CIFAR-10 Image Classification

Summary

---

• Dataset: CIFAR-10 (50,000 training images, 10,000 test images; 32×32×3 color images across 10 classes)
• Languages & Libraries: Python 3.x, NumPy, Matplotlib

Implementations

k-Nearest Neighbors (kNN)

• Three distance-matrix methods: two-loop, one-loop, and fully vectorized.
• Fully vectorized version achieved a 435× speedup (0.22 s vs. 96.9 s) over the two-loop baseline.
• Performed 5-fold cross-validation over k ∈ {1,3,5,8,10,12,15,20,50,100}; k = 10 yielded the best test accuracy (28.2%, up from ~27%).

Support Vector Machine (SVM)

• Implemented both naïve and fully vectorized loss and gradient computations.
• Verified gradients with numerical checks, then trained via SGD to reach 37.7% train and 38.0% validation accuracy.

Softmax Classifier

• Built looped and vectorized versions of Softmax loss/gradient.
• Tuned learning rate and regularization to achieve 34.3% validation and 33.6% test accuracy.

Two-Layer Neural Network

• Modular implementation with affine, ReLU, and “sandwich” layers.
• Hyperparameter grid search led to >52% validation and ~58% test accuracy on CIFAR-10 .
• Visualized learned weight matrices and misclassified examples for model interpretation.

Visualization

• Training curves, confusion matrices, and weight visualizations are generated via Matplotlib and NumPy scripts.