ICCM Posters

There is significant utility in determining individual memory characteristics, as these might affect learning strategy. This study aimed to assess how well estimated model parameters that reflect individual cognitive characteristics, predict learning strategy. A weighted mixture model based on ACT-R, of the stimulus-response learning task (RLWM, Collins, 2018) that contained declarative (LTM) and procedural memory (RL) components was used. Learning strategy was measured by estimating what percentage of trials were learned with LTM vs RL. The ACT-R LTM parameters speed of forgetting, (SoF, which measures memory decay rate, Pavlik & Anderson, 2008) and spreading activation (S, working memory capacity analog) and RL learning rate parameter (α) were estimated for each individual by selecting the best-fitting set of parameters to behavioral data. We hypothesize that the mostly-LTM group would have significantly higher S values and lower SoF values compared to the mostly-RL group. We expected that the mostly-RL group would have higher learning rate values. This would suggest that ultimate learning strategy choices might rely on available individual memory characteristics. Model fits were remarkably good, achieving low root-mean-squared-error (M:0.064, SEM=0.0012). To test our hypothesis, participants were grouped into ‘mostly-LTM’ (greater than 50% of trials were performed with the LTM model) and ‘mostly-RL’ (greater than 50% RL engagement). We found that the RL parameter learning rate was a significant predictor of learning strategy (p < 0.0001) but not the declarative SoF (p=0.196) and S parameters (p=0.424). Our results suggest that individual differences might be best captured by RL models, compared to models of declarative memory.

This is an in-person presentation on July 20, 2024 (17:00 ~ 20:00 CEST).

As people engage in tasks over extended periods, their psychological states change (e.g., due to practice effects or boredom). Despite this, the most common methods for modelling cognitive processes, such as evidence accumulation models, only consider a single estimate of a process across the duration of an experiment, therefore failing to account for important time-varying factors such as learning. In this study, we describe a simple method for modelling time-varying changes to diffusion model parameters by assuming that rather than being constant across time, their estimates follow theoretically informed trial-varying or block-varying functions (e.g., exponential functions). Focusing on two parameters, drift-rate (task efficiency) and threshold (caution) and a number of candidate time-varying functions, we assessed 1) the measurement properties of this framework, 2) the extent to which these models could describe empirical data from three typical experimental psychology paradigms over and above the standard diffusion model, and 3) how much the standard diffusion model was mislead by time-varying processes in these data.

This is an in-person presentation on July 20, 2024 (17:00 ~ 20:00 CEST).

To properly capture interindividual variability in cognitive processes, cognitive modelers increasingly employ hierarchical Bayesian models in which subjects are treated as random effects. Additional random effects may be added to the model for stimuli, for example, when subjects are crossed with social target stimuli as in social cognition experiments (Judd et al., 2012). To date, few simulation studies have comprehensively investigated the estimation performance of such more complex hierarchical cognitive models. In our simulation study, we sought to close this gap for the crossed random effects variant of the Drift Diffusion Model (DDM; Ratcliff, 1978; Vandekerckhove et al., 2010). We used a simulation design with two crossed random effects - mirroring subjects and targets as in social cognition experiments - and we varied design settings in ways realistic to such experiments. Specifically, we manipulated the variance of subject and target population distributions, mirroring homo- vs. heterogeneous populations, as well as the number of draws from each population, mirroring subject and trial numbers. Additionally, we manipulated model complexity by inducing constraints on the estimated random effect structure (crossed vs. single vs. no random effects). All models were estimated in JAGS and their estimation was evaluated based on different performance criteria (e.g., bias). Importantly, performance evaluation considered both the individual and the population level for both subjects and targets, providing novel insights into the interplay of multiple random effects on cognitive parameter estimation across levels.

This is an in-person presentation on July 20, 2024 (17:00 ~ 20:00 CEST).

Transfer learning, the reuse of newly acquired knowledge under novel circumstances, is a critical hallmark of human intelligence, yet the underlying computations have been little investigated in humans. We argue that successful transfer learning upon task acquisition is ensured by updating inductive biases and transfer of knowledge hinges upon capturing the structure of the task in the inductive bias. To explore this, we trained participants on a non-trivial visual stimulus sequence task (Alternating Serial Response Times, ASRT). During the training, participants were trained in two distinct sequences successively, while the underlying structure of the task remained the same. We analyzed the acquired knowledge by recovering individual internal models of the task using infinite Hidden Markov Models. Our results show that beyond the acquisition of the stimulus sequence, our participants were also able to update their inductive biases. Acquisition of the new sequence was considerably sped up by earlier exposure, but this enhancement was specific to individuals showing signatures of abandoning initial inductive biases. Enhancement of learning was reflected in the development of a new internal model. Additionally, our findings highlight the ability of participants to construct an inventory of internal models and alternate between them based on environmental demands. Further investigation of the behavior during transfer revealed that it is the subjective internal model of individuals that can predict the transfer across tasks. Our results demonstrate that even imperfect learning in a challenging environment helps learning in a new context by reusing the subjective and partial knowledge about environmental regularities.

This is an in-person presentation on July 20, 2024 (17:00 ~ 20:00 CEST).

Recent advances in self-supervised deep learning computer vision algorithms have resulted in significant improvement across several benchmark tasks including image classification and object detection. Similar to humans, these deep learning models are able to perform well on new tasks absent of task-specific guidance in the form of new labeled training data. The seemingly human-like ability that deep learning-based computer vision systems possess to learn robust representations that transfer between a variety of tasks raises two key questions: 1. How similar are learned representations from modern deep learning models and computational cognitive models of human visual perception? 2. Can we introduce a new objective during the training of a deep learning model to guide the model towards learning more human-like representations? We address the first question by characterizing the relation between modern deep learning vision transformer models and models of human visual perception. To address the second question we introduce a novel training regime for deep learning models that encourages representational alignment to cognitive models of human perception. We compare different deep learning and cognitive models and show that our Human Aligned Vision Transformer (HuViT) training objective results in learned representations that are more similar to those produced by models of human perception over an equivalent unmodified deep learning computer vision model while maintaining a similar level of performance on computer vision benchmark tasks.

This is an in-person presentation on July 20, 2024 (17:00 ~ 20:00 CEST).

Sensorimotor learning is defined as one’s ability to interact with the environment by interpreting the sensory world and responding to it with the motor system. Sensorimotor learning in speech has been shown to be influenced by cognition, but most models of speech mechanisms, including the DIVA model (Tourville & Guenther, 2011), do not include cognitive factors like attention or memory. Nonetheless, speech motor control is observably disrupted in individuals with impaired memory function (e.g., people with Alzheimer’s disease, Liu et al., 2012), and even in typical speakers, there appears to be a relationship between speech motor learning and memory capacity (Lametti et al., 2012), that can vary quite widely. We explore the plausibility of incorporating existing cognitive modeling paradigms, like ACT-R, with models of speech production, to better define the role of memory in speech tasks. Specifically, we consider the hypothesis that incorrect responses in a pitch perception task result from a failure to retrieve the memory of the target pitch. The ranges of time scales and decay rates that result in retrieval failure can guide the implementation of a more complete model that integrates elements of both cognitive models and models of speech production.

This is an in-person presentation on July 20, 2024 (17:00 ~ 20:00 CEST).

Disinhibition is a prominent feature of syndromes associated with frontotemporal lobar degeneration (FTLD), encompassing impulsive behaviours and difficulty suppressing inappropriate or habitual responses. Being disinhibited in these syndromes has been linked to higher caregiver burden, earlier institutionalisation, and poorer prognosis (Murley et al., 2021). There are currently no treatments for disinhibition in FTLD. However, an avenue for potential treatment is that of neurotransmitter deficits. Gamma-aminobutyric acid (GABA) and noradrenaline deficits in FTLD are well established and are correlated with disinhibition in isolation (Murley et al., 2020; Ye et al., 2023). To develop and validate treatment strategies for disinhibition, we need to understand the delicate balance of neurotransmitter deficits in these syndromes and their link to disinhibition. Here we use a manual stop-signal task to quantify inhibitory control in Progessive Supranuclear Palsy (PSP, Richardson’s syndrome, n= 5), behavioural variant frontotemporal dementia (bvFTD; n = 9) and age- and sex-matched healthy adults (n=14). The stop-signal task is a well-established tool to quantify inhibitory control, with trans-species and trans-diagnostic utility. We confirm that patients with PSP and bvFTD are impaired on the stop signal task (SSRT; M = 301.38, SD = 98.87) compared to controls (M = 187.38, SD = 32.78, p = 0.0003). Ongoing work is analysing the contribution of GABA-ergic and noradrenergic deficits to these deficits in inhibitory control. Understanding the variance of inhibitory control has implications for timing of symptom onset, prognostication, and the development of pharmacological interventions to mitigate the behavioural challenges faced by affected individuals and their caregivers.

This is an in-person presentation on July 20, 2024 (17:00 ~ 20:00 CEST).

Model quality is generally determined by some metric of fit that represents how well the model captures the target behavior. However, as models become more complex, it becomes increasingly difficult to determine how well a given model is doing, and to compare it to alternates. In recent work comparing different accounts of high level cognitive structure in the brain using Dynamic Causal Modeling (DCM), the Common Model of Cognition (CMC) was determined to be the most plausible model configuration using Bayesian Model Selection (BMS). This paper explores some lower level comparison metrics in an effort to gain a better understanding of what contributes to a model's overall "fit" to complex data, with the goal of creating additional methods for evaluating model quality.

This is an in-person presentation on July 20, 2024 (17:00 ~ 20:00 CEST).

Kanerva (2014) suggested that it would be possible to construct a complete Lisp out of a vector-symbolic architecture. We present the general form of a vector-symbolic representation of the five Lisp elementary functions, lambda expressions, and other auxiliary functions, found in the Lisp 1.5 specification (McCarthy, 1960), which is near minimal and sufficient for Turing-completeness. Our specific implementation uses holographic reduced representations (Plate, 1995), with a lookup table cleanup memory. Lisp, as all Turing-complete languages, is a Cartesian closed category (nLab authors, 2024), unusual in its proximity to the mathematical abstraction. We discuss the mathematics, the purpose, and the significance of demonstrating vector-symbolic architectures’ Cartesian-closedness, as well as the importance of explicitly including cleanup memories in the specification of the architecture.

This is an in-person presentation on July 20, 2024 (17:00 ~ 20:00 CEST).

This paper reports an experiment investigating learning and retention in a complex task over multiple sessions across an extended period of time. The primary aim of the experiment is to evaluate the Predictive Performance Equation (PPE: Jastrzembski & Gluck, 2009) a model of learning and forgetting that predicts retention based on past performance. The second aim is to test a taxonomy for knowledge, skills and attitudes and a competence retention analysis technique developed to improve competence retention in military training (Cahillane, Launchbury, MacLean, & Webb, 2013). Participants were trained over 16 weeks on the Multi-Attribute Task Battery (MATB: Comstock Jr & Arnegard, 1992), a computer-based task analogous to piloting an aircraft. The study reveals significant variation in learning profiles for the MATB sub-tasks and demonstrates the PPE’s ability to make accurate predictions of human performance over intervals ranging from 27 to 111 days.

This is an in-person presentation on July 20, 2024 (17:00 ~ 20:00 CEST).

Humans have cognitive vulnerabilities that can be leveraged to influence individuals. One such vulnerability is the continued influence effect (CIE), where misleading information can have a lasting effect even after corrections or factual discrediting information is presented. The CIE has been addressed experimentally and memory-based explanations exist. However, no current cognitive models specify cognitive mechanisms for prediction, simulation, and detailed testing of hypotheses. Here, we discuss relevant literature and propose a novel cognitive model to investigate memory mechanisms underlying the CIE.We demonstrate the utility of the model using simulations which show how the CIE emerges from memory processes and discuss plans for future research.

This is an in-person presentation on July 20, 2024 (17:00 ~ 20:00 CEST).

Sensorimotor grounding of cognitive processes may be the key to why humans exhibit efficient goal-directed behavior in a variety of dynamic environments. Modeling such behavior computationally poses a challenge as the model has to exhibit equally dynamic motor control in order to ground cognitive processes in it. Once the computational model has been developed, the next challenge lies ahead: how to evaluate the model behavior using human data? Here we present an eye tracking experiment to investigate action control in dynamic environments in which fixational eye movements reflect cognitive processes of action selection. Slightly increased uncertainty in motor control leads to more cautious action selection shown by fixations being initiated closer to a reference point, whereas strongly increased uncertainty leads to the need to monitor the environment for potential threats and thus greater distances to the reference point. We equip a computational model with the hypothesized action selection processes and single out the central parameter within its structure. In the last section, a likelihood method is discussed that could be used to evaluate the model based on human eye- movement behavior and to infer the parameter value.

This is an in-person presentation on July 20, 2024 (17:00 ~ 20:00 CEST).

Affect dynamics, or the study of changing patterns of emotional responses across time, has emerged as a key field of research in Mathematical Psychology. Traditionally, Affect dynamics research has relied on the Experience Sampling Method (ESM), a data gathering technique in which participants describe their feelings, thoughts, and behaviors at various times throughout the day. This technique studies Intensive Longitudinal Data (ILD) using Mixed Linear or Nonlinear Models (MLM) or Vector Autoregressive Models (VARs) (VAR). These theories characterize emotion in terms of time and complexity. However, they fail to recognize the underlying unity of emotional dynamism: the transition between affects. Although emotions occur in a sequential sequence, the transition between them takes into account the previous state in comparison to the current one. Individuals can experience and describe many emotions at the same time, but one feeling often gains precedence, influencing or being compared to the previous one. In this paper, we will show how to use and implement discrete Markov chains to evaluate each transition between past and current emotional states, while neglecting earlier transitions in the same way that a Markov chain does. Researchers may use Markov chains to quantify the odds of migrating between distinct emotional states across time, allowing for a better understanding of affect dynamics. This method not only overcomes the constraints of traditional data gathering and processing approaches, but it also allows for a more sophisticated investigation of the processes driving emotional variations.

This is an in-person presentation on July 20, 2024 (17:00 ~ 20:00 CEST).

It is widely accepted that there are five senses. There, however, appear to be several more. This paper attempts to provide and describe a comprehensive list of sensors that might be found in a complete cognitive architecture. We also briefly note how widely used these senses have been and which ones could yet be implemented in an architecture.

This is an in-person presentation on July 20, 2024 (17:00 ~ 20:00 CEST).

We posit that early attachment to kith and kin has a marked influence on later life reasoning and especially on generosity toward others. We present a series of computational experiments showing that the final (adult) levels of generosity differ depending on such early life exposure, and this independently of the otherwise homogeneous reasoning behavior. The benchmark experiment defines a developmental progression of three stages, from attachment only to perception of cooperative or non-cooperative actions of others in a controlled social environment to finally complex environments of arbitrary participants. We use as a behavioral basis the well-known Iterated Prisoner’s Dilemma game and its classic strategy Tit-For-Tat in a simulated society of individuals. We show that in the final stage of an arbitrary complex society the social scores obtained by the developing individuals are consistently higher than the reference, undeveloped (adult) individuals, and that this is due to the developed degree of nonzero generosity. We also show that individuals with a disturbed understanding of others’ emotional behavior (thus of attachment) but with intact reasoning tend to be more reciprocal and less generous in the end. On the other hand, individuals with the opposite disturbance of the understanding of reasoning but with intact understanding of others’ emotional behavior tend to be far more impulsive and behave as driven by attachment only. A succession of generations of typical, developing agents stabilizes the levels of generosity in the society. The effect of various parameters on the developed behaviors is also studied. Further implications of this developmental model are finally given.

This is an in-person presentation on July 20, 2024 (17:00 ~ 20:00 CEST).

This paper explores the integration of Virtual Reality (VR) into behavioral experiments, addressing the technical challenges that researchers face due to the necessity of advanced programming and game engine knowledge. By developing a VR environment tailored for conducting computer-based experiments and pairing it with VR Analysis Tool (VRAT) for data analysis and visualization, we facilitate a more accessible entry into VR-based research. The advantage that our tool provides is that researchers can conduct their traditional computer-based experiments in an environment with superior eye tracking and high experiment validity due to a high level of control over environmental factors. We then showcase the advantages of VR eye-tracking systems over traditional screen-based counterparts in terms of accuracy and precision, highlighting their consistency across various screen sizes and user demographics which is only one of the many superiorities of VR over the conventional methods. Due to the applicability of the 3D design, we believe that the future of behavioral research will increasingly pivot toward VR, with tools like VisiTor pioneering this transition by enabling models to effectively interact within VR spaces. Our tool is a step toward enabling behavioral studies to immigrate from traditional methods to VR.

This is an in-person presentation on July 20, 2024 (17:00 ~ 20:00 CEST).

The psychological literature has put forth several auto-associative memory models of attitude formation and change. The status of frequency effects in such models is not well understood. We compare frequency effects in auto-associative memory models of attitudes to the well-established frequency effects found in the ACT-R cognitive architecture. We found striking differences between the model classes, but only under some conditions. We discuss future directions that might stem from this provisional work.

This is an in-person presentation on July 20, 2024 (17:00 ~ 20:00 CEST).

The field of Artificial Intelligence (AI), particularly in the area of computer vision, has experienced significant advancements since the emergence of deep learning models trained on extensively large labeled datasets. However, reliance on human labelers raises concerns regarding bias, inconsistency, and ethical issues. This study aimed to address these concerns by exploring the feasibility of replacing human labelers with an interactive cognitive model. We investigated human behavior in a two-phase image labeling task and developed a model using the VisiTor (Vision + Motor) framework within the ACT-R cognitive architecture. This study was designed based on a real labeling task of identifying different crystals in optical microscopic images after various treatments for inhibiting the formation of the crystals. The outcomes from the image labeling experiment, which included both learning and testing phases, revealed meaningful observations. The observed decrease in task completion times for all participants during the learning phase suggests an increased familiarity with the image features, facilitated by the reference images presented in all four consecutive example tasks. During the testing phase, despite initial confusion caused by shaded zones in microscopic images, participants were able to correctly identify targets, highlighting the potential for cognitive models to learn and adapt. It was also discovered that the subtle distinctions between classes led to confusion in making decisions about labels. The developed interactive cognitive model was able to simulate human behavior in the same labeling task environment, while the model achieved high accuracy, it still relies on pre-defined features therefore limited its application to seen data only. Future work will expand the number of participants and task complexity and refine the ACT-R model to enhance its decision-making capabilities. Our findings suggest that interactive cognitive modeling offers a promising avenue for replacing human labelers with robust, consistent, and unbiased labeled datasets. This research can help mitigate ethical concerns and ultimately move us toward the goal of automating the labeling process in AI.

This is an in-person presentation on July 20, 2024 (17:00 ~ 20:00 CEST).