• German
German

Main Navigation

Solving Large Scale Learning Tasks, OH 14, E23

Event Date: October 20, 2016 16:15


In celebration of Prof. Dr. Moriks 60th birthday, the Festschrift ''Solving Large Scale Learning Tasks'' covers research areas and researchers Katharina Morik worked with.

Articles in this Festschrift volume provide challenges and solutions from theoreticians and practitioners on data preprocessing, modeling, learning and evaluation. Topics include data mining and machine learning algorithms, feature selection and creation, optimization as well as efficiency of energy and communication.

Bart Goethals: k-Morik: Mining Patterns to Classify Cartified Images of Katharina

When building traditional Bag of Visual Words (BOW) for image classification, the k-Means algorithm is usually used on a large set of high dimensional local descriptors to build a visual dictionary. However, it is very likely that, to find a good visual vocabulary, only a sub-part of the descriptor space of each visual word is truly relevant for a given classification problem. In this paper, we explore a novel framework for creating a visual dictionary based on Cartification and Pattern
Mining instead of the traditional k-Means algorithm. Preliminary experimental results on face images show that our method is able to successfully differentiate photos of Elisa Fromont, and Bart Goethals from Katharina Morik.

Arno Siebes: Sharing Data with Guaranteed Privacy

Big Data is both a curse and a blessing. A blessing because the unprecedented amount of detailed data allows for research in, e.g., social sciences and health on scales that were until recently unimaginable. A curse, e.g., because of the risk that such – often very private – data leaks out though hacks or by other means causing almost unlimited harm to the individual.
To neutralize the risks while maintaining the benefits, we should be able to randomize the data in such a way that the data at the individual level is random, while statistical models induced from the randomized data are indistinguishable from the same models induced from the original data.
In this paper we first analyse the risks in sharing micro data – as statisticians tend to call it – even if it is anonymized,  discretized, grouped, and perturbed. Next we quasi-formalize the kind of randomization we are after and argue why it is safe to share such data. Unfortunately, it is not clear that such randomizations of data sets exist. We briefly discuss why, if they exist at all, will be hard to find. Next I explain why I think they do exist and can be constructed by showing that the code tables computed by, e.g., Krimp are already close to what we would like to achieve. Thus making privacy safe sharing of micro-data possible.

Nico Piatkowski: Compressible Reparametrization of Time-Variant Linear Dynamical Systems

Linear dynamical systems (LDS) are applied to model data from various domains—including physics, smart cities, medicine, biology, chemistry and social science—as stochastic dynamic process. Whenever the model dynamics are allowed to change over time, the number of parameters can easily exceed millions. Hence, an estimation of such time-variant dynamics on a relatively small—compared to the number of variables—training sample typically results in dense, overfitted models.

Existing regularization techniques are not able to exploit the temporal structure in the model parameters. We investigate a combined reparametrization and regularization approach which is designed to detect redundancies in the dynamics in order to leverage a new level of sparsity. On the basis of ordinary linear dynamical systems, the new model, called ST-LDS, is derived and a proximal parameter optimization procedure is presented. Differences to l1 -regularization-based approaches are discussed and an evaluation on synthetic data is conducted. The results show, that the larger the considered system, the more sparsity can be achieved, compared to plain l1 -regularization.

Marco Stolpe: Distributed Support Vector Machines: An Overview

Support Vector Machines (SVM) have a strong theoretical foundation and a wide variety of applications. However, the underlying optimization problems can be highly demanding in terms of runtime and memory consumption. With ever increasing usage of mobile and embed ded systems, energy becomes another limiting factor. Distributed versions of the SVM solve at least parts of the original problem on different networked nodes. Methods trying to reduce the overall running time and memory consumption usually run in high performance compute clusters, assuming high bandwidth connections and an unlimited amount of available energy. In contrast, pervasive systems consisting of battery-powered devices, like wireless sensor networks, usually require algorithms whose main focus is on the preservation of energy. This work elaborates on this distinction and gives an overview of various existing distributed SVM approaches developed in both kinds of scenarios.



SFB-876 NEWSLETTER
Newsletter RSS Twitter

NEWEST TECHREPORTS