Hierarchical Representation Learning

Hierarchical representation learning studies how models can encode data at multiple levels, from local details such as pixels, atoms, or signal segments to broader objects, concepts, and tasks. Common approaches use multiscale pooling and attention, self-supervised prediction or contrastive learning across levels, and explicit taxonomies or graphs; hyperbolic geometry is often used to represent tree-like relationships. These methods are applied to images, language, graphs, molecules, medical data, and long time series, often to improve transfer, fine-grained prediction, retrieval, or interpretability. A central tension is whether useful hierarchies should be imposed from labels and domain knowledge or discovered from data, and how to preserve fine detail without making representations costly or overly constrained.