Dimensionality Reduction to Dynamically Reduce Data

Abstract

We perform experiments using dynamic data reduction on datasets of moderate complexity, with focus on classification of a Micro-PCB image dataset. As deep learning models increase in complexity, the data that they use increases at a rate we can't keep up with. The result of this is often slight improvements to the model's accuracy, at the improportional cost of computational runtime, which increases the electricity used, and ultimately carbon emissions. By using data reduction techniques, we attempt to identify the least critical data to be excluded from training, which in turn cuts the environmental cost. We show the effect of data reduction techniques on moderately complex image data, including PCB images, to reduce runtime by 2% and improve the accuracy by 0.013%.