Seminar series

Enrique Valero Leal is a final-year ELLIS PhD student at both the Technical University of Madrid and Ludwig-Maximilians University of Munich. His research focuses on interpretable machine learning,  probabilistic models, and causality, with technical experience in Industry 4.0 applications. 

Existing methods for group counterfactual explanations often fail to generalize, rely on strict assumptions, or distort group geometry. I explore how to overcome this by learning an optimal transport map that efficiently generates geometry-preserving counterfactuals for any instance without re-optimization, outperforming current baselines across both linear and non-linear models. I also provide insight on future lines of research and applications of (group) counterfactual explanations with respect to the broader state-of-the-art, such as LLMs or industrial applications.

Abdullah Oğuz Kızılçay is an Assistant Professor at Zonguldak Bülent Ecevit University. His research focuses on electromagnetics and microwave antenna design, with emphasis on optimization techniques and machine learning/deep learning for modeling and analysis of electromagnetic systems. 

Machine learning is becoming an important tool for microwave data analysis and interpretation. This talk explores its role in microwave sensing by reviewing common sensing approaches and discussing how machine learning can support data analysis, feature engineering, and decision-making processes. Representative applications and example case studies are used to illustrate different data-driven solution strategies. The seminar concludes with a discussion of current challenges and future directions in machine learning-based microwave sensing.

Antonio Guillén Teruel is a Computer Science PhD candidate with a degree in Mathematics, specializing in Machine Learning. He develops methods to mitigate class imbalance and Concept Drift, applying them to medical data (Parkinson’s, COVID-19) in collaboration with UCL and Jagiellonian University. 

This talk focuses on a novel three-stage pipeline to generate synthetic transcriptomic (RNA-seq) data conditioned on clinical covariates like sex, age, and tissue type to address sample scarcity and dataset imbalance. We introduce an approach that combines a Transformer Autoencoder for dimensionality reduction with Normalizing Flows to map complex gene expression distributions into a standardized Gaussian space. An Ordinary Least Squares (OLS) model is then used to condition the generation of new profiles. Through validations on the GTEx dataset, we demonstrate that these methods achieve high biological fidelity, preserve gene co-expression networks, and maintain distinct signals of the targeted covariates.

Vinícius R. Cota is an Assistant Professor in the Dept. of Electronic Engineering at Maynooth University (Ireland) working in neuroengineering. His research focuses on closed-loop neuroprostheses for neurological disorders, including novel temporal complex electrical stimulation for epilepsy. 

Electrical stimulation is an established therapy for neurological disorders, but most current protocols remain open-loop and temporally simplistic, with limited sensitivity to the dynamical structure of brain activity. In this talk, I will present our work on neuromodulation strategies that explicitly incorporate spatiotemporal brain dynamics, including non-periodic electrical stimulation for epilepsy and pathological anxiety, as well as sleep-coordinated stimulation for post-stroke rehabilitation. I will then discuss how AI/ML can help bridge these approaches to the next generation of precision neuroprostheses by enabling robust brain-state decoding, prediction of therapeutic responses, and adaptive closed-loop control. Together, these efforts point toward intelligent neuroengineering systems in which stimulation is not only delivered, but continuously shaped by ongoing neural dynamics to achieve more precise, state-aware, and personalized interventions.

PhD student in Computer Science specializing in collaborative robotics, human–robot interaction, and intelligent manufacturing. Research focuses on repurposing industrial robots into cobots using AI, digital twins, and multimodal perception, with experience in academia and industry. 

This talk presents ongoing research on the use of Retrieval-Augmented Generation (RAG) and open-source Large Language Models (LLMs) to support the conversion of conventional industrial robots into collaborative robots. The proposed framework combines structured robotic datasets, semantic retrieval, and expert knowledge to identify missing collaborative and safety requirements for legacy robotic systems. The presentation will also discuss the development of an industrial safety audit framework based on ISO 10218-1:2025, ISO 10218-2:2025, and ISO/TS 15066, which is being applied to a real industrial robotic cell as the case study for validating the proposed methodology. Finally, the talk will outline current progress, ongoing challenges, and future research directions toward AI-assisted collaborative retrofit assessment.

PhD researcher in systems engineering and informatics at University of Hradec Králové. Conducting systematic review on Agentic AI governance, safety, and deployment. Integrates agent-based modeling, statistical analysis, and sustainability frameworks for evidence-based decision support in AI systems.

Agentic Artificial Intelligence systems, autonomous agents capable of multi-step reasoning, planning, and goal-directed action are rapidly transitioning from research laboratories into production environments across healthcare, finance, customer service, and cybersecurity. Yet this transition occurs without standardized governance frameworks, validated benchmarking approaches, or evidence-based deployment guidelines. This seminar presents a PRISMA 2020 systematic literature review addressing this critical gap, synthesizing 90+ peer-reviewed empirical studies (2015–2026) to establish: (1) definitional clarity and architectural taxonomies distinguishing Agentic AI systems; (2) real-world deployment contexts and sector-specific implementation patterns with success/failure factors; (3) a five-dimensional benchmarking framework addressing the 83% technical-to-sustainability measurement imbalance.

Jia (Justin) Liu is a PhD candidate at the University of Waikato, New Zealand, supervised by Prof. Albert Bifet. His research focuses on streaming machine learning and anomaly detection, with particular emphasis on isolation forest methods for evolving data streams. His work addresses core challenges in stream learning, including concept drift adaptation, semi-supervised learning under label scarcity, and evaluation methodologies for imbalanced streaming data. His recent contributions include the Adaptive Isolation Forest (Discovery Science 2025) and the Streaming Isolation Forest (PAKDD 2025), alongside ongoing work on semi-supervised extensions of isolation-based detectors. He has submitted his doctoral thesis and expects to complete his degree in 2026. 

Anomalies are rare, high-impact events, and detecting them in continuous data streams remains a persistent robustness challenge for deployed AI systems. Streaming settings break the assumptions most detectors rely on: each instance is seen once under bounded memory, the underlying distribution drifts over time, and labels are unavailable or expensive. My research develops four isolation-based detectors designed for these conditions. In the unsupervised domain, one preserves a global view of the cumulative distribution via reservoir sampling and incremental subtree regrowing, while the other tracks the recent local distribution using a sliding window and selectively replaces the weakest ensemble members. Building on these, two semi-supervised extensions use a small labelling budget, with label information guiding scoring, ensemble maintenance, and active learning.

PhD candidate in Computational Modeling at FURG, Brazil, researching cognitive architectures for cooperative multi-robot coordination. She is Technical Project Manager of the Hydrone project, developing autonomous air-water robotic systems for offshore inspection.

Robotic operations that span air, water surface, and underwater environments pose challenges that rarely appear in single-domain systems. In this talk, I will present the Hydrone Project, a decade-long effort at the Federal University of Rio Grande (FURG) to build a hybrid unmanned aerial–underwater vehicle (HUAUV) that cooperates with an autonomous surface vehicle (the WhiteBoat USV) to inspect and monitor offshore environments. I will walk through the platform from the ground up: the 6-DOF dynamic modeling of both vehicles and its experimental validation, the finite-state autonomy architecture that runs the missions under real hardware and communication constraints, and the engineering decisions behind safe air-to-surface-to-dive transitions. I will conclude with the open research question that emerged from this work — how two robots can coordinate reliably when the communication link between air and water fails — which motivates my current doctoral research.

Hamza Zaidi is a PhD student at Otto-von-Guericke University Magdeburg, working at the Knowledge Management & Discovery Lab (KMD) and TACTIC Graduate School. His research focuses on spatio-temporal deep learning, wearable sensor systems, health signal processing and interpretable graph neural networks aimed at bridging the gap between AI research and clinical practice. 

This talk focuses on a graph based approach for multi sensor time series, where breathing motion, hand movement, and similar signals are distributed across many coupled sensors rather than captured by one. We introduce the Physio Digital Temporal Graph, representing sensors as nodes and their anatomical or functional coupling as edges, combined with temporal graph neural networks that quantify how a local change at one node influences its neighbors. This influence factor is paired with attribution based explanation to localize evidence rather than only produce a prediction. Through case studies on smart vest breathing data and instrumented glove hand kinematics, we demonstrate that this approach localizes abnormalities and movement strategies more clearly than channel wise or purely global models.

Bruno Veloso is an associate professor at Faculty of Economics at the University of Porto, Portugal, and a Senior Researcher at LIAAD – INESCTEC (Laboratory of Artificial Intelligence and Decision Support). He received his BSc degree in Electronics and Computer Engineering in 2010, and in 2012 he finished his MSc in Electronics and Computer Engineering in Telecommunications at Instituto Superior de Engenharia do Porto, Portugal. He has an International PhD in Telematics Engineering from the University of Vigo. He develops scientific research in Artificial Intelligence, which involves Multiagent Systems and Machine Learning. Bruno has authored over 80 publications in peer-reviewed international journals, book chapters and conference proceedings. During the last years, he has served as an expert/reviewer of several conferences and journals, and he was associate editor of two journals.

This talk focus on a novel AutoML approach for dynamic data streams, where optimal hyperparameters must evolve to handle concept drift. We introduce self-parameter-tuning algorithms based on Nelder-Mead optimization, micro-evolutionary algorithms, and swarm intelligence. These techniques use exploration and deployment phases to continuously adapt model parameters on the fly. Through case studies on diverse ML tasks, we demonstrate with synthetic and real-world data that these methods achieve lower prediction error, successfully adapt to drift, and consistently outperform static AutoML techniques.