Neural Network from Scratch
Pure NumPy implementation achieving 99.6% MNIST accuracy through optimized gradient descent, backpropagation, and regularization techniques.
Pure NumPy neural network achieving 99.6% MNIST accuracy. Hand-coded gradients, optimized SGD, and regularization — no frameworks, complete algorithmic transparency.
Training Results
Convergence Analysis: Loss decreases smoothly, gradients decay exponentially, learning rate schedule prevents overshooting.
Performance Breakdown: Per-digit accuracy analysis reveals systematic patterns. Digits 8/9 most challenging due to visual similarity.
Activation Function Comparison: ReLU eliminates vanishing gradients, enabling 10× faster training than sigmoid while achieving superior accuracy.
Related projects
Deep Learning from Scratch
Backpropagation, BatchNorm, Dropout, and CNNs implemented from first principles in NumPy — then PyTorch deployment achieving 74.8% on CIFAR-10.
Decision Trees & Ensemble Methods
From-scratch implementation of decision trees with pruning, random forests, and AdaBoost. Comprehensive analysis of overfitting, feature selection, and ensemble performance on real datasets.
Image Geolocation with k-NN & Linear Regression
Computer vision system predicting photo locations from visual features. Combines k-nearest neighbors with regression models, achieving 127km median error on global street-view dataset.