Submitted: September 2020
Abstract
In the last decades, the advancement of processing power has greatly benefited the field of machine landing and especially deep learning. New powerful and affordable hardware has opened ways to use formerly computationally expensive, if not infeasible methods of processing data.
In this thesis, we use machine learning techniques to classify images based on the mathematical properties of their numeric encodings, comparing how accurately they can predict those abstract properties. In order to gather large enough datasets for the training process, these images are generated from numeric encodings fulfilling such properties. The goal is to find out if such techniques are viable in this scenario and identify the challenges of this task rather than maximizing accuracy for any given model.
More specifically, the influence of the existence and difficulty of patterns of certain mathematical properties on the machine learning process is explored. Properties with simple patterns like even numbers are expected to be detected easily while properties without any clear patterns like prime numbers are expected to be hard or impossible to detect. To find out how these expectations hold up in practice, multiple machine learning models with different complexity and hyperparameter setups are compared.