Publications
The work of our research team, and external collaborators, is often published in major scientific journals. This research is the foundation of our learning system.

Adaptive fact-learning systems typically rely on accuracy or correct-response latency, ignoring the timing of errors. In a spaced-repetition recognition task, we found that fast and slow errors differ in their implications for learning. In a following experiment, fast errors were self-corrected more often than slow errors when prompted to reconsider the response, and successful correction of fast errors led to improved subsequent accuracy. Prompting slow errors, by contrast, did not benefit later performance.

Learners often struggle to sustain retrieval practice during self-study. We examined whether gamified feedback could increase motivation without harming learning. Gamified feedback enhanced motivation, however, learning behavior during practice and delayed recall were unaffected

Memory impairment is common but understudied in kidney transplant recipients. We present the Seattle-Groningen Memory Assessment (SGMA), a remote 8-minute screener that estimates long-term retention from paired-associate learning. The SGMA provided reliable memory scores and revealed greater forgetting in transplant recipients than in (potential) donors. Scores were related to age, education, sleep quality, fatigue, quality of life, and physiological markers including blood cell counts and iron status.

In this study, we present a data-driven method for automatically generating plausible orthographically related distractors, based on a combination of common incorrect responses from learners doing open-answer retrieval practice, and rule-based generation of common misspellings. This flexible and scalable approach to generating distractors that align with genuine misconceptions, can make MCQs a more effective tool for vocabulary learning.

We propose a machine-learning approach to detect anomalous behaviour in an online retrieval practice task using keystroke dynamics and performance features. The method achieves high sensitivity and specificity, and requires only a small amount of training data. Classification remains robust even when some feature types are missing.

We investigated whether combining pretesting and posttesting improves learning under realistic digital learning conditions.
Pretesting initially increased retrieval accuracy and reduced response times across both fixed and adaptive schedules, but this
advantage disappeared after three repetitions, limiting its benefit in spaced practice. However, removing items answered
correctly during the pretest improved learning efficiency, particularly for learners with moderate to high prior knowledge, without
harming learners with low prior knowledge.

In this study, we focus on learners with dyslexia. We compare their performance to the performance of typical learners in our adaptive fact-learning system using both typing-based and speech-based response conditions. We find that typical learners outperform learners with dyslexia when they are asked to respond by typing, but that this difference is much smaller when learners can respond by speech.

We propose an application for learning multiplication problems with our adaptive algorithm mode for primary school students. We demonstrate that our application helps learners automatise multiplications through repeated practice over three levels of difficulty.

This study explores how people can quickly learn new tasks with minimal instruction by applying the concept of rapid composition of cognitive skills, as modeled in ACT-R and PRIMs. The research uses unsupervised machine learning to analyze arithmetic learning data from Dutch vocational students, creating a skill graph to support a tutoring system.

In this research, participants studied French-Dutch vocabulary in three learning conditions: one with unrelated foils, another with open questions, and a third using multiple-choice questions with related foils. The related foils were either semantically or orthographically similar to the correct answer. The results showed no significant difference between the open questions and the related foils condition, indicating comparable effectiveness.

Memory consolidation strength (Speed of Forgetting; SoF) was assessed using the MemoryLab adaptive learning system for
15 healthy controls and 11 patients with migraine. The RoF was significantly increased in the migraine group compared to
controls. The outcomes of these studies may help to evaluate whether SoF can serve as a low-cost sensitive behavioral
marker during the trajectory of migraine and based on these findings as well for other common headache disorders or
neurological disease.

Failing to address the “cold-start” problem – where the difficulty of the material for a specific learner is undetermined – leads to suboptimal learning and potential disengagement from the system. Using predicted parameters from a large learning dataset as starting estimates for the adaptive learning system yielded a more accurate model of learners’ memory performance than using default values. We found that predictions based on the difficulty of the fact (“what”) generally outperformed predictions based on the ability of the learner (“who”), though both contributed to better model estimates.

In this paper, we present a quantitative assessment of memory function spanning an age range of five to 85 years that is based on a model-based memory assessment. This approach yields a uniform metric that is directly interpretable and can be compared across different tasks and materials that are appropriate for different age groups. The results show a robust U-shape function, with long-term memory function at age 5 being comparable to that of cognitively impaired elderly individuals. These results and the method utilized could provide a new foundation for future studies on memory development across life stages.

We demonstrate that an online, remote model-based memory assessment, can identify individuals with Mild Cognitive Impairment (MCI) with an accuracy rate exceeding 84% in a single 8-minute session. Furthermore, the test can be repeated multiple times with increasing accuracy over multiple assessments. Our findings show that traditional methods to assess memory decline could be replaced by adaptive, precise, and patient-friendly online tools based on computational modeling techniques.

This study tested whether dropping items answered correctly on the first attempt improves learning outcomes.In a real-world within-subjects study with Dutch middle school students, a drop algorithm led to higher short- and long term retention than a non-drop algorithm. Dropping initially known items did not harm retention, suggesting it helps learners focus on unmastered material and increases learning gains

This paper describes three levels at which insights from human cognition can be integrated in user models. By analysing examples in the context of education, the paper argues that user models that are more deeply rooted in models of cognition, as is the case in MemoryLab, offer more valid and more fine-grained adaptations to an individual user.

This study introduces a new set of paired-associate stimuli consisting of 112 country outline-name pairs, designed to better reflect real-world learning by varying levels of prior knowledge. The prior knowledge for each item ranged from very high to none. This diverse dataset allows researchers to choose materials based on participants' prior knowledge.

In this paper, we aim to compare the efficiency of typing- and speech based vocabulary learning. Furthermore, we explore the possibilities of improving such speech-based learning using our adaptive algorithm based on a cognitive model of memory retrieval. We show that typing- and speech-based learning result in similar learning outcomes.

Here, we explore the possibility of improving our adaptive learning model by using information present in speech signals during spoken retrieval attempts. We extracted high-level prosodic speech features, such as pitch dynamics, speaking speed, and intensity from over 7,000 utterances. We demonstrate that some prosodic speech features are associated with accuracy and response latency for retrieval attempts, and that speech feature-informed memory models make better predictions of future performance relative to models that only use accuracy and response latency.

The goal of this project was to predict and track memory decline in subjectively or mildly cognitively impaired (MCI) individuals by using our model-based, adaptive fact-learning system. Here we present data demonstrating that these tools can diagnose mild memory impairment with over 80% accuracy after a single 8-minute learning session.

This study maps ACT-R’s declarative memory onto the Linear Ballistic Accumulator (LBA) model, which describes memory retrieval as a process of accumulating evidence. This mapping allows us to estimate ACT-R parameters more efficiently than traditional methods, which require extensive computational work. We validate this approach with simulated data and then apply it to real data, showing that the LBA can accurately estimate ACT-R parameters.

Real-life learning sessions tend to be relatively short, and requiring errors to update the model is therefore suboptimal. Here we show that our adaptive learning system can be calibrated to a learner’s proficiency in relatively short learning sessions by taking accuracy and response latency into account. In this paper, we highlight the benefit of using response latencies along with accuracy.

Adaptive learning systems have been successfully applied to word learning using keyboard-based input, but they have seen little application in spoken word learning. We show that typing and speech-based learning result in similar behavioral patterns that can reliably estimate individual memory processes. This is particularly beneficial for individuals with impaired typing skills, such as elderly learners.

In this paper, we focus on predicting university students’ exam performance using our model-based adaptive fact-learning system. The model’s estimate of a learner’s rate of forgetting predicts overall grades and performance on individual exam questions. In practice, this information could be used as an early warning signal to students and their instructors that additional study is likely required for a passing grade.

This paper investigates how the switch to distance learning affected study activity and performance in our online retrieval practice web app used for language learning in Dutch secondary education. We report insights from a data set consisting of over 115 million retrieval practice trials completed by more than 133 thousand students over the course of two consecutive school years. We found that, in this particular learning task, learners’ performance did not slow down and accuracy and response time on open-answer questions increased, perhaps due to students being more focused at home.

In many cases, there is an initial mismatch between the adaptive learning system model and the learner’s actual performance on the presented items, causing a “cold start” during which the system is poorly adjusted to the situation. In this study, we implemented several strategies for mitigating this cold start problem in our adaptive fact-learning system and experimentally tested their effect on learning performance. We found that cold start mitigation improved learning outcomes.

Here, the individual speed of forgetting in long-term memory is correlated with a readily available, task-free neuroimaging measure: the resting-state EEG spectrum. Statistical analyses revealed that individual rates of forgetting were significantly correlated across verbal and visual materials. These findings suggest that model parameters that reliably characterize an individual's performance, such as speed of forgetting can be observed in that individual's neurophysiological activity at rest.

This study investigates whether domain-general individual differences, such as working memory capacity (WMC) and general cognitive ability (GCA), can inform the selection of initial parameters in adaptive fact-learning systems. These systems typically begin with default parameters, which are adjusted based on learners' responses during the learning process. The goal was to determine if WMC and GCA, measured prior to learning sessions, could improve the accuracy of initial model parameters, specifically those that affect repetition schedules. The study found no significant relationship between WMC, GCA, and learning outcomes, suggesting that executive functioning and attentional control do not significantly predict delayed recall.

Our adaptive algorithm model is tailored to individual learners and outperforms traditional systems by continuously updating the estimated speed of forgetting for each item based on learners' accuracy and reaction time. In this paper, we investigate whether the speed of forgetting remains stable over time and across different materials. We demonstrate that they are stable over time but not across materials.

This paper explores how pupil dilation reflects memory processing during the retrieval of paired associates. By measuring pupil response as participants retrieved items from memory, the study found that pupil dilation decreases with stronger memory, indicating that it correlates with the retrieval effort. These findings suggest that pupil dilation can serve as an online marker of memory strength, even before an overt response is given.