Towards ARSPI-Net: development of an efficient hybrid deep learning framework

In this paper, we implement and experiment with a portion of our overall hybrid deep learning framework, An Affective hYbrid SPIking Neural Network [ARSPI-NET]. In order to motivate and show the usage of liquid state machines as an efficient, feature extraction framework we perform experiments on connection architecture as well as neuron model and their effect on overall classification performance. We perform our initial experimentation on the MNIST dataset and achieve a 87% classification using a liquid state machine and logistic regression classifier. Our results suggest that our framework can compare with current models in terms of accuracy, however, we outperform traditional deep learning methods in terms of energy consumption and the potential to move to energy-efficient neuromorphic platforms. In addition, our framework has the advantage of being more interpretable, as it allows us to model the spatiotemporal dynamics of signals through the usage of a liquid reservoir and an interpretable readout vector. This paper sets precedence for future experimentation and development of ARSPI-Net as a hybrid deep learning framework.