Ing. Radek Mařík, CSc.

The research is focused on the intersection of active learning and deep neural networks within the realm of natural language processing. The study explores and develops advanced uncertainty representation and measurement methodologies for deep neural networks. Contribution aims to enhance the efficiency and performance of active learning strategies across multiple natural language processing tasks.

The application of artificial intelligence methods in telecommunications and social-humanities poses diverse challenges. In many cases, these domains provide either huge amounts of data or only very small fragmented and uncertain samples. In various cases, metrics are not designed to produce results consistent with human judgment. So far, modern techniques do not sufficiently document the resulting decisions and thus reduce the guarantees under which the methods are functional. The dissertation will focus on one of the open problems mentioned above.

Applications of complex network analysis methods in telecommunications and socio-humanities are based on building relational relations with non-trivial topologies. The specific conditions of these domains allow defining the measures of relationships that lead to alternative views of the solved problems based on the investigation of the properties of weighted graphs. Many techniques, such as the detection of overlapping bipartite network communities, are still not satisfactorily resolved. There is also a lack of a set of suitable metrics that enable solving tasks with nodes described by very heterogeneous attribute sets, especially if the attributes are binary or nominal. Problems operating with the dynamic development of these networks achieve even greater complexity.

Graph Neural Networks (GNN) is a field of machine learning models designed for the analysis of data expressible in graph structure used for classification, regression or anomaly detection at the node, edge or whole graph level. The combination of graph neural networks with signal or image data processing methods will be used to improve the interpretation of semantic content symbols by applying information propagation using a topology of injected complex networks. The structure of the data with elements of hierarchy and dependency aspects of the problem to be solved can be captured by an appropriate graph neural network and subsequent processing of topological information combined with state information and properties of neighboring nodes. There are different categories of GNNs, e.g., recurrent and convolutional graphical neural networks, graphical autoencoders, and spatio-temporal graphical neural networks and attention networks. GNNs appear in modeling physical or chemical systems, image recognition applications, anomaly detection or problems subject to adversarial attacks, as well as in industrial applications such as knowledge graphs, transportation networks and infrastructure security. Special attention should be paid to the explainability of GNNs and their use in critical infrastructure.