Research themes that connect cognitive modeling and modern AI

Key research areas include robust model development, model optimization, learning from imbalanced data, and applied machine learning.

Question. How can machine learning and deep learning models be developed and evaluated to remain reliable when working with limited, noisy, heterogeneous, or highly imbalanced data?

Method. Various models are trained and optimized on the laboratory’s GPU infrastructure, or with national high-performance computing (HPC) resources used when larger-scale experiments are required. Evaluation goes beyond reporting a single performance score. We examine how consistent the results are across repeated experiments, different data splits, and model configurations, and pay particular attention to problems caused by limited or imbalanced datasets. We treat reproducibility as a foundational stone throughout the experimental process.

Research on image, video, signal, and multimodal data for detection, classification, representation learning, and decision support.

Question. How can computer vision and multimodal models move beyond simple recognition and capture more complex aspects of what is happening in a scene, such as the quality, context, or progression of an observed action?

Method. We use image, video, and multimodal models, including vision-language methods, temporal modeling, and multimodal data fusion. Our work includes problems such as action-quality assessment, where the aim is not only to recognize an action but also to assess how well it is performed. We also examine model robustness, computational efficiency, and whether suitable models can be deployed on edge devices.

We work with language models, retrieval-supported systems, domain adaptation, and methods for evaluating how these models behave in practical settings.

Question. How can language models be adapted, supported by retrieval, and evaluated so they remain reliable in practical settings, including when sensitive content must stay on local infrastructure?

Method. We develop retrieval-supported and locally deployed language-model systems, with particular attention to privacy, traceability, and reliable evaluation. Work includes domain adaptation and checking documents against clearly defined requirements, so that a model’s answers can be traced to source text and the workflow does not depend on sending content to an external service.

Computational perspectives on affect, physiological signals, and human behaviour in interaction with intelligent systems.

Question. How can physiological and behavioural signals be modelled computationally so that affect, cognitive load, or related states can inform the design and evaluation of intelligent systems?

Method. We use signal processing and machine learning on physiological and behavioural data, from wearable and laboratory recordings. Evaluation looks at whether the derived measures are stable enough to support system design, not only at a single classification score.

Computational models of learning, decision-making, perception, interaction, and problem-solving.

Question. Which computational accounts of learning, decision-making, and interaction actually help when we build and evaluate an intelligent system?

Method. We develop and test models of behaviour and interaction, drawing on cognitive science and experimental psychology. The aim is to use those models as working accounts of intelligent behaviour, not only as post-hoc explanations of a deployed system.

Integration of models, tools, and decision steps into supervised intelligent workflows and autonomous prototypes.

Question. How can separate models, tools, and a human decision step be composed into a workflow that can be supervised, evaluated, and trusted on a stated task?

Method. We build multi-component prototypes and evaluate them on the task they are meant to support, rather than as demonstrations only. Particular attention goes to the human role in the loop, to what must remain local, and to whether the assembled system is inspectable.