AI@JSI Seminar | Department of Knowledge Technologies

From Understanding People to Securing AI: A Human-Centered Research Journey Through Large Language Models

Anja Glusic — Tue, 10 Mar 2026 08:38:53 +0000

The talk presents a connected view of recent research on large language models (LLMs) through a human-centered lens, moving from what these systems can infer about people to how they interact with them, where they present security risks, and why such vulnerabilities matter in socially consequential settings. It begins with work on how LLMs can infer personality from short texts and on the role of communication style in shaping user experience and task outcomes, showing both the potential of LLMs for personalization and the importance of designing interaction carefully across contexts. It then broadens to questions of trustworthiness and toxicity, covering security risks in prompt-based interaction, attack surfaces, and the growing challenge posed by multilingual, multimodal, and autonomous jailbreaks. Further, it examines representational harms through methods for measuring gender bias in gendered and under-resourced languages. The talk is concluded with a high-stakes application of mental health crisis response, where clinically informed evaluation reveals that increasing model capability does not automatically translate into safe, appropriate, or context-aware behavior. Across these topics, the unifying theme is that progress in LLMs should be matched by rigorous work on evaluation, safety, fairness, and responsible deployment.

Zoom link: Click me. (Audio issues from last time are being actively addressed.)

The seminar is also organized under the ELLIOT project.

AI-enabled Mammography

Anja Glusic — Thu, 05 Mar 2026 12:18:05 +0000

Breast cancer is the most common malignancy and leading cause of cancer deaths in women around the world. The preferred diagnostic and prevention method used to fight breast cancer is mammography, being effective, cheap, reliable and suitable for screening large populations. The analysis of mammograms requires considerable effort from qualified radiologists, who have been increasingly relying on computer-aided diagnosis and the development of tools that require little supervision could save the lives of hundreds of thousands of women across the world.

In recent years, the integration of Artificial Intelligence (AI) in the field of medicine has brought about a significant transformation in healthcare delivery. Computer vision technology finds itself at the forefront of this proces. Its applications to mamography data represent an active research field, stimulated by the fact that among the clinical imaging modalities, mammography stands out in terms of high spatial resolution (full-field digital mammography systems usually produce images at resolutions ranging from 1920×2304 to 4708×5844 pixels.

I will discuss research work stemming from projects conducted at the Institute for Artificial Intelligence Research and Development of Serbia, aimed at implementing AI-assisted mamography screening at the national level. Viewed through the lens of medical applications, the talk will cover the evolution of state-of-the-art research in the domain of computer vision and multimodal AI over the last 5 years, from the applications of “classical” convolutional neural networks and visual transformer models, explainable AI extensions, generative (diffusion) models, to self supervised and multiple instance learning approaches to address specific challenges and (downstream) tasks such as: image classification, lesion detection and synthetic mammogram generation.

Zoom link: Click me. (Audio issues from last time are being actively addressed.)

Diffusion Language Models: Problem Solving and Reasoning

Anja Glusic — Tue, 03 Feb 2026 13:08:09 +0000

Masked diffusion models (MDMs) offer a compelling alternative to traditional autoregressive language models. They generate strings by iteratively refining partially masked inputs in parallel. This makes them efficient, but their computational capabilities and the limitations inherent to the parallel generation process remain largely unexplored.
In this talk, I will talk about what types of reasoning problems MDMs can provably solve and how efficiently they can do it. We will describe the relationship between MDMs and the well-understood reasoning frameworks of chain of thought (CoT) and padded looped transformers (LTs): We will see that MDMs and polynomially padded LTs are, in fact, equivalent, and that MDMs can solve all problems that CoT-augmented transformers can. Moreover, we will showcase classes of problems (including regular languages) for which MDMs are inherently more efficient than CoT transformers, where parallel generation allows for substantially faster reasoning.

Zoom link: Click me.

Learning Non-Stationary Dynamics: Oscillations, Chronotaxicity, and Time-Varying Structure in Complex Systems

Nina Omejc — Mon, 26 Jan 2026 10:09:48 +0000

Many real-world systems—from physiology and neuroscience to climate and finance—are inherently non-stationary, operating far from equilibrium with structure that evolves in time. This talk presents a physical and data-driven framework for understanding such systems through time-varying oscillatory dynamics. Central to this approach is the concept of chronotaxic systems: non-autonomous oscillators with dynamically stable, time-dependent characteristic frequencies, providing a principled way to model robustness and adaptability without assuming stationarity or time-asymptotic behaviour.

I will discuss how instantaneous frequency, nonlinear mode decomposition, and multiscale ridge extraction enable the learning of interpretable latent dynamics from complex, noisy time series where classical spectral methods fail. These methods form the basis of the MODA toolbox and support reliable separation, tracking, and interpretation of interacting components with evolving amplitudes and frequencies.

The framework further enables inference of time-varying interactions and causal structure using wavelet-based coherence, information-theoretic measures, and Bayesian dynamical inference, including in networks with non-stationary coupling. Applications will be drawn primarily from physiology and neuroscience—such as neurovascular coupling, anaesthesia-induced phase transitions, and ageing—but the methods generalise across domains, including physical, engineered, and socio-economic systems.

The talk highlights how physically grounded models of non-stationary dynamics complement machine learning by providing interpretable representations, principled constraints, and reliable inference in settings where purely data-driven approaches struggle.

Zoom link: Click me.

Towards leveraging multi-modal data to enhance single-modality classification models: An application to image analysis

Anja Glusic — Tue, 09 Dec 2025 13:30:22 +0000

Cross-modal knowledge distillation (CMKD) refers to the scenario in which a learning framework must handle training and test data that exhibit a modality mismatch, more precisely, training and test data do not cover the same set of data modalities.

Traditional approaches for CMKD are based on a teacher/student paradigm where a teacher is trained on multi-modal data with the aim to successively distill knowledge from a multi-modal teacher to a single-modal student. Despite the widespread adoption of such paradigm, recent research has highlighted its inherent limitations in the context of cross-modal knowledge transfer.

Taking a step beyond the teacher/student paradigm, in this talk I will introduce a new collaborative framework for cross-modal knowledge distillation, named DisCoM-KD (Disentanglement-learning based Cross-Modal Knowledge Distillation), that explicitly models different types of per-modality information with the aim to transfer knowledge from multi-modal data to a single-modal classifier. To this end, DisCoM-KD effectively combines disentanglement representation learning with adversarial domain adaptation to simultaneously extract, for each modality, domain-invariant, domain-informative and domain-irrelevant features according to a specific downstream task.

Unlike the traditional teacher/student paradigm, our framework simultaneously learns all single-modal classifiers, eliminating the need to learn each student model separately as well as the teacher classifier. We evaluated DisCoM-KD on standard multi-modal benchmarks and compared its behaviour with recent state of the art knowledge distillation frameworks. The findings clearly demonstrate the effectiveness of DisCoM-KD over competitors considering mismatch scenarios involving both overlapping and non-overlapping modalities. These results offer insights to reconsider the traditional paradigm for distilling information from multi-modal data to single-modal neural networks.

Zoom link: https://zoom.us/j/98326976345?pwd=TlV9MGPKNNIe4qnBmV83KtCJqA9sCC.1

Reliable Machine Learning – Methods and Applications in Environmental Sciences

Anja Glusic — Thu, 22 May 2025 14:20:56 +0000

I will discuss reliability of Machine Learning (ML) algorithms, from the perspective of adversarial learning and its implications in terms of model quality and robustness given new data. The talk will highlight how adversarial learning can be used as a valuable tool to measure and improve reliability, particularly in the domain of environmental sciences.
Highlighting the central role of reliable ML in environmental sciences, I will showcase our work in various applications in environmental sciences, particularly atmospheric and environmental chemistry.

For an overview of my ongoing research projects visit my lab’s webpage at: https://wickerlab.org.

Address of the Zoom meeting: https://zoom.us/j/93704860483?pwd=6XAVHx5guys9a94XTNlUaXkHTozdbl.1

The Blueprints of Self-Organisation

Anja Glusic — Fri, 09 May 2025 16:39:50 +0000

Self-organisation sculpts the topology and physicochemical properties of biological, organic, inorganic, and hybrid materials. Its explainability emerges from understanding the core principles and preconditions that dictate synthetic outcomes. A formal, machine-actionable representation of these blueprints is crucial for developing explainable AI systems that explore both immediate and deep chemical spaces. In this talk, we first show how molecular modelling uncovers elusive self-organisation patterns in cluster materials, some of which can be used as scaffolds for designing extended material systems. Next, we examine the development of dynamic knowledge-graph systems that accelerate the discovery of self-organised reticular material architectures.

Generating Point Sets of Small Star Discrepancy

Anja Glusic — Fri, 07 Mar 2025 07:59:00 +0000

Uniformly distributed point sets of low discrepancy are heavily used in experimental design and across a very wide range of applications such as numerical integration, computer graphics, and finance. Recent methods based on Graph Neural Networks [Rusch et al., PNAS 2024] and solver-based optimization [Clément et al., accepted in Proceedings of the AMS] identified point sets having much lower discrepancy than previously known constructions. We show that further substantial improvements are possible by separating the construction of low-discrepancy point sets into (i) the relative position of the points, and (ii) the optimal placement respecting these relationships. Using tailored permutations, we construct point sets that are of 20\% smaller discrepancy on average than those proposed by Rusch et al. In terms of inverse discrepancy, our sets reduce the number of points in dimension 2 needed to obtain a discrepancy of 0.005 from more than 500 points to less than 350. For applications where the sets are used to query time-consuming models, this is a significant reduction.
The presentation is based on joint work with François Clément (University of Washington, US), Kathrin Klamroth (University of Wuppertal, Germany), and Luís Paquete (University of Coimbra, Portugal).

Label-efficient panoptic segmentation

Anja Glusic — Tue, 14 Jan 2025 15:10:23 +0000

Panoptic segmentation provides a comprehensive understanding of visual scenes by assigning each pixel a semantic class label and, for objects, an instance ID. While highly effective, traditional methods rely heavily on large-scale annotated datasets, posing significant challenges for scalability and adaptability. Additionally, the process of annotating images with panoptic labels is both labor-intensive and time-consuming, highlighting the importance of label-efficient approaches. In this talk, I will present an overview of various label-efficient methods, focusing on my recent work in unsupervised domain adaptation and semi-supervised learning. I will also discuss the emerging field of open-vocabulary panoptic segmentation, which extends recognition capabilities to categories beyond the training taxonomy.

Zoom link:
https://zoom.us/j/95847029023?pwd=hYqMPhZ4K2yVttbCl5gSCDvYNIGBZN.1

Multi-Objective AutoML: Towards Accurate and Robust Neural Networks

Anja Glusic — Mon, 02 Dec 2024 13:31:45 +0000

Automated machine learning has been successful in supporting data scientists in selecting appropriate machine learning architectures, as well as optimizing hyperparameters. By doing so, data scientists can focus their attention on more important tasks. Partially thanks to the TAILOR project, in which Leiden University and JSI have successfully collaborated, we have seen a demand on AutoML techniques to not only provide solutions that are accurate, but also those that are trustworthy according to several relevant criteria. In particular neural networks are known to be vulnerable to adversarial attacks, whereas robustness (against such attacks) is an important criterion of trustworthiness. In this talk, I will summarize various projects we have done through this collaboration, that envision AutoML solutions that specifically address robustness of neural networks.

Zoom link:
https://zoom.us/j/91879263877?pwd=9sxadIBnOrSxnyZdipwLiNNurcta3O.1