GeNeX: Genetic Network eXperts Framework for Addressing Validation Overfitting.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computer Vision

Background:

• Validation overfitting (VO) occurs when models perform well on validation data but poorly on test data.
• This issue is exacerbated in low-data scenarios and under distribution shifts, compromising model reliability.
• Current ensemble methods often rely on validation scores, making them susceptible to VO.

Purpose of the Study:

• To introduce Genetic Network eXperts (GeNeXs), a novel framework designed to mitigate validation overfitting.
• To enhance the robustness and generalization capabilities of ensemble models.
• To improve the reliability of machine learning models in real-world deployment.

Main Methods:

• GeNeXs employs a dual-path strategy: gradient-based training and genetic model evolution for robust model generation.
• Candidate networks are clustered by prediction behavior to identify complementary model spaces for ensemble construction.
• Prototype networks are formed by fusing diverse experts at the weight level, with final ensemble predictions optimized using sequential quadratic programming (SQP).

Main Results:

• GeNeX demonstrated consistent outperformance against state-of-the-art ensembles across four real-world image classification tasks.
• The framework showed strong generalization capabilities under limited-data and shift-aware conditions.
• Experiments confirmed minimal validation overfitting gaps, highlighting GeNeX's resilience.

Conclusions:

• GeNeXs effectively addresses validation overfitting in both model generation and ensemble construction.
• The proposed VO-aware evaluation protocol simulates realistic deployment challenges.
• GeNeX offers a reliable solution for building robust and generalizable machine learning ensembles.