ETD

• deep learning
• hierarchical classification
• lexicographic

Hierarchical image classifiers assign a label to an input at each level of a taxonomy. Predictions should be accurate at every level and form a coherent path through the hierarchy. This thesis formulates hierarchical classification as a lexicographically ordered multi-objective problem, assigning higher priority to coarser objectives than to finer ones. This order is assumed on the grounds that the taxonomy makes errors unequally costly: along a consistent path, an error is repeated at every level below the one where it is first made, so a coarse mistake is the expensive one.
The study examines four mechanisms: LexiGrad, a lexicographic orthogonalisation of the level gradients in the parameter update; an LH-projection layer; supervision of taxonomy-subspace scores; and output correction through the Hierarchical Constraint Cascade. Their comparison assesses how the choice of method and the point at which constraints are applied affect accuracy and hierarchical consistency. All four are evaluated in selected hierarchical classification architectures with alternative readout and decoding rules. Experiments on CIFAR-100, CUB-200-2011 and FGVC-Aircraft use a common protocol to assess per-level accuracy, complete-path correctness, hierarchical error and prediction consistency.