DiTASK: Multi-Task Fine-Tuning with Diffeomorphic Transformations

Keyphrases

• Multi-task Learning 100%
• Weight Matrix 100%
• Diffeomorphic Transformation 100%
• Task-specific Fine-tuning 100%
• Multi-task 50%
• Computer Vision 50%
• Diffeomorphic 50%
• Geometric Structure 50%
• Singular Vectors 50%
• Full Rank 50%
• Singular Values 50%
• Force Task 50%
• Degradation Performance 50%
• Hidden Representation 50%
• Vision Transformer 50%
• LoRa 50%
• Fine-tuning Strategy 50%
• Feature Modulation 50%
• Constrained Subspaces 50%
• Low-rank Update 50%
• State-of-the-art Performance 50%
• Pre-trained Features 50%
• Fewer Parameters 50%
• PASCAL 50%
• Parameter-efficient 50%
• Dense Prediction Tasks 50%
• Task-specific Adaptation 50%
• Task-specific Features 50%
• Pre-trained Representations 50%

Computer Science

• Multitask Learning 100%
• Multiple Task 50%
• Singular Value 50%
• Art Performance 50%
• Singular Vector 50%
• Vision Transformer 50%
• Computer Vision 50%