Improving Soybean Disease Prediction by Performing Late-Stage Re-Training Using Fireworks Algorithm

This paper proposes a novel method of training and optimizing a model to get as optimal a result as possible. This study aims to shed light on alternative ways of using multiple optimizers with different properties to get better results, achieving an incremental increase in accuracy. This is done by switching the optimizer after training once, hence performing Late-Stage Re-Training (LSRT). The present study shows the use of 2 different optimizers, Adam and Fireworks, for their unique properties to gain better results. Adam is used first for most of the training process and is later switched out to Fireworks for re-training this model. This methodology is tested on the Soybean dataset. The above process leads to an average of 2-3% increase, thus obtaining a training accuracy of 98.9%.