Relational Retrieval: Leveraging Known-Novel Interactions for Generalized Category Discovery
Researchers from the computer vision community have introduced a novel approach to Generalized Category Discovery (GCD) titled Relational Pattern Consistency (RPC). Published on arXiv, this study addresses the limitations of existing methods that treat labeled and unlabeled data separately. Instead, RPC explicitly couples these data sources through bidirectional knowledge transfer to enable mutual enhancement. The method employs One-vs-All classifiers for soft in-distribution/out-of-distribution decomposition. It introduces two key mechanisms: semantic behavioral alignment to preserve known classes and relational pattern matching to discover novel categories by leveraging invariant relationships with known-class prototypes. This strategy transforms unreliable pseudo-labeling into robust relational pattern matching. Extensive experiments indicate that RPC achieves state-of-the-art performance on both generic and fine-grained benchmarks. By allowing labeled data to guide unlabeled learning while identifying new categories through collective relational signatures, this work represents a significant advancement in machine learning algorithms for category discovery without extensive manual labeling.
Wire timeline
Relational Retrieval: Leveraging Known-Novel Interactions for Generalized Category Discovery
Researchers from the computer vision community have introduced a novel approach to Generalized Category Discovery (GCD) titled Relational Pattern Consistency (RPC). Published on arXiv, this study addresses the limitations of existing methods that treat labeled and unlabeled data separately. Instead, RPC explicitly couples these data sources through bidirectional knowledge transfer to enable mutual enhancement. The method employs One-vs-All classifiers for soft in-distribution/out-of-distribution decomposition. It introduces two key mechanisms: semantic behavioral alignment to preserve known classes and relational pattern matching to discover novel categories by leveraging invariant relationships with known-class prototypes. This strategy transforms unreliable pseudo-labeling into robust relational pattern matching. Extensive experiments indicate that RPC achieves state-of-the-art performance on both generic and fine-grained benchmarks. By allowing labeled data to guide unlabeled learning while identifying new categories through collective relational signatures, this work represents a significant advancement in machine learning algorithms for category discovery without extensive manual labeling.
cs.AI updates on arXiv.org