DESPINE: NAS generated Deep Evolutionary Adaptive Spiking Network for Low Power Edge Computing Applications

In this work, we investigate a Recurrent Spiking Convolution Neural Network (RSCNN) with Adaptive Spiking Neurons (ASNs) that is enabled with adaptive threshold, spiking frequency, and surrogate gradient (ASG) slope, evolved from neuro-evolutionary scheme adopted spiking neural architecture search (NAS), to classify three primary emotional states: angry, happy and sad. The proposed neuro-evolved NAS search in the form of Adaptive Deep Evolutionary SPIking NEtwork (DE-SPINE) configuration is designed for edge deployment, seeking an energy efficiency of $9.4 \times$ and yet attain a near state-of-the-art (SOTA) accuracy of 85.36% (SOTA is 86.3% on a GPU, consuming ~35 W on a P100 GPU and ~2.8 W on an edge-GPU) for the model trained on FER-13 dataset with 64 timesteps. The EDP gain of $19.42 \times$ was achieved over the SOTA methods. The adaptive neuron model scheme with the NAS run extracted Recurrent Spiking Convolution Neural Network (RSCNN) model offers $3.13 \times$ energy savings over the corresponding NAS evolved RSCNN model comprising of non-adaptive neuron model. The proposed DESPINE framework was evaluated for MNIST, CIFAR-10 and Neuromorphic-MNIST (NMNIST) datasets towards achieving SOTA accuracy, in a view to showcase the suitability of the approach for diverse datasets. The low power costs of the SNN is beneficial to healthcare-bound diagnostic solutions, where the device and the process are continuously running in the background for a long duration. All the scripts and extracted RSCNN model files are made freely available for further usage to the researchers and designers’ community.