Instance Selection-Based Surrogate-Assisted Genetic Programming for Feature Learning in Image Classification.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Genetic programming (GP) shows promise in image classification feature learning.
• GP-based methods are often computationally expensive due to extensive fitness evaluations, especially with large datasets.

Purpose of the Study:

• To develop a computationally efficient GP approach for image classification feature learning.
• To reduce the high computational cost associated with traditional GP methods.

Main Methods:

• Proposes an instance selection-based surrogate-assisted GP algorithm.
• Utilizes multiple small surrogate training sets of varying sizes to progressively reduce computational load.
• Evaluates the population on small surrogate sets and best individuals on the full training set each generation.
• Learned features are classified using linear support vector machines.

Main Results:

• Significantly reduces computational cost compared to using the entire training set.
• Improves generalization performance across 11 diverse image datasets.
• Outperforms baseline GP and non-GP methods.
• Employing multiple surrogate training sets yields better results than single sets or random instance selection.

Conclusions:

• The proposed instance selection-based surrogate-assisted GP offers an effective and efficient solution for image classification feature learning.
• The method balances computational cost reduction with enhanced generalization capabilities.