surface electromyography (sEMG) data in an event-driven manner. At the same time, the network was designed by implementing a simple-structured and hardware-friendly Physical Reservoir Computing (PRC) framework called Rotating Neuron Reservoir (RNR) within the domain of Spiking neural network (SNN). The spiking RNR (sRNR) is promising to pipeline an innovative solution to compact embedded wearable systems, enabling low-latency, real-time processing directly at the sensor level. The proposed system was validated by an open-access large-scale sEMG database and achieved an average classification accuracy of 74.6\% and 80.3\% using a classical machine learning classifier and a delta learning rule algorithm respectively. While the delta learning rule could be fully spiking and implementable on neuromorphic chips, the proposed gesture recognition system demonstrates the potential for near-sensor low-latency processing.

Figures (8)

• Figure 1: Architecture of the proposed classification system. The raw sEMG signals are encoded into SNN-compatible spike trains by an event-based encoding scheme. An SNN consisting of physical reservoirs is used to generate transient responses to a higher dimensional feature space. The collected dynamical states are trained in the readout layer only by machine learning algorithms for classifying gestures.
• Figure 2: (a) Examples of one channel of normalized sEMG signals and encoded spike trains for Gesture 1, Gesture 5 and Gesture 8 respectively. (b) The process of spike encoding. After encoding, the original 12 channels of sEMG signals are encoded to 48 channels of spike trains.
• Figure 3: The description of reservoir topologies. The weights of input masks is represented by different colors, dash types and line thicknesses. (a) A classical reservoir topology. The input weight mask $W_{in}$ is fixed and random from a uniform distribution [-1,1]. In addition, the connections among neurons are also fixed and random. (b)&(c) An RNR topology. A 3D description in (b) and a 2D sketch expanded by time in (c). The connections of the input-to-reservoir layer and reservoir-to-output layer are circularly shifted at each time step. (b) exhibits a binary input mask, randomly selected numbers from {-1,1} in conventional eRNR. (c) demonstrates a binary input mask {0,1} in a three-neuron sRNR topology proposed in our work that incorporates spike trains as inputs. The resulting outputs are also spike trains.
• Figure 5: (a) Each 10-neuron reservoir processes an input spike train. The parallel reservoirs project the input spike trains to a higher dimension (from 48 to 480). (b) Two examples of input spike patterns (left) and output spike patterns (right).
• Figure 6: The effect of network size on the classification accuracy of a representative subject by performing exercise B (17 gestures).