Gradient Projection for Continual Parameter-Efficient Tuning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Parameter-efficient tunings (PETs) offer efficient training for large models but struggle with the trade-off between learning new information and retaining old knowledge, causing generalization collapse and cross-modal hallucination.
• Existing PET methods like Adapter, LoRA, Prefix-tuning, and Prompt-tuning face challenges in continual learning scenarios.

Purpose of the Study:

• To propose a unified framework, Parameter Efficient Gradient Projection (PEGP), that addresses the knowledge forgetting problem in PETs.
• To theoretically and empirically demonstrate the effectiveness of orthogonal gradient projection in mitigating forgetting within large-scale models.

Main Methods:

• Reformulating existing PETs (Adapter, LoRA, Prefix-tuning, Prompt-tuning) from a gradient projection perspective.
• Introducing orthogonal gradient projection into various PET paradigms.
• Theoretically analyzing how orthogonal gradient projection resists forgetting by minimizing impact on the old feature space.

Main Results:

• The proposed PEGP framework effectively reduces forgetting across various continual learning settings, including class-incremental, domain-incremental, task-incremental, and multi-modality continual learning.
• Experiments on ViT and CLIP backbones demonstrate the efficiency of PEGP in reducing forgetting without significant increases in memory or training time.
• Orthogonal gradient projection was theoretically shown to be effective in resisting forgetting even in large-scale models.

Conclusions:

• PEGP offers a unified and effective approach to mitigate catastrophic forgetting in parameter-efficient tuning methods for large models.
• The orthogonal gradient projection mechanism provides a principled way to balance learning new knowledge with preserving existing knowledge.
• This framework has broad applicability across diverse continual learning tasks and model architectures.