The final state exam of Jonas Daniel Nienhaus takes place semi-distantly on November 21 from 14:30 via MS Teams. Jonas will be defending his Master's Thesis titles "Neural Network Cascades to Incorporate Domain Knowledge for Hematopoietic Cell Classification" (available in DSpace) supervised by Jan Havlík.
In case you want to watch the exam online, please request access at email@example.com. Recording of the exam will be available later upon request.
In this work, classification of hematopoietic cells is performed hierarchically, employing cascades of deep neural networks. Two strategies are defined to obtain combined predictions from the cascades, a probabilistic approach and a greedy, deterministic method. Although different in theory, both strategies lead to very similar outcomes in practice. Hierarchical cascades can be trained either separately for each network or end-to-end as a whole. It is shown that individual training is preferable both in terms of achieved performance as well as flexibility. Different hierarchies are defined to guide the classification process, either oriented on the biological cell lineages, on certain features and characteristics, or in order to compensate for class imbalance in the dataset. Several methods and optimization techniques are evaluated, including the possibility to incorporate regressors instead of classifiers for certain sub-tasks where the classes have an ordinal character. In general, it is shown that, for appropriate hierarchies, similar classification performances as with single networks can be achieved. Especially pretraining of the individual networks on the target dataset yields improvements. The main advantage of the cascades is increased modularity as well as additional information compared to single networks. With feature forwarding, a method introducing embedding vectors into the cascades and passing them to subsequent networks, the learned feature space can be visualized not only for the entire model but also at intermediate levels.