January 1, 2025Open Access

Exploring the Intersection of Information Theory and Machine Learning

YJYousef JaradatAl-Zaytoonah University of Jordan MMMohammad MasoudAl-Zaytoonah University of Jordan AMAhmad ManasrahAl-Zaytoonah University of Jordan

Key Points

Improved classifier accuracy was observed at 95.1% by utilizing mutual information for feature selection, indicating the effectiveness of information theory.
Cross-entropy loss in neural networks led to high accuracy of 0.98 during training, suggesting faster convergence compared to traditional loss metrics.
Application of KL-divergence in structuring latent spaces effectively promoted compact representations in generative models, proving its utility in machine learning.
The information bottleneck principle yielded competitive test accuracy while reducing training time by over 90 seconds, highlighting its role in efficient learning.

Abstract

This study addresses the need for a unified framework demonstrating Information Theory’s (IT) pervasive impact across diverse Machine Learning (ML) tasks. We investigate how IT principles-including entropy, Mutual Information (MI), cross-entropy, KL-divergence, and Information Gain (IG)-rigorously guide ML model design, optimization, and interpretability. Our approach combines theoretical elucidation with empirical validation on standard benchmarks. IT enhances feature selection; for instance, MI-ranked features in the breast cancer dataset improved classifier accuracy to 95.1% (top 20) and 93% (top 5), outperforming F-score selection. It also improves model training; cross-entropy loss in Neural Networks (NNs) for Iris classification led to faster convergence and high accuracy (0.98 training, 0.95 validation), surpassing MSE loss. For generative models, KL-divergence effectively structures Variational Auto-Encoder (VAE) latent spaces from Modified National Institute of Standards and Technology (MNIST) data, promoting compact, continuous representations ideal for generation. Finally, the Information Bottleneck (IB) principle, applied to Canadian Institute For Advanced Research (CIFAR-100), yielded competitive test accuracy (51% vs. 50% for baseline Convolutional Neural Network. (CNN)) and reduced training time (925.02s vs. 1015.75s), highlighting its efficacy in learning compressed, predictive representations. These findings collectively underscore its continued crucial role as a unifying paradigm for addressing fundamental challenges in the evolving ML ecosystem, providing solutions for feature selection, model robustness, and generalization

KI fragen

Bookmark

View Full Paper