Content uploaded by Erik Alexander Billing
Author content
All content in this area was uploaded by Erik Alexander Billing on Sep 09, 2015
Content may be subject to copyright.
PROCEEDINGS+OF+THE+!
2015!SWECOG'CONFERENCE!
!
!
!
!
!
!
!
!
!
!
SKÖVDE,!JUNE!15-16,!2015!
BY!THE!SWEDISH!COGNITIVE!SCIENCE!SOCIETY!
!
!
!
!
!
!
EDITORS!
Erik!Billing,!!
Jessica!Lindblom,!and!
Tom!Ziemke!
Copyright c
2015 The Authors
Sk¨
ovde University Studies in Informatics 2015:3
ISBN 1653-2325
ISSN 978-91-978513-8-1
PUBLISHED BY THE UNIVERSITY OF SK¨
OVDE
Revised version, August 2015
Preface
Welcome to SweCog 2015!
The aim of the Swedish Cognitive Science Society is to support networking among
researchers in Sweden, with the goal of creating a strong interdisciplinary cluster of
cognitive science oriented research.
This little booklet contains the abstracts of the invited talks as well as all oral and poster
presentations at the 2015 SweCog conference (in alphabetical order). In addition to the
usual interdisciplinary mix of research that is typical for cognitive science meetings,
this year’s conference also seems to challenge some of the fundamentals of mathemat-
ics: One of the invited speakers claims that 1 + 1 = 3, and as organizers we would
like to assert that 2 = 11, given that this is both the second and the eleventh SweCog
conference. If you doubt that, there’s something to discuss in the coffee breaks...
Last, but not least, we would like to thank the many people that have contributed to this
conference, including many colleagues at the University of Sk¨
ovde, and in particular
of course all authors and reviewers.
The reviewers were: Rebecca Andreasson, Alexander Alm´
er, Anna-Sofia Alklind Tay-
lor, Anton Axelsson, Christian Balkenius, Erik Billing, Nils Dahlb¨
ack, Karl Drejing,
Boris Duran, Benjamin Fonooni, Paul Hemeren, Lars-Erik Janlert, Erik Lagerstedt,
Gauss Lee, Robert Lowe, Jana Rambusch, Fredrik Stjernberg, Tarja Susi, Henrik Svens-
son, Serge Thill, Niklas Torstensson, David Vernon, Annika Wallin, and Tom Ziemke.
1
2
Conference Program
Conference chairs: Jessica Lindblom and Erik Billing.
Monday 15th of June
page
10:00 — 10:25
Registration, G-building (G111)
10:25 — 10:30
Conference opening
10:30 — 11:30
Invited speaker: Linda Handlin, University of Skövde
11
11:30 — 12:00
Ginevra Castellano
8
12:00 — 12:30
Henrik Siljebråt
19
12:30 — 13:30
Lunch
13:30 — 14:00
Claes Strannegård
20
14:00 — 14:30
Abdul Rahim Nizamani
16
14:30 — 15:00
Coffee break
15:00 — 16:00
Invited speaker: Tony Belpaeme, Plymouth University
7
16:15 — 18:45
Bus departures to visit Varnhem Church and Monastery ruins
19:30
Dinner
Tuesday 16th of June
page
09:00 — 10:00
Invited speaker: David Vernon, University of Skövde
24
10:00 — 11:15
Poster session: Rebecca Andreasson (p. 5); Giuseppe, Innamorato (p. 9); Mattias
Kristiansson (p. 12); Ludvig Londos (p. 14); Sara Nygårdhs (p. 17); Sam Thellman
(p. 21); Niklas Torstensson and Tarja Susi (p. 23)
11:15 — 11:45
Trond Arild Tjøstheim
22
11:45 — 12:15
Joel Parthemore
18
12:15 — 13:15
Lunch
13:15 — 13:45
Annika Wallin
25
13:45 — 14:15
Christian Balkenius
6
14:15 — 15:00
SweCog annual meeting
15:00 — 15:05
Final words
3
4
User experience of affective touch in human-robot interaction
Rebecca Andreasson
School of Informatics, University of Skövde
rebecca.andreasson@his.se
Robotic technology is quickly advancing and robots are entering both professional and domestic settings. An
increased application of robots in elderly care and in therapy shows a shift towards social robots acting in human
environments, designed to socially interact with humans. Socially interactive robots need to act in relation to
social and emotional aspects of human life, and be able to sense and react to social cues. Touch, as one of the
most fundamental aspects of human social interaction (Montagu, 1986) has lately received great interest in
human-robot interaction (HRI) research (e.g. Dahiya et al., 2010; Silvera-Tawil et al., 2015) and the
interpretation of touch in robotics has been presented as an unresolved research area with a crucial role in further
development of HRI (Silvera-Tawil et al., 2015). It has been argued that the communicative distance between
people and robots would be shortened and that the interaction would be more meaningful and intuitive if robots
were able to “feel”, “understand”, and respond to touch in accordance with expectations of the human (Silvera-
Tawil et al., 2015). However, this reasoning takes the notion of user experience (UX) for granted. The concept of
UX embraces both pragmatic and hedonic aspects of interaction with technology in a particular context (Hartson
& Pyla, 2012). In the field of human-computer interaction, UX has been acknowledged as a key term in the
design of interactive products, but UX has not been emphasized in HRI. Accordingly, this research argues that it
is important to study not only the robotic technology aspect of tactile interaction but also the user’s experience of
the interaction, i.e. taking on the human-centered HRI approach presented by Dautenhahn (2007). Research on
human-human interaction has showed that humans are able to communicate emotions via touch, and that specific
emotions are associated with specific touch behaviors (Hertenstein et al., 2009). As a starting point for narrowing
the distance between UX and HRI, the present research suggests a study where subjects are instructed to convey
specific emotions to a humanoid robot. The study aims at investigating the role of affective touch in HRI with a
focus on touch behaviors (e.g. stroking, grasping) for specific emotions, touch locations on the robot, and user
experience of interacting with the robot via touch. The intended contributions of this study are an increased
understanding of the necessary properties of tactile sensors enabling affective touch in human-robot interaction,
the relevant placements of the sensors on the robot, and how the robot’s “look and feel” affects the user’s
experience of the interaction. The proposed research embarks on a new track of HRI research and will, contrary
to prior research on tactile interaction in HRI, emphasize the user experience of affective touch, highlighting that
a positive user experience has to be systematically and consciously designed in order for the social robots to
achieve the intended benefits of being socially interactive. Accordingly, the proposed study is believed to give
new insights about the understudied dimension of UX in HRI, with the potential to enrich interaction between
humans and social robots.
References
Dahiya, R. S., Metta, G. & Valle, M. (2010). Tactile sensing – from humans to humanoids. IEEE Transactions
on Robotics, 26(1), 1-20.
Dautenhahn, K. (2007). Socially interactive robots: dimensions of human-robot interaction. Philosophical
Transactions of the Royal Society B: Biological Sciences, 362(1480), 679-704.
Hartson, R. & Pyla, P. S. (2012). The UX Book: Process and Guidelines for Ensuring a Quality User
Experience. Amsterdam: Morgan Kaufmann.
Hertenstein, M. J., Holmes, R., McCullough, M. & Keltner, D. (2009). The communication of emotion via touch.
Emotion, 9(4), 566-573.
Montagu, A. (1986). Touching: The Human Significance of the Skin (3 ed.). New York: Harper & Row.
Silvera-Tawil, D., Rye, D. & Velonaki, M. (2015). Artificial skin and tactile sensing for socially interactive
robots: a review. Robotics and Autonomous Systems, 63, 230-243.
5
Developing Flexible Skills
Christian Balkenius1Birger Johansson1
1Lund University Cognitive Scienca
christian.balkenius@lucs.lu.see
We present an novel developmental architecture motivated by findings from neuroscience and learning theory.
Through interaction between many different cognitive subsystems, a surprisingly large number of flexible goal-
directed behaviors evolve over time. The architecture includes mechanisms for goal-directed bottom-up and top-
down attention, reward driven and anticipatory learning as well as goal-setting and navigation. There are also
a number of memory systems in the architecture serving different functions, including a working memory that
stores spatial-feature bindings. The architecture allows a robot to show great flexibility by being able to adapt to
an arbitrary environmental layout and to manipulate objects and combine them into any configuration that it has
previously seen. The architecture has been tested on a visually guided mobile robot with a manipulator that allows
it to stack blocks.
One of the cornerstones of the architecture is the idea of attention as selection-for-action (Allport, 1990), which
leads to a view of action execution as consisting of two stages (Balkenius, 2000). In the first stage, a target object
is selected by the attention system (Balkenius, Morén, & Winberg, 2009). At this stage the particular action to
perform is not yet determined. In the second stage, one of potentially several actions compatible with the target
object is selected.
On the lowest level, actions consist of small attention controlled goal-directed behavior fragments that increase the
probability that randomly selected actions will lead to useful consequences. The behavior produced at this level
is subsequently used to train a context sensitive reinforcement learning component (Balkenius & Winberg, 2008).
The final control level depends on outcome-action associations that can be used for situated simulation of actions.
When the attention system has selected a target object, it is able to look ahead to evaluate the utility of different
actions and future targets before selecting the action to execute.
A target object can also be selected from memory. If the object is not directly visible, the spatial memory system
will be used to recall the location of the object and a place-field-based navigation system is invoked to construct a
path to the location of the remembered object.
The robot adapts to a changing or novel environment by relearning or by restructuring the environment to match its
expectations by moving objects around (Friston, Daunizeau, Kilner, & Kiebel, 2010). The architecture also allows
for human interaction even though no part of the architecture was designed specifically for this. The robot can
interactively be taught to build novel designs, a human can help the robot, and the robot can even help a human
builder if it recognizes what she is building. As there is currently no explicit social component in the architecture,
these abilities all depend on learning to predict regularities in the environment.
References
Allport, A. (1990). Visual attention. In M. I. Posner (Ed.), Foundations of cognitive science. MIT Press.
Balkenius, C. (2000). Attention, habituation and conditioning: toward a computational model. Cognitive Science
Quarterly,1(2), 171-214.
Balkenius, C., Morén, J., & Winberg, S. (2009). Interactions between motivation, emotion and attention: From
biology to robotics. In L. Cañamero (Ed.), Proceedings of the ninth international conference on epigenetic
robotics.
Balkenius, C., & Winberg, S. (2008). Fast learning in an actor-critic architecture with reward and punishment. In
Proceedings of scai (Vol. 173). IOS Press.
Friston, K. J., Daunizeau, J., Kilner, J., & Kiebel, S. J. (2010). Action and behavior: a free-energy formulation.
Biological cybernetics,102(3), 227–260.
6
Invited contribution
Social Robots: Moving From the Quirky to the Useful
Tony Belpaeme
School of Computing, Electronics and Mathematics, Plymouth University
tony.belpaeme@plymouth.ac.uk
Robots that interact with people using one or several communicative modalities have been around for almost 20
years. The technological challenges of creating robust human-robot interaction are huge, and progress in
building the artificial intelligence required to make autonomous social robots has been unsteady. But even
though the social performance of robots is far from that of humans, the gaps in the robot's social cognition are
often plugged by humans' gregarious social cognition. As such we are now at a time where the science and
technology of social robots is mature enough to be useful. This talk will give a brief overview of the current state
of the art in social robots, and will show how the cognitive sciences are central to building social robots and
understanding how our behaviour towards social robots. In a second part, the talk will dwell on the applications
of social robots, and will show how they can be used as hospital companions and teachers.
7
Affective robotic tutors
Ginevra Castellano
1Department of Information Technology,
Uppsala University, Sweden
ginevra.castellano@it.uu.se
Recent research on personal robots shows that robots are increasingly being studied as partners that collaborate
and interact with people [1]. Robot companions [2], for example, are envisioned to play an important role in
several applications, such as providing assistance for the elderly at home, serving as tutors for children by
enriching their learning experiences, acting as therapeutic tools for children with autism or as game buddies for
entertainment purposes.
This paper presents ongoing work in the EMOTE project (www.emote-project.eu) aiming to develop personal
robots that act as robotic tutors. EMOTE is building a new generation of robotic tutors that have perceptive
capabilities to engage in empathic interactions with learners in a shared physical space. The project proposes to
build tutors that enrich learning experiences by (a) monitoring the learner's abilities and difficulties throughout
the learning process; (b) modelling affect-related states experienced by the learner during the learning task and
the interaction with the tutor; (c) providing appropriate feedback to the learner by means of contextualised
empathic reactions, adaptive dialogue and personalised learning strategies.
In order to build an intelligent artificial tutor with affective capabilities, an appropriate computational model
needs to be developed and properly trained to automatically recognize and classify the emotional state of the
user. For training purposes, representative data is required. A Wizard-of-Oz (WoZ) study was performed to
collect multimodal data from school pupils aged between 10 and 13 interacting with a Nao robot acting as a tutor
while performing a map reading task on a multi-touch table. During the study the robot was controlled by an
experienced teacher using a bespoke networked Wizard control interface which allowed full control over the
various parts of the system. The robot has been endowed with a large collection of flexible utterances and
predefined naturalistic behaviours to help scaffold an interaction. The creation of Nao's behaviours and
utterances are based on an extensive literature review and inspiration drawn from our previous human-human
studies with teachers and students from different European cities using a mock-up prototype of the educational
activity. During each session, we manually captured three videos via digital camcorders. Simultaneous video
feeds from the cameras, the Q sensor from Affectiva (i.e., electro-dermal activity and temperature), OKAO from
Omron (i.e., facial characteristics such as expression estimation and smile estimation, eye gaze information and
blink estimation) and the Kinect sensor (i.e., head gaze information, depth and facial action units) were recorded
during the tutor-learner interaction. The interaction between the tutor and the learner in terms of tutor dialogue
actions, utterances and learner responses in terms of button presses was also logged. Videos and data are
currently being analysed in order to build a user model that accounts for the affect of the learner.
Acknowledgement
The work of the author is partially supported the European Commission (EC) and funded by the EU FP7 ICT-
317923 project EMOTE. The author is solely responsible for the content of this publication. It does not represent
the opinion of the EC, and the EC is not responsible for any use that might be made of data appearing therein.
References
[1] Breazeal, C. (2009). Role of expressive behaviour for robots that learn from people. Philosophical
Transactions of the Royal Society B, vol. 364, pp. 3527–3538, 2009.
[2] Tanaka, F., Cicourel, A., and Movellan, J. R. (2007). Socialization between toddlers and robots at an early
childhood education center. Proceedings of the National Academy of Science, vol. 194, no. 46, pp. 17 954–17
958, 2007.
[3] Castellano, G., Paiva, A., Kappas, A., Aylett, R., Hastie, H., Barendregt, W., Nabais, F., and Bull, S. (2013).
Towards Empathic Virtual and Robotic Tutors. Proceedings of the 16th International Conference on
Artificial Intelligence in Education (AIED’13), Memphis, USA, July 2013.
8
Did the practice of Partible Paternity select for Emotional Intelligence? A systematic
review of Ritual Couvade in lowland Indigenous Amazonian societies
Giuseppe, Innamorato
Department of Psychology, Umeå University
giin0001@gapps.umu.se
Abstract
Evolutionary psychology is based on the assumption that psychological traits must have been selected as a result
of natural and sexual selection. Amotz Zahavi (1979) suggested that some traits that appear to be non-adaptive,
by inflicting great costs on the individual expressing them, are nevertheless beneficial for its progeny, and are
therefore selected for by the opposite sex. The purpose of the present thesis is to explain an unusual behavior
called ritual couvade that is common amongst a large part of the Amazonian Indigenous populations. It is
related to their belief that a child is the result of the mother engaging in multiple copulations with different men,
and their recognition of each of these men as co-fathers, which is known as Partible Paternity (Beckerman et al.,
1998; Beckerman et al., 2002; Beckerman & Valentine, 2002) in the anthropological literature. Given this
inability to determine paternity, Indigenous males remain oblivious as to whether their parental investment is
directed towards their biological child. I propose that the ritual couvade constitutes a symbolic re-enactment of
birth-giving, which attempts to portray an empathic commitment towards the pregnant mother, and whose
ultimate purpose is to become selected for the genetic father role during subsequent pregnancies in exchange for
committing increased paternal responsibility towards the present child of indeterminate genetic origin.
By considering the present-day Amazonian foragers as a working model for our evolutionary past, I suggest that
paternity confusion is the optimal stressor condition for the emergence of the cognitive traits at the base of male
enhanced empathic behavior display. Specifically, I will compare the feasibility of three hypotheses. First, that
in patrilineages ritual couvade is an exclusive formal sanctioning of the primary father, which provides no
biological advantages to the performer. Second, that in matrilineages, secondary fathers are encouraged to
perform ritual couvade in order to gain indirect genetic benefit by displaying emotional commitment. Third, that
in ambilateral groups women form philopatric residential clusters to enter in non-fraternal polyandrous
marriages. The latter marriage arrangement generates the highest degree of paternity confusion, and induces
thereby a higher level of paternal care within the realm of the cultural and environmental conditions of
Amazonian Indigenous populations.
To these ends, I carry out a systematic review of published research with the keywords Partible Paternity and
Ritual Couvade using the JSTOR database. The search result consisted of 13 mainly anthropological journal
articles of a descriptive nature, showing that in ambilateral groups believing in Partible Paternity, women tended
to enlist kinsmen as husbands depriving ritual couvade of the evolutionary force required for the emergence of
trait emotional intelligence (ref,. e.g. Goleman, 1995). Putting these findings into the context of Life History
Theory (Figueredo et al. 2006; Rushton 1985), it seems likely that female sexual promiscuity among Amazonian
lowland Indigenous populations (Crocker, 1990, 1994) increases women’s fitness and the survival of their
offspring by increased paternal support (Beckerman et al., 2002, Beckerman & Valentine, 2002; Hill & Hurtado,
1996; Pollock, 2002).
References
Beckerman, S., Lizarralde, R., Ballew, C., Schroeder, S., Fingelton, C., Garrison, A., & Smith, H. (1998). The
Bari Partible Paternity Project: Preliminary Results. Current Anthropology, 39(1), 164–168.
http://doi.org/10.1086/204706.
Beckerman, S., Lizarralde, R., Lizarralde, M., Bai, J., Ballew, C., Schroeder, S., .. & Palermo, M. (2002). The
Barí PP project, phase one. Cultures of multiple fathers: The theory and practice of PP in lowland South
America, 27-41. Gainesville. University Press of Florida.
Beckerman, S., & Valentine, P. (Eds.). (2002). Cultures of multiple fathers: The theory and practice of partible
paternity in lowland South America. University Press of Florida.
9
Crocker, W. H. (1990). The canela (eastern timbira), I: an ethnographic introduction (Vol. 1). Washington,
DC: Smithsonian Institution Press.
Crocker, W. H., & Crocker, J. (1994). The Canela: Bonding through kinship, ritual, and sex. Harcourt College
Pub.
Figueredo, A. J., Vásquez, G., Brumbach, B. H., Schneider, S. M. R., Sefcek, J. a., Tal, I. R., … Jacobs, W. J.
(2006). Consilience and Life History Theory: From genes to brain to reproductive strategy.
Developmental Review, 26(2), 243–275. http://doi.org/10.1016/j.dr.2006.02.002
Goleman, D., & Sutherland, S. (1996). Emotional intelligence: Why it can matter more than IQ. Nature,
379(6560), 1–34. http://doi.org/10.1016/j.paid.2003.12.003
Hill, K. R., & Hurtado, A. M. (1996). Ache life history: The ecology and demography of a foraging people.
Aldine de Gruyter, New York.
Pollock, D. (2002). Partible paternity and multiple paternity among the Kulina. In: Beckerman S, Valentine P,
editors. Cultures of Multiple Fathers: Theory and Practice of Partible Paternity in Lowland South
America. Gainsville, FL: University Press of Florida; 2002. pp. 42–61.
Rushton, J. P. (1985). Differential K theory: The sociobiology of individual and group differences. Personality
and Individual Differences. http://doi.org/10.1016/0191-8869(85)90137-0
Zahavi, A., & Zahavi, A. (1997). The Handicap Principle: A Missing Piece of Darwin’s Puzzle. Evolution and
Human Behavior (Vol. 117). Oxford University Press. Retrieved from
http://www.myilibrary.com/?id=83121
10
Invited contribution
1+1=3 When It Comes to Interaction With Animals
Linda Handlin
School of Health and Education, University of Skövde
linda.handlin@his.se
Throughout history, pets have lived in close contact with humans and have now become central to family life,
providing companionship and pleasure, and are often considered as family members. During the last decades,
research has emerged that shows health benefits associated with interactions with companion animals. For
example, pet ownership has been shown to improve cardiovascular health, animal contact have positive effects
on empathy and can reduce the subjective feeling of anxiety and promote calmness. Some animals have the
potential to reduce depression and to improve the mood of people who receive treatment for mental health
problems or patient in long-term care. These animals may also influence trust toward other humans. In addition,
children having a dog present in their classroom display increased social competence. Interaction between dogs
and their owners have been shown to induce oxytocin release in both the dogs and the owners and it seems as if
oxytocin is a major player when it comes to orchestrating the effects of human-animal interaction. Both the
physical contact with the dog and the attachment of the owner to the dog seems to play important roles in
generating these effects.
11
Interacting with the environment for remembering intentions: from normal to atypical
cognitive aging
Mattias Kristiansson
Department of Computer and Information Science, Human-Centered Systems
mattias.kristiansson@liu.se
In the poster I compare normal cognitive aging and people with aging-related neurocognitive diseases in terms of
prospects and issues for utilizing the physical environment (see for instance Clark, 2005) for forming and
executing intentions at an appropriate point in time and space (known as prospective memory, PM). The
comparison is based on (a) a previously conducted cognitive ethnography on normal older adults and (b) an
ongoing literature review on, and recently initiated observations of, and interviews with people with a dementia
diagnose. Such comparison can for instance be important for future developments of practices and artefacts for
both a normal and atypical population.
Cognitively normal older adults (+65) are known to perform better in real-life experiments measuring PM than
what is predicted from their performances on laboratory-based experiments (known as the age-prospective-
memory-paradox, see for instance Kvavilashvili & Fisher, 2007; Phillips, Henry, & Martin, 2008; Rendell &
Thomson, 1999; Uttl, 2008). Some studies explain this by suggesting that older adults are more efficient than
younger adults in their utilization of the physical environment to remember (Maylor, 1990, see Uttl, 2008 for
another explanation). Almost no studies have used observations in real-life to describe the mechanisms of the
efficient uses (see Palen & Aaløkke, 2006 for an exception). I have throughout my cognitive ethnography on
older adults and PM observed several inter- and intra-individual differences of practices used to more or less
efficiently couple with the environment to deal with activities such as leaving home with intended objects. One
such group of beneficial practices deals with the deliberate or more automatic practices of retrieving information
from an environment (an environment that can be more or less shaped to deal with specific PM situations,
Kristiansson, Wiik, & Prytz, 2014). From these observations I have concluded that a reason for why people
manage PM situations in real life is because they most of the time efficiently and rather quickly perceive and
interpret features in the physical environment as cues for previously formed intentions; and thereby are reminded
of what to do when in an upcoming or ongoing activity.
For the study of people with a dementia diagnose some observations of real-life situations exist in previous
literature. For instance it is to some extent known that a good shaping of the physical environment can be
absolutely crucial for them being able to keep track of near-future intentions (see for instance Vikström, Borell,
Stigsdotter-Neely, & Josephsson, 2005 on the activity making tea). A good shaping can for instance be
characterized by reducing the demand of attentional resources by the creation of more salient features. But it is
also known that for a dementia diagnose, for instance in the case Alzheimer’s disease the progression is
characterized by a neurocognitive deterioration of areas related to perceptual abilities (Braak & Braak, 1995).
Therefore, despite that the shaping of the physical environment is important for supporting attentional resources,
perceptual and interpretive issues of an external feature can still result in an inefficient coupling with the
environment to remember intentions. It seems that people with dementia disease, more than the general older
population, have to rely on more explicit cues and physical constraints to manage PM situations in real life. A
part of the inter-individual differences in practices among the typically developed older adults I have observed
deals with the variation of how explicit cues they create, and how much they physically constrain the likelihood
of being reminded by their environment. It can therefore be the case that some older adults use practices that are
more adapted to an everyday life with neurocognitive diseases, or adapted to living more cognitively challenging
everyday lives for other reasons. Since a large majority of research on dementia is based on the perspective of a
significant other, and also how the significant other shapes the physical environment for the person with
dementia disease, I aim to add to the field by empirically describing the practices the person with dementia use
to utilize the physical environment to keep track of future intentions.
12
References
Braak, H., & Braak, E. (1995). Staging of Alzheimer’s Disease-Related Neurofibrillary Changes. Neurobiology
of Aging, 16(3), 271–278.
Clark, A. (2005). Beyond the flesh: some lessons from a mole cricket. Artificial life, 11(1-2), 233–244.
Kristiansson, M., Wiik, R., & Prytz, E. (2014). Bodily orientations and actions as constituent parts of
remembering objects and intentions before leaving home: the case of older adults. Sensoria: A Journal of
Mind, Brain & Culture, 10(1), 21–27.
Kvavilashvili, L., & Fisher, L. (2007). Is time-based prospective remembering mediated by self-initiated
rehearsals? Role of incidental cues, ongoing activity, age, and motivation. Journal of Experimental
Psychology. General, 136(1), 112–132. doi:10.1037/0096-3445.136.1.112
Maylor, E. A. (1990). Age and Prospective Memory. Quarterly Journal of Experimental Psychology, 42A, 471–
493.
Palen, L., & Aaløkke, S. (2006). Of Pill Boxes and Piano Benches : “ Home-made ” Methods for Managing
Medication. In Computer Supported Cooperative Work Banff, Alberta, Canada (pp. 79–88).
Phillips, L. H., Henry, J. D., & Martin, M. (2008). Adult Aging and Prospective Memory: The Importance of
Ecological Validity. In M. Kliegel, M. A. McDaniel, & G. O. Einstein (Eds.), Prospective Memory:
Cognitive, Neuroscience, Developmental, and Applied Perspectives (pp. 161–185). London: Lawrence
Erlbaum Associates.
Rendell, P. G., & Thomson, D. M. (1999). Aging and prospective memory: differences between naturalistic and
laboratory tasks. The Journals of Gerontology. Series B, Psychological Sciences and Social Sciences,
54(4), P256–69. Retrieved from http://www.ncbi.nlm.nih.gov/pubmed/12382595
Uttl, B. (2008). Transparent meta-analysis of prospective memory and aging. PloS One, 3(2), e1568.
doi:10.1371/journal.pone.0001568
Vikström, S., Borell, L., Stigsdotter-Neely, A., & Josephsson, S. (2005). Caregivers’ Self-Initiated Support
Toward Their Partners With Dementia When Performing an Everyday Occupation Together at Home.
OTJR: Occupation, Participation and Health, 25(34), 149–159.
13
Measuring the noticing of an unexpected event in Magical Garden with a Teachable Agent
using Eye-Tracking
Ludvig Londos1
1Div. of Cognitive Science, Lund University
ludvig.londos@gmail.com
Not developing number sense in childhood can have dire consequences: failing early mathematics and
developing learning disabilities later on (Griffin, Case, & Siegler, 1994; Gersten, 1999; Chard et al, 2005). How
do you catch children’s attention and promote learning? The importance of play as a pedagogical tool for
teaching and to get children motivated in their acquisition for new abilities has been known for many years
(Griffin, Case, & Siegler, 1994; Geary, 1995; Gee, 2003). With new technology, a genre of educational games
for mathematics has emerged. Utilizing the motivational and captivating power of computer games, educational
games for mathematics have shown an effect on both learning and motivation (Schwartz, 2004; Moreno, 2005).
The educational game Magical Garden has all the prerequisites for training and testing number sense, and the
help of a Teachable Agent (TA). Axelsson et al. (2013) requested further research on preschooler’s social
interaction with TA. Schneider et al. (2008) emphasized the validity and utility of using eye-tracking as a
measure of developing number sense. The close connection between top-down control and eye movements
(Henderson, 2003; Deubel & Schneider, 1996), as well as Smith (2012) provide grounds for considering that
noticing something unexpected could be manifested as visual attention towards an Area of interest (AOI).
In the present study, eye-tracking was used as method to record if children noticed an unexpected event in
Magical Garden. The unexpected event was designed in a way that only the children who had a sufficient level
of number sense would react and notice the unexpected event. The unexpected event was a tree elevator
malfunction; the elevator passed the correct level and crashed in the tree top. A model of detection was
proposed: Looking back at the AOI of the correct level. The corresponding hypothesis was: “Looking back”
would correlate with the performance in Magical Garden. Performance was the rate of correct answer in the eye-
tracking session. Other eye-movements such as anticipation, and looking at the elevator button were collected.
In this study, 40 preschoolers participated (21 girls, M=4.6, SD=0.72), from three preschools, in the south of
Sweden. The study consisted of two phases; first a training phase, and then an eye-tracking experiment. The eye-
tracking experiment was conducted at the preschools, by having children play the specially designed version of
Magical Garden. The child and the TA took turns being in charge in the game. Will the children look at the TA
during an unexpected event and is there a difference in “look at TA” depending on who was in charge?
A significant result was found in that “looking back” correlated to performance, p = 0.018, 95% CI of [0.067 –
0.679]. A significant difference was found in that children looked at a higher rate at the TA when the TA was in
charge, p < .001, 95% CI of [0.064-0.223]. With an explorative look at the eye-movement data, “looking at the
elevator button” correlated strongly with performance r(38) = .50, p < .001.
This study introduces the noticing of an unexpected event as novel way of getting children to expose their level
of number sense without being in a test situation. The proposed model of noticing, a look back, did not account
for the whole notion of detecting an unexpected event. However, a better model of noticing could be constructed
by combining measurements of verbal, non-verbal detections, as well as eye movements such as “look back” and
“look at button”. Future research could learn from this study and examine the possibility of creating a better
model of noticing.
References
Axelsson, A., Anderberg, E., & Haake, M. (2013, January). Can preschoolers profit from a teachable agent based
play-and-learn game in mathematics?. In Artificial Intelligence in Education (pp. 289-298). Springer Berlin
Heidelberg.
Chard, D. J., Clarke, B., Baker, S., Otterstedt, J., Braun, D., & Katz, R. (2005). Using measures of number sense
to screen for difficulties in mathematics: Preliminary findings. Assessment for Effective Intervention, 30(2),
3-14.
14
Deubel, H., & Schneider, W. X. (1996). Saccade target se-lection and object recognition: Evidence for a
common attentional mechanism. Vision research, 36(12), 1827-1837.
Geary, D. C. (1995). Reflections of evolution and culture in children's cognition: Implications for mathematical
development and instruction. American Psychologist, 50(1), 24.
Gee, J. P. (2003). What video games have to teach us about learning and literacy. Computers in Entertainment
(CIE), 1(1), 20-20.
Gersten, R., & Chard, D. (1999). Number Sense Rethinking Arithmetic Instruction for Students with
Mathematical Disabilities. The Journal of special education, 33(1), 18-28.
Griffin, S.A., Case, R., & Siegler, R. S. (1994). Rightstart: Providing the central conceptual prerequisites for first
formal learning of arithmetic to students at risk for school failure. In K. McGilly (Ed.). Classroom lessons:
Integrating cognitive theory and classroom practice (pp. 25-49). Cambridge, MA: MIT Press
Henderson, J. M. (2003). Human gaze control during real-world scene perception. Trends in cognitive sciences,
7(11), 498-504.
Moreno, R., & Mayer, R. E. (2005). Role of Guidance, Reflection, and Interactivity in an Agent-Based
Multimedia Game. Journal of educational psychology, 97(1), 117.
Schneider, M., Heine, A., Thaler, V., Torbeyns, J., De Smedt, B., Verschaffel, L., ... & Stern, E. (2008). A
validation of eye movements as a measure of elementary school children's developing number sense.
Cognitive Development, 23(3), 409-422.
Schwartz, D. L., & Martin, T. (2004). Inventing to prepare for future learning: The hidden efficiency of
encouraging original student production in statistics instruction. Cognition and Instruction, 22(2), 129-184.
Smith, B. (2012). Eye tracking as a measure of noticing: A study of explicit recasts in SCMC. Language
Learning & Technology, 16(3), 53-81.
15
Perceiving, learning, and reasoning in arbitrary domains
Claes Strannegård1,Abdul Rahim Nizamani2, Jonas Juel, Ulf Persson3
1Department of Philosophy, Linguistics and Theory of Science, University of Gothenburg, Sweden and Department of
Applied Information Technology, Chalmers University of Technology, Sweden
2Department of Applied Information Technology, University of Gothenburg, Sweden
3Department of Mathematical Sciences, Chalmers University of Technology, Sweden
abdulrahim.nizamani@gu.se
When Alice in Wonderland fell down the rabbit hole, she entered a world that was completely new to her. She gradually
explored that world by perceiving, learning, and reasoning. This paper presents a system
that operates analogously. We model Alice’s Wonderland via a general notion of domain and Alice herself with
a computational model including an evolving belief set along with mechanisms for perceiving, learning, and reasoning.
The system operates both manually with human intervention, and autonomously by learning from random streams of
facts from arbitrary domains. It has proven able to challenge average human problem solvers in such domains as
propositional logic and elementary arithmetic.
The paper improves and extends on our previous work (Strannegård, Nizamani, & Persson, 2014; Strannegård, Niza-
mani, Juel, & Persson, in press 2015), The earlier versions of this model were able to learn and reason in arbitrary
domains, albeit when fed with carefully selected examples. The current system can learn from arbitrary streams of
observations, with or without human intervention. The computational complexity of the system is restricted by using
a simple cognitive model with bounded cognitive resources. This cognitive model is not a psychologically plausible
model of human thought, but is useful nonetheless in reducing the computational complexity in an artificial reasoning
system. It forms the basis for deductive reasoning, and learning of inductive rules is improved by a refined method
of abstraction and satisfiability. Introspection enables the system to check for soundness of the potential rules before
updating the belief set. The system is constructed with formal definitions of involved concepts, and examples are used
to illustrate its basic usage.
Most artificial reasoning systems are narrow and only understand a single domain. This system is designed to achieve
general intelligence, although limited to unambiguous symbolic domains. This seems to be a severe restriction, how-
ever, achieving general intelligence in even such domains is a way forward to achieving the larger goal of realizing
artificial agents with a broader general intelligence. This model may also be useful in understanding formal and math-
ematical properties of reasoning.
References
Strannegård, C., Nizamani, A. R., & Persson, U. (2014). A general system for learning and reasoning in symbolic
domains. In Artificial general intelligence (pp. 174–185). Springer.
Strannegård, C., Nizamani, A. R., Juel, J., & Persson, U. (in press 2015). Bounded cognitive resources and arbitrary
domains. The Eighth Conference on Artificial General Intelligence, Berlin, 2015.
16
How do car drivers make decisions?
Sara Nygårdhs1,2
1the Department of Computer and Information Science, Linköping University
2the Swedish National Road and Transport Research Institute
sara.nygardhs@vti.se
A car simulator study with 80 participants (aged 55-75 years) has been carried out at the Swedish National Road
and Transport Research Institute (VTI). Drivers in the study were confronted with varying traffic events that
“normally” occur in regular traffic. One example is roadworks in the opposing lane, where oncoming cars
neglect their obligation to give way and continue to drive past the roadworks instead. Another example is when a
child runs out in front of a bus at a bus stop in the driver’s lane. A third example is when a parked car suddenly
starts to drive out in front of the driver. The events were used for creating safety marginal measures, such as
time-to-collision (TTC) to a car ahead for instance. Before the driving session, the participants filled in
questionnaires, one of them concerning the likelihood of different factors being the cause of accidents in traffic
(i.e. Traffic Locus of Control, T-LOC). Based on results from the T-LOC questionnaire, the drivers have been
categorized in one out of three groups; as either finding that the likelihood for other drivers to cause an accident
is much larger than for themselves (Others), that the likelihood of causing an accident is about equal for
themselves and other drivers (Equals), or somewhere in between Others and Equals statistically (Betweeners).
The aim is to examine whether or not there is any difference in tactical decision-making between the different
driver categories discussed. The main goal of the approach is to understand drivers’ different decision making
strategies and how they affect traffic safety.
Until now there have for instance been studies in simulators where the participants could make subjective ratings
of e.g. feelings of risk (Lewis-Evans & Rothengatter, 2009) or how certain they are that they would be able to
avoid a collision if they encountered a deer on the road (Schmidt -Daffy, 2014). There have also been studies on
on-road behaviour of drivers using an on-road driving evaluation scoring system for finding out which cognitive
abilities and personality traits are related to driving performance among older drivers (Adrian, Postal, Moessinger,
Rascle, & Charles, 2011). However, connecting views of control to driver behaviour in a car simulator has, to the
best of our knowledge, not been attempted before.
References
Adrian, J., Postal, V., Moessinger, M., Rascle, N., & Charles, A. (2011). Personality traits and executive functions
related to on-road driving performance among older drivers. Accident Analysis & Prevention, 43, 1652-1659.
Lewis-Evans, B., & Rothengatter, T. (2009). Task difficulty, risk, effort and comfort in a simulated driving task –
Implications for Risk Allostasis Theory. Accident Analysis & Prevention, 41, 1053-1063.
Schmidt-Daffy, M. (2014). Prospect balancing theory: Bounded rationality of drivers’ speed choice. Accident
Analysis & Prevention, 63, 49-64.
17
Dualistic Thinking and Investigations into Consciousness:
Will the “Right” (Non)-dualism Please Stand Up?
Joel Parthemore
Department of Neuroscience and Philosophy, University of Skövde
joel.parthemore@his.se
It is common to treat all dualism as Cartesian substance dualism and all dualism as consequently bad. However,
dualism comes in a number of flavours, not just the well-known alternative of property dualism, where “mental”
and “physical” reflect not distinct substances but distinct sets of properties of a common substance; or David
Chalmers’ naturalistic dualism, which many dismiss – wrongly, I think – as nothing more than substance
dualism.
The call is frequently made to resist dualistic thinking in any form – but that cannot be right. In a crucial sense,
“dualistic” thinking is fundamental to our conceptual nature, deriving ultimately from the way that our ability to
identify something as an X depends on our ability to divide the world into Xs and not-Xs. While the various
schools of philosophical dualism take this practical conceptual requirement and elevate it to the metaphysical
stage, the idea of getting rid of dualisms altogether is hopelessly naïve and logically impossible.
In assessing varieties of dualism, it is useful to distinguish between those whose claims are essentially
ontological – thinking here primarily of variations on substance and property dualism – and those whose claims
are more epistemological. In the latter category, I place my own preferred form of dualism, which I prefer to call
perspectival dualism, although it is closely related to the position known as neutral monism, with such well-
known advocates as Spinoza, Bertrand Russell, and William James.
According to perspectival dualism, “mental” and “physical” reflect competing, complementary, mutually
necessary (each requiring the other), and yet ultimately irreconcilable views on one and the same world –
perspectives that we glide for the most part effortlessly and un-self-consciously between to the extent that the
two perspectives appear to blur into one. This key insight is my reason for preferring perspectival dualism over
other expressions of neutral monism and related positions.
Such philosophical conundrums as the so-called mind/body problem and explanatory gap arise because of a
largely unchallenged belief that there should and can be one, single, “correct” way of looking at the world. The
reality may well be that most if not all sufficiently complex phenomena simply do not have one, single, “correct”
explanation. The irony in that case is that all the “bad” forms of dualistic thinking arise precisely because of
resistance to perspectival pluralism, grounded in perspectival dualism.
18
Seeingred:PickingflowersinMinecraftwithQlearning
HenrikSiljebråt
1
&ChristianBalkenius
1
1
LundUniversityCognitiveScience
henrik@siljebrat.se
Robotsarestupid.Thoughwehavemadethemdoamazingthings,westillcannottellthegardenrobottopicka
nicebouquetofflowersasitneedstolearntoproperlydistinguishweedfromflower.Tobefair,thisparticular
problemisonesharedwithmanyuntrainedhumans,buthowcanweexpecttoteachthedifferencetoarobotwhen
wedon’tunderstandhowhumangardenerslearn?Recentdevelopmentsincognitivesciencepointstoaviewofthe
brainasapatternpredictorwheretopdownconnectionstrytopredicttheincomingsensoryinput(Clark,2013).In
therelatedfieldofmachinelearning,thereiscurrentlyalargehypearoundsocalleddeepneuralnetworks(Bengio,
2009).Suchnetworksuseastructureofhierarchicallayerssimilartothecorticallayersinthebrain.
Inrelationtohumanandanimalbehavior,theresultsofMnihetal.(2013,2015)areespeciallyinteresting.They
traineda“DeepQnetwork”toplay49twodimensionalvideogamesaswellasorbetterthananexperthuman
player.Thiswasaccomplishedwithonlypixelvaluesandgamescoreasinput.Themethodwasbasedonthe
Qlearningalgorithm(Watkins&Dayan,1992)combinedwithdeeplearningmethods.Thistypeofreinforcement
learningresemblesoperantconditioningasusedforanimallearning;foreveryactionthereisrewardorpunishment
(Staddon&Niv,2008).
Inspiredbytheseresults,weattemptasimilarbutsimplifiedalgorithmforteachinganagentwithQlearning.Can
thisapproachbeusefultorobotsandvirtualagentsseenasembodiedcreaturesinathreedimensionalenvironment?
Inordertosimulateanenvironmentmoresimilartotheonearobotoranimalwouldnavigate,thethreedimensional
videogameMinecraftwaschosen.Itsgameworldconsistsofcubesthatarecombinedtocreatefields,forests,hills,
plantsandanimalsmuchlikehowLegoworks.Theplayerseesthisworldinafirstpersonviewandcanmove
aroundbyusingthemouseandkeyboard.Interactionmainlyconsistsof“chopping”(hitting)blockstocollect
materialswhichallowsforplacingblockstobuildstructures.
Thegoalistoteachanagenttofindandpickasmanyredflowersaspossible,usingonlygamescreenpixelsas
inputandtheamountofflowerspickedasrewardsignal.Aninnatebehaviorwasadded,causingtheagentto
“chop”ifitseesredinthecenterofthescreen.Eachtrainingepisodetakesplaceinaflatrectangular“pasture”
filledwithredflowersongrassandsurroundedbywallstocontaintheagent.Asnotallchopsaresuccessful(the
flowermightbetoofaraway),theagenthastolearnhowtonavigateitsenvironmenttomaximizethesizeofits
flowerbouquet.
Usingthenumberofpickedflowersasaperformancemeasure,weexploredifferenttypesoflearningmechanisms,
comparingtheperformanceofourQlearnertoarandomflowerpicker.
References
Clark, A. (2013). Whatever next? Predictive brains, situated agents, and the future of cognitive science. Behavioral
andBrainSciences
,36
(03),181204.
Bengio,Y.(2009).LearningdeeparchitecturesforAI.Foundationsandtrends®inMachineLearning
,2
(1),1127.
Mnih, V., Kavukcuoglu, K., Silver, D., Graves, A., Antonoglou, I., Wierstra, D., & Riedmiller, M. (2013). Playing
Atariwithdeepreinforcementlearning.arXivpreprintarXiv:1312.5602
.
Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., ... & Hassabis, D. (2015).
Humanlevelcontrolthroughdeepreinforcementlearning.Nature
,518
(7540),529533.
Watkins,C.J.,&Dayan,P.(1992).Qlearning.Machinelearning
,8
(34),279292.
Staddon,J.E.,&Niv,Y.(2008).Operantconditioning.Scholarpedia
,3
(9),2318.
19
Versatile Systems Based on Reinforcement Learning
Claes Strannegård1,2
1Department of Applied Information Technology, Chalmers University of Technology
2Department of Philosophy, Linguistics and Theory of Science, University of Gothenburg
claes.strannegard@chalmers.se
The goal of artificial general intelligence is to create general intelligence at the human level or beyond (Goertzel &
Pennachin, 2007). To get anywhere near that goal one needs to construct versatile agents that can adapt to a wide range
of environments without any human intervention. In natural nervous systems, reinforcement learning is a powerful
mechanism that enables organisms to adapt to different environments and survive there (Niv, 2009). In artificial
systems, reinforcement learning has been used as the basis of relatively versatile agents, e.g. for robotic locomotion
across different anatomies and for gaming across different arcade games (Mnih et al., 2015).
In this talk I will discuss how still more versatile systems might be constructed, e.g. systems that can both learn how
to move like a snake and how to do simple mathematics. The formalism used is a variation of the transparent neural
networks (Strannegård, von Haugwitz, Wessberg, & Balkenius, 2013). The long-term memory is a network of this
kind that develops dynamically, partly based on factors relating to reward. The working memory is also a developing
network of the same kind, but with strict limitations imposed on its size.
The actions are stored in the long-term memory and they are of two types: physical actions that activate motor se-
quences, and mental actions that transform the working memory. The physical actions are used for generating motion
and the mental actions are used for generating computations with bounded cognitive resources, e.g. in simple mathe-
matics.
References
Goertzel, B., & Pennachin, C. (2007). Artificial general intelligence (Vol. 2). Springer.
Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A. A., Veness, J., Bellemare, M. G., .. . others (2015). Human-level
control through deep reinforcement learning. Nature,518(7540), 529–533.
Niv, Y. (2009). Reinforcement learning in the brain. Journal of Mathematical Psychology,53(3), 139–154.
Strannegård, C., von Haugwitz, R., Wessberg, J., & Balkenius, C. (2013). A cognitive architecture based on dual
process theory. In Proceedings of AGI 2013 (pp. 140–149). Springer.
20
Social attitudes toward robots with different degrees of human-likeness
Sam Thellman
Department of Computer and Information Science, Linköping University
samth549@student.liu.se
The physical appearance of robots that are supposed to interact with people is important because it helps establish
social expectations. Just as in the case of humans and animals, the external features of robots can function as indicators
for the internal mechanisms that govern their behavior. Employing degrees of human-likeness in the physical design
of robots is associated with both positive and negative effects on people’s attitudes toward them. Doing so can serve as
a design strategy to facilitate meaningful social human-robot interaction, but it can also give rise to “uncanny” feelings
or frustration by failing to meet social expectations. The causes of these effects are poorly understood.
In an experimental setup, 164 Swedish university students answered a questionnaire concerning their attitudes toward
robots. The questionnaire design was based on previous research on the Negative Attitudes toward Robots Scale
(NARS; Nomura, Suzuki, Kanda & Kato, 2006). NARS includes 14 questionnaire items divided into three subordinate
scales, covering different kinds of negative attitudes. The scale has previously been used in a range of scenarios to
identify several factors which affect people’s negativity toward robots, including gender, age, prior experience and
cultural differences (Tsui et al., 2011). Each participant was randomly assigned a questionnaire displaying one of
three robot images: a non-, semi- or a highly anthropomorphic robot type. The results suggest that employing an-
thropomorphism in robot design affect attitudes toward robots negatively primarily when the robot is intended for
social interaction. The semi-anthropomorphic robot type was perceived as less socially appealing—more unnerving,
dangerous, less dependable and less suitable for interaction with children—when compared to the other robot types.
There have been several proposed explanations to the negative effects associated with anthropomorphic robot design.
Some of them are based on the idea that negative effects are elicited by incongruent social expectations which arise
when people fail to understand or predict a robot’s social behavior. For example, MacDorman and Ishiguro (2006)
proposed that such effects “may be symptomatic of entities that elicit a model of human other but do not measure up
to it”. Ferrey, Burleigh and Fenske (2015) demonstrated that the mid-point between images on a continuum achored
by anthropomorphic and non-anthropomorphic entities produced a maximum of negative effect. The fact that the
semi-anthropomorphic robot type gave rise to more negativity than the non- and highly anthropomorphic types can be
explained by a failure to understand or predict the behavior of the robot and how to (or indeed whether to at all) interact
with it socially. Further research on the effects of robot anthropomorphism on attitudes toward robots is needed to
establish whether the proposed explanation is correct.
References
Ferrey, A. E., Burleigh, T. J., & Fenske, M. J. (2015). Stimulus-category competition, inhibition, and affective devalu-
ation: a novel account of the uncanny valley. Frontiers in psychology, 6.
MacDorman, K. F., & Ishiguro, H. (2006). The uncanny advantage of using androids in cognitive and social science
research. Interaction Studies, 7(3), 297-337.
Nomura, T., Suzuki, T., Kanda, T., & Kato, K. (2006). Measurement of negative attitudes toward robots. Interaction
Studies, 7(3), 437.
Tsui, K. M., Desai, M., A Yanco, H., Cramer, H., Kemper, N. (2011). Measuring attitudes towards telepresence robots.
International Journal of Intelligent Control and Systems, 16.
21
What difference does it make?
A computational model of sameness
Trond Arild Tjøstheim1& Christian Balkenius1
1Lund University Cognitive Science
kog13ttj@stud.lu.se, christian.balkenius@lucs.lu.se
"What’s the difference?" is probably one of the most common questions heard by teachers when students are trying
to learn new concepts. Unknown things have a sameness about them, we do not discriminate our actions towards
them. In a fundamental sense, to know is to be aware of differences.
This work is about building a computational model of a simple form of abstraction: the classification of a signal as
same as, or different from another signal. The ability to detect difference and sameness is fundamental to abstract
thinking, and indeed to human thinking in general (Farell, 1985). It is the basis for mathematical equivalence, and
is also necessary for language, since a word can be seen as a symbol for things that are somehow the same.
The models were built using the Ikaros framework (Balkenius, Morén, Johansson, & Johnsson, 2010), which
provides a way of efficiently setting up system level simulations, and has particularly good support for timing.
The core of the models are convolutional self organizing maps, or CSOMs (Balkenius, forthcoming). These can
be stacked, and do bottom up organization of inputs. Perceptron networks do supervised learning on temporal
differences of the CSOM’s activation, and learn to detect similarities or differences.
Three experiments have been carried out. The first replicates the basic elements of Wright and Katz (2006), where
capuchins and pigeons were taught to differentiate between pictures of natural scenes. The results indicate that the
model can learn to respond appropriately in less time and with higher accuracy than the animal subjects, and can
do so in the presence of random transformations like translation, rotation, scaling, and Gaussian noise.
The second experiment employs the same model as the first, but is given simple letter shapes as input. The
magnitude of the aforementioned transformations were also increased. The increase was 66% and 3000%, for
scale and rotation, and 50% and 600% for translation in x and y direction respectively. The results indicate that
the model performs above chance, but much lower than when transformations have smaller magnitude. It achieved
70% correctness for the larger magnitudes vs. 100% correctness for the smaller ones.
The third experiment adds separate learning units for each of the transformations, to investigate whether the model
can learn specific differences. Here, shape is most readily learnt, followed by rotation, while translation and scale
have the lowest correctness. Transformations are considered same when they are exactly equal.
What the models actually learn in all these cases are specific patterns of difference. In the first experiment, the
perceptron learns to distinguish the relatively small difference in activation due to the image transformations from
the much larger difference that occur when two different images are subtracted from one another. The inputs to the
perceptrons are matrices which can be interpreted as images. A small difference will leave only a few pixels with
high values, whereas a large one will yield several.
The results indicate that the models can make general same/different distinctions, and also that it is possible for
them to learn specific differences. These results may contribute to the realization of more complex categorization
abilities. Specifically, the models are meant to be integrated with attentional processes to do explicit, or rule based
categorization.
References
Balkenius, C., Morén, J., Johansson, B., & Johnsson, M. (2010). Ikaros: Building cognitive models for robots.
Advanced Engineering Informatics,24(1), 40–48.
Farell, B. (1985). Same-different judgments: A review of current controversies in perceptual comparisons. Psy-
chological Bulletin,98(3), 419.
Wright, A. A., & Katz, J. S. (2006). Mechanisms of same/different concept learning in primates and avians.
Behavioural Processes,72(3), 234–254.
22
Online Sexual Grooming and Offender Tactics – What can We Learn from
Social Media Dialogues?
Niklas Torstensson & Tarja Susi
The School of Informatics, University of Skövde
niklas.torstensson@his.se, tarja.susi@his.se
While online social networking sites and other digital media provide a means for positive online experiences,
they are also being misused for offences like online sexual grooming. Attempts have been made to analyse and
model online grooming in order to understand this kind of predator behaviour (e.g., O’Connell, 2004; Williams
et al., 2013). This research, and the resulting models of the grooming process, is however, invariably based on
material where adult decoys (e.g., researchers, law enforcement officers, adults trained to entrap offenders) pose
as children in the interaction with potential offenders. We argue that such material, i.e., decoy-offender chat logs,
does not reflect real grooming processes; Decoys have an underlying agenda to make prosecutable cases against
offenders, which entails decoys resorting to manipulation tactics otherwise typical for offender behaviour. In all
essence, this often leads to a dialogue with two adults using grooming tactics on each other, and the resulting
models do not capture the patterns of child-offender dialogues.
Contrary to previous research, we have analysed real-world child-offender chat logs from closed forums. Our
data set, selected dialogues (ca. 500 pages) from a corpus of ca. 12 000 A4-pages was thematically analysed and
categorised using NVivo 10 software. The coding was done by both authors for inter-rater reliability. Where
coding differed, the authors explored the categorisation until agreement was reached (cf., Whittle et al., 2013).
The material was also compared to decoy-offender chat logs (ca. 100 pages, publically available on perverted-
justice.com).
The analysis of the different data sets reveal quite different pictures of the grooming process. While previous
models describe the grooming process as sequential (O’Connell, 2004) or thematic (Williams et al., 2013), our
findings suggest a far more complex behavioural pattern – significantly diverse dialogue patterns with different
tactics emerge, depending on whether the respondent is a decoy or a child, and their respective responses. The
(preliminary) results show differences in both dialogue and process structure. Dialogues with decoys commonly
show what can best be described as “artificial compliance”, presumably due to their underlying agenda of
generating prosecutable cases. Furthermore, decoys tease out personal information from the offenders, and also
share “personal” information about themselves, even when not asked for it.
Child-offender dialogues instead show patterns of reluctance or objections to offender requests for personal
information, suggestions of sexual nature, etc. Another offender tactic is threats to obtain compliance, which was
not found in any of the analysed decoy-offender dialogues. Other deviations include differences in dialogue
length, number of dialogue turns, and complexity, with regard to changes in topics and offender tactics. Further
research is necessary for a more thorough understanding of online grooming, and new models are needed that
reflect real-world grooming processes. This includes offender behaviours, reasoning, decisions, and tactics used
in grooming. Further, such knowledge is of outmost importance for risk awareness measures for young people so
they can better cope with online challenges and risks, and make sensible judgements and decisions in online
interactions.
References
O’Connell, R. (2003). A typology of child cybersexploitation and online grooming practices. UK: Cyberspace
Research Unit, University of Central Lancashire. Retrieved from http://www.northern-
care.co.uk/assets/O%27connell.pdf
Whittle, H.C., Hamilton-Giachritsis, C. & Beech, A.R. (2013) Victim’s voices: The impact of online grooming
and sexual abuse. Universal Journal of Psychology, 1(2), 59-71.
Williams, R., Elliott, I. A. & Beech, A. R. (2013). Identifying sexual grooming themes used by Internet sex
offenders. Deviant Behavior, 34(2), 135-152.
23
Invited contribution
The Importance of Joint Episodic-Procedural Memory - What a Cognitive Architecture
needs for Effective Interaction
David Vernon
School of Informatics, University of Skövde
david.vernon@his.se
Prospection lies at the core of cognition: it is the means by which an agent - a person or a cognitive robot - shifts
its perspective from immediate sensory experience to anticipate future events, be they the actions of other agents
or the outcome of its own actions. Prospection, accomplished by internal simulation, requires mechanisms for
both perceptual imagery and motor imagery. While it is known that these two forms of imagery are tightly
entwined in the mirror neuron system, we do not yet have an effective model of the mentalizing network which
would provide a framework to integrate declarative episodic and procedural memory systems and to combine
experiential knowledge with skillful know-how. Such a framework would be founded on joint perceptuo-motor
representations. In this talk we examine the case for this form of representation, contrasting sensory-motor
theory with ideo-motor theory, and we discuss how such a framework could be realized by joint episodic-
procedural memory.
24
How information availability interacts with visual attention
Annika Wallin1, Philip Pärnamets1, Roger Johansson2 & Kerstin Gidlöf1
1Lund University Cognitive Science
2Department of Psychology Lund University
annika.wallin@lucs.lu.se
Decisions in front of a supermarket shelf probably involve a mix of visually available information and associated
memories – and interactions between those two. Several cognitive processes, such as decision-making, search
and various judgments, are therefore likely to co-occur, and each process will influence visual attention. We
conducted two eye-tracking experiments capturing parts of these features by having participants make either
judgments or decisions concerning products that had been previously encoded. Half the time participants made
their choices with full information about the available products and half the time with crucial task-relevant
information removed. By comparing participants’ use of visual attention during decisions and search- and
memory-based judgments, respectively, we can better understand how visual attention is differently employed
between tasks and how it depends on the visual environment. We found that participants’ visual attention during
decisions is sensitive to evaluations already made during encoding and strongly characterized by preferential
looking to the options later to be chosen. When the task environment is rich enough, participants engage in
advanced integrative visual behavior and improve their decision quality. In contrast, visual attention during
judgments made on the same products reflects a search-like behavior when all information is available and a
more focused type of visual behavior when information is removed. Our findings contribute not only to the
literature on how visual attention is used during decision-making but also to methodological questions
concerning how to measure and identify task-specific features of visual attention in ecologically valid ways.
25
26
