PhD Oral Exam - Vitaliy Tayanov, Computer Science

When studying for a doctoral degree (PhD), candidates submit a thesis that provides a critical review of the current state of knowledge of the thesis subject as well as the student’s own contributions to the subject. The distinguishing criterion of doctoral graduate research is a significant and original contribution to knowledge.

Once accepted, the candidate presents the thesis orally. This oral exam is open to the public.

Abstract

This research is devoted to the problem of overfitting in Machine Learning and Pattern Recognition. It should lead to generalisation ability and accuracy boosting in the case of small and/or difficult classification datasets. The aforementioned two problems have been solved in two different ways: by splitting the entire datasets into functional groups depending on the classification difficulty using consensus of classifiers, and by embedding the data obtained during classifier stacking into nonlinear spaces i.e., Riemannian and Grassmann manifolds. These two techniques are the main contributions of the thesis. The insight behind the first approach is that we are not going to use the entire training subset to train our classifiers but some of it which approximate the true geometry and property of classes. In terms of Data Science, this process can also be understood as Data Cleaning. According to the first approach, instances with high positive (easy) and negative (misclassified) margins are not considered for training as those that do not improve (or even worsen) the evaluation of the true geometry of classes. The main goal of using Riemannian geometry consists in embedding our classes in nonlinear spaces where the geometry of classes in terms of easier classification has to be obtained. Before embedding our classes on Riemannian and Grassmann manifolds we do multiple Data Transformations using different variants of Classifier Stacking. Riemannian manifolds of Symmetric Positive Definite matrices are created using the classifier interactions while Grassmann manifolds are built based on Decision Profiles. The shared idea behind the two aforementioned approaches consists in Data Complexity reduction. According to general opinion among researchers, Data Complexity reduction should lead to an overfitting decrease as well as to classification accuracy enhancement.

We carried out our experiments on the datasets of general characters from the UCI Machine Learning repository. We also tested our approaches on two datasets related to the Video Analysis problem. The first dataset is a Phase Gesture Segmentation dataset taken from the UCI Machine Learning repository. The second one is the Deep Fake detection Challenge (DFDC) dataset. In order to apply our approach to solve the second problem, some image processing has been carried out. Numerous experiments on datasets of general character and those related to Video Analysis problems show the consistency and efficiency of the proposed techniques. We compared our techniques with state-of-the-art ones too. The obtained results show the superiority of our approaches for most of the cases. The significance of carried out research and obtained results consists in better representation and evaluation of the geometry of classes which may overlap only in feature space due to some improper measurements, errors, noises, or by selecting features that do not represent well our classes. Carried out research is a pioneering one both in terms of Data Cleaning and Classifier Ensemble Learning in Riemannian geometry.