Description
Designed and trained a custom image classification model to identify dog breeds, achieving 80% test accuracy.
Process
Combined techniques like ReLU activation, max-pooling, dropout, early stopping, and manual weight initialization to achieve 80% test AUROC. Improved accuracy on unseen images by applying transfer learning, freezing pre-trained convolutional layers and fine-tuning fully connected layers.
Learnings
Learned how to balance model complexity and generalization to improve performance on unseen data, and developed intuition for when to apply transfer learning versus training from scratch. Gained a deeper understanding of how different techniques like dropout, early stopping, and layer freezing affect model stability and convergence.