Improving CNN Generalization with PDE Preprocessing and the Variational Information Bottleneck

This paper presents a hybrid approach that integrates pre-defined convolutional layers based on Partial Differential Equations (PDEs) with the Variational Information Bottleneck (VIB) framework to improve image classification performance.While PDE-based layers enhance low-level structural features in the input, they may also introduce redundant information.To address this, we introduce a VIB module after the PDE layers, which learns compressed latent representations that retain only task-relevant information.The proposed architecture is evaluated on the CIFAR-10 dataset using various CNN backbones, including ResNet and VGG.Experimental results show that our method improves classification accuracy while reducing representational complexity.This demonstrates that combining physics-inspired priors with information-theoretic compression offers an effective strategy for enhancing deep neural networks.