Posters: Applied & Meta-Science

Recognising the intention of a human partner is a key challenge for collaborative systems in human-robot interaction. However, existing studies of intention recognition abilities in AI system mostly focus on data-driven approaches and the recognition of direct action intentions (low-level intentions). We propose an artificial intention recognition approach that is implemented as a cognitive model in the theory-based ACT-R architecture and that infers superordinate action goals (high-level goals). We tested our approach for the recognition of cocktails from mixing sequences performed by human participants in an experimental study. Intention recognition speed of the model was evaluated and compared to human intention recognition performance. Our results indicate that the implemented model successfully recognises high-level intentions and tends to be substantially faster than humans.

I am broadly interested in applied statistics with particular emphasis on cognitive psychology. In particular, my current project explores how doctors and patients use the qualitative and quantitative probability language to communicate the chances of pregnancy in fertility treatment contexts. Doctors commonly use probabilistic language when communicating the evidence-based chances that a treatment will be successful. When asked about terms used by doctors --- such as that pregnancy is “likely” or has “little chance” --- doctors and patients showed heterogeneity in both their behavior and their ordinal rankings of terms and their quantitative ratings. Specifically, the wide variety between and within individuals in terms of their judgement of probability estimates across contexts; the fatigue from participants causing the individuals fluctuation among a long list of identical probability terms (e.g., likely, probably, probable, better than even, unlikely etc.). Furthermore, a pairwise correlation test between participants’ ranking responses explicitly showed three clusters within the dataset. The clusters illustrated three types of participants in the dataset, who ranked the terms in the requested order; a small proportion of participants ranked the terms in a reversed order; and some compiled the task partially. I am interested in formally modeling the probability judgements to understand both differences across people within a group (individual differences) and differences between groups (doctors). In the results of the data analysis, I am looking forward to providing insightful and referable guidance for doctors so that they have a certain level of knowledge regarding if using probability language actually facilitates the clearness of a conversation. If that is helpful, what specific forms or certain words are recommended in the fertility conversations. By studying the unique characteristics fertility patients might have in terms of understanding probability language, they are expected to benefit from both emotionally and physically because more understandable information can be derived from the effective communication facilitating their decision-making process. Therefore, all sources of heterogeneity should be taken seriously in the data analysis in expecting a meaningful result.

Computational psychology is a growing field that uses computer simulations and mathematical modeling to explain and predict complex behavior in psychology, psychiatry, and neuroscience. It is now one of the priority areas for research funders, scientific journals, and faculty hiring. However, anecdotal evidence has always pointed to longstanding diversity issues in the field with a lack of representation among women and black individuals. We sought to move beyond anecdotal accounts and examine the extent of these disparities in awards given and authorship of peer-reviewed articles. Our goal was to highlight the need for increased diversity, equity, and inclusion (DEI) within computational psychology. Name-based classifiers were used to estimate individuals’ gender and race based on first and last names. From these classifications, we examined demographic trends among N=27,163 authors and N=390 award-winners in preeminent journals/societies across computational psychology (including computational psychiatry, mathematical psychology, and computational neuroscience), psychology/neuroscience, and computational science. Results indicated that women represented just 23% of authors and 15% of award-winners within computational psychology—markedly lower than computational science (authors: 29%) and psychology/neuroscience (authors: 40-47%; award-winners: 38%). Black individuals were underrepresented among authors (2-4% of authors in computational science and psychology/neuroscience), but representation was lowest in computational psychology, where black individuals represented a mere 0.8% of authors. Taken together, these findings highlight major gender and racial disparities among computational psychology researchers. Evidence suggests that diverse teams tend to show better performance, more creativity, and produce more novel, high-impact science. Therefore, these disparities emphasize the necessity of targeted efforts (e.g., outreach, mentorship programs) to increase DEI within computational psychology to ensure equitable access to resources and promote scientific advancement.

We aim to create an explanatory formal model for addiction. We deem earlier attempts to create this type of model for addiction too complex and thus we propose to use just one ordinary differential equation: dN/dt = r * N * (1 - N/K) - (B * N^2) / (A^2 + N^2) This equation has been studied extensively since originally proposed by Ludwig et al. (1978) to model the outbreak of spruce budworms. From only the first term it would follow that N grows with rate r until limit K is reached. However, the second term controls the growth of N. The larger B in the second term the more the growth of N is controlled, with maximum control of the growth at A. How these parameters are interpreted in addiction depends on the specific type of addiction. In general, N can be thought to represent the consumption of an addictive substance or the frequency with which addictive behavior occurs. The r can be interpreted as the rate at which consumption leads to more consumption, which could for example be influenced by brain processes and peer pressure. B could represent the upper limit of self-control, which is reached if consumption gets so high that behavioral control is lost. 1/A can be thought of how fast self-control starts to influence behavior, which can for example be influenced by the social environment and beliefs about the consequences of the addictive behavior. The equilibrium states of the model we propose can be described by a cusp catastrophe model. In the cusp, there are two stable states, which could correspond to problematic or non-problematic behavior in terms of addiction. The behavior of the cusp catastrophe model can reproduce some of the important phenomena that are present in addiction. For one, quitting in addiction is hard, which corresponds to the hysteresis effect that we see in the cusp. Moreover quitting and relapsing are often sudden phase transitions just as can occur in a cusp catastrophe model. The cusp model also allows for more gradual changes which can be more appropriate for the initial transition to problematic behavior or substance use.

Working memory, a cognitive system involved in the retention and manipulation of stored information, plays a critical role in many cognitive processes and in cognitive development. Understanding how WM processing develops over time and how it interacts with education is critical for improving cognitive outcomes in children. In this study, we analyzed data from a large online adaptive learning platform to examine the development of working memory (WM) processing in children. Data were collected from elementary school students between grades 3 and 8 who played two different working memory games, a verbal WM task, and a visuospatial WM task. Using item response theory, multilevel modeling, and cognitive modeling, we examined classic WM benchmarks to gain insight into the dynamic developmental trajectory of WM processing. We found that item characteristics, particularly set size, affect item difficulty across the age range. We also investigated primacy and recency effects and found that position effects vary across age groups, suggesting that there are dynamic changes in WM processing as children grow older. Finally, we analyzed different types of errors and found that children were more likely to forget an item than to add or repeat it. However, as children matured, we observed a decreased tendency to forget items but an increased tendency to erroneously repeat them. Our findings provide an understanding of the dynamic development of WM processing in children and highlight the robustness of classical WM findings.

Cognitive impairment is an often-overlooked issue that cancer survivors face, with a third of non-CNS cancer survivors reporting memory problems. Memory, however, is complex and consists of various underlying cognitive processes. The objective of this research is to investigate memory problems more thoroughly in cancer patients. This was done through an adapted Hierarchical Bayesian cognitive model from the Alzheimer’s Disease literature, which splits memory into several processes relating to either learning or retrieving words from any of three memory states (unlearned, partially learned, learned). Participants were cancer survivors (n=187) of various non-CNS tumors (breast, prostate, and others) who received various cancer treatments (chemo-, endocrine-, radio- and immunotherapy) and no-cancer controls (n=204). The participants completed the Amsterdam Cognition Scan (ACS), in which classical neuropsychological tests are digitally recreated for online at-home administration. The specific test used to investigate verbal memory was the ACS equivalent of a Verbal Learning Test, in which participants are tasked with recalling a list of 15 words five times. Later in the test battery the participant is asked to recall these 15 words again, as a delayed recall trial. A traditional analysis of the sum of trials 1-5 indicated a small effect size difference between patients and controls, t(385.23)=2.81, p <.01, d=.28. There was no significant difference between patients and controls in the delayed recall trial. For the underlying memory processes, significant differences were found in the immediate retrieval process parameters, both retrieval from a partially learned state (t(378.47)=2.6, p <.01, d = .26) and retrieval from a learned state (t(381.57)=2.44, p =.02, d = .25). No differences were found in any parameters related to learning processes, nor in the delayed retrieval process parameter. The results indicate that the memory problems in cancer survivors are likely due to selective impairment of memory retrieval processes, rather than through learning impairment or a general impairment.

We present the first steps towards a processing model to understand transitions of control in semi-automated vehicles. In a transition of control, a human takes over the control from a (semi-) automated vehicle. Based on a recent theoretical model, we describe this process as interruption handling. In an interruption handling process, various distinct processing steps can be identified. We then take the data from a recent meta-review on transitions of control to map response times to specific processing stages of interruption handling. We then estimate the response time distribution for each stage. The model can then be used to identify what response distributions might look like for different scenarios, such as different alert modalities. Initial findings highlight how for example bi-modal alerts mostly speed-up initial processes of the interruption handling, but later processes less so.

In this study, we contrasted six different models to show the effects of different interventions by adaptive systems designed to prevent mind-wandering while driving. Although cognitive load associated with secondary tasks tends to affect driving negatively (e.g., Unni et al., 2017; Salvucci & Macuga, 2002; Ito et al., 2001), sometimes a simple secondary task can improve driving performance when the situation is mundane (e.g., Engström et al., 2017; Nijboer et al., 2016). Nijboer and colleagues (2016) have hypothesized that if the driving task is simple, people might start mind wandering, which interferes with driving (Yanko & Spalek, 2013, 2014; Martens & Brouwer, 2013). A simple secondary task, which imposes less workload than mind-wandering, could prevent this from happening. Automation system that adapt to the cognitive state of the driver could leverage this effect by inducing mild cognitive load during mundane driving scenarios with the goal to improve driving performance. To test suitable interventions, we combined an existing driver model (Salvucci, 2006) with an existing model of mind wandering in the cognitive architecture ACT-R (van Vugt et al., 2015) and tested different interventions that impose cognitive workload in different amounts during specific times of the simulation. Using these different models we, firstly, show how mind-wandering harms driving performance, secondly, show that mild cognitive load can mitigate this effect and, lastly, show that adapting to the cognitive state of the model incurs a significant processing cost that adaptive automation systems have to account for.

We tend to make judgments of our peers in many ways, including their level of competence, and we are more likely to interact with those who we have judged as competent (Fiske, 2007). Such judgments may partly be affected by cognitive factors within us, such as our internal control or our quest for knowledge, and it is possible that our judgments might also be attenuated by the extent to which we believe that absolute truth and morality is subjective (Forsyth, 1980). The present study hypothesized a structural model to evaluate the above ideas. The outcome variable of judged peer competency was a measured variable using the competency scale. Cognitive factors was the latent predictor variable indicated by need for cognition, symbolic immortality, and internalized control. Holding relativistic beliefs was the measured mediator variable based on the relativism subscale. Inventories were completed by 193 undergraduate at California State University, Sacramento students. The chi-square value for model fit was not statistically significant, 8.105 (4, N = 193), p = .08, and most of the other indexes also suggested a good fit: GFI = .98, NFI = .93, CFI = .96, RMSEA = .07. Of the three paths, only two were statistically significant. The path from abstract definitions to perceptions of peer competence was not significant, but the paths from our cognitive factor to relativistic beliefs (path coefficient = -.18) and the path from relativistic beliefs to perceptions of peer competence (path coefficient = .18) were each significant, with approximately 18% of the variance in perceptions of peer competence explained. Because the direct path path from our cognitive factor to perceptions of peer competency was not significant in the mediated model, there was the possibility that mediation might have been obtained. We evaluated that by evaluating the unmediated model. However, in the unmediated model the direct path from the cognitive factor to perceptions of peer competency was not statistically significant (path coefficient = -.03). It therefore appears that our cognitive factor was not directly related to perceived peer competency, but did act through relativistic beliefs in affecting judged peer competence.