Student Projects in LAS Group

Are you interested in doing a semester project, a Bachelor thesis or a Master thesis in our group? We offer both topics with a strong research emphasis, as well as more applied projects. The M.Sc. projects have often led to publications in leading conferences. We offer projects in:
  • Large-Scale Machine Learning
  • Network Analysis
  • Recommender Systems
  • Probabilistic Inference
  • Crowdsourcing
  • Active Learning
  • Reinforcement Learning
  • Learning Theory

To learn about the research done in the group consult the research page. Our publications provide more details. Current topics can be downloaded only from the ETH domain.

Current Topics

Robust Bayesian Truth Serum
Robust Bayesian Truth Serum                                                                                           

Previous Master Theses

Incentivizing Users for Balancing Bike Sharing Systems
In this thesis, we tackle the imbalance problems in bike sharing systems such as the unavailability of bikes or parking docks at stations. We design a crowdsourcing mechanism that incentivizes the users in the bike repositioning process by providing them with alternate choices to pick or return bikes in exchange for monetary incentives. We deploy the proposed system through a smartphone app among users of a large-scale bike sharing system in Mainz, Germany. Results of this thesis are published as a conference paper at AAAI 2015.
Coresets for the DP-Means Clustering Problem
Coresets are weighted subsets of an original data set with the property that solutions found on the coreset are competitive with solutions obtained on the full data set. In this thesis, we show the existence of coresets for DP-Means clustering, a prototypical nonparametric clustering problem. We propose a practical coreset construction algorithm and demonstrate its effectiveness on real-world data sets. A continuation of this work was presented and published at ICML 2015.

Higher-Order Inference for Multi-class Log-supermodular Models
Although shown to be a very powerful tool in computer vision, higher-order models are mostly restricted to computing MAP configurations for specific energy functions. In this thesis, we propose a multi-class model along with a variational marginal inference formulation for higher-order log-supermodular models. We evaluate the scalability and the effectiveness of our approach in a natural scene image segmentation task, demonstrating state-of-the-art performance for both marginal and MAP inference. Results of this thesis will appear at ICCV 2015.
High-Dimensional Gaussian Process Bandits
In this thesis we analyze the optimization of high dimensional functions from expensive and noisy samples. We will assume that the variables of the function are very correlated so that it effectively varies only along a low dimensional subspace, and that it is well behaved with respect to our prior beliefs (has a low norm in some reproducible kernel Hilbert space). We introduce a strategy, called SI-BO, that enjoys a theoretical bound on the cumulative regret, and performs well on numerical experiments against several baselines. Results of this thesis appeared at NIPS 2013.

Active Learning for Level Set Estimation
In this thesis, we considered the problem of estimating a specified level set of an unknown function using sequential measurements. We proposed an algorithm that models the function as a Gaussian process, and strikes a balance between exploring new regions of the space, and improving the accuracy of the estimated level set near already-explored regions. We both theoretically analyzed this algorithm, and simulated its performance on limnological data sets from lake Zurich. Part of this work was published at IJCAI 2013.
Reinforcement Learning for Spacecraft Maneuvering using Optic Flow and Time-To-Contact
In this thesis we tackle the task of spacecraft maneuvering near small celestial bodies, like asteroids or comets. We develop two robust non-linear controllers for body-fixed hovering in unknown gravitational environments based on Reinforcement Learning. In particular, we obtain a hovering controller than outperforms the state-of-the-art. This thesis was done in collaboration with the European Space Agency.