Supported by
Athanasios Davvetas has studied fundamental definitions and concepts in the areas of Probability Theory, Statistics and Machine Learning, which are directly related to the field of study of his thesis (deep learning). Furthermore, he studied research methods concerning the collection and utilisation of results from experiments, writing practices, as well as organising and presenting said results in scientific papers. He conducted a study of relevant literature on neural networks, which are fundamental for the implementation of his working hypothesis. In addition, a literature study was conducted on combining heterogeneous data and their utilisation by neural networks, as well as, combining different relation database schemas (schema matching). By studying the relevant literature, he concluded to his working hypothesis “External heterogeneous data evidence improves deep representations”, which evaluated experimentally. He designed a deep learning method of learning representations according to external categorical evidence, utilised to improve a primary task. His proposed method called “Evidence Transfer” has been evaluated using text and image datasets with the primary task of learning representations for clustering, while introducing multiple sources and various kinds of external evidence. Moreover, he provided a probabilistic interpretation of the effects of evidence transfer method on the latent representations by comparing his method to a well-received method of information theory called “Information Bottleneck”. To verify the hypothesis of his probabilistic interpretation he conducted an empirical study on the latent feature relevance inspired by feature selection and feature ranking methods
