Residual Tuning: Toward Novel Category Discovery Without Labels

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Discovering novel visual categories from unlabeled images is key for intelligent vision systems.
• Current methods pretrain on labeled data then fine-tune for clustering, facing a feature representation tradeoff.
• This limits adapting to new data while preserving existing knowledge.

Purpose of the Study:

• To propose a novel residual-tuning approach for unsupervised visual category discovery.
• To overcome the limitations of unified feature representations in existing methods.
• To enable intelligent systems to learn new concepts without human annotation.

Main Methods:

• A residual-tuning approach is proposed, estimating and adding residual features to basic features.
• This disentangled representation adjusts visual features for unlabeled data.
• It avoids forgetting knowledge from labeled data and does not require replaying labeled images.

Main Results:

• Consistent and considerable gains over state-of-the-art methods on three benchmarks.
• Reduced performance gap compared to fully supervised learning.
• Demonstrated effectiveness in extended scenarios with fewer labeled classes and continual discovery.

Conclusions:

• Residual-tuning offers an efficient solution for unsupervised visual category discovery.
• The approach effectively adapts representations while preserving prior knowledge.
• It significantly advances the capabilities of intelligent vision systems in learning new concepts.