Low latency inference is a highly desirable metric to deploy any autonomous control systems. To aid in this field, we propose and investigate multiple spike decoders for Spiking Neural Networks (SNN) to perform forearm angle prediction by taking Surface Electromyography (sEMG) signal as input, as part of an autonomous post stroke motor rehabilitation system. The proposed regression decoders are compared with the existing implementations and conventional ANNs in terms of accuracy and relative efficiency. The models introduced here exhibited an accuracy of 90% compared to the baseline’s 82% while also offering 10× to 12× of higher energy efficiency and EDP gains 24×. These models also exhibited relatively lower loss in performance with varied input data parameters compared to the baselines. These SNNs are a low latency alternatives for implementing complex regression models like the reference rehabilitation system with potential usability in other Reinforcement Learning (RL) problems.

Long short-term memory (LSTM) has emerged as a definitive network model for analyzing and inferring time series data since their introduction. LSTM has the capability to not only extract spectral features similar to convolutional-neural-network (CNN) models but also a mixture of temporal features. Due to this distinguished advantage, similar feature extraction method is explored for the spiking counterpart of the neural network, targeted for time-series data. Though LSTMs perform well in the spiking form of neural network, they tend to be compute and power intensive. Addressing this issue, the work proposes multicompartment leaky (MCLeaky) neuron as a viable alternative for efficient processing of time series data. The MCLeaky neuron, introduced as a derivative of the leaky integrate and fire (LIF) neuron model, contains multiple memristive synapses interlinked to form the memory component of the neuron, by emulating Hippocampus’ structure of brain as reference. The proposed MCLeaky neuron-based spiking neural network (SNN) model and its quantized variant were benchmarked against state-of-the-art (SOTA) spiking LSTMs to perform human stress detection by comparing computing requirements, compute-latency, and real-world performances on freshly acquired unseen data with models that is acquired by employing neural architecture search (NAS). Results show that the networks with MCLeaky activation neuron managed a superior accuracy of 98.8% to detect stress based on electrodermal activity (EDA) signals, better than any other investigated model, while using 20% less parameters on average. MCLeaky neuron was also investigated for different modality of signals, including EDA Wrist, EDA Chest, Temperature, electrocardiogram signal, and combination of them. Quantized MCLeaky model was also derived and validated to forecast their performance on hardware aware architecture, which resulted in 91.84% accuracy. The neurons were evaluated for multiple modalities of data toward stress detection, which resulted in energy savings of$25.12\times - 39.20\times $and EDP gains of$52.37\times - 81.9\times $over the artificial neural network model, besides offering the best accuracy of 98.8% when compared with the remainder of the SOTA implementations.

Reliable and low-power stress-detector ‘at the edge’ is extremely beneficial for continuously monitoring hospitalized patients. In this context, feed-forward spiking neural networks (SNNs) for stress-detection using physiological time-series signals of electrodermal activity (EDA), body temperature, and a multi-modal signal comprised of both, are designed and evaluated.

Reliable, low-power stress-detection ‘at the edge’ is important in the context of monitoring post-stroke patients. In this context, we have developed feed-forward spiking neural networks (SNNs) for stress-detection using physiological timeseries signals of electrodermal activity (EDA), body temperature, and a multi-modal signal comprised of both. Execution of the SNNs on Intel Loihi-2 (a neuromorphic research chip) showed 5× to 83×, and 9× to 123× better energy-delay product (EDP) compared to equivalent ANNs executed on a low-power edgeGPU, and FPGA respectively. We report that the largest EDP gains (83×) are obtained for the multi-modal SNN, which has „9× less number of parameters and is „9× faster in inference latency than a conventional ANN executed on an edge-GPU.