Automatic Kernel Selecion We present an algorithm (Infinite Kernel Learning,IKL) which is capable to learn linear combinations of kernels from very general classes. In fact these classes can be even infinite dimensional, e.g. all Gaussian kernels with p.d. covariance matrix. The algorithm solves a problem directly generalized from the Multiple Kernel Learning framework. project details.

Colour Constancy is the tendency to perceive surface color consistently, despite variations in ambient illumination. In this work we revisted the work on Bayesian Color Constancy and investigated the use of more precice reflectance priors. We collected a new dataset for Color Constancy consisting of different scenes, both indoor and outdoor taken under different illuminations. The code and the dataset are available online. project details.

Industrial Surface Inspection We have designed a method for detecting arbitrary defects in textured surfaces, for which we received the first prize at the 2007 DAGM competition on weakly supervised learning for industrial optical inspection. The code is available through Max Planck Innovation GmbH, Munich

Multiple Instance Learning is a learning framework which can cope with ambiguous data. Label information is not available on a data point level but rather for a bag of possibly more than one data point. We are designing learning algorithms which are able to cope with this ambiguity. We are also exploring how tasks such as object recognition or named entity recognition can be cast as MIL problems and solved within this framework. project details.

Rate Adapting Poisson Model for Information retrieval The Rate Adapting Poisson Model is an undirected probabilistic graphical model. It is designed to model count data by mapping it to a lower dimensional representation also understood as a topic representation. In experiments on benchmark text data we compare this model to its directed counterparts such as PLSI. project details.

Product of Edgeperts denoising. Product of Edgeperts are a probabilistic model for higher order dependencies of wavelet transformed images. We present an efficient denoising algorithm which yields state-of-the-art denoising results on a benchmark set of images. The algorithm is "blind" i.e. the parameters are estimated on the noisy version of an image alone. Matlab code and more can be found here: project details.

Matlab Code and Datasets Here you can find several code snippets I wrote during several projects. You can find an implementation of pLSI, ePCA, and fast k-means clustering code code website.