InfBA: Interference-Free Bottleneck Adaptation for Continual Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Continual learning enables models to sequentially learn multiple tasks.
• Key challenges include maintaining performance on old tasks (stability) and adapting to new ones (plasticity).
• Parameter-Efficient Fine-Tuning (PEFT) methods are increasingly popular for continual learning but often struggle with task interference.

Purpose of the Study:

• To propose a novel PEFT method for continual learning that addresses the trade-off between stability and plasticity.
• To introduce Interference-Free Bottleneck Adaptation (InfBA) to eliminate interference between new and old tasks.

Main Methods:

• InfBA utilizes a bottleneck architecture to reduce and then restore embedding dimensionality.
• It constrains model updates within a subspace designed to prevent interference.
• InfBA can be integrated with existing PEFT techniques like Adapter, LoRA, and Prefix-tuning.

Main Results:

• InfBA demonstrates superior performance compared to existing PEFT-based continual learning methods.
• Experimental results across multiple datasets confirm the method's effectiveness.
• The approach successfully balances model stability and plasticity.

Conclusions:

• InfBA offers an effective solution for the stability-plasticity dilemma in continual learning.
• The proposed method represents a significant advancement in PEFT for sequential task learning.
• InfBA provides a versatile framework compatible with various PEFT approaches.