Multitask Feature Learning as Multiobjective Optimization: A New Genetic Programming Approach to Image...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Feature learning is crucial for image classification but faces challenges with high image variations and small datasets, leading to overfitting.
• Multitask feature learning can enhance generalization, yet existing methods struggle with partially conflicting tasks.

Purpose of the Study:

• To propose a novel approach for multitask feature learning by framing it as a multiobjective optimization problem.
• To develop a genetic programming method with a new representation for effective feature learning in image classification.

Main Methods:

• A genetic programming approach with a new, compact program representation was developed.
• All tasks share a common solution space, with solutions evaluated across multiple tasks simultaneously within a single population.
• An ensemble of nondominated solutions is created to mitigate overfitting and improve classification accuracy.

Main Results:

• The proposed approach significantly outperforms numerous benchmark methods across 15 image classification datasets.
• The method effectively optimizes multiple, potentially conflicting, tasks simultaneously.
• The ensemble strategy demonstrably reduces the risk of overfitting.

Conclusions:

• Solving multitask feature learning as a multiobjective optimization problem enhances generalization capabilities.
• The novel genetic programming approach with a unique representation offers a powerful solution for complex image classification tasks.