Martin Dallinger

Machine Learning and Pattern Classification

Held by: Univ.-Prof. Dr. Gerhard Widmer

From very basic ideas of (weighted) classification over unsupervised methods to Hidden Markov Models - this course offered a broad variety of common ML methods with deep explanations about their inner-workings.

In the exercise, my team consisting of David Radonic, Lisa Krimbacher, Patrick Schilcher and me won the challenge which was about building an AI voice assistant from a provided dataset.

Unsupervised Techniques (Lecture and Exercise)

Held by: Univ.-Prof. Dr. Sepp Hochreiter

This course covers advanced techniques to find structure in data from a deeply technical and mathematical perspective. This ranges from methods like Factor Analysis (basically PCA/SVD with noise explanations) to bi-clustering methods like FABIA.

Supervised Techniques

Held by: Dr. Johannes Kofler

Here we started with simpler linear models like SVMs (non-linear with the Kernel-Trick) to then unveil a multitude of different models like gradient boosted trees and neural networks with a big focus on detailed mathematical foundation and derivation.

Audio and Music Processing

Held by: Dr. Jan Schlüter

We discovered modern music processing methods which ranged from extracting adequate features (like Mel-Spectrograms or decorrelated MFCCs) to applying algorithms of different difficulty to find the required metric (like onsets, beat, tempo, fundamental frequency). Furthermore, the course covered more advanced concepts which are still not considered solved in research, like some special score following tasks, defining music similarity and multi-pitch estimation.

The exercise part of this course was a competition about tempo, beat and onset estimation which I won with my team consisting of Daniel Schmid, Lukas Gandler and me.

Security Infrastructures

Held by: Univ.-Prof. Priv.-Doz. DDI Dr. Stefan Rass

The lecture part of this course covered the following topics:

• Public-Key Infrastructures
• Certificate Management
• Cryptographic Key Management
• Drive Encryption
• Anonymous Communication
• Secret Key Infrastructures: Quantum Key Distribution

Additionally, I held a presentation about adversarial machine learning, which can be found here.

Programming/Project in Python II

Held by: Dr. Andreas Schörgenhumer

This was a project-based course where in the beginning all students created a big dataset consisting of images, which were then gray-scaled and split to a train/test-dataset by the tutors. The task was image inpainting with rectengular masks.

I chose a Wasserstein-GAN-based approach which yielded the close second place in the competition - the trick I found was focussing on the edges of the mask for the loss (spatial discounting), since they should be easier to predict.

Mathematics for AI III

Held by: Dr. Oliver Roche-Newton

Covering the basis for understanding advanced Machine Learning algorithms in detail: Vector/Normed spaces, hilbert spaces, fourier series and convergence theorems as well as basic measure theory mostly underlined with rigorous proofs and in the complex domain.

Social Media Mining and Analysis

Held by: Univ.-Prof. Dr. Markus Schedl

Group Topic: Investigating on country bias of different recommender systems

This was also a project-base course on which I worked together with Benjamin Bergmann, Jonas Geiger and Max Walder. We found flaws in the implementation of a Bachelor-Thesis regarding de-biasing, corrected these and investigated further: We found indications for some recommendation-algorithms like NeuMF to over-represent US music.

The final publication was then thankfully continued after the next semester by the three other members, as I was forced to opt out for time-reasons, but the findings roughly stayed the same.

Credit: Lesota, O., Geiger, J., Walder, M., Kowald, D., and Schedl, M. Oh, Behave! Country Representation Dynamics Created by Feedback Loops in Music Recommender Systems, Submitted to RecSys 2024