An Empirical Evaluation of Deep Architecture Classification Using Simulated Datasets

Abstract

I present an empirical evaluation of deep architectures used for supervised learning and classification. Deep learning methods have recently become a popular topic, which many researchers have claimed to have better performance on complicated vision datasets. So far, comparisons of deep architectures with other classical methods only utilize the MNIST dataset of handwritten digits (and permutations thereof) to evaluate their performance. In this thesis, I expand this evaluation of deep learning methods by generating datasets with varying size, dimension, number of classes, and distribution to fully stress these methods across a variety of datasets. From these tests, I find that deep learning methods have equal performance or, in some cases, worse performance than classical methods on certain datasets.