Enhancing Efficiency and Regularization in Convolutional Neural Networks Strategies for Optimized Dropout
Mehdi Ghayoumi · Preprints.org · 2025
This study explores dropout optimization in Convolutional Neural Networks (CNNs), aiming to beat traditional approaches in regularization and efficiency. We intro- duce dynamic, context-aware strategies, embodied by Probabilistic Feature Importance Dropout (PFID). This method modifies dropout rates to the unique learning phase of CNNs, integrating adaptive, structured, and contextual dropout techniques. Experimentation, benchmarked against current state-of-the-art methods, demonstrates improvements in network performance, particularly in generalization and training efficiency. The findings represent a more adaptable and robust CNN model for complex datasets and computational landscapes.