ML Image Classification
ResNet-50 Fine-Tuning & Interpretability
Improved model accuracy from 81% to 94.2% through systematic fine-tuning and data augmentation.
Context
A from-the-fundamentals computer-vision project: fine-tune ResNet-50, then push accuracy the disciplined way — one controlled experiment at a time — and use Grad-CAM to see what the model was looking at, not just whether it was right. Going from 81% to 94.2% came from reading the failures, not throwing data at it.
The problem
Baseline model accuracy was 81%, insufficient for production use. The challenge was improving accuracy through principled techniques while maintaining interpretability — understanding not just whether the model was right, but why.
How I built it
Implemented targeted data augmentation based on failure analysis of misclassified samples
Applied cosine learning-rate scheduling for stable convergence
Used class-weighted loss to handle class imbalance without oversampling
Added strategic dropout regularization to prevent overfitting
Built Grad-CAM visualization pipeline for model interpretability and failure diagnosis
Why these choices
Grad-CAM for interpretability
Accuracy metrics alone don't reveal whether a model is learning meaningful features. Grad-CAM visualizations showed where the model was attending, enabling targeted improvements.
Transfer learning with selective unfreezing
Full fine-tuning risks catastrophic forgetting on small datasets. Selective unfreezing of later layers preserved learned features while adapting to the target domain.