Elastic Multi-Gradient Descent for Parallel Continual Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Optimization

Background:

• Multi-Task Learning (MTL) assumes fixed tasks, limiting adaptability.
• Serial Continual Learning (SCL) has delayed responses due to sequential processing.
• Parallel Continual Learning (PCL) faces task conflict and catastrophic forgetting with dynamic task arrivals.

Purpose of the Study:

• To introduce a novel paradigm for Parallel Continual Learning (PCL) in dynamic multi-task scenarios.
• To address challenges of task conflict and catastrophic forgetting in PCL.
• To propose an effective optimization method for PCL.

Main Methods:

• Formulated PCL as a dynamic multi-objective optimization problem.
• Introduced Elastic Multi-Gradient Descent (EMGD) with task-specific elastic factors.
• Developed a gradient-guided memory editing mechanism for rehearsal data.

Main Results:

• EMGD guides gradient updates towards Pareto-optimal directions for balanced learning.
• The memory editing mechanism mitigates interference from rehearsal data.
• Theoretical analysis confirms EMGD's Pareto criticality.

Conclusions:

• EMGD significantly outperforms existing PCL, MTL, and SCL methods on image classification benchmarks.
• The proposed method effectively handles dynamic and asynchronous task arrivals.
• EMGD offers a robust solution for continual learning in complex, evolving environments.