Andreas Müller

Diplommathematiker (Diploma in Mathematics, equiv. M.Sc)
Member of the scientific staff at Autonomous Intelligent Systems Group
Member of the Deep Learning Group

Address: 

   Andreas Müller
   Rheinische Friedrich-Wilhelms-Universität Bonn
   Institut für Informatik VI
   Friedrich-Ebert-Allee 144
   53113 Bonn 

Tel:  +49 (0) 228 73-4175

Email: amueller _at_ ais.uni-bonn.de

Office: N910

Teaching

Currently I am tutoring for Prof. Behnke's lecture on Robot Learning.
Since some students wanted to have the scribbles from the tutorials, I put them online here, the filename refecting date and time of the tutorial. They are probably full of mistakes and inconcistencies. Also they probably don't make much sense by themselves. The book by Sutton and Barto should always be your standard reference and can be trusted to be well reviewed.

Short CV

I received my MS degree in Mathematics (Dipl.-Math.) in 2008 from the Department of Mathematics at the University of Bonn. Since 2009, I am working on my PhD Thesis at the Institute of Computer Science VI at the University of Bonn. Since April 2010, I hold a scholarship of the B-IT Research School.

Research Interests

My research focuses on supervised and unsupervised learning in image data. I am interested in extracting features for modeling image data and for discriminative purposes using Probabilistic Graphical Models. I am also interested in general learning and inference algorithms for Graphical Models and Bayesian approaches to learning.

Current Work

My current work concerns image segmentation and object clustering with nonparametric Bayesian methods. I am trying to build a model that uses a superpixel-based bag-of-words representation to build a hierarchical segmentation into objects and object parts. Whenever I have time I write about my current work and interesting developments in my blog: peekaboo-vision.blogspot.com

Software

CUV library Together with Hannes Schulz and others, I am working on this library, which provides a framework for matrix routines in CUDA and C++, especially designed towards neural networks and Restricted Boltzmann Machines. Using the NVidia Cuda framework it is possible to optain speedups of 20-100 times compared with a optimized single CPU implementation. There are python bindings available for very fast development. Often an existing Python or Matlab project can easily be converted to the cuv library to obtain huge times speedups without any further optimization.

Restricted Boltzmann Machine and Annealed Importance Sampling Release of our RBM code using CUV for large scale RBM learning with an easy interface. Includes annealed importance sampling and exact calculation of the likelihood for small problems.

Python wrappers for VLfeat I continued work of Mikael Rousson to make the vlfeat library by Andrea Vedaldi and Brian Fulkerson usable with python. In particular I included quickshift wrappers that can be used to create image segmentations and superpixels.

Publications

Andreas Müller and Sven Behnke:
Multi-Instance Methods for Partially Supervised Image Segmentation
Accepted for First IAPR Workshop on Partially Supervised Learning (PSL), Ulm, to appear September 2011.

Hannes Schulz, Andreas Müller, and Sven Behnke:
Exploiting Local Structure in Boltzmann Machines
Accepted for Neurocomputing, special issue on ESANN 2010, Elsevier, to appear 2011

Hannes Schulz, Andreas Müller, and Sven Behnke:
Investigating Convergence of Restricted Boltzmann Machine Learning
NIPS 2010 Workshop on Deep Learning and Unsupervised Feature Learning Whistler, Canada, December 2010

Dominik Scherer, Andreas Müller, and Sven Behnke:
Evaluation of Pooling Operations in Convolutional Architectures for Object Recognition
20th International Conference on Artificial Neural Networks (ICANN), Thessaloniki, Greece, September 2010.

Andreas Müller, Hannes Schulz, and Sven Behnke:
Topological Features in Locally Connected RBMs
in the International Joint Conference on Neural Networks (IJCNN 2010)

Hannes Schulz, Andreas Müller, and Sven Behnke:
Exploiting local structure in stacked Boltzmann machines
in European Symposium on Artificial Neural Networks, Computational Intelligence and Machine Learning (ESANN), Bruges, Belgium

Diploma Thesis

I received my Diploma in Mathematics from Rheinische Friedrich-Wilhelms-Universität Bonn.
My Diploma thesis belongs to the field of Arithmetic Algebraic Geometry. It is about classifying a certain kind of singularities in Affine Grassmannians of Linear Algebraic Groups over fields of arbitrary characteristic. Its title is Singularities of Minimal Degenerations in Affine Grassmannians (pdf). Introductions are in both, German and English, but the main part of the thesis is English only.