Dynamic Consolidation for Continual Learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Training deep learning models on nonstationary data is crucial for general artificial intelligence.
• Continual learning (CL) enables models to learn new information without forgetting past knowledge.
• Existing CL methods limit model exploration and use suboptimal fixed parameter importance.

Purpose of the Study:

• To address limitations in current continual learning methods.
• To enhance the exploration ability of deep learning models during training.
• To develop a more dynamic and adaptive approach to parameter importance in CL.

Main Methods:

• Relaxed vicinity constraints on model parameters using a global definition of parameter importance.
• Defined parameter importance as the sensitivity of the global loss function to model parameters.
• Proposed adaptive adjustment of parameter importance to align with dynamic parameter updates.

Main Results:

• The proposed method allows for broader exploration of the parameter space.
• Adaptive adjustment of parameter importance proved more effective than fixed strengths.
• Experimental results showed performance improvements of up to 24% in average accuracy over strong baselines.

Conclusions:

• The novel CL approach effectively overcomes limitations of existing methods.
• The method demonstrates significant potential for building more robust and adaptable intelligent systems.
• This work advances the field of continual learning for nonstationary data.