Dynamic Data Inclusion with Sliding Window

Due to the ever growing data requirements of our new machine learning models, it is increasingly vital to understand how we can use our data efficiently. We propose a method to dynamically apply data to our neural network for the task of image classification, in order to observe which data is most useful to train with at different stages of training. Our results show that specific data is better for training at earlier stages of training, and that training may be reduced to lessen the resources required to train, with a marginal effect on accuracy.