prof. Ing. Václav Šmídl, Ph.D.

The thesis should contain: i) a survey of active learning methods as an extension of semi-supervised learning, ii) survey of structured data representations with special focus on deep neural networks and their uncertainty, iii) a novel method of active learning for structured data such as text documents, JSON files, or graphs, iv) an analysis of suitability of uncertainty representations for these data, v) an algorithm using specific properties of these data for improved active learning, vi) detailed comparison of all proposed methods with existing state of the art.

Active learning is a well-known class of algorithms for supervised learning aiming to minimize the training effort of the human annotator of the cost of training. However, active learning strategies are almost nonexistent for weakly supervised learning that aims to relax the need for strict input-output relation of the trained model. Weak supervision is used, e.g., in interactive explainable learning or metric learning. The aim of this thesis is to introduce active strategies in this domain. The challenge is to design probabilistic models that are capable of expressing uncertainty over actions that are used in this domain, such as sets of unknown cardinality. We have recently proposed a novel class of tractable probabilistic models that can be used for this task.

Physical experiments are often very costly either in terms of financial cost or simulation time. Minimization of the experimental cost and time has the benefit of significant speedup of the scientific progress. In principle, the task can be addressed by Bayesian methods, such as Bayesian optimization. However, conventional Gaussian processes are not suitable, e.g., due to the problem dimensionality or the lack of likelihood function (i.e., likelihood-free inference). The aim is to explore non-standard probability models in Bayesian experiment design in application physical experiment design. Collaboration with the Institute of Plasma Physics and the Institute of Astronomy of the Czech Academy of Science is established.

Data records stored in the form of a hierarchical structure of heterogenous records are commonly used in internet services (XML, JSON, etc. ), in physical experiments, or environmental science. The co-supervisor of this thesis has recently proposed a theoretically justified approach to learn a discriminative classifier using raw hierarchical data as its input. The aim of this thesis is to extend this approach beyond discriminate learning. For example, investigate unsupervised or semisupervised learning (such as clustering), probabilistic approaches for calibration of the uncertainty of model predictions, or search for a model structure that would be interpretable by humans.