ICCM: Poster session
Prof. Krishna Miyapuram
Differential payoffs can bias simple perceptual decisions. Drift Diffusion models (DDM) have been successfully used to simultaneously model for response times (RTs) and accuracy of binary decisions. The DDM allows for identification of latent parameters that represent psychological processes underlying perceptual decisions. These parameters characterize decision making as a noisy process that accumulates evidence towards one of the two boundaries. Previous research in two alternative forced choice (2AFC) experiments has found that asymmetric payoffs result in a bias towards those decisions that result in higher payoff. We manipulate the reward structure resulting in symmetric and asymmetric payoffs for a simple orientation discrimination task and test for the differences in parameters of drift diffusion model that might relate to reward-induced bias in perceptual decisions. To understand the mechanisms of how reward information might be integrated with perceptual decisions, we altered the relative timing i.e. processing order of reward information and perceptual stimuli.Computational modelling using a hierarchical DDM revealed starting point bias towards stimuli oriented in the direction of higher rewards in asymmetric as well as symmetric rewards.The drift rate reflected the average reward expectation when reward information was presented before, but not after the perceptual stimulus. Our results suggest that integration of rewards with perceptual decisions is mediated by modulating motivation for evidence accumulation over time and prior bias in starting point.
Tiffany (Jastrzembski) Myers
Research of mathematical models of learning and retention have focused on accounting for an individual’s performance across a variety of learning schedules (i.e., spaced and massed). The attempted goal of such research is to develop a model which can adequately predict human performance across a range of learning scenarios. However, little attention of this model development has focused on the interpretation of a model’s best fitting parameters given the structure of a model’s equations and its predicted performance values. The effect of this can lead to the development of models where the parameter values are correlated hindering a theoretical interpretation of performance. Here we examine the structure of the Predictive Performance Equation (PPE) and highlight portions of PPE’s equations that lead to correlations across its free parameters. We propose a fix for these issues (Modified PPE) and conduct a formal model comparison showing the Modified PPE is simpler, has less parameter correlation and its best fitting parameters map on to identifiable aspects of an individual’s performance.
Despite of strong historical connections between information theory and the study of perceptual independence and separability, few modern approaches take advantage of these connections. We revive Garner and Morton’s (1969) classic Mutual Uncertainty Analysis (MUA), complement it with Partial Information Decomposition (PID, Williams & Beer, 2010), and apply both to a sample of data from contemporary studies. While existing theories can dissociate between perceptual and decisional separability and identify dependencies at the level of individual stimuli, PID can provide analogous diagnostics for identifying the existence of perceptual independence and separability, decompose them into their constituents, and provide a measure for their strength.
Dr. Mary Kelly
We present how a cognitive architecture can be built from the neural circuit models proposed under the frameworks of holographic memory and neural generative coding. Specifically, we draw inspiration from well-known cognitive architectures such as ACT-R, Soar, Leabra, and Nengo, as well as the common model of cognition, to propose the kernel that might drive a complex, modular system that would prove useful for developing intelligent agents that tackle statistical learning tasks, as well as for answering questions and testing hypotheses in cognitive science and computational neuroscience.
We propose a retrieval interference-based explanation of a prediction advantage effect observed in Stone et al. (in press). They reported two dual-task eye-tracking experiments in which participants listened to instructions involving German possessive pronouns, e.g. ‘Click on his blue button’, and were asked to select the correct object from a set of objects displayed on screen. Participants’ eye movements showed predictive processing, such that the target object was fixated before its name was heard. Moreover, when the target and the antecedent of the pronoun matched in gender, predictions arose earlier than when the two genders mismatched — a prediction advantage. We propose that the prediction advantage arises due to similarity-based interference during antecedent retrieval, such that the overlap of gender features between the antecedent and possessum boosts the activation level of the latter and helps predict it faster. We report an ACT-R model supporting this hypothesis. Our model also provides a computational implementation of the idea that prediction can be thought of as memory retrieval. In addition, we provide a preliminary ACT-R model of changes in visual attention as a result of language processing.
Intelligence is fundamentally the ability for an agent to infer causal dependencies in its environment. However, the precise conceptualization across systems and scales is a polemical question. The concept of “Intelligence” may as well refer to a quantitative measure of formal cognitive ability than to a qualitative property of skilled agency. This difficulty in understanding the concept compounds when we try to scale to descriptive and predictive models of collective behavior. While it is self-evident that groups may leverage pairwise interactions or their collective resources to tackle complex problems, is that process only the sum of individual intelligences or is the group intelligent in its own right? If the latter, what does it mean for the classical internalist conception of intelligence and agency? If the former, then what is the proper scale of analysis in systems of nested organization, such as human societies? This question can be approach rigorously through a non-reductive account of the physical processes underlying intelligence. Here I propose that the latent model framework(with active inference as intrinsic reward mechanism) framework is a promising approach that could live up to the multiple dimensions of adeptness required by any framework that would attempt to generalize cognition across scales. A statistical state model for mathematical state transitions can be built and can be used to define cognitive models like causation and correlation.
Psychometric methods have been argued to not be able to test the assumption that the underlying latent scale is really an interval scale. More specifically, the Rasch model was accused to provide an interval scale only because it fits measurement error, an issue known as the "Rasch paradox". Regardless of whether the Rasch paradox is real or not, it would be interesting to be able to derive interval, or even ratio, scales from ordinal data. The aim of the present study is to propose a procedure that combines the probabilistic Guttman scaling with Goode’s method to obtain either an interval or a ratio scale from dichotomous psychometric data. We present how the procedures are combined to derive the metric scales and how fit to the data can be calculated using RMSE. Final considerations note the limitations of the procedure, but also value its potentials.
Dr. Aaron Bornstein
In patch leaving problems, foragers must decide between engaging with a currently available, but depleting, patch of resources or foregoing it to search for another, potentially better patch. Overharvesting, or staying in the patch longer than what is optimally prescribed, is widely observed in these problems. Most previous explanations for this phenomenon focus on how foragers’ mis-estimations of the environment could produce overharvesting. They suggest that if the forager correctly learned the environment’s quality, then they would behave according to Marginal Value Theorem (MVT). However, this proposal rests on the assumption that the forager has full knowledge of the environment’s structure. Rarely does this occur in the real world. Instead, foragers must learn the structure of their environment. Here, we model foragers as pairing an optimal decision rule with an optimal learning procedure that allows for the possibility of heterogeneously-structured (i.e. multimodal) reward distributions. We then show that this model can appear to produce overharvesting, as measured by the common optimality criterion, when applied to the usual tasks, which employ homogeneous reward distributions. This model accounts for behavior in a previous serial stay/leave task, and generates novel predictions regarding sequential effects that agree with participant behavior. Taken together, these results are consistent with overharvesting reflecting optimality with respect to a different set of conditions than MVT and suggests that MVT’s definition of optimality may need to be adjusted to account for behavior in more naturalistic contexts.
Prof. Cleotilde (Coty) Gonzalez
Dr. Leslie Blaha
Dr. Erin McCormick
In dynamic decision tasks, the situations we confront are never the same: the world is constantly changing. Generally, our ability to generalize learned skills depends on the similarity between the learned skills and the situations in which we will apply those skills. However, in dynamic tasks, the situations we are trained in will most likely be different from the situations in which we need to apply skills. For example, in the face of emergencies, one could be trained to handle hypothetical disaster scenarios, but remain unprepared for the emergency that is actually experienced. How can we best prepare for the unexpected? Cognitive Science research suggests that heterogeneity during training helps people’s adaptation to unexpected situations. However, evidence for a general diversity hypothesis is limited. In this research, we investigate this Diversity Hypothesis using a cognitive model of learning and decisions from experience based on Instance-Based Learning (IBL) Theory. We focus on the concept of decision complexity to investigate whether confronting decisions of diverse complexities results in improved adaptation to unexpected decision complexities, compared to situations of consistent decision complexity. We conduct a simulation experiment using an IBL model in a Gridworld task, and expose agents to learning various degrees of diversity; we then observe how these agents transfer their acquired knowledge to a novel decision complexity situation. Our results support the Diversity Hypothesis and the benefits of diversity on adaptation.
Dr. Aaron Bornstein
In behavioral economic experiments with randomized, or unstructured choice sets, trial-level sequential dependencies at the level of choice behavior or reaction time are usually assumed not to be present in behavior, and thus not explicitly accounted for. We present a flexible Bayesian hierarchical model that allows us to test for the presence or absence of linear stimulus-driven sequential effects on parameters of interest and subsequent choice. We apply this model to two data sets: one intertemporal choice and one risky decision making. We demonstrate sequential effects on risk tolerance inference and on deliberative evaluation of discounted value. Our results show that data collected in sequence cannot, without first verifying this assumption, be treated as if it were collected independently.
Prof. Cleotilde (Coty) Gonzalez
Traditional anti-phishing training is often non-personalized and does not typically account for human experiential learning. However, to personalize training, one requires accurate models and predictions of individual susceptibility to phishing emails. The present research is a step toward this goal. We propose an Instance-Based Learning model of phishing detection decision-making, constructed in the ACT-R cognitive architecture. We demonstrate the model’s ability to predict behavior in a frequency training study, and its generality by predicting behavior in another phishing detection study. The results shed additional light on human susceptibility to phishing emails and highlight the effectiveness of modeling phishing detection as decisions from experience. We discuss the implications of these results for personalized anti-phishing training.
In this paper we present the cognitive modeling library txt2actr, which facilitates the specification of an ACT-R environment through simple text files and partially automates the construction of certain components within a cognitive model. Our general purpose goes beyond this library and aims at promoting the modular parametrization and systematic evaluation for cognitive models. In particular, we suggest to establish benchmarks that allow (i) the competition among models with respect to classical tasks in experimental psychology, and (ii) the evaluation of possibly new or more applied tasks with respect to benchmark models. Such benchmarking proposals can be found in various other disciplines and usually serve as an incentive to improve existing theories and eventually converge towards a common language. Yet, txt2actr is far from providing a solution to the associated challenges. It rather serves as a proof of concept by illustrating how two model components for very specific cognitive phenomena in situation awareness can be applied in three different environments.
Dr. Tomi Silander
Inspired by Masip et al.'s (2016) test of ADCAT model's decision component, we wanted to see if we could replicate their findings using different data from a similar scenario-based study. They found that expected value of telling the truth predicted the decisions to lie or tell the truth more accurately than the expected value of lying, and even better than the motivation to lie, which they defined as a difference between these two expected values. In contrast, in our modeling study the motivation to lie was the best predictor of choices for both actual liars and truth tellers in conditions involving gains and large losses, whereas only in the condition involving large losses the expected value of telling the truth outperformed the expected value of lying. We conclude that whether the participants could gain something or avoid losing something by deceiving determined if they focused on benefits of lying or costs of telling the truth.
The study of knowledge representations and reasoning problems faced by a cognitive agent interacting with a dynamic and incompletely known world is relevant to cognitive robotics and understanding complex cognition and related fields. The paper introduces four cognitive agents that were modeled in a student project with specific requirements. The cognitive architecture ACT-R was used to model flexible agents that interact with objects in a grid field with only a limited field of view. Long-term planning is not possible here: the meaning of objects needs to be discovered and the field explored to find the goal as quickly as possible. The project demonstrates how the four agents learn from interactions and what information needs to be kept available to flexibly decide in unpredictably occurring situations. All four agents are shortly described in more detail. The project covers on a small scale some aspects that are crucial for autonomous agents in a simple game environment. The four agents are faced with 15 challenge environments that need to be explored and managed. The challenge performance results show that a higher number of productions does not necessarily lead to better performance.
The development of automated vehicles is accompanied by the question of how this technology will interact with vulnerable road users (VRUs; e.g. pedestrians, cyclists). Especially in shared spaces, implicit communication signals, such as vehicle deceleration, proved to be crucial. However, previous studies on the parameterization of vehicle deceleration indicated that human detection of vehicle deceleration may depend on various situational and individual factors. This research has two aims: (1) We want to investigate how the detection and perceptual decision-making on vehicle deceleration can be formally described using a cognitive model. For this, we discuss the applicability of a drift-diffusion model (DDM). (2) Further, we will follow up on previous research regarding the influence of different situational and individual factors on the detection performance and discuss how these factors could be related to the DDM parameters. With this research, we would like to contribute to a better understanding and a consistent, formal description of different factors influencing the detection of vehicle deceleration. This could be associated with improved interaction between automated vehicles and VRUs.
Dr. Megan Brianne Morris
Dr. Glenn Gunzelmann
Fatigue is a problematic factor in many workplace environments, resulting in safety and health risks that require monitoring and management. One means to monitor and manage fatigue is through the use of tools implementing biomathematical fatigue models to create assessment and predictions of operator fatigue based on sleep habits. Unfortunately, these models tend to provide assessments and predictions for an “average” operator given work schedules, lacking individualization. One way in which these models can be individualized is through the use of at-the-moment performance data that can modulate the model estimates. In the current effort, we describe an initial attempt at developing an algorithm to individualize fatigue assessments and predictions from a widely-used biomathematical fatigue model with performance data from a common attention task. We discuss the sleep dataset used for the effort, scaling procedure, and model fitting using a genetic algorithm. We then discuss future directions we will take to further increase the effectiveness and efficiency of the individualization capability and its implications.
While it is widely accepted that children use distributional information to acquire multiple components of language, the underpinnings of these achievements are unclear. The goal of the current work is to investigate the role of linguistic context in the acquisition of nouns and verbs. In particular, we use a Distributional Semantic Model (DSM) to predict the age of acquisition of nouns and verbs, and we analyse the hyperparameters of the model to find out how much context is helpful for the acquisition of these words. DSMs have been extensively evaluated against human adult ratings on semantic associations, but less so against children’s emerging semantic representations. For reasons of space, we limit our review of prior work to the most recent study that is closest to our goals. In that study, Alhama et al. (2020) propose two methods to evaluate DSMs for children’s acquisition of nouns. Their results suggest that the Skipgram version of word2vec (Mikolov et al., 2013) is most successful in predicting the Age of Acquisition (AofA) of nouns. In our work, we look more in-depth into the hyperparameters of Skipgram that best predict AofA, to find out more about the influence of context in acquisition. In addition, we extend the study to verbs.
Ms. Tanja Stoll
Evaluating criticality in driving is of utmost importance, especially in dynamic driving scenarios such as lane changing. Current theories assume that drivers' evaluation process is based on perception of time-to-collision (TTC). We argue that determining whether a situation is critical or not is guided by retrieving memories containing the perceived situation elements. This memory retrieval helps drivers build up situation awareness and it takes place regardless of whether participants possess a memory which is a perfect match to the scenario at hand, or only a partially matching one, including some of the situation’s elements. Therefore, memory retrieval leads to a more or less reliable situation model (SM) and anticipation about how the scenario is going to develop. Furthermore, we assume that drivers’ SM also includes the SM of a potentially relevant road user (i.e., one that might interfere with the driver) to some extent as well. We are creating an ACT-R model in order to explore perception and memory retrieval which result in a perceptual decision participants make about the situation’s criticality in a highly dynamic lane-changing scenario.
Gerrit Jan Kootstra
Prof. Rob Schoonen
Bilingual speakers are more likely to use a syntactic structure in one language if they have recently encountered that same structure in another language. This cross-language structural priming effect is predicted to be positively modulated by second language proficiency according to a developmental account by Hartsuiker and Bernolet (2017). We propose to extend this account from sequential bilinguals to simultaneous bilinguals. In this latter group, syntactic structures develop in parallel and can integrate from the onset. Therefore, we do not expect proficiency or other measures of development, such as exposure, to modulate cross-language structural priming in these bilinguals. In simulated cross-language structural priming experiments, we explored how proficiency affects priming of transitives. We use an implicit learning model of sentence production to model the simultaneous English-Spanish bilinguals in these simulations. Furthermore, we investigated whether the priming effect is modulated by exposure to the non-dominant language, which only Kutasi et al. (2018) also analyzed. We found no evidence for any modulating effects for either proficiency or exposure, which is in line with the previously reported behavioral result of Kutasi et al. (2018). Together, our modeling results and Kutasi et al.’s (2018) behavioral results support an extended version of the developmental account of cross-language structural priming that predicts a modulating effect of proficiency in sequential bilinguals, but not in simultaneous bilinguals.
Dr. Jelmer Borst
Dr. Anirudh Unni
Prof. Jochem Rieger
In an effort towards improving the safety in everyday traffic, adaptive automation has emerged as a promising technology in recent years. A key step in this approach is the accurate prediction of momentary cognitive workload while driving. Previous research has found an interaction between working memory load and visuospatial attention complicating the accurate predicition for these cognitive concepts. We have developed an ACT-R model to investigate the nature of the interaction and improve the prediction accuracy for working memory load and visuospatial attention while driving. This ACT-R model is driving on a multi-lane highway with concurring traffic and alternating lane-widths while doing a secondary n-back task using speed signs. Furthermore, it is able to handle complex driving situations like overtaking traffic and adjusting its speed according to the n-back task. The behavioral results show an increase in error rates in the secondary task with increasing n-back difficulty as well as a decrease in driving performance with increasing difficulty in the n-back task. The results of the model indicate an interaction at a common task-unspecific level.
The weak completion semantics is a three-valued, non-monotonic theory which has been shown to adequately model various cognitive reasoning tasks. In this paper we extend the weak completion semantics to model disjunctions and exclusive disjunctions. Such disjunctions are encoded by integrity constraints and skeptical abduction is applied to compute logical consequences. We discuss various examples and relate the approach to the elimination of disjunctions in the calculus of natural deduction.
Dr. Robert Goldstone
Prof. Cleotilde (Coty) Gonzalez
Many of the decisions we make in day-to-day life are made on the basis of incomplete information. The experiences of members of our social network are often an important source of decision-relevant information. In a 2008 experiment, Mason, Jones, and Goldstone showed that a person’s social network structure can have an impact on their success at identifying the optimal decision given incomplete information: Members of more interconnected networks excelled at easier tasks, while members of more dispersed networks did comparatively well when the task was more difficult. Drawing on these results, we synthesize work from various areas of cognitive science into a computational cognitive model of search in a social context: the Social Interpolation Model (SIM). The SIM incorporates three avenues for individual difference, or free parameters: breadth of generalization, degree of optimism, and degree to which personal experience is weighted more heavily than the experiences of others. We report the results of simulations of interacting agents who are embedded in the same task structure as the one designed by Mason et al. (2008) and whose behavior is determined by the SIM. Based on these simulation results, we discuss qualitative effects of varying each of the SIM’s free parameters in the context of different social network structures. Our work highlights interaction effects between information-processing biases, social context and task structure on agents’ success at identifying the optimal solution.
Dr. Jana Jarecki
This work compares two types of psychological similarity in categorization. Similarity is a central component of categorization theories. Exemplar theories, for instance, assume that people categorize new exemplars based on their similarity to previous category members. Traditionally, the underlying psychological similarity is based on the sum of two exemplars' squared feature value differences (Euclidean similarity). The Euclidean similarity, however, ignores the distribution of exemplars within categories by assuming uncorrelated features within categories. The Mahalanobis similarity, in turn, extends the Euclidean similarity by accounting for within-category feature correlations. Results from machine learning have shown that in categorization problems involving correlated features within categories, the Mahalanobis similarity can outperform the Euclidean similarity. On the empirical side, results from psychology indicate that people can be sensitive to within-category feature correlations: Some findings suggest a general sensitivity for within-category feature correlations, yet others have argued that this sensitivity depends on the category structure, task format, and amount of training. The present work rigorously tested the correlation-insensitive Euclidean similarity against the correlation-sensitive Mahalanobis similarity to investigate if people use within-category feature correlations in categorization.