Determining the optimal number of hidden neurons in Multilayer Perceptrons (MLPs) is a critical challenge in machine learning, with current methods relying heavily on manual tuning, heuristics, or computationally expensive algorithms. These traditional approaches often lead to overfitting, inefficient models, and limited adaptability across dataset issues that hinder real-world deployment in domains like healthcare, finance, and IoT.
Our technology presents an Automated Hidden Neuron Optimization framework that integrates hybrid feature selection (Mutual Information + Random Forest), k-fold cross-validation with fixed initialization, and early stopping to dynamically identify the ideal MLP structure for classification tasks. Tested on seven real-world datasets from the UCI repository, our approach consistently outperformed baseline models—achieving up to 95.07% accuracy with just 2 hidden neurons in the Breast Cancer dataset and reducing input features by over 80% in some cases. Unlike conventional techniques, this solution offers a scalable, adaptive, and computation-efficient pipeline, making it a game-changer for industries seeking high-performance AI with minimal overhead.
Flowchart of the proposed model