cnn_ablation

Understanding CNN Generalization Through Systematic Ablation

This repository contains a small-scale experimental study analyzing how common regularization techniques affect CNN generalization.

Experiments

• Baseline CNN (no regularization)
• CNN with Batch Normalization and Dropout
• CNN with Early Stopping

Dataset

• CIFAR-10

Key Findings

• Baseline models overfit quickly
• Batch normalization stabilizes training
• Early stopping prevents overfitting without architectural changes

Usage

Install dependencies:

Motivation

This project focuses on understanding training dynamics rather than maximizing benchmark accuracy.