Semantic Mismapping in Generalized Category Discovery: Evaluation and Mitigation
Global label matching can inflate new-category accuracy by up to 25.63 points, the authors report.
In generalized category discovery, supervised outputs for known categories already have fixed meanings, but global Hungarian matching can reassign them and conceal prediction errors. The authors propose Semantics-Preserving Evaluation, which keeps those meanings fixed and matches only new-category outputs. They also introduce training losses to improve routing between known and new categories and encourage distinct new-category prototypes. They report that the method is effective under both the existing and proposed evaluations.
ArXiv · AI/CL/LG's note
In generalized category discovery, supervised outputs for known categories already have fixed meanings, but global Hungarian matching can reassign them and conceal prediction errors. The authors propose Semantics-Preserving Evaluation, which keeps those meanings fixed and matches only new-category outputs. They also introduce training losses to improve routing between known and new categories and encourage distinct new-category prototypes. They report that the method is effective under both the existing and proposed evaluations.
ArXiv · AI/CL/LG's note
score 4