Pareto Self-Paced Learning Based on Differential Evolution.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Optimization Algorithms

Background:

• Self-paced learning (SPL) ranks data from easy to complex, assigning weights.
• Current SPL methods use pace parameters that are difficult to tune and lack backward adjustment capabilities.
• Existing SPL regimes struggle with iterative parameter tuning and cannot revert to previous states when model performance degrades.

Purpose of the Study:

• To propose a novel Pareto self-paced learning (PSPL) approach to overcome limitations of existing SPL methods.
• To address the challenges of pace parameter tuning and the inability to backtrack in current SPL regimes.
• To introduce a method that optimizes the self-paced function and weighted loss simultaneously without pace parameters.

Main Methods:

• Framing the SPL problem as a multiobjective optimization task.
• Utilizing a Pareto-based differential evolution algorithm for simultaneous optimization.
• Designing a new representation for nonpositive sample weights and a modified mutation operator inspired by SPL ranking and weighting.

Main Results:

• The proposed PSPL approach eliminates the need for pace parameters.
• PSPL obtains an entire solution spectrum, enabling automatic adjustment of the solution path.
• Demonstrated effectiveness through rigorous studies on matrix factorization and multiclass classification problems.

Conclusions:

• PSPL offers an effective alternative to traditional SPL methods by treating it as a multiobjective optimization problem.
• The absence of pace parameters and the ability to explore the solution spectrum enhance flexibility and performance.
• PSPL shows significant potential for improving machine learning tasks like matrix factorization and classification.