This project builds a robust CNN from scratch for CIFAR-10 image classification, integrating custom learning rate scheduling, Monte Carlo Dropout for uncertainty estimation, and Temperature Scaling for calibration. Achieved 85.28% test accuracy with well-calibrated confidence scores and strong generalization.
python benchmark deep-learning interactive tensorflow numpy scikit-learn keras cnn seaborn classification matplotlib regularization cifar10 fine-tuning simple-cnn project04
-
Updated
Aug 17, 2025 - Jupyter Notebook