Upcoming events

On Rationality of Nonnegative Matrix Factorization

James Worrell
University of Oxford
SWS Distinguished Lecture Series
02 May 2017, 10:30 am - 1:30 pm
Saarbrücken building E1 5, room 002
simultaneous videocast to Kaiserslautern building G26, room 112
The nonnegative rank of a nonnegative m x n matrix M is the smallest number d such that M can be written as the product M = WH of a nonnegative m x d matrix W and a nonnegative d x n matrix H.  The notions of nonnegative rank and nonnegative matrix factorization have a wide variety of applications including bioinformatics computer vision communication complexity document clustering and recommender systems. A longstanding open problem is whether when M is a rational matrix the factors W and H in a rank decomposition M=WH can always be chosen to be rational.  In this talk we resolve this problem negatively and discuss consequences of this result for the computational complexity of computing nonnegative rank.

This is joint work with Dmitry Chistikov Stefan Kiefer Ines Marusic and Mahsa Shirmohammadi.

Digital Knowledge: From Facts to Rules and Back

Daria Stepanova
MPI-INF - D5
Joint Lecture Series
03 May 2017, 12:15 pm - 3:15 pm
Saarbrücken building E1 5, room 002
Knowledge Graphs (KGs) are huge collections of primarily encyclopedic facts which are automatically extracted from the Web. Prominent examples of KGs include Yago DBPedia Google Knowledge Graph. We all use KGs when posing simple queries like "capital of Saarland" to Google. Internally such queries are translated into machine readable representations which are then issued against the KG stored at the backend of the search engine. Instead of syntactically relevant Web pages the actual answer to the above query "Saarbrücken" is then output to the user as a result.

However since KGs are automatically constructed they are often inaccurate and incomplete. In this talk I will investigate how deductive and inductive reasoning services could be used to address these crucially important issues. More specifically first I will present an approach for repairing inconsistencies in hybrid logical systems that can be built on top of KGs. Second I will describe a method for inductive learning of rules with exceptions from KGs and show how these are applied for deriving missing facts.

Digital Knowledge: From Facts to Rules and Back

Daria Stepanova
MPI-INF - D5
Joint Lecture Series
03 May 2017, 12:15 pm - 3:15 pm
Saarbrücken building E1 5, room 002
Knowledge Graphs (KGs) are huge collections of primarily encyclopedic facts which are automatically extracted from the Web. Prominent examples of KGs include Yago DBPedia Google Knowledge Graph. We all use KGs when posing simple queries like "capital of Saarland" to Google. Internally such queries are translated into machine readable representations which are then issued against the KG stored at the backend of the search engine. Instead of syntactically relevant Web pages the actual answer to the above query "Saarbrücken" is then output to the user as a result.

However since KGs are automatically constructed they are often inaccurate and incomplete. In this talk I will investigate how deductive and inductive reasoning services could be used to address these crucially important issues. More specifically first I will present an approach for repairing inconsistencies in hybrid logical systems that can be built on top of KGs. Second I will describe a method for inductive learning of rules with exceptions from KGs and show how these are applied for deriving missing facts.

Towards an Approximating Compiler for Numerical Computations

Eva Darulova
Max Planck Institute for Software Systems
Joint Lecture Series
07 Jun 2017, 12:15 pm - 3:15 pm
Saarbrücken building E1 5, room 002
Computing resources are fundamentally limited and sometimes an exact solution may not even exist. Thus when implementing real-world systems approximations are inevitable as are the errors introduced by them. The magnitude of errors is problem-dependent but higher accuracy generally comes at a cost in terms of memory energy or runtime effectively creating an accuracy-efficiency tradeoff. To take advantage of this tradeoff we need to ensure that the computed results are sufficiently accurate otherwise we risk disastrously incorrect results or system failures. Unfortunately the current way of programming with approximations is mostly manual and consequently costly error prone and often produces suboptimal results.

In this talk I will present our vision and efforts so far towards an approximating compiler for numerical computations. Such a compiler would take as input exact high-level code with an accuracy specification and automatically synthesize an approximated implementation which is as efficient as possible but verifiably computes accurate enough results.

Your Photos Expose Your Social Circles - Social Relation Recognition from 5 Social Domains

Qianru Sun
MPI-INF - D4
Joint Lecture Series
05 Jul 2017, 12:15 pm - 3:15 pm
Saarbrücken building E1 5, room 002
Social relations are the foundation of human daily life. Developing techniques to analyze such relations in visual data such as photos bears great potential to build machines that better understand people at a social level. Additionally through better understanding about such hidden information in exposed photos we would like to inform people about potential privacy risks. Social domain-based theory from social psychology is a great starting point to systematically approach social relation recognition. The theory provides a coverage of all aspects of social relations and equally is concrete and predictive about the visual attributes and behaviors defining the relations in each domain. Our work provides the first photo dataset built on this holistic conceptualization of social life that is composed of a hierarchical label space of social domains and social relations and contributes the first models to recognize such domains and relations and find superior performance for attribute based features. Beyond the encouraging performances we have some findings of interpretable features that are in accordance with the predictions from social psychology literature. Our work mainly contributes to interleave visual recognition and social psychology theory that has the potential to complement the theoretical work in the area with empirical and data-driven models of social life.