Avoiding overfitting in multilayer perceptrons with feeling-of-knowing using self-organizing maps

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Overfitting is a significant challenge in multilayer perceptron (MLP) training, leading to poor generalization on unseen data.
• Current on-line learning methods for MLPs often struggle to effectively mitigate overfitting.
• The feeling-of-knowing (FOK) phenomenon offers a potential mechanism for regulating learning processes.

Purpose of the Study:

• To develop and evaluate a novel method for avoiding overfitting in on-line MLP training.
• To investigate the efficacy of incorporating a feeling-of-knowing (FOK) mechanism, guided by self-organizing maps (SOMs), into MLP learning.
• To improve the generalization performance of MLPs in on-line learning scenarios.

Main Methods:

• Proposed a modified MLP architecture that integrates a FOK mechanism calculated using SOMs.
• The learning process in the proposed MLP dynamically adjusts based on the computed FOK degree.
• Employed standard MLP training protocols for comparative analysis.

Main Results:

• The proposed MLP with FOK using SOMs demonstrated a significantly lower mean square error on the test set compared to conventional MLP methods.
• The FOK-guided learning process effectively controlled the model's complexity and prevented overfitting.
• Empirical results indicate superior performance and generalization capabilities of the novel approach.

Conclusions:

• The integration of FOK, calculated via SOMs, provides an effective strategy to combat overfitting in on-line MLP training.
• The proposed method offers a promising alternative to existing techniques for improving the robustness and reliability of MLPs.
• This research contributes to the development of more stable and generalizable artificial neural network models.