Unsupervised Learning Techniques

By Sacha Medaer and Soroosh Tayebi Arasteh | سروش طیبی آراسته

This project contains programming exercises of the Pattern Analysis course (SS 2019, Dr.-Ing. Christian Riess) offered by the Pattern Recognition Lab (LME) of the Computer Science Department at University of Erlangen-Nuremberg (FAU).

Prerequisites

To get the repository running, you will need several packages such as NumPy/SciPy, Matplotlib or scikit-learn.

You can obtain them easily by installing the conda environment file included in the repository. To do so, run the following command from the Conda Command Window:

$ conda env create -f environment.yml
$ activate PatternAnalysis

Note: This step might take a few minutes

Contents

The main goal of the project is to illustrate different algorithms and to try them out in some real-world applications.

Overview of the project:

• Density Estimation: Implementation of the Parzen Window using different Kernels for probability distribution estimations.

• HMM Signature Verification: Application of Hidden Markov Models to perform human signature verification in order to distinguish original from fake signatures.

• K-Means and Gap Statistics: Implemenation of the K-Means clustering algorithm, overview and explanation of the model selection problem, and implementation of Gap Statistics to find out the optimal K for the algorithm.

• Unsupervised as Supervised: Here I illustrate how density estimation (inherently unsupervised) can be performed as a regression task (supervised learnning) using an auxiliar distribution and a Random Forest Regressor.

Acknowledgment

Part of this README file is written based on the Angel Villar's project.