Energy Efficiency Evaluation of Neural Network Architectures on the Neuromorphic-MNIST Dataset

With the ever-growing need for artificial intelligent applications, the demand for energy-efficient neural networks is more critical than ever given the significant environmental and economic challenges posed by high energy consumption. Inspired by synaptic spikes in the brain mechanism, SNNs with their event-driven processing, are proposed as a promising alternative due to their potential for lower power consumption. This study presents a comparative analysis of Artificial Neural Networks (ANNs), Convolutional Neural Network (CNNs), Spiking Neural Networks (SNNs), and Convolutional Spiking Neural Networks (SCNNs) to evaluate their energy efficiency and performance. Using the N-MNIST dataset, a neuromorphic adaptation of the traditional MNIST dataset, we trained and tested those four NN architectures. The ANN model, trained using a backpropagation algorithm, achieved an accuracy of 97.68%. The CNN model, LeNet-5 architecture, produced a slightly better accuracy of 97.68%. The SNN model, utilizing Leaky Integrate-and-Fire (LIF) dynamics and rate-coding, reached 96.56% accuracy. And the SCNN,using surrogate gradient descent, reported 97.66% accuracy. Despite the expectations for SNN and SCNN, the ANN not only achieved higher accuracy but also demonstrated greater energy efficiency, consuming the least amount of energy at approximately 0.27 watt-hours. %compared to the SNN’s 3.93 watt-hours. These findings also indicate that ANNs and CNNS currently outperform SNN variations in both performance and energy consumption. This study underscores the need for further research into optimizing SNNs to bridge the performance and energy efficiency gap with traditional ANNs, highlighting the challenges and potential directions for improvement in neuromorphic computing applications.