Skip to ContentSkip to Navigation
Over onsFaculty of Science and EngineeringAgenda

Bachelorsymposium KI

dinsdag 30 juni 2015

Sessie 1 (289) Taalwetenschap &  Kunstmatige Intelligentie


13:00 Ontvangst in zaal 289

13:15 Older Adults' Comprehension of Distributive and Collective Quantification — Maarten van der Velde (Begeleiding: Jennifer Spenader & Ken Drozd (Center for Language and Cognition))
13:30 Pronoun Interpretation by Older Adults — Hayo Ottens (Begeleiding: Jennifer Spenader & Margreet Vogelzang)
13:45 Investigating the interpretation of Dutch personal pronouns and demonstratives using a self-paced reading experiment — Niels Visscher (Begeleiding: Jennifer Spenader)

14:00 Pauze

14:15 Pronoun Interpretation in Direct and Indirect speech by Older Adults — Leonoor Ellen (Begeleiding: Jennifer Spenader & Emar Maier)
14:30 Pronoun Interpretation in a Direct and Indirect Speech Environment by Older Adults — Olaf Visker (Begeleiding: Jennifer Spenader)
14:45 Robots writing stories — Paul-Luuk Profijt - Jennifer Spenader - Robots writing stories (Begeleiding: Jennifer Spenader)

15:00 Pauze

15:15 Deep Belief Network on Context of Handwritten Words — Jonne Engelberts (Begeleiding: Lambert Schomaker)
15:30 Detecting humans from a top-down perspective using an unmanned aerial vehicle — Xeryus Stokkel (Begeleiding: Marco Wiering & Felipe Nascimento Martins)
15:45 Deep convolutional neural networks and support vector machines for gender recognition — Jos van de Wolfshaar (Begeleiding: Marco Wiering & Mahir Faik Karaaba)
16:00 Bag of Visual Words classifier for gender recognition — Ariane Meijer - van de Griend (Begeleiding: Marco Wiering & Mahir Faik Maraab)

16:15 BORREL

Sessie 2 (293) Cognitiewetenschap en Baysiaanse netwerken

13:00 Ontvangst in zaal 293

13:15 Modelling mind-wandering — Roald Baas & Joram Koiter (Begeleiding: Marieke van Vugt)
13:35 Increasing attentional blink using a strategy-based training — Tjesse Riemersma (Begeleiding: Jelmer Borst & Trudy Buwalda)

14:00 Pauze

14:15 Modelling legal cases in Bayesian networks — Siebert Looije samen met Steven Bosch (Begeleiding: Charlotte Vlek & Bart Verheij)
14:30 Modelling legal cases in Bayesian networks — Steven Bosch (Begeleiding: Charlotte Vlek)
14:45 Modelling a murder case in probability theory — Rene Mellema (Begeleiding: Bart Verheij)

15:00 Pauze

15:15 Predicting theory of mind-orders using agent-based modeling — Denny Diepgrond (Begeleiding: Harmen de Weerd)
15:30 Effects of Feedback on Human Performance in Negotiation with a Metacognitive ACT-R Model — Tom Renkema (Begeleiding: Christopher Stevens & Niels Taatgen)
15:45 Comparing a meta-cognitive ACT-R model of negotiation with human players — Jordi Top (Begeleiding: Cristopher Stevens & Niels Taatgen)

16:00 Pauze
16:15 Borrel!

_____________________________________

Abstracts Sessie 1: zaal 289

13:15 Older Adults' Comprehension of Distributive and Collective Quantification

Maarten van der Velde
Begeleiding: Jennifer Spenader & Ken Drozd (Center for Language and Cognition)
Sentences containing numerically quantified expressions (NQEs), such as 'Three people are holding two lamps',, can be interpreted in multiple ways. Inserting 'each' reduces the ambiguity of the sentence by strongly suggesting a distributive interpretation. Previous studies found that five-year old children were more likely to accept certain interpretations of such sentences that young adults deemed wrong. However, none of these studies included older adults. This study investigated how adults over the age of 65 interpret NQEs. The hypothesis was that older adults would be more child-like in their interpretations due to age-related declines in working memory capacity and inhibition. 24 older adults (mean age 74) participated in a truth-value judgement task with a 2 x 2 design. The factors were image type (distributive or collective) and sentence type (3/2 or 3/each/2). Additionally, working memory capacity and inhibition were measured. The results did not confirm the hypothesis. In fact, older adults' performance equalled or exceeded that of young adults. A mixed effects model contained significant effects of image type (p < 0.001) and sentence type (p < 0.001), and a significant interaction of these factors (p < 0.001). No effects of working memory capacity or inhibition were found.

13:30 Pronoun Interpretation by Older Adults

Hayo Ottens
Begeleiding: Jennifer Spenader & Margreet Vogelzang
As you might know, pronouns are used to replace parts of sentences. But the use of pronouns can cause ambiguity in the listeners interpretation. "The Pig is making bread. Yesterday the Pig asked the Elephant how to make the bread. While he saw him at the supermarket." Now, who was seen? By (sub)consiously going through different available discourse factors, people are often able to resolve the right antecedent, in this case the Elephant. Previous research shows that for children and young adults the more ambiguity in sentences, the harder the resolving process. Because aging is considered to correlate with lower working memory, but also with more experience in language, it would be interesting to test older adults too. Therefore, in this research, I'm testing people of 65 years and older on their interpretation of stories like this. Results will be presented at the symposium.

13:45 Investigating the interpretation of Dutch personal pronouns and demonstratives using a self-paced reading experiment

Niels Visscher
Begeleiding: Jennifer Spenader
This research investigates whether the choice of the Dutch male pronoun `hij' or the demonstrative `die', when refering to either a subject or an object antecedent in the last sentence, affects reaction times in a self-paced reading experiment. Contrast and grammatical role have been shown to contribute to the preferred interpretation of a referent in earlier research, and similar research in the German language has shown a significant effect of referent-bias conditions on reaction times in another self-paced reading experiment. A word-by-word self-paced reading experiment was set up in a two-by-two design, with stories in which the referent in the last sentence was either `hij' or `die', refering to either a subject or an object antecedent. We hypothesized that in the `die' condition when refering to a subject antecedent reaction times will be higher. 16 non-dyslectic native speakers of Dutch participated voluntarily. Reaction times were measured on two target positions, and spillover positions. Results will be presented at the symposium.

14:00 Pauze

14:15 Pronoun Interpretation in Direct and Indirect speech by Older Adults

Leonoor Ellen
Begeleiding: Jennifer Spenader & Emar Maier
Children aged up to eleven mix pronouns in indirect and direct speech in a referent-selection task. Direct speech is interpreted as indirect speech. No such pattern is found in adults. This study investigates the performance of older adults, aged 65+, on the same task. Since previous result suggest that the interpretation of the pronouns is influenced by executive functions, which are not as well developed in children, it was expected that the older adults show similar results. The task consisted of two parts, each with sentences in Dutch. In the no report condition the pronouns referred to the current context, without a direct/indirect speech distinction. In the report condition the pronouns referred to a previous context, with a direct/indirect speech distinction. In addition, a digit span and Stroop task were used. Older adults interpret direct speech less accurately than indirect speech, similar to children. The pronoun he has a lower accuracy than the other pronouns (I and you) in the no report condition, similar to children and adults. The Stroop task was a significant predictor of accuracy. This suggest that interpretation of pronouns is influenced by executive functions.

14:30 Pronoun Interpretation in a Direct and Indirect Speech Environment by Older Adults

Olaf Visker
Begeleiding: Jennifer Spenader
This study investigates the interpretation of pronouns in a direct speech (Zoidberg said, “I am a doctor”) and indirect speech (Zoidberg said that he was a doctor) environment for a single source narrator type, results obtained from older adults. This study also investigates if there is a correlation between the performance of interpreting direct speech and the ability to engage in cognitive inhibition. We hypothesize that older adults will perform better in a direct speech environment compared to an indirect speech environment. We also predict that cognitive inhibitory performance will positively correlate with the performance of direct speech interpretation. We tested 14 healthy Dutch-speaking older adults between 64 and 83 in a Stroop task, Digit-span task and a direct and indirect speech interpretation task. We found that older adults perform better in a direct speech environment, supporting our first hypothesis. In addition we have found differences between referents, regarding interpretation performance.

14:45 Robots writing stories

Paul-Luuk Profijt - Jennifer Spenader - Robots writing stories
Begeleiding: Jennifer Spenader
There are currently two methods to generate stories automatically: The first is the agent-based approach, which puts autonomous agents in a predefined world, and writes down what the agents do. This approach has the benefit that it allows for unpredicted events to occur, but the downside is that these events are often very uninteresting and chaotic. The second is the text-based approach, which traverses possible storylines through state transitions, much like a choose-your-own-adventure book. This approach has the benefit that it is very interesting and structured, but the downside is that everything must first be predefined, allowing for no new content to be generated. My research proposes an alternative approach, the plot-driven approach, that has the benefits of both approaches, but the downsides of neither. It does so by first generating a plausible "plot" using the text-based approach, sampling from a huge amount of available data and randomizing a lot of factors. It then generates a random, unique world that is populated by random, unique agents that are all free to do as they will, within the constraints of the plot. The result is a story that is unique, interesting and entirely random.

15:00 Pauze

15:15 Deep Belief Network on Context of Handwritten Words

Jonne Engelberts
Begeleiding: Lambert Schomaker
In this study feature vectors of word zones from the handwriting recognition system Monk were used to test whether deep learning on trigrams of these feature vectors results in semantic knowledge of these words. To deal with the lack of labeled data a Deep Belief Network was used, because this network allows unsupervised learning. In the first layer the dimensionality of the feature vectors were reduced with almost 99% from 4356 features to 50 features, without a big loss in Nearest Neighbor finding word zones with the same label: 75% correct for the 4356-sized raw data and 73% correct for the 50-sized compressed data. The compressed feature vectors of three succeeding word zones were padded together. These trigram feature vectors were used in the next layers of the network. Using Nearest Neighbor on the results, the network was able to find word zones with equal labels, but didn't show any semantic knowledge by finding semanticly close words.

15:30 Detecting humans from a top-down perspective using an unmanned aerial vehicle

Xeryus Stokkel
Begeleiding: Marco Wiering & Felipe Nascimento Martins
Search and rescue is often time and labour intensive, large groups of people search vast areas for missing people. In this paper we present a system that uses a drone to make search and rescue less resource intensive. The system uses a downward facing camera on the drone to detect people in open areas. The detector uses a sliding window to extract histogram of oriented gradients features that are classified using a Linear Support Vector Machine. Several preprocessing methods and models are compared for their classification and runtime performance. We also introduce a method to dynamically determine whether positive windows are true positives by looking at how they overlap. By doing this we hope to bring down the amount of false positives when the detector is used to count the number of people in an image. The developed method shows good performance on classifying frames as containing persons, when the system is used to determine the amount of people in a frame then the performance deteriorates. Although the detector shows great promise it is too slow to be of practical use.

15:45 Deep convolutional neural networks and support vector machines for gender recognition

Jos van de Wolfshaar
Begeleiding: Marco Wiering & Mahir Faik Karaaba
Social behavior and many cultural etiquettes are influenced by gender. There are numerous potential applications of automatic face gender recognition such as human-computer interaction systems, content based image search, video surveillance and more. The immense increase of images that are uploaded online has fostered the construction of large labeled datasets. Recently, impressive progress has been demonstrated in the closely related task of face verification using deep convolutional neural networks. In this paper we explore the applicability of deep convolutional neural networks on gender classification by fine tuning a pretrained neural network. In addition, we explore the performance of dropout support vector machines by training them on the deep features of the pretrained network as well as on the deep features of the fine tuned network. We evaluate our methods on the color FERET data collection and the recently constructed Adience data collection. We report cross-validated performance rates on each dataset. For the color FERET dataset we adopt our own partitioning. We further explore generalization capabilities of our approach by conducting cross-dataset tests. It is demonstrated that our fine tuning method exhibits state-of-the-art performance on the Adience dataset.

16:00 Bag of Visual Words classifier for gender recognition

Ariane Meijer - van de Griend
Begeleiding: Marco Wiering & Mahir Faik Maraab
Gender recognition is one of the first steps towards facial recognition. In this research we propose that the bag of visual words method is also a viable method for gender recognition. We will discuss the basic setup of this research, however it might be possible that no results will be available yet to present.

16:15 BORREL

_____________________________________

Abstracts Sessie 2: zaal 293

13:15 Modelling mind-wandering

Roald Baas & Joram Koiter
Begeleiding: Marieke van Vugt
Mind-wandering is the process of having task-unrelated thoughts. These thoughts can have both positive and negative effects during certain tasks. For these two cases, we will look at two types of mind-wandering: more adaptive mind-wandering (thinking about other things during a task) and executive failures (failing to attend to a task). The contents of task-unrelated thoughts can vary between lightweight daydreaming and the more emotionally loaded rumination which we also take into consideration. Two experiments were performed to look at the differences between these two types of mind-wandering: a Working Memory (WM) task and a Choice Reaction Time (CRT) task. The WM task was set up in a way that constant attention was required, while the CRT task didn't need as much continuous involvement. Preliminary results showed a significant difference in how much participants were on-task between WM and CRT. On average, participants paid more attention during the WM task. Building upon an ACT-R model by van Vugt, we model the two types of  mind-wandering, to determine whether it is necessary for there to be two different mechanisms to be able to distinguish between them. Results from the models compared to the human research will be presented at the symposium.

13:35 Increasing attentional blink using a strategy-based training

Tjesse Riemersma
Begeleiding: Jelmer Borst & Trudy Buwalda
De attentional blink is een paradigma dat al lang gebruikt wordt om limitaties in het verwerken van visuele stimuli te onderzoeken, dit omdat mensen over het algemeen suboptimaal presteren op een dergelijke taak. Recentelijk is gebleken dat door middel van specifieke training het effect van de attentional blink te verminderen is. Het huidige onderzoek breidt voort op een theorie door Taatgen et al.. Er wordt in deze theorie onderschijdt gemaakt tussen twee verschillende strategieën die het brein kan gebruiken tijdens de attentional blink en er wordt beargumenteerd dat het fenomeen ontstaat doordat mensen een suboptimale strategie toepassen. In dit experiment is gekeken of het mogelijk is om, door middel van training, mensen de suboptimale strategie aan te leren. Als dit het geval is zou het effect van de attentional blink groter moet worden. In twee verschillende onderzoeksopzetten zijn 24 mensen getest. Analyse van de data laat zien dat geen van de trainingstaken resulteerde in een vergrote attentional blink.

14:00 Pauze

14:15 Modelling legal cases in Bayesian networks

Siebert Looije samen met Steven Bosch
Begeleiding: Charlotte Vlek & Bart Verheij
In het recht moeten moeilijke beslissingen genomen worden door de rechters of juryleden zonder dat er een hulpmiddel gebruikt wordt. Om te helpen met het nemen van de beslissingen, is er een methode geïntroduceerd waarbij scenario's gecombineerd worden in een Bayesiaans netwerk. Dit zou de communicatie tussen rechters of juryleden met een expert kunnen verhogen. De methode zet de scenario's in een Bayesiaanse netwerk door vier stappen te volgen. De methode is al eerder onderzocht aan de hand van de “Anjum-zaak”. In dit artikel wordt onderzocht of het mogelijk is om met de methode de “schietpartij te Brielle” te modelleren en vervolgens wordt de methode geëvalueerd met behulp van de drie criteria uit het vorige onderzoek en drie nieuwe criteria. Het was mogelijk om deze strafzaak te modelleren aan de hand van de methode en de verschillende criteria te kunnen beantwoorden.

14:30 Modelling legal cases in Bayesian networks

Steven Bosch
Begeleiding: Charlotte Vlek
There are several ways of analyzing the facts of a legal case. In the court room, the most common approaches judges and juries use are argumentative and narrative. Forensic experts on the other hand, are more inclined to use a probabilistic way of analyzing their findings. In a recent paper, Vlek et al. (2014) discuss a method that aims to improve the communication between judges and forensic experts by modelling evidence and relevant scenarios of the case in a Bayesian network, specifically focussing on the concept of scenarios. This research analyzed this method by modelling the case of the 'mud murder' and evaluating it using a number of criteria. Results will be presented at the symposium.

14:45 Modelling a murder case in probability theory

Rene Mellema
Begeleiding: Bart Verheij
Redeneren met bewijs is een belangrijke taak voor mensen. Het is vooral belangrijk in de rechtspraak, waar de uitkomst van een redenatie het lot van mensen bepaalt. Hiervoor zijn drie normative raamwerken gecreëerd, de normative, de argumentative en probabilistische raamwerken. Een nieuwe methode om deze drie raamwerken te combineren was geïntroduceerd in Verheij (2014). Deze methode is getest met behulp van een case study van de Anjum zaak. Het is ook vergeleken met een Bayesiaanse methode die de normative en probabilistische raamwerken combineert. De nieuwe method kan de zaak modelleren.

15:00 Pauze

15:15 Predicting theory of mind-orders using agent-based modeling

Denny Diepgrond
Begeleiding: Harmen de Weerd
People use their beliefs, emotions and intentions to predict the possible outcomes of their decisions. Decisions made in a social context can lead to attributing these mental states to others. This phenomenon is called theory of mind. When people engage in social interactions, they use recursive thinking of the sort “I think that you think that I think”. The sophistication of this recursion reflects the theory of mind-order of the agent. In this study we will use data of earlier research on theory of mind in the zero-sum game “hide and seek” between participants and Bayesian agents. The aim of this study is to see if we can predict what order theory of mind humans use in simple social interactions. Agents based on reinforcement learning will be used to predict the theory of mind-behavior of both the agents and the human participants. The predicting abilities will first be tested on the Bayesian agents. These predictions can be evaluated with their known theory of mind-orders. After this we will predict the theory of mind-behavior of the human participants. Results will be shown and discussed at the symposium.

15:30 Effects of Feedback on Human Performance in Negotiation with a Metacognitive ACT-R Model

Tom Renkema
Begeleiding: Christopher Stevens & Niels Taatgen
In this study, an experiment was conducted in which human subjects played a mixed-motive negotiation game, called Game of Nines, against a metacognitive ACT-R agent. In this game, two players have to divide up 9 points between them. The goal for both players is to get as many points as possible in order to get a higher score. It was studied whether showing visual feedback about the agent's current negotiation strategy had an effect on the human player's performance. The results showed an improvement over time across all participants, but no overall effect of feedback was found. However, splitting up participants into medians sorted on their obtained score did show a positive effect of feedback for the best performing subjects. These participants showed a large improvement in score over time, whereas the best performing subjects who didn't receive feedback did not show any further improvement. This suggests that only the best performing people in Game of Nines benefit from the feedback and adapt their own strategy according to the agent's strategy, which leads to a higher score and thus to a better performance.

15:45 Comparing a meta-cognitive ACT-R model of negotiation with human players

Jordi Top
Begeleiding: Cristopher Stevens & Niels Taatgen
In this paper we investigate the question: "Can a meta-cognitive model for a mixed-motive bargaining task outperform humans when contesting human players?". In our experiment two parties had to negotiate, with a computer model taking the role of one of the negotiators on half of our trials. Participants were asked to rate their counterpart's agreeability, without knowing whether this was a human or the model. No signif cant difference in agreeability and absolute score was found, yet the model gained a sign cantly higher relative score. Thesefindings suggest that even if it only helps to improve relative gains, teaching people the meta- cognitive strategy can help them become better negotiators, and will not impair their performance on other relevant performance measures.

16:00 Pauze

16:15 BORREL


Last modified:01 February 2017 01.29 a.m.