Relational Retrieval: Leveraging Known-Novel Interactions for Generalized Category Discovery
Researchers have introduced Relational Pattern Consistency (RPC), a novel approach to Generalized Category Discovery (GCD). RPC explicitly links labeled and unlabeled data through bidirectional knowledge transfer, a departure from existing methods that treat these sources separately. The method uses One-vs-All classifiers for soft ID/OOD decomposition and incorporates mechanisms for known-class preservation and category discovery by analyzing invariant relationships between samples and known-class prototypes. Experiments show RPC achieves state-of-the-art performance on various benchmarks. AI