Research Humanities

Entrainment to Others in Conversational Speech

Prof. Julia Hirschberg

http://www1.cs.columbia.edu/~julia/

14-01-2009

15.30

TZ7

When people engage in conversation, they adapt the way they speak to the speaking style of their conversational partner in a variety of ways. For example, they may adopt a certain way of describing something based upon the way their conversational partner describes it, or adapt their pitch range or speaking rate to a conversational partner's. They may even align their turn-taking style or use of cue phrases to match their partner's. These types of entrainment have been shown to correlate with various measures of task success and dialogue naturalness. While there is considerable evidence for lexical entrainment from laboratory experiments, much less is known about other types of acoustic-prosodic and discourse-level entrainment and little work has been done to examine entrainments in multiple modalities for the same dialogue. I will discuss research in entrainment in multiple dimensions in the current literature and in our own research on the Columbia Games Corpus and the Switchboard Corpus. Our goal is to understand how the different varieties of entrainment correlate with one another and to determine which types of entrainment will be both useful and feasible for Spoken Dialogue Systems. (This is joint research with Agustin Gravano, Columbia, and Ani Nenkova, University of Pennsylvania.)

http://lyrawww.uvt.nl/~iflesch/Colloquium/Presentation/Hirschberg15-01-2009.ppt

New Developments in Computer Chess

Prof. H.Jaap van den Herik

http://www.tilburguniversity.edu/webwijs/show/&anr=286943&lan=nl?anr=286943&lan=nl

28-01-2009

15.30

TZ7

New techniques potentially usable in chess are: (i) Monte-Carlo Tree Search, (ii) UCT (Upper Confidence bounds applied to Trees), (iii) Supercomputers, such as IBMp6 (with 3328 processors), and (iv) Grid Technology. Conditions on and predictions for these techniques will be discussed in the lecture as well as their impact. An optimistic date for solving chess is 2035 and a pessimistic one (assuming that we can solve the game) is 2065. The solving time in 2035 (optimistic prediction) will be between 37 days and 4 months. Chinchalkar, S. (1996). An Upper Bound for the Number of Reachable Positions. ICCA Journal, Vol. 19, No. 3, pp. 181-183. Schaeffer, J. et al. (2007). Checkers is Solved. Science, Vol. 317, No. 5844, pp. 1518-1522.

http://lyrawww.uvt.nl/~iflesch/Colloquium/Presentation/vandenHerik28-01-2009.ppt

Complexity vs. predictability: on the role of causality in cognition

Prof. Leo Noordman

http://www.nieuwsbank.nl/inp/2005/01/28/R023.htm

04-03-2009

15.30

TZ7

Our cognitive system is particularly bad in dealing with randomness. We have difficulty with uncertainty and high informativity in the stimuli we perceive. We "prefer", organization, coherence, and meaningfulness. I will show that the inclination to perceive coherence and causal explanation manifests itself in quite different cognitive domains. In visual perception we perceive Gestalts. In statistical reasoning we tend to interpret correlations in terms of causal relations. In interpreting every day social behavior of ourselves and others, we attribute causes. In discourse processing sentences that are highly connected and causally connected are easier to process than sentences that are less connected. In information- theoretic terms, a coherent text is more redundant and therefore easier to be reproduced than a less coherent text. Complexity of a stimulus should be defined not only in terms of characteristics of the stimuli, but also in terms of properties of the human information processing system. Predictability is probably a more useful concept in explaining behavior than complexity.

Friendship, Deception and Retaliation in Negotiations

Per van der Wijst

http://www.tilburguniversity.edu/webwijs/show/&uid=per.vanderwijst?uid=per.vanderwijst

18-03-2009

15.30

TZ7

Negotiations are all about reaching one's goals. In distributive bargaining, a negotiator's first aim is to serve one's own interests. However, in repeated bargaining, it is important to maintain a good relationship with the other. When negotiating with a friend, the impact of that relational factor becomes even more important and it may influence the strategies that will be selected. In bargaining situations with a zero sum character ("my gain is your loss"), withholding or misrepresenting the background information is a common strategy in order to gain more power. Two studies will be presented in which the negotiating friends will be compared to negotiating strangers. Study 1 addresses the question of how friends cope with that double task of the negotiator, and if they areprepared to deceive the other in order to get a better result. A repeated ultimatum game experiment indicated that competing friends are as willingto deceiveeach other as strangers are. Deception leads generally to more rejections, but friends reward honesty more than strangers. Study 2 goes one step further and examines the consequences of deception: will it lead to retribution? Or will friendship form asecurity factor that prevents the deceiving counterpart to retaliate? In a second ultimatum game friends and strangers were all mislead by their counterpart, and they were given the opportunity to get back on their counterparts. A majority of subjects took that opportunity. Results of this study will be presented and discussed.

http://lyrawww.uvt.nl/~iflesch/Colloquium/Presentation/vanderWijst18-03-2009.pdf

Domain Knowledge Integration in Evolutionary Optimization: Case Studies in Medical Image Analysis and Drug Design

Michael Emmerich

http://www.liacs.nl/~emmerich/

01-04-2009

15.30

TZ7

Evolutionary Optimization is a flexible and robust tool for automatically searching for optimal or improved solutions in complex design spaces. The efficiency and effectiveness of the approach depends very much on whether and how domain knowledge is incorporated into the search procedure. Different artificial intelligence techniques can be used for this purpose. However, it needs to be considered carefully, how the domain knowledge is used. For instance bias of the mutation operator should be avoided, as well as under- and overrepresentation of the search space. In the talk the design of tailored, metric-based evolutionary algorithms for different non-continuous problem domains will be discussed, including the optimization of image analysis software for intravascular ultrasound movies and the computer aided search for drug molecules with optimized pharmaceutical properties.

http://lyrawww.uvt.nl/~iflesch/Colloquium/Presentation/Emmerich01-04-2009.pdf

Formal and Empirical Language Modelling

Menno van Zaanen

http://ilk.uvt.nl/~menno/main

15-04-2009

15.30

TZ7

In the context of computational linguistics, there are several tasks that are typically tackled with the help of language models. These models provide a rough description of the language under consideration, however they are not good enough to fully describe the language. Taking the limitations of the language models into account, they are often used to filter the output of other components by removing sentences that are clearly incorrect. A completely different field of research, grammatical inference, deals with finding good representations of languages, often in the shape of grammars, given a set of example sentences. This field is mainly interested in formal classes of grammars and aims to show which classes of grammars can be learnt efficiently under certain conditions. I this talk, I will describe language models and their use in computational linguistics, as well as results in grammatical inference. Finally, I will show some work that tries to bridge the two fields.

http://lyrawww.uvt.nl/~iflesch/Colloquium/Presentation/vanZaanen15-04-2009.pdf

Visualizing High-Dimensional Data using t-SNE

Laurens van der Maaten

http://homepage.tudelft.nl/19j49/Home.html

29-04-2009

15.30

TZ7

Over the last decade, many new techniques have been developed that visualize high-dimensional data by giving each datapoint a location in a two-dimensional map. The aim of these techniques is to represent the pairwise distances of the datapoints by similar pairwise distances of the corresponding points in the map. As the pairwise distances cannot be represented perfectly in the map, the emphasis of these techniques is on representing small pairwise distances accurately. In the talk, I will present a new technique for data visualization, called t-SNE, that converts the pairwise distances between the datapoints into probabilities of selecting pairs of data points. The selection probability of a pair of datapoints is proportional to a Gaussian function of their pairwise distances, as a result of which the probabilities measure the local structure of the data. If the distances between points in the map are converted into pairwise probabilities in the same way, any given arrangement of map-points can be evaluated by measuring the divergence between the probability distributions obtained from the data points and the probability distributions obtained from the map-points. A good arrangement of map-points can then be found by performing gradient descent to minimize this divergence. Unfortunately, if the probabilities of pairs of map-points are computed using a Gaussian function of their pairwise distance, the difference between the distributions of pairwise distances in high-dimensional and low-dimensional spaces causes the map-points to be crowded together in the center of the map. This crowding problem can be overcome by using a heavy-tailed Student t-distribution in the computation of the selection probabilities of pairs of map-points. The resulting technique, called t-SNE, constructs maps that reveal much more of the structure of the data than maps that are constructed by other recent visualization techniques. In particular, t-SNE is very good at preserving clusters in the data at many different scales simultaneously. The talk describes joint work with Geoffrey Hinton (University of Toronto). References: - L.J.P. van der Maaten and G.E. Hinton. Visualizing High-Dimensional Data Using t-SNE. Journal of Machine Learning Research 9(Nov):2579-2605, 2008 - L.J.P. van der Maaten. Learning a Parametric Embedding by Preserving Local Structure. To appear in Artificial Intelligence and Statistics, 2009.

http://lyrawww.uvt.nl/~iflesch/Colloquium/Presentation/vanderMaaten29-04-2009.pdf

Touchspeek. A computerised communication aid for people with aphasia

Mieke van de Sandt

http://www.rijndam.nl/scrivo/asset.php?id=279848

13-05-2009

15.30

TZ7

Aphasia is an acquired language disorder caused by brain damage. The most frequent cause of aphasia is a stroke. People with aphasia have lost their ability to communicate verbally, either partly or completely. To support the communication of people with severe aphasia, several other communication channels can be used, either verbal (writing, alphabet board, choice from written words/messages) or nonverbal (gestures, mimic, drawing, pictures, photographs). So far, the development of high-tech AAC solutions for communication in aphasia has been somewhat disappointing. One of the problems is the heterogeneity of the aphasic population in their communicative needs and abilities. In this presentation, I will present TouchSpeak, a modular communication system that was designed to support aphasic communication. TS can be individually adapted. Its use in everyday communication was investigated in a group of 33 patients with aphasia.

http://lyrawww.uvt.nl/~iflesch/Colloquium/Presentation/vandeSandt13-05-2009.pdf

Using Persuasive Technology to Promote Sustainable Behavior in Smart Home Environments

Prof. Cees Midden

https://venus.tue.nl/ep-cgi/ep_detail.opl?fac_id=98&rn=19950248&taal=US&hash=4xt2eTXybqErrezVdy2Nzygunh

27-05-2009

15.30

TZ7

Sustainable living is to a large extent the outcome of how consumers use the technology surrounding them. Seen from this perspective the rather strict separation of technological and behavioral solution is not only artificial but also detrimental to finding real sustainable solutions. Persuasive technology aims to intervene in these user-system interactions by using intelligent agents to change human attitudes and behavior. Embodied agents like robots and avatars go beyond the function of a simple tool by adopting social behavior that allows for social influence on human users. However, people often lack motivation or cognitive capacity to consciously process such relative complex information (e.g., numerical consumption feedback). Here, we argue that Ambient Persuasive Technology can be employed to provide feedback that needs less cognitive resources, can persuade the user without receiving the user�s conscious attention, and in general be more influential than more focal forms of persuasive technology.

From Data Mining to Knowledge Fitting

Joost Kok

http://www.liacs.nl/~joost

10-06-2009

15.30

TZ7

A classical view on data mining is expressed by the equation "Data Mining = Data Search using a Knowledge Bias". Using domain knowledge makes the search for patterns more efficient (or even feasible). Patterns and models need to be explained and made probable to the user. To this end, the patterns and models in the data need to be complemented with knowledge that convinces the user. There is so now so much knowledge available (from various sources for example on the Web and due to data mining) that we are witnessing a transition: instead of mining (only) for knowledge in data, we (also) mine the knowledge for knowledge patterns that describe/fit our data: "Knowledge Fitting = Knowledge Mining using a Data Bias". In our presentation we will describe a scenario for knowledge fitting that finds subgroups in relational data. This scenario deals with well-structured data, as well as knowledge in diverse public web resources, including structured taxonomic information and well as free text describing individual nodes in the ontology, with links to other data sources. We will apply the scenario to a number of medical data sets.

http://lyrawww.uvt.nl/~iflesch/Colloquium/Presentation/Kok10-06-2009.ppt

Experiments with a Creativity-Support System based on Perceptual Similarity

Prof. Bipin Indurkhya

http://www.iiit.net/people/faculty/bipin

16-09-2009

12time"> 12.30

DZ8

Researchers who study creativity in real-world situations have found that a primary hurdle facing human creativity is in stepping outside of our habitual conceptual association. A few techniques have been suggested to aid this process: making-the-familiar-strange and de-conceptualization, for example. The main objective of these techniques is to help the cognitive agent break the bond of conceptual association it has acquired culturally and through a lifetime of experiences. We have argued, and demonstrated in our past research, that computer-based systems can be very helpful in this task, as computers do not have any conceptual associations to begin with. For instance, we found that incorporating one familiar but unrelated object in a picture setting, or presenting unrelated pairs of objects to a cognitive agent, and asking them to make a story out of it, or to make sense of it in some other way, stimulates their creativity and imagination. In our current research, we are exploring the hypothesis that low-level perceptual similarity that is, similarity based on low-level perceptual features like color, shape and texture plays a key role in creation of novel conceptual associations. In other words, we are claiming that if the unrelated object that is introduced into a picture or paired with another object bears some low-level perceptual similarity with other objects inthepicture, or with the paired object, it is likely to be more effective in stimulating creativity than a random unrelated object. We willpresent here the results of some of our preliminary experiments toexplore thishypothesis. As the similarity with respect to low-level perceptual features is determined algorithmically using image-based pattern matching, our approach can be used to design more effective computer-based creativity-support systems. We will presentan outline of such architecture and mention a number of possible application domains for such systems.

The Body-Specificity of Language and Thought

Daniel Casasanto

http://www.casasanto.com/Site/home.html

07-10-2009

12.30

DZ7

If the content of our minds depends in part on the structure of our bodies, thenpeoplewith different types of bodies should think differently. I will review evidence that action verbs semantics and action imagery are differently lateralized in right- and left-handers� brains, consistent with the way they perform actions with their dominant hands. Handedness also shapes how adults and children represent abstract ideas like goodness and intelligence, and has visible consequences for the way right- and left-handers communicate about these ideas. Changing how people use their right and left hands in the laboratory can cause them to think differently, suggesting that handedness is not merely correlated with cognitive differences. Rather,body-specific patterns of physical experience shape the way we think.

http://lyrawww.uvt.nl/~iflesch/Colloquium/Presentation/Casasanto07-10-2009.pdf

Research activities at DANS and TiCC

Peter Doorn

http://www.dans.knaw.nl/nl/over_dans/organisatie/peter_doorn

15-10-2009

14.30

PZ005

14.30 - 14.40 Introduction by Jaap van den Herik 14.40 - 15.40 Presentations by DANS Peter Doorn: "Introduction" Rutger Kramer: "Past GRID experiences ..." Ren� van Horik: "Smart migration: the MIXED project" Maarten Hoogerwerf: "Persistent identifiers: challenges and applications" 15.40 - 16.00 Coffee/Tea 16.00 - 16.45 Presentations by TiCC Emiel Krahmer: "Data, Communication and Cognition" Jeroen Janssens: "The odd-one-out" Antal van den Bosch: "Text busters" 16.45 - Drinks and snacks

Accessing Natural History

Marieke van Erp

http://ilk.uvt.nl/~merp/Home.html

28-10-2009

12.30

DZ7

The Dutch National Museum for Natural History (Naturalis) harbours over 12 million animal and plant specimens that are important for national and international biodiversity research. Besides the specimens themselves, the provenance of the specimens is used by the researchers to investigate habitats and spreading patterns of species. The MITCH project (a collaboration between Tilburg University and Naturalis) was aimed at improving the accessibility of the textual data that accompanies each specimen in the Naturalis collection. In order to achieve this, (semi-)automatic approaches were developed to turn analogue data resources from the natural history domain into a high quality, enriched and easily accessible digital resource. I will highlight three key aspects of my work in MITCH, namely, data cleaning, data structuring and data retrieval.

http://lyrawww.uvt.nl/~iflesch/Colloquium/Presentation/vanErp28-10-2009.pdf

Universal Natural Shapes

Johan Gielis

http://www.genicap.com/Site/

18-11-2009

12.30

DZ8

To study forms in living organisms a unifying geometrical approach for describing various abstract, natural and man-made shapes has been proposed. Starting from Lam�s superellipses and supercircles, a large variety of shapes can be described by a single geometrical equation, based on a generalized Pythagorean Theorem. As a geometrical concept Gielis curves and surfaces are being applied in a variety of other scientific fields and technological applications. �Universal Natural Shapes� relates to the fact that these shapes occur at all levels in nature. This provides a unifying, descriptive basis for a geometrical theory of shape of "what dwells in the world and the world wherein we dwell". It is also in line with Gabriel Lam�s strive for a Unique Rational Science.

http://lyrawww.uvt.nl/~iflesch/Colloquium/Presentation/Gielis18-11-2009.pdf

Perception of synchrony between the senses

Prof. Jean Vroomen

http://spitswww.uvt.nl/~vroomen

09-12-2009

12.30

DZ7

For most multisensory events, observers perceive synchrony between the various senses (vision, audition, touch) despite naturally occurring lags in arrival and processing times of the different information streams. A substantial amount of research has examined how the brain accomplishes this. I will review several key issues about intersensory timing, and identify four mechanisms of how intersensory lags might be dealt with, namely: 1) by ignoring lags up to some point (a wide window of temporal integration), 2) by compensating for predictable variability, 3) by adjusting the point of perceived synchrony, and 4) by shifting one stream towards the other.

Complex but not complicated: the stopping power of visual advertising

Prof. Rik Pieters

http://www.tilburguniversity.edu/nl/webwijs/show/&uid=f.g.m.pieters?uid=f.g.m.pieters

16-12-2009

12.30

DZ8

The visual complexity of advertisements influences their attention attraction ability, i.e., their stopping power, yet in unexpected ways. We distingusih two types of visual complexity and show how these have different effect on attention. Ads can be visually complex because they contains detailed perceptual features ( feature complexity ) and/or because they have an intricate creative design ( design complexity ). We propose objective measures for these two types of visual complexity, and a measure for the visual complicatedness of ads, which often confused with visual complexity. An analysis of almost 250 ads that were tested with infra-red eyetracking methodology shows that the two types of visual complexity influence the stopping power and likeability of advertising differently. On the one hand, feature complexity of ads hurts ad attitude and attention to the brand. On the other hand, design complexity of ads helps ad attitude and attention to the pictorial, and thereby to the ad as a whole, as long as such complex ads are not visually complicated. Our metrics can be used to reveal how visual complexity can improve the stopping power and likeability of advertising, and perhaps other forms of visual communication.

http://lyrawww.uvt.nl/~iflesch/Colloquium/Presentation/Pieters16-12-2009.ppt

Mine your own business

Bernard Veldkamp

03-02-2010

12.30

DZ8

http://lyrawww.uvt.nl/~iflesch/Colloquium/Presentation/deVries03-02-2010.ppt

Modeling of Strategic Behavior and Goals in Large Organizations

Prof. Aske Plaat

http://plaat.nl/

17-03-2010

12.30

DZ8

Public and private sector organizations find it hard to achieve policy goals on time and on budget. Large organizations employ large numbers of people, whose behavior is determined in part by their own interests. Achieving organizational goals requires effectively dealing with conflicting interests and strategic behavior. Managers need skills in addressing this task. The goal of this project is to increase the effectiveness of organizations by better understanding goals, behavior and cooperation. We will use artificial intelligence techniques to model social relations between people. Creating such a model requires a multidisciplinary approach. Simulated agents need to understand social conventions and power, and need to understand the social significance of actions. In this talk I will give an overview of our research plans.

http://docs.google.com/present/view?id=dfrbp4kr_146fsnk3jfr

From PhD to Post-doc

Kristian Spoerer

24-03-2010

12.30

DZ5

In this talk I will present two main parts. First I will discuss the research from my PhD thesis and closely related studies on the game of Lemmings. In the second part I will provide an overview of the research being carried out here at JAIST where I am a Post-doc researcher, along with some initial thoughts on a new project.

http://lyrawww.uvt.nl/~iflesch/Colloquium/Presentation/Spoerer24-03-2010.ppt

Parts and places: Object-centered geometries

Jürgen Bohnemeyer

http://www.acsu.buffalo.edu/~jb77/

21-04-2009

12.30

DZ6

The studies to be presented probe the role of linguistic resources for reference to object parts in the linguistic and cognitive representation of regions of space. The background of these studies is the still-growing controversy over the use spatial frames of reference (FoRs) in language and cognition. FoRs are cognitive coordinate systems used to identify places and directions. Levinson (1996, 2003) and Pederson et al. 1998 show that speakers� selection of types of FoRs in discourse predicts their use of the same types of FoRs in recall memory and spatial inference. They advance a relativistic interpretation according to which language determines FoR selection in internal cognition. Li and Gleitman (2002) instead argue that the alignment is the result of cultural factors such as literacy, education, and the adaption to topography and ecology. The lecture focuses on a case study conducted as part of the project Spatial language and cognition in Mesoamerica ('MesoSpace' for short; NSF Award # BCS-0723694, PI J. Bohnemeyer), which investigates 15 indigenous languages of the Mesoamerican (MA) linguistic and cultural area. MesoSpace examines two unusual traits of spatial reference in MA: i) the widespread absence or paucity of use of relativeFoRs and ii) the highlyproductive use of 'meronymic' terminologies that identify object parts in terms oftheir shape. Competing accounts forthe unusually high productivity of MA meronymies invoke global analogical domain mappings(MacLaury 1989) and local shape-analytical algorithms (Levinson 1994). The overarching hypothesis informing MesoSpace is the idea that the availability of productive geometric meronymies may disfavor the use of relative FoRs in both language and internal cognition. If confirmed, this 'meronymy-allocentrism pattern' would represent evidence for a purely linguistic determinant of reference frame selection. Part descriptions and placement descriptions were collectedfrom five pairs of adult native speakers of Yucatec, a Mayan language of Mexico, with a referential communication task involving objects of unfamiliar shape. These data are complemented by part and location descriptions elicited from seven speakers using line drawings and photographs. The Yucatec meronym system is distinct from those proposed by MacLaury for Ayoquesco Zapotec and Levinson for Tseltal Maya. Yucatec meronymy distinguishes between three semi-autonomous subsystems, which label surfaces, volumes, and curvature extremes (edges and points), respectively. The subsystems for surface and curvature extreme naming are fully productive. In contrast, the use of the volume part terms appearsto be conventionalized. Spatial descriptions that are to be interpreted in intrinsic or relative FoRs must be formed with a surface meronym in Yucatec. The meronym may head either the 'ground phrase' itself - the place-denoting co-constituent of the verb - or the complement of a semantically nearly empty 'generic' preposition. If the ground phrase is formed without a meronym, or with a volume or edge/point meronym, it will be interpreted 'topologically', i.e., perspective-free. A second referential communication task, this one involving picture-to-picture matching, examined preferences for FoRs in reference to indoor-scale spatial configurations. The task involved four sets of 12 pictures each, all featuring a ball and a chair in varying configurations, and was conducted with five pairs of speakers. The results confirm the preference for intrinsic over relative FoRs predicted by the meronymy-allocentrism hypothesis: of the total 240 locative descriptions, 45% were intrinsic, 23% absolute, 22% topological, and just 10% relative. In Yucatec, the use of both intrinsic and relative FoRs presupposes the ability to consistently reference object geometry. But the relative use of meronyms requires speakers to assign these disregarding the geometry of theground, superimposing it with a projection of the observer's body instead. I hypothesize that Yucatec speakersdispreferthis because the frequent use of geometric meronyms habituates them to mentally encoding object geometry.

http://lyrawww.uvt.nl/~iflesch/Colloquium/Presentation/Bohnemeyer21-04-2010.pptx

Computer mediated communication and making friends, social effects of the internet, and social media marketing

Marjolijn Antheunis

28-04-2010

12.30

DZ5

The internet and ICT technology in general, is becoming a more and more social phenomenon. We share, befriend (and unfriend), sell, buy and negotiate via email, chat, twitter and social network sites such as hyves and facebook. In this talk I will give an overview of three area's of research that investigate the effect of these technologies on friendship formation, social relations and marketing. First I will present a general overview of my work, after which several studies will be discussed in depth.

http://lyrawww.uvt.nl/~iflesch/Colloquium/Presentation/Antheunis28-04-2010.pptx

On the distributed nature of mutual understanding

Dale Barr

19-05-2010

12.30

DZ8

Mutual understanding is one of the most important topics in the study of language use; it is also one of the most perpelexing. How is it that people understand one another given the fundamentally ambiguous nature of communication? Why do language users find conversation so effortless and unproblematic, while theories of language use suggest that it requires inordinately complex processes and representations? I will suggest that such paradoxes arise out of a tendency to localize processes of mutual understanding in the minds of individual language users. Instead, I will suggest that the work of mutual understanding is distributed more broadly, over individual, interactional, and cultural levels of language use. This approach offers a new way of understanding of the functional significance of certain psycholinguistic phenomena, such as apparent failures of perspective taking in referential communication.

http://lyrawww.uvt.nl/~iflesch/Colloquium/Presentation/Barr19-05-2010.pdf

Surgical pathology, humanity and computing.

Marius Nap

29-09-2010

12.30

E3

Surgical pathology is a medical specialty dealing with knowledge of disease processes, diagnostic classification of the status of either human single cells, (cytology), cells in syncitial arrangement, (histology), complete or partial organs, (gross inspection) and post mortem full body examination (autopsy). Traditionally this is done by describing visual or palpatory observations in texts that are presented to the treating physician who will use this information while planning further treatment options with his patients. These human observations as well as the structure of reporting will be affected by variable levels of subjectivity, influenced by knowledge, experience and discipline. Increasing demands are posed to health care providers to formalize their work process, reduce subjectivity and improve their auditability. However, strict protocols also risk to create loss of human interaction that characterises the expert. A combination of speech recognition as data input for a report, storage of large volume text files, directly linked to an individual observer with a high degree of repetitive phrasing combined with the availability of comprehensive textbook knowledge on diagnostic criteria may offer an ideal opportunity to apply computer support for an individualised report generation and control. In recent years “whole slide imaging” has been introduced both in research as well as in clinical setting. Digital images from complete histological or cytological specimens are available and can be used to apply computer vision techniques, image analysis and identification of normal components of the tissue in combination with the presence of abnormal structures as rare events in the specimen or the comparison of different grading characteristics. Combining computational analysis of text and image may stimulate the acceptance of individualised or humanised “protocols” and improve the correct identification and use of diagnostic hallmarks. This approach may effectively link computers, humanity and surgicalpathology with each other, resulting in a higher level of quality in health care.

Finding the connectome of the visual system: Computational systems neuroscience weighs in

Michael Capalbo

27-10-2010

12.30

DZ5

How the human brain is wired has always been of great interest to the scientific community, yet we are still a long way from understanding human brain connectivity. Recently the pace of the quest has been picking up and has been coined the ‘ search for the human connectome’ . In our opinion, understanding the human connectome requires state-of-the-art experimental methods, advanced mathematical tools and sophisticated computational models. In this talk I will discuss some of the advances we have been making. First I will discuss the work we have done to model the large-scale structure of the visual system combining both functional and structural brain data. I will argue that although the hierarchy of the system is not fixed our method reduces the indeterminateness in the model and highlights the importance of sub- cortical routes. Second I will discuss our attempts to visualize the connections between the functional modules of the human brain in vivo. Using a combination of Diffusion Weighted Imaging and functional Magnetic Resonance Imaging we can visualize the major pathways in the brain. Using tensor tracking and graph theory we can arrive at a first approximation of the connectome of the human visual system.

\

"Multiple-instance learing", or "set classification".

David Tax

10-11-2010

12.30

PZ50

Standard pattern recognition and machine learning most often assumes that it is possible to represent objects by a (single) feature vector. For some classification problems this is too restrictive, in particular in situations where the objects are complex and compounded of different subparts. This can happen in general images, time series, videos, etc. For these situations multiple-instance learning (MIL) can be used. In MIL it is assumed that an object is represented by a set of feature vectors. If at least one feature vector is labeled 'positive', the whole object is labeled positive, and otherwise it is labeled negative. In this talk I would like to discuss the basic assumptions of multiple instance learning, show some algorithms, and give some criticism of the model. Finally I will show some applications in the classification of images and timeseries.

Surprise!

Marret Noordewier

8-12-2010

12.30

DZ3

Surprise is widely recognized as a fundamental emotion that helps people deal with unexpected events. Interestingly, however, the scientific knowledge of its nature is still very limited. One of the most salient lacunas in our knowledge of surprise is the valence of this emotion. With this research, we hope to fill this void. Based on the reasoning that surprise interrupts ongoing activities and frustrates people’s need for predictability, we argue that surprise has a negative valence. Once the unexpected stimulus is understood, surprise dissipates and is followed by other emotions. This account is tested in four studies, ranging from personal experience to facial expression and attention.

Hidden-Unit Conditional Random Fields

Laurens van der Maaten

22-12-2010

12.30

In this talk, I will introduce a new Conditional Random Field (CRF) model, called the hidden-unit CRF. The most prominent feature of the hidden-unit CRF is that it has stochastic binary hidden units that are conditionally independent given the data and the label sequence. The hidden units drastically improve the complexity of the decision boundaries that can be learned by the CRF model. In fact, it can be shown that the individual predictors (i.e., predictors that ignore the temporal correlations) are universal approximators for discrete data. At the same time, the conditional independence properties allow us to efficiently compute (1) the exact gradient of the conditional log-likelihood, (2) the most likely label sequence for a given time series, and (3) marginal distributions over label sequences. We develop and investigate various training algorithms for hidden-unit CRFs. Our experiments reveal the strong performance of hidden-unit CRFs on a wide range of tasks, including optical character recognition, sentence labeling, protein secondary structure prediction, and part-of-speech tagging. Hidden-unit CRFs outperform the respective state-of-the-art techniques on three of the four tasks. This talk describes joint work with Max Welling (University of California Irvine) and Lawrence Saul (University of California San Diego).

Gesture as an embodied and/or communicative act-Evidence from brain and behavior.

Asli Ozyurek

22-2-2011

12.30

DZ5

Gesturing during language use, a form of meaningful action, is an integral part of everyday communication. People use gestures in all ages, cultures and in all kinds of communicative contexts and even in various communicative impairments (e.g., autism, aphasia etc.). Yet we know little about the processes underlying use of gestures in production and/or comprehension. In this talk I will compare evidence for two currently debated views on gesture use; namely what I call as “Gesture as a window into thought” versus “Gesture as a communicative act” view. By focusing on a subset of gestures, namely the representational ones that accompany speech (e.g., making a POURING gesture while saying “pour”), I will argue that such gestures that speakers use are better understood better understood not simply as a window into what goes on in producer’s mind (e.g., embodied actions or images) but rather in terms of what representation the speaker wants to create on “the other’s” mind. Thus I will emphasize the main role of gesture as a communicative act integrated with speech/language processing. I will provide evidence for the latter claim from both production and comprehension, from crosslinguistic investigations of gesture use , behavioral experiments, and finally from how the brain processes cospeech gestures.

Generating medical narratives from raw data: From summaries to stories

Albert Gatt

2-3-2011

12.30

DZ5

Data-to-text systems are a subclass of Natural Language Generation (NLG) systems which produce textual summaries of raw data. The BabyTalk Project was concerned with building a family of data-to-text NLG systems to summarise large, high-density datasets collected in a Neonatal Intensive Care Unit (NICU). One of these systems, BT-Nurse, summarized twelve hours of patient data to produce a nursing shift summary. Patients in the NICU are continuously monitored and data is collected both through sensors (such as ECG and blood pressure monitors) and via manual input. Thus, the available information is heterogeneous and includes numerical, symbolic and raw textual information. Moreover, the generated summaries can be fairly lengthy and report on a diversity of events which need to be adequately linked together if the summary is to serve its purpose, namely, to support clinical decision-making. This talk will first give an overview of the NLG architecture adopted in BabyTalk and identify some of the key challenges in generating coherent text from NICU data. Foremost among these challenges is the construction of an adequate narrative which allows a clinician to (a) identify the clinically salient events that occurred in relation to a patient during a given period; (b) understand the temporal and causal relationships between these events. Meeting these two requirements has consequences at all stages of the generation architecture. The second part of the talk focuses on how the data-to-text technology developed in the project was evaluated. An initial prototype system, BT-45, was evaluated in a controlled setting, in which it was compared to graphical presentations and human-written summaries. The larger-scale BT-Nurse system was deployed on-ward over a number of months, and used by nurses to generate summaries of live, previously unseen data. The latter evaluation, involving deployment of a system in its target environment, was one of the first of its kind in NLG. Both evaluations gave positive and encouraging results. They also served to highlight some new research challenges. Foremost among these is the problem of creating good computational models of narrative that adequately handle its complex structural and temporal aspects.

PoliticalMashup

Maarten Marx

13-4-2011

12.30

DZ3

We present a large scale data integration project, PoliticalMashup, in which we connect all kind of political data. We show several applications of structure-extraction from textual data, and list a number of natural language processing and machine learning challenges.

This page will provide you with an overview of all TiCC colloquia. Please use the slider to scroll through the different seminars. By default, the next seminar is visible, followed by the previous four colloquia. If you have questions or suggestions for the TiCC colloquium, please contact us. Archive Upcoming Seminars

Towards a better understanding of human emotional tears

Ad Vingerhoets

11-5-2011

12.30

DZ5

Crying (with tears) is a unique and typically human emotional expression, that has hardly received the attention of researchers. Nevertheless in the popular literature there are two major claims. First, crying is said to bring relief and it is considered to be healthy. Second, it has been speculated that tears in addition have strong inter-personal effect and stimulate social bonding. But what is the scientific status of these claims? In my presentation, I will provide an overview of the state-of-the-art with respect to research on (adult) crying. In addition, I want to launch some hypotheses about the development of crying over the life span and about the possible relationship between crying and morality.

Seeing Signs. How do people recognize meaningful hand movements?

Jeroen Arendsen

1-6-2011

12.30

DZ8

In this lecture dr ir Jeroen Arendsen will talk about his dissertation entitled ‘Seeing Signs’. It contains the results of a series of studies on the appearance of manual movements in gestures. The main goal of this research is to increase our understanding of how humans perceive signs and other gestures (and how this is related to the perception of other human behavior such as fidgeting with the hands). Generated insights from human perception may aid the development of technology for recognizing gestures and sign language automatically with cameras and computers. One example of an application of automatic gesture recognition that has played a role in shaping the research in this dissertation is ELo, an Electronic Learning environment for deaf and hearing impaired children to practice Sign Language of the Netherlands (SLN) signs. The questions addressed in the research focus on a number of aspects including temporal processing of signs, discrimination of gestures from other human behaviour, and how humans handle variation in signs (this last item will be addressed in more detail, focusing on the question ‘how can the acceptability of variation be explained or predicted, for example using SLN phonology (e.g. Van der Kooij, 2002), and what is the role of iconicity?’)

Dynamic and task-dependent encoding of speech and voice in the auditory cortex

Milene Bonte

15-6-2011

12.30

DZ5

In speech perception, extraction of meaning and speaker identity from complex streams of sounds is surprisingly fast and efficient. This efficiency depends on the crucial capability of our speech recognition system to deal with the acoustic variability of the input signal and to form invariant abstract representations. Furthermore top-down cues such as linguistic context and task demands may bias and facilitate this process and can be used to predict incoming information. In this talk, using examples from EEG, MEG and fMRI studies, I will illustrate temporal as well as spatial neural mechanisms enabling the adaptive decoding of speech signals into abstract representations of words, speech sounds and speaker identity.

Numbers in Space

Diane Pecher

22-6-2011

12.30

DZ5

People often use spatial metaphors for number magnitude, as in expressions like “prices are going up” or the use of a horizontal historical timeline. This implies that numbers might be mentally understood by grounding in spatial orientation, where small numbers are represented as low or to the left and large numbers are represented as high or to the right. In a first series of experiments we investigated whether number magnitude affects spatial attention. Central presentation of a relatively small (1,2) or large (8,9) number was followed by a target presented at the left or right. Only when number magnitude was task relevant did it affect spatial attention: targets were detected faster on the left after small numbers and on the right after large numbers. In a second series we investigated both vertical and horizontal spatial attention. In addition, we compared concrete and abstract magnitude. We presented numbers in concrete (7 shoes in a shoe shop) or abstract (29 – 7) contexts and asked participants to make relative magnitude judgments. Following the judgment a target letter was presented at the top or bottom, or left or right of the visual field. Participants were better at identifying letters at congruent than incongruent locations, but this effect was obtained only when numbers were presented in concrete contexts. We conclude that numbers do not automatically direct spatial attention. In addition, spatial grounding might have a smaller role for numbers in abstract than in concrete context.

Language learning as a unified probabilistic process

Afra Alishahi

28-9-2011

12.30

PZ15

I will present computational models that simulate various aspects of language learning from naturalistic input data. In particular, I will discuss the process of learning word meanings, lexical categories, and argument structure constructions from a combination of linguistic and perceptual input in the presence of noise and uncertainty. The proposed computational models exploit statistical regularities of the input in a probabilistic framework to learn the structural properties of language. They demonstrate that robust knowledge of language can be acquired from usage data, and provide explanations for the observed behavioural patterns in humans.

From an Image to a Description

Margaret Mitchell

12-10-2011

12.30

PZ5

In this talk, I will be discussing preliminary work on building a vision-to-language generation system. The system generates captions of images from computer vision output by leveraging knowledge of visible features and word co-occurrence statistics. It employs a new kind of generative grammar, specifically developed to generate semantically and syntactically well-formed structures from the seeds provided by computer vision detections. This follows work my group did at the Johns Hopkins Workshop on visually descriptive text (http://www.clsp.jhu.edu/workshops/ws11/groups/ehlvdt/), where we focused on characterizing what it means to describe an image, and how to filter computer vision output in the face of heavy noise.

Dialogue Reference in a Visual Domain

Jette Viethen

12-10-2011

12.30

DZ3

In this talk I will give an overview of my recent work on generating dialogue reference in a visual domain, which was done incollaboration with colleagues at Macquarie University, Sydney. When people refer to visually available objects, they are likely to take into account the visual context. however, work in psycholinguistics has shown that reference produced as part of an ongoing dialogue is also influenced by the discourse context. We used a simple machine learning approach in order to investigate the impact of these visual and discourse factors on the content of referring expressions in a corpus of dialogues. The dialogues were collected in a controlled route description task involving schematic maps. The aspects we explored include the impact of priming/alignment compared to identification, interdependency between the use of different attributes that can be used for identification, and the size of the visual context that people take into account when they refer to objects on a map.

Aldebaran's NAO Humanoid robot

Pascal Morel, Aldebaran Robotics

23-11-2011

12.30

PZ17

NAO is a research platform used by more than 350 prestigious universities and research labs around the world. NAO is a versatile platform used to explore a great variety of research topics in the robotics field as well as in computer science, man-machine interaction and social sciences. The fully programmable robot comes with a programming environment suitable even for researchers with no programming experience. NAO’s many sensors and actuators, convenient size and attractive appearance combined with sophisticated embedded software makes it a unique humanoid robot ideal for many research fields. NAO has for example capacity for face and object recognition, automatic speech recognition, text to speech capacity in seven languages and whole body motion. In this presentation, NAO's capacities and suitability for research will be demonstrated. For more information: http://www.aldebaran-robotics.com.

Does avatar appearance matter? The effects of similarity and self-identification on virtual team performance.

Alexander Schouten

30-11-2011

12.30

DZ5

I will present (part of) the results of a recent study we conducted on the effects of similarity and self-identification on virtual team performance. Effective communication and collaboration for a large part depends on how well people can get along and their willingness to share information. It is not clear, however, how information sharing may be facilitated in virtual teams. On the one hand, identifiability of self and others may lead to enhanced information sharing because being identifiable reduces group dynamics such as social loafing. Moreover, avatar identification has shown to lead to better performance in virtual interactions. On the other hand, being similar to other leads to perceptions of common group membership resulting in a higher willingness to share information. In this study, we manipulate both similarity and self-identification by using avatars. Specifically, we compare 3-person teams that need to solve a hidden profile task while being represented by avatars that are either similar or not (different vs. similar avatars for the team members) and either self-identifiable or not (avatar that resembles self vs avatar that not resembles self). Results show that groups that are both similar as well as personally identifiable outperform teams in the other three conditions. However, the exact mechanisms through which this happens is still unclear.

Investigating syntactic representations using structural priming.

Roger P.G. van Gompel

23-02-2012

12.30

DZ6

The structural priming method has been a very fruitful method for investigating the syntactic representations that people access during language production (Pickering & Ferreira, Psychological Bulletin, 2008). I will present a series of structural priming experiments that investigated whether structural representations are lexically based (e.g., part of the verb's lemma information) or more abstract in nature. Research by Van Gompel, Arai, and Pearson (Journal of Memory and Language, in press) suggests that representations are lexically based, except monotransitive structures, because this is the most common, default structure in English. I will also present recent work that shows that lexically-based structures are represented as part of the head noun's lemma, but not as part of the lemmas of non-heads, consistent with most linguistic theories.

Gaming in higher and professional education

Rens Kortman

14-03-2012

12.30

DZ3

Gaming is increasingly used in an educational setting. In my talk I will provide some background notions on the use of gaming in higher and professional education. The theory will be illustrated though a TU Delft teaching case study and projects. For instance, we have used Biggs’ theory on constructive alignment to design a game-based course on high-tech management. For Rijkswaterstaat, the Department of Waterways and Public Works in The Netherlands, we are currently developing a game-based professional training on leadership development. Finally, as part of the European network of excellence ‘Games and Learning Alliance’ (GALA) we participate in the development of a game-based curriculum for teaching entrepreneurship competencies to students of technical universities in Europe. The talk will provide ample room for discussion and exchange of ideas.

Self-organisation of quality dimensions in conceptual spaces

Paula Roncaglia-Denissen

31-05-2012

12.30

DZ8

In this event-related potential (ERP) study we investigated the role of rhythm in the disambiguation of syntactic structures during auditory sentence processing. We presented participants with syntactically ambiguous German sentences embedded in regular and irregular rhythmic patterns. Rhythmically regular pattern was ensured by constant inter-stress interval which created rhythmic groups with constant size. Accuracy rates reveal that participants made less errors in rhythmically regular sentences than in rhythmically irregular ones. ERP results show a reduction in the P600 mean amplitude for object-first sentences in rhythmically regular context in comparison with their rhythmically irregular counterparts. This reduction suggests a facilitation effect through the decrease of processing costs for the less-preferred structure (object-first sentences). Our findings suggest that rhythmic regularity may be used as cue to facilitate the processing of syntactic complexity in auditory sentence processing.