10-601 Homework Repository

Contains code for SVM, Logistic Regression and Neural Networks. Written for the Fall 2014 instance of 10-601 Machine Learning at CMU.

Directory structure

Folder	Description
AE	Autoencoder code. Somewhat messy and probably not useful for use in a homework.
Kernels	Contains code to run Kernelized versions of LR or SVM. Both are trained using gradient methods (either SGD or L-BFGS).
LR	Logistic Regression code. Can be used to train and test a LR classifier (or SVM).
NN	Neural Network code.
data	Contains some datasets used by the different classifiers.
shared	Some helper functions used all over the code.

Getting Started

Setup minFunc.

This tool is used for solving the optimization problems. Open Matlab or Octave in the root directory, then type the following:

cd ./shared;
addpath ./minFunc;
mexAll ./minFunc;
rmpath ./minFunc;
cd ..

The output should look like:

Compiling minFunc files (octave version)...
mcholC compiled
lbfgsC compiled
lbfgsAddC compiled
lbfgsProdC compiled
Done.

Train and test sample data

Most of the directories contain a function runX, such as runLR or runNN which will generate some data, train the model, and display the decision boundary. Look at the README in Kernels, for some detailed examples.

Train and test MNIST

Most of the directories contain a function runDigits, which will train and test the relevant classifer on the MNIST dataset and display the misclassified digits.