Accelerating LSTM model training through biologically inspired NeuroAI

Modern Deep Neural Networks, have an ever increasing number of parameters that need to be optimised on enormous amounts of data during training. Even with advanced hardware, the training process for these models can take months and can consume large amounts of energy. This prolonged use of computing resources drives up cost significantly. In this paper we explore how one uses Spiking Neural Network (SNN) in LSTM via the NeuroAI toolkit. Compared against the traditional LSTM, the training time of our adapted LSTM vastly decreased by as much as 75% but with an increase in accuracy. We also demonstrate that this decreases the amount of CPU used by as much as 14% thereby being more efficient and sustainable.