This project implements one-vs-all (OvA) classification using logistic regression and a pre-trained neural network to recognize handwritten digits (0-9) from the MNIST dataset.
For K classes, K separate binary classifiers are trained. Each classifier distinguishes one class from all others. Prediction selects the class with the highest probability.
A two-layer neural network with pre-loaded weights is used for forward propagation (prediction only). The network architecture is:
- Input layer: 400 units (20x20 pixel images)
- Hidden layer: 25 units with sigmoid activation
- Output layer: 10 units (digits 0-9)
| File | Description |
|---|---|
sample3.m |
Main script: one-vs-all logistic regression |
sample3_nn.m |
Main script: neural network prediction with pre-trained weights |
lrCostFunction.m |
Regularized logistic regression cost function |
oneVsAll.m |
Trains K binary classifiers |
predictOneVsAll.m |
Predicts using one-vs-all classifiers |
predict.m |
Neural network forward propagation |
displayData.m |
Displays digit images in a grid |
fmincg.m |
Conjugate gradient optimization |
ex3data1.mat |
MNIST handwritten digit dataset |
ex3weights.mat |
Pre-trained neural network weights |
- One-vs-All: Training set accuracy of ~95.0%
- Neural Network: Training set accuracy of ~97.5%
Top-left: Sample handwritten digits. Top-right: Per-class accuracy. Bottom-left: Confusion matrix. Bottom-right: Learned weights for each digit class.
Exercises from Andrew Ng's Machine Learning course on Coursera, completed by Keivan Hassani Monfared.