Reproduce “AdaptGCD: Multi-Expert Adapter Tuning for Generalized Category Discovery” on OpenResearch

About this paper

Yuxun Qu, Yongqiang Tang, Chenyang Zhang, Wensheng Zhang

AdaptGCD presents a multi-expert adapter tuning framework for Generalized Category Discovery (GCD), enhancing the classification of known categories and the discovery of novel ones. The method utilizes parameter-efficient adapters and a category-aware routing constraint to effectively balance pre-trained knowledge with task-specific adaptation, yielding consistent performance improvements across various benchmarks.

Key insights

• Adapter tuning offers a parameter-efficient and effective alternative to traditional fine-tuning methods for GCD, successfully balancing pre-trained knowledge preservation with task-specific adaptation.
• Employing a multi-expert architecture with explicit category-aware routing constraints effectively mitigates the inherent bias towards labeled 'old' classes, leading to improved discovery and classification of 'new' categories.
• The integration of non-linear layers within adapter bottlenecks and their strategic placement in higher Transformer blocks contribute to superior adaptability compared to other parameter-efficient fine-tuning methods.

Summary by alphaXiv.