Location: University of Bremen, Cartesium, Rotunde
Fri 15 Sep 2017, 15:15
Parallelization of agent-based models for scalable micro-simulation and data sciences
Prof. Munehiro Fukuda, University of Washington Bothell
Agent-based models (ABMs) are intended for executing agent micro-interaction to observe their emergent collective group behavior as their computational results. This talk discusses about two ABM research opportunities: (1) ABM parallel simulation - how to scale up ABM simulations and ease their programmability and (2) ABM application to data sciences - how to use agents for scientific data analysis and machine learning.
For ABM simulation, NetLogo and Repast Symphony facilitate sequential execution platform on top of JVM, which attract many scientists with their user-friendly GUI but suffer from scaling up ABM simulation. Parallelization is a solution to enable CPU and memory scalability. While RepastHPC implemented such a C++ based parallel execution environment, the system burdens users with its very steep learning curve. We are addressing its programing difficulty by implementing our own parallel multi-agent spatial simulation (MASS) library in C++ and applying MASS to more scalable and practical applications.
We are also applying MASS to data sciences by porting it to Java that is of course slower than C++ due to its nature of interpretation but yet can provide users with distributed memory space over a cluster system and facilitate good GUI, which would benefit users in data science. To see the MASS applicability, we have parallelized climate analysis and biological network motif search using MASS-Java. We are also planning to use MASS for gradient descend, K-means clustering, and other optimizations such as ant colony optimization ad grasshopper optimization algorithm.
Wed 18 Jan 2017, 16:00
What's not on the map (and what that means for intelligent technologies)
Prof. Brent Hecht, Northwestern University
Intelligent technologies often rely heavily on large datasets of crowdsourced information. Focusing on the important domain of crowdsourced geographic information (e.g. geotagged social media, citizen science observations, OpenStreetMap contributions), I will show that due to under-coverage biases in these large crowdsourced datasets, intelligent technologies are often less effective and accurate for some people (e.g. people in poor and rural areas) than for others (e.g. people in urban and rich areas). I will also discuss work that points to an exciting potential solution to this problem: social science theory. Specifically, I will show that by operationalizing several well-known theories from the discipline of human geography and combining these with approaches based on crowdsourced datasets, we were able to create an intelligent technology (a semantic relatedness algorithm) that has state-of-the-art performance and is more robust against under-coverage concerns. I will close by discussing how addressing under-coverage biases can be an important role for traditional social science theory in our “big data” era, as well as by providing an overview of several related projects.
Fri 16 Dec 2016, 15:15
What is a cognitive map? - Unraveling its mystery using robots
Prof. Wai Yeap, Auckland University of Technology
The theory of cognitive mapping in spatial cognition posits that many different species learn a map of its environment. However, in recent years, the theory has been seriously challenged while a 2014 noble prize has been awarded for the discovery of the neural substrate for such a map. In this talk, I hope to unravel the mystery of cognitive mapping using robots. First, I present a novel parsimonious view-based model of cognitive mapping and second, empowering a robot with it, I show what is computed is not a single global metric map which is problematic but rather a series of disjointed global maps. The resulting process bears many aspects of cognitive mapping and provides insights into how cognitive mapping could have evolved in different species, from insects to humans, and why our perception of the world is stable.
Fri 25 Nov 2016, 15:15
Design to encourage environmentally conscious consumer behavior: Affordances for product use and reuse
Prof. Li Shu, University of Toronto
Much effort in sustainability has focused on developing more resource-efficient technologies. However, human behavior is clearly relevant to resource conservation. Technically efficient products may even cause consumers to be more complacent about their use, such that overall resource consumption continues to rise after an initial decline, a phenomenon reported as the rebound effect. That is, resource-efficient devices may be used longer and remain left on unnecessarily more so than their less-efficient predecessors. Such behavior offsets at least part of the anticipated gains in resource efficiency intended by resource-saving modes and technologies. Therefore, the described work approaches sustainability by designing products that encourages and enables more resource-efficient behaviors.
Interventions for pro-environmental behavior are often described with respect to the degree of user versus product control. At the informing end, information / feedback allows the user to be in full control. At the determining end, forcing / automatic performance of desired actions gives control to the product. In the middle, persuading includes enabling, encouraging, guiding, and steering, often through physical affordances. This talk focuses on the role of affordances on product use and reuse.
Fri 26 Aug 2016, 15:15
Two non-deductive spatial reasoning problems
Dr. Jae Hee Lee, University of Technology Sydney
In this talk I want to present approaches for two spatial reasoning problems that I worked on recently:
The first problem is concerned with finding the best hypothesis about the spatial structure of a moving complex region, when only partial information about the region is available through different coarse sensor measurements.
The second problem is about predicting the change of an evolving spatial region when some past observations of the region boundary are available.
Detailed descriptions of the problems and approaches can be found in http://ijcai.org/Abstract/15/133 and http://ijcai.org/Abstract/16/156.
Fri 12 Aug 2016, 15:15
Spatiotemporal Bayesian networks for malaria prediction
Prof. Peter Haddawy, Mahidol U
Malaria remains a global public health problem with an estimated 214 million cases of malaria globally in 2015 and 438,000 malaria deaths. Since malaria is prevalent in less developed and more remote areas in which public health resources are often scarce, prediction and targeted intervention are essential elements in effective malaria control. While a diversity of modeling technique have been used to create predictive models of malaria, no work has made use of Bayesian networks. Bayes nets are attractive due to their ability to represent uncertainty, model time lagged and nonlinear relations, and provide explanations of inferences. In this talk I describe our work on spatiotemporal Bayes nets for malaria prediction. I demonstrate the approach with a village level model with weekly temporal resolution for Tha Song Yang district in northern Thailand. The network is learned using data on confirmed cases and environmental covariates. Out of sample evaluation shows the model to have high accuracy for one and two week predictions. While models for individual villages can be built by hand, modeling of some aspects of transmission such as spatial autocorrelation can introduce complexity that results in models with hundreds of nodes that are too large to build manually. I describe a solution to this in which we store model fragments representing local relationships in a library in the form of temporal probability logic sentences and automatically construct networks in response to features in a GIS. Finally, I describe our ongoing work to use these predictive models to interpret crowdsourced data.
This work is carried out in collaboration with the Center of Excellence for Biomedical and Public Health Informatics, Mahidol University.
Mon 04 Jul 2016, 16:00
Intelligent simulation environments for training of dental surgery
Prof. Peter Haddawy, Mahidol U
Surgical skill training has been traditionally based on the Halstedian apprenticeship model. But shortened training programs, reduced working hours for residents and limitations on available operating room time have strained this model and spurred interest in use of simulation for training. Surgical simulation has the potential to provide students with increased training time outside the operating room, objective assessment of procedures, and formative feedback without the need for close direct supervision of experts in limited supply. While numerous simulators exist for training of technical skills and some for non-technical skills, realizing the full potential of such environments requires development of more sophisticated analysis and feedback mechanisms than exist in most simulators.
In the first part of my talk, I describe our work on developing an intelligent VR environment for training of technical skills in dental surgery, focusing on design of formative feedback mechanisms. Effective formative feedback requires assessment of outcome and correlation of procedure to outcome, coupled with the ability to communicate in a language natural to dental students. I describe our haptic VR dental simulator, the outcome scoring component, as well as the initial steps in development of the component that relates procedure to outcome. I present the results of evaluation studies of some feedback mechanisms as well as of the outcome scoring component.
In the second part of my talk, I describe our work on building a simulator for teaching decision making skills in dental surgery. The work began with an observational study of teaching in the OR to identify the types of decisions made as well as the teaching strategies used by human experts. Findings of the study have driven the design of a prototype system called Desitra which integrates a simulation environment with a classical intelligent tutoring architecture. The work addresses teaching of routine decisions as well as handling of emergency situations. We make use of tablets and low cost VR components such as Google cardboard in order to make the developed system maximally accessible. I discuss our ongoing work on modeling the surgical process in order to automate the generation of tutorial intervention. I conclude my talk with a discussion of outstanding challenges and directions for future research.
Thu 02 Jun 2016, 14:15
A method to bridge the existing gap between usability and access control
Prof. Delphine Reinhardt, U Bonn
Access control is a key principle to protect user privacy online. The combination of both the wealth of user-generated data in online social networks and overly complex user interfaces lead to a high user burden for privacy control, hence making the observance of the above principles difficult. We investigate how communication metadata on smartphones can facilitate providing tailored suggestions for restricted audience groups, thus limiting the sharing of data to the intended users only. To this end, we have performed a user study collecting a dataset including contact names, calls, SMS, MMS, and e-mail on personal smartphones in everyday use. We further perform an explorative questionnaire-based study with 42 participants. Our results confirm that users are overtaxed with existing schemes. We identify the expectations and preferences of users, thus facilitating the design of improved solutions.
Wed 04 May 2016, 14:15
What is the future trend of QSR? - Current topics from my QSR research
Prof. Kazuko Takahashi, Kwansei Gakuin University
Qualitative Spatial Reasoning (QSR) is a promising approach that treats spatial data with little computational complexity. Lots of representations or ontologies have been proposed focusing on various aspects of spatial data. I have studied QSR for these years and found lots of interesting results. Among these I'd like to introduce three current topics.
1. Formalization of PLCA PLCA is a framework for QSR focusing on connected patterns of regions. Spatial data is represented using four objects: Point, Line, Circuit and Area, and their incident relations. First, we show the condition of planarity of PLCA expression. Then, we define its inductive construction, and the obtained one satisfies the planarity condition. It corresponds to a subdivision of a finite two-dimensional plane allowing a region with a hole as a piece of subdivision. Formalization and proofs are given by a proof assistant Coq.
2. Superposition of rectangles We consider a treatment of qualitative rectangles. Each unit rectangle has two parts: white region is a part which should be visible and black region is a part which may be hidden. We propose an algorithm of superposition of two rectangles so that all the white parts are visible and black parts are hidden as far as possible. The algorithm is sound and complete, and efficiently produce a best solution.
3. Event extraction from video data We show a representation of relative positional/directional relationship of objects extracted in a rectangle form from a video data. Focusing on specified objects, we make a sequence of such representations. We also define events as sequences of such relationships. and extract what events occurred in the video. We evaluated our method using a video of football games.
Fri 29 Apr 2016, 15:30
Semantic annotation for medieval cartography: The example of the Behaim Globe of 1492
Prof. Günther Görz, University of Erlangen-Nuremberg
The Behaim Globe ("Erdapfel") of 1492 is the oldest extant globe of the earth. It is an early masterpiece of various scientific and technological achievements; nowadays it is a famous exhibit of the Germanic National Museum in Nuremberg. It is a manifestation of the city republic's interest in an encyclopedic image of the world in the spirit of humanism. The globe's map image is primarily Ptolemaic, including elements of medieval universal cartography and of portulans. Its luxurious decoration shows 100 pictorial illustrations plus 60 banners and coats of arms, more than 2000 place names, and more than 50 long legends. The Behaim Globe is one of the very few existing cartographical works where different traditions of late medieval mapmaking are bound together.
We report on an ongoing research project, aiming at a digital as well as a printed edition of the globe on the basis of a new (2011) high-resolution digital photographic record and a 3D reconstruction. Furthermore, there is a database of digitized analog images of its surface taken in 1990 and black-and-white images from 1940. Based on these image data, a comprehensive catalogue of all visually relevant places including text fields is built up by means of a domain ontology for medieval cartography in description logics (OWL-DL). Up to now, nearly 3000 instances (catalogue entries and comments) have been created. These instances are the core of the planned edition. Generic concepts and properties for objects, time and space, events, actors, processes, etc., are represented in a reference ontology, such that domain specific concepts are derived from the generic ones. The ICOM/CIDOC's Conceptual Reference Model CRM (ISO 21227) is such a reference ontology; our implementation "Erlangen CRM" provides a semantic base for the cartographic domain ontology, suitable for automatic reasoning. Furthermore, the globe provides a georeferential organization of contemporary knowledge which opens up a new dimension for semantic disclosure. Historical maps are cognitive maps in the first place, which require a formal qualitative representation of (abstract) regions, their relative positions, but also direction, orientation, and distance. The integration of such representations within a general logical framework leads to a system of hybrid reasoning for processing complex spatial queries over the Behaim globe instance base. The application of formal Semantic Web methods allows for new access paths to the rich information on the globe and its context.
Wed 13 Apr 2016, 14:15
Supporting consistencies in multi-language knowledge sharing
Amit Pariyar, Kyoto University, Japan
The goal of this research is to support multi-language knowledge sharing with a focus on content consistency among local, regional and international communities. Due to the rapid growth of online collaboration, the probability of sharing inconsistent content in various languages and communities is increased. This is problematic for multi-language knowledge sharing system because (i) it is not practical to explicitly state consistency rules in advance for each content shared among communities and (ii) there exists constraint to support content consistency from diverging view among communities.
To this end, this research proposes techniques and guidelines on content consistency to support knowledge sharing among multi-language communities. To leverage knowledge equally a process-based approach is proposed to detect inconsistent content shared among communities. The proposed approach is not language specific as it is based on the concept of synchronizing user editing actions. To support customization in knowledge sharing we suggested guidelines for web managers to promote consistency while sharing specific content categories among communities within and beyond geographic regions. For this purpose we proposed an approach based on propagation to qualitatively compare the content belonging to specific categories in webpages of country-specific websites and examined their propagation in (a) website graph and (b) website pairs. Further, we used the same approach to qualitatively compare webpages of country-specific websites from several geographic regions and examined their propagation (a) within the regions and (b) among the regions. Preferences among communities for sharing specific content categories and for specific geographic regions were revealed which are useful to avoid regional discrepancies. Later we proposed a simple approach to propagate content updates consistently among communities with pattern of sharing.
From the techniques that are simple and applicable in variety of languages along with the guidelines for web managers to deal with inconsistencies in (a) multilingual content and (b) cross-site content, this research contributed to the design of a multi-language knowledge sharing system that caters to the goals of communities which require both for leveraging knowledge equally and customization in knowledge sharing.
Tue 15 Mar 2016, 11:15 am
Some key features of evolution of minds
Prof. Aaron Sloman, University of Birmingham
I suspect that if Alan Turing had not died two years after publication of his paper on The Chemical Basis of Morphogenesis (in 1952) he might have worked on a project to explain how minds were produced from molecules on a lifeless planet: the Meta-Morphogenesis project. He didn't, so I am trying. As a partial progress report, I'll present some ideas about conjectured strands in the evolutionary process linking requirements for animal intelligence --- including visual perception, intention formation, planning, intelligent action, collaboration, learning, remembering, being puzzled, and more --- the use of *internal* languages, with structural variability, potentially unbounded complexity, and compositional semantics. Such internal languages seem to be required for many forms of non-human intelligence (e.g. in squirrels, weaver birds, elephants, orangutans, cetaceans, as well as pre-verbal human toddlers). I'll discuss some of the implications for evolution of language and vision and, if there's time, evolution of minds like Euclid's. There are implied deep criticisms of current AI theories and models (whether in or out of fashion). But progress can be made if we don't keep promising too much too soon, and if we educate researchers more broadly.
Mon 14 Mar 2016, 04:15 pm
CoInvent: A computational model of creativity based on concept blending and analogy
Prof. Enric Plaza, IIIA - Artificial Intelligence Institute, CSIC
Concept blending, a cognitive process which allows for the combination of certain elements (and their relations) from originally distinct conceptual spaces into a new unified space combining these previously separate elements and allowing the performance of reasoning and inference over the combination, is taken as a key element of creative thought and combinatorial creativity. This talk presents the development of a computational-level and algorithmic-level account of concept blending and analogy based on the notion of amalgams, and their relation to inductive generalization and learning.
Fri 18 Dec 2015, 03:30 pm
The aspect of structure in GIScience and related disciplines
Franz-Benjamin Mocnik, Vienna University of Technology
Natural science, e.g. physics, is highly successful in describing nature and predicting the future. One reason behind becomes apparent when we investigate the history of natural science: the aspect of structure has been highly relevant for the development of laws and theories, e.g. Maxwell’s equations, and the aspect of structure has also been decisive for the continuity of formal descriptions under paradigm shifts, e.g. for the transition from Fresnel’s aether theory to Maxwells’s theory of the electrodynamic field. It is yet unclear how to lift the scientific discussion in GIScience and related disciplines to a structural level.
The talk approaches the question of how to lift the discussion to a structural level, by outlining two examples of structural discussions in GIScience: first, the structure of maps and texts is discussed. Both modes of representation emphasize different aspects and thus offer different advantages. Texts can, for example, tell stories, but telling stories and transporting emotions by maps is much harder. The talk argues how these differences between maps and texts can be explained by their structure.
The second example approaches the question of how to lift concepts of space, at least in parts, to a structural level. Different concepts of space and thus of spatial information exist, and it is not clear whether these different concepts share a common structure. The talk argues that such a common structure, called the spatial structure, seems to exists. A model of this spatial structure is introduced, and the talk discusses how this structure can be used as a prototypical model of spatial information. Finally, an outlook is given on how to approach algorithmic improvements for spatial information by the use of this model.
Tue 29 Sep 2015, 10:30 am
Different points of view on the spatio-temporal analysis of human activities
Dr. Eris Chinellato, University of Leeds
New, information rich datasets regarding human activities at different space and time scales are produced daily but, despite the rapidly increasing availability of data, and the many potential applications this envisages, there seem to be no surveys or general purpose studies summarising and comparing the various types of data analysis that can be applied to human activity data.
We aim at covering such a gap by building a comprehensive collection of measures and analysis techniques for studying human activity data, able to cover as many different type of datasets, sensors, activities, and time and space ranges.
The talk will include a review of the research developed by the Human Activity Analysis Group of the University of Leeds, such as learning from human walking trajectories, clustering the motion of multiple people, learning and recognition of short-duration activities.
Mon 14 Sep 2015, 2:00 pm
Wayfinding - where Cognition meets Architectural Design
Prof. Christoph Hölscher, ETH Zurich
Orientation and navigation can be a challenging task in the built environment, especially in large-scale public buildings such as hospitals, conference centers, train stations or airports. This can limit the functional qualities of a building as well as patron's well-being. Understanding how environmental features as well as individual spatial abilities shape orientation as well as movement behaviors can help architects improve the usability and user experience of their building designs. We employ spatial analysis tools such as Space Syntax, behavior observation and targeted user experiments, eye-tracking studies and Virtual Reality simulation as part of human-centered design support and evaluation. Such behavioral user studies are complemented by interviews and design experiments with architectural designers in order to clarify how architects anticipate building usage and human cognition in their own design process. I will provide an overview of these studies and hope to discuss how this can extend to movement behavior of groups and how multi-agent simulation can be integrated.
Wed 02 Sep 2015, 2:00 pm
Humanoide Roboter - Wie nehmen sie die Umgebung wahr und wie planen sie ihre Bewegungen?
Prof. Maren Bennewitz, Rheinische Friedrich-Wilhelms-Universität Bonn
Ein zentrales Ziel der Robotik liegt in der Entwicklung von Robotern, die komplexe Aufgaben übernehmen und Menschen durch verschiedenste Services unterstützen. Dies erfordert, dass die Roboter ihre Umgebung wahrnehmen, autonom Navigations- und Manipulationsaufgaben durchführen, mit Menschen interagieren und von ihnen lernen können. In dieser Vorlesung werden innovative Lösungen für die in diesem Kontext auftretenden Probleme vorgestellt. Ausgehend von einer geeigneten Umgebungsrepräsentation werden Techniken zur Zustandsschätzung, Hinderniserkennung und Bewegungsplanung in hochdimensionalen Suchräumen präsentiert. Ein Schwerpunkt der Vorlesung liegt auf humanoiden Robotern, die einen menschenähnlichen Körperbau haben und sich auf zwei Beinen fortbewegen, wodurch sie besonders geeignet für den Einsatz in für Menschen geschaffenen Umgebungen sind.
Fri 30 Jan 2015, 3:00 pm
Plan-Based Robot Control in the RACE Project
Prof. Joachim Hertzberg, University of Osnabrueck and German Research Center of Artificial Intelligence (DFKI)
Using action plans for controlling or guiding robot behavior is a classical idea in AI with a great number and a large variety of examples. In the recently finished EU FP7 project RACE (Robustness by Autonomous Competence Enhancement), an attempt was made to employ it as integrated into a robot architecture for learning from experiences.
The talk reports about the approach towards plan-based robot control in RACE and about lessons learned. As foreseen in the RACE project plan, we started with a quick-and-dirty solution to planning (using the off-the-shelf HTN planning system SHOP2), to provide a fully integrated system and allow real robot traces to be generated early for fueling the learning work in RACE. In the end, we drafted a hierarchical hybrid planner that suits much better the needs in robotics to cope with causal, temporal, spatial, and resource reasoning in an integrated way, using classical HTN ideas to save search and to provide domain structure for helping representation and learning.
Mon 26 Jan 2015, 1:00 pm
Are you sure the library is that way? Metacognition in spatial reasoning
Dr. Christopher Stevens, University of Groningen
Imagine you are on your way to a meeting in an unfamiliar place. How do you know that you are heading in the right direction? How do you know when you can find your own way and when you need more information (e.g. from a map or GPS)? Despite much research on spatial memory and navigation, extremely little is known about how people monitor the accuracy of their spatial memories and judgments.
Previous research has shown that the accuracy of a spatial judgment is dependent on one's viewpoint. For instance, viewpoints that are aligned with the intrinsic structure of a space result in more accurate judgments than misaligned viewpoints. In this talk, I will discuss a series of experiments investigating whether viewpoint characteristics also affect metacognitive monitoring. Participants were asked to assume a particular viewpoint within a known space and point to a target landmark. After giving their directional judgments, they were asked to indicate the uncertainty of their judgment by giving a confidence interval. The results revealed that confidence judgments are sensitive to viewpoint characteristics. However, people are also more overconfident when using more difficult viewpoints. This suggests that people can detect when a viewpoint is not optimal, but the difficulty of the spatial inference prevents them from fully accounting for its effect on performance. Finally, the direction of the target with respect to one's body (i.e. in front or behind) is a salient confidence cue. This factor affects confidence even in cases when it does not affect performance. The results suggest that people do use viewpoint characteristics as confidence cues when making spatial inferences.