A Comprehensive Study On ANN-To-SNN Conversion for Energy-Efficient Neural Network Deployment

Abstract

These days deep neural networks are being integrated into each and every device that exists, from smartphones to portable home devices that run on batteries. Even though these models are smarter, they are also very energy-consuming. This is very true in the case of standard Artificial Neural Networks (ANNs). In order to perform accurately, they never really turn off. They constantly process data in large amounts, which drains the batteries. If we want AI to work on portable edge devices without draining their batteries, we need a much more efficient way to handle the computing workload. This makes Spiking Neural Networks (SNNs) a strong alternative. They are different from regular neural networks because they are event-driven . It only generates signals when necessary to avoiding unnecessary activity, Which is similar to how neurons function in the human brain. This makes them far better for saving energy. One of the main challenges is that SNNs are difficult to train from scratch. Because of this, a practical solution is to train an Artificial Neural Network (ANN) first and then convert it into a Spiking Neural Network (SNN). In this paper, we will look into the total conversion of the ANN to SNN and the model performance. Our main goal is to implement high performance models in small devices without frequent failures and low power consumption. Also this research paper contains the current challenges and future research ideas for creating energy-efficient models