Posters presented at the 4th TAILOR conference

Poster number #	Presenter	Affiliation	Author(s)	Title	Abstract
1	Meelis Kull	University of Tartu	Meelis Kull	Uncertainty in Text-Generating Foundation Models	Understanding and managing uncertainty is crucial for achieving trustworthy reasoning. In supervised learning, uncertainty is commonly divided into aleatoric and epistemic uncertainty, representing inherent randomness and model knowledge deficiencies. We will discuss which forms of uncertainty are important to consider when reasoning with text-generating foundation models. Additionally, we will examine the challenges in estimating these uncertainties.
2	Andra Cristiana Minculescu	TNO	Andra Cristiana Minculescu, Wico Mulder, Harmen de Weerd	Towards Human-Centric AI Companions in Healthcare: Integrating Theory of Mind for Efficient and Trustworthy Collaborative Decision-Making	Although Artificial Intelligence (AI) has demonstrated promising results in medical applications, the limited focus on its integration into human-centered workflows means that current models are still largely relegated to the role of mere tools. To transition AI into a diagnostic companion within healthcare decision-making, current models must go beyond reliability (high prediction accuracy) and should gain trust through adaptability, explainability, and efficiency. Firstly, AI entities should incorporate real-time feedback to adjust to the dynamic collaborative decision-making scenarios found in medical settings. Secondly, for doctors to consider AI outputs, the systems must transparently explain the reasoning behind their decisions, a key element of explainability. Moreover, the exchange between AI and doctors must be efficient, optimizing how and what information is conveyed to avoid disrupting standard healthcare practices. Ideally, this exchange mirrors interactions between medical professionals who focus on sharing pivotal information that clarifies their diagnostic reasoning to one another. To achieve these goals one of the ingredients that AI systems need to embody human social capabilities is Theory of Mind (ToM). ToM enables AI agents to reason about the unobservable states and beliefs of others, possibly enhancing social interactions. The hypothesis that I address in my research is that an AI system equipped with ToM can thoughtfully consider what information is necessary for doctors to understand its conclusions and vice versa. This not only involves determining the essential data to convey but also understanding and addressing any gaps in its own knowledge by seeking further information from the medical team, ensuring a balanced, two-way interaction. This research project employs a game theory framework to model real-life decision-making processes found in a medical context, specifically in the field of dermatology and Mohs Surgery. The game facilitates a two-way communication channel, enabling both the AI and the doctor to efficiently exchange information and communicate the underlying reasons for classifying ambiguous cells in histological samples. This exchange mirrors real-life consultations between medical professionals, where the primary diagnostic indicator for ambiguous tissue often hinges on their neighboring cells. The effectiveness of this model is evaluated based on the ability of different ToM levels to select the most convincing information to support decisions, the efficiency of information exchange, and the accuracy of the collaborative decision-making process.
3	Peter Anthony	Comenius University in Bratislava/Slovak.AI	Peter Anthony, Francesco Giannini, Michelangelo Diligenti, Martin Homola, Marco Gori, Stefan Balogh, and Jan Mojzis	Tailored Logic Explained Network for Explainable Malware Detection	Malware detection is a constant challenge in cybersecurity due to the rapid development of new attack techniques. Traditional signature-based approaches struggle to keep pace with the sheer volume of malware samples. Machine learning offers a promising solution, but faces issues of generalization to unseen samples and a lack of explanation for the instances identified as malware. However, human-understandable explanations are especially important in security-critical fields, where understanding model decisions is crucial for trust and legal compliance. While deep learning models excel at malware detection, their black-box nature hinders explainability. Conversely, interpretable models often fall short in performance. To bridge this gap in this application domain, we propose the use of Logic Explained Networks (LENs), which are a recently proposed class of interpretable neural networks providing explanations in the form of First-Order Logic (FOL) rules. This paper extends the application of LENs to the complex domain of malware detection, specifically using the large-scale EMBER dataset. In the experimental results we show that LENs achieve robustness that exceeds traditional interpretable methods and that are rivaling black-box models. Moreover, we introduce a tailored version of LENs that is shown to generate logic explanations with higher fidelity with respect to the model’s predictions.
4	Igor Farkaš	slovak.AI (Comenius University Bratislava)	Štefan Pócoš, Iveta Bečková, Igor Farkaš	RecViT: Enhancing Vision Transformer with Top-Down Information Flow	We propose and analyse a novel neural network architecture — recurrent vision transformer (RecViT). Building upon the popular vision transformer (ViT), we add a biologically inspired top-down connection, letting the network ‘reconsider’ its initial prediction. Moreover, using a recurrent connection creates space for feeding multiple similar, yet slightly modified or augmented inputs into the network, in a single forward pass. As it has been shown that a top-down connection can increase accuracy in case of convolutional networks, we analyse our architecture, combined with multiple training strategies, in the adversarial examples (AEs) scenario. Our results show that some versions of RecViT indeed exhibit more robust behaviour than the baseline ViT, on the tested datasets yielding ≈18 % and ≈22 % absolute improvement in robustness while the accuracy drop was only ≈1 %. We also leverage the fact that transformer networks have certain level of inherent explainability. By visualising attention maps of various input images, we gain some insight into the inner workings of our network. Finally, using annotated segmentation masks, we numerically compare the quality of attention maps on original and adversarial images.
5	Andrea Passerini	University of Trento	Emanuele Marconato, Stefano Teso, Antonio Vergari, Andrea Passerini	Reasoning shortcuts in neuro-symbolic models	The reliability of a neuro-symbolic model crucially depends on the quality of the concepts that the model extracts from the data. In this work we show how existing neuro-symbolic frameworks are prone to learning reasoning shortcuts, i.e., concepts that are consistent with the given knowledge but have an unintended semantics, thus compromising the reliability of the learned predictor. We discuss the conditions behind the occurrence of reasoning shortcuts and derive several natural mitigation strategies, in both standard and continual learning settings.
6	Nicolo’ Brandizzi	Fraunhofer IAIS, Sapienza University of Rome	Nicolo’ Brandizzi	R2Net AI Research & Mental Well-Being	This poster detailes the outcomes of the one-day workshop presented as part of the TAILOR-ESSAI Summer School, in collaboration with CLAIRE Rising Research Network (R2Net). The event focused on fostering a closer-knit community of young AI researchers in Europe, including supporting AI researchers and promoting mental well-being for Ph.D. students. The event featured presentations, discussions, and interactive activities, as well as provided networking opportunities for participants, showcasing the value of R2Net’s collaborative efforts in the AI community.
7	Anders Jonsson	Universitat Pompeu Fabra	David Kuric, Guillermo Infante, Vicenç Gómez, Anders Jonsson, Herke van Hoof	Planning with a Learned Policy Basis to Optimally Solve Complex Tasks	Conventional reinforcement learning (RL) methods can successfully solve a wide range of sequential decision problems. However, learning policies that can generalize predictably across multiple tasks in a setting with non-Markovian reward specifications is a challenging problem. We propose to use successor features to learn a policy basis so that each (sub)policy in it solves a well-defined subproblem. In a task described by a finite state automaton (FSA) that involves the same set of subproblems, the combination of these (sub)policies can then be used to generate an optimal solution without additional learning. In contrast to other methods that combine (sub)policies via planning, our method asymptotically attains global optimality, even in stochastic environments.
8	Krzysztof Krawiec	Poznan University of Technology	Krzysztof Krawiec, Piotr Wyrwiński	Neuro-Guided Graph Expansion for Symbolic Regression	This study introduces a novel neurosymbolic approach to symbolic regression that works by gradually constructing formulas according to a predefined expression grammar, ensuring so syntactic correctness. The resulting formulas can be traced back to their constituents, enhancing transparency in both the final symbolic artifact and the synthesis process itself. We propose an iterative search algorithm that builds multiple models in parallel, represented as a directed graph. This graph is expanded iteratively, guided by a graph neural network until a satisfactory solution is found or computational resources are exhausted. The results confirm the method’s feasibility, as it consistently outperforms several baseline algorithms.
9	Jerzy Stefanowski	Poznan University of Technology	Jerzy Stefanowski, Mateusz Lango and Jakub Raczynski	Natural language explanations of recommendations using coherent text opinions	Providing explanations for predictions of complex machine learning algorithms, including recommender systems is still challenging. One form of such explanation that is particularly useful from the perspective of a non-expert user is an explanation expressed in natural language. Several methods for providing such explanations have recently been proposed for the recommendation task, however they do not take into account properly the coherence between generated text and predicted rating, which is a necessary condition for an explanation to be useful. In our recent studies we have attempted to improve this issue of explanation and prediction coherence by 1) presenting results from a manual verification of explanations generated by one of the state-of-the-art approaches 2) proposing a method of automatic coherence evaluation 3) introducing a new transformer-based method that aims to produce more coherent explanations than the state-of-the-art approaches 4) performing an experimental evaluation which demonstrates that this method significantly improves the explanation coherence without affecting the other aspects of recommendation performance.
10	Jan van Rijn	LIACS, Leiden University	Matthias Konig, Annelot Bosman, Jan van Rijn, Holger Hoos	Multi-Objective AutoML: Towards Accurate and Robust Neural Networks	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, 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 the TAILOR project, that envision AutoML solutions that specifically address robustness of neural networks.
11	Anna Vettoruzzo	Halmstad University	Anna Vettoruzzo, Joaquin Vanschoren, Mohamed-Rafik Bouguelia, Thorsteinn Rögnvaldsson	Learning to learn without forgetting using attention CF	Continual learning (CL) refers to the ability to continually learn over time by accommodating new knowledge while retaining previously learned experience. While this concept is inherent in human learning, current machine learning methods are highly prone to overwrite previously learned patterns and thus forget past experience. Instead, model parameters should be updated selectively and carefully, avoiding unnecessary forgetting while optimally leveraging previously learned patterns to accelerate future learning. Since hand-crafting effective update mechanisms is difficult, we propose meta-learning a transformer-based optimizer to enhance CL. This meta-learned optimizer uses attention to learn the complex relationships between model parameters across a stream of tasks, and is designed to generate effective weight updates for the current task while preventing catastrophic forgetting on previously encountered tasks. Evaluations on benchmark datasets like SplitMNIST, RotatedMNIST, and SplitCIFAR-100 affirm the efficacy of the proposed approach in terms of both forward and backward transfer, even on small sets of labeled data, highlighting the advantages of integrating a meta-learned optimizer within the continual learning framework.
12	Zekeri Adams	Comenius university	Zekeri Adams	INVESTIGATING META MODELLING LANGUAGES WITH THE AIM TO BETTER CHARACTERIZE THE HIDDEN SEMANTICS IN KNOWLEDGE GRAPHS. CF	Knowledge graphs have emerged as a powerful tool shaping the landscape of artificial intelligence and intelligent systems. Operating on a graph-based data model, they play a pivotal role in integrating, managing, and deriving value from diverse datasets on a large scale. The semantics of knowledge graphs goes beyond the factual structuring of objects and relationships in nodes and edges. Most real-life domains are complex where classes are instances of other classes. As a result of modelling by different experts in these complex domains, it has resulted in entity ambiguity where the same entity is modelled differently at different places. In the light of this, it is important of have a unified framework for modelling in such a complex domain, as this will enhance semantics interoperability, data integration and visualization. Our investigation delves into two meta-modelling languages, ML2 and PURO, both equipped with higher-order constructs tailored for modelling subject domains with intricate features. These languages are formalized in first-order logics, providing a structured framework to accommodate complexities often absent in graph-based models within such domains. Our objective is to synthesize the unique attributes of PURO and ML2 to create a comprehensive framework for modelling in knowledge graphs across complex domains. Through the utilization of the PURO Modeler tool, we demonstrate the practical significance of these languages by identifying issues in multi-level taxonomic structures, using segments from a known knowledge graph as illustrative examples.
13	Ivica Obadic	Technical University of Munich, Munich Center for Machine Learning	Ivica Obadic, Dmitry Kangin, Plamen Angelov, Xiaoxiang Zhu	Graph Neural Networks for Remote Sensing_CF	Remote sensing imagery is becoming a popular data source to tackle pressing problems related to climate changes, extreme events, or urbanization. In recent years, graph neural networks have become increasingly popular in computer vision. Their representation of the input image in terms of objects and relations between them can be instrumental in capturing the complex relationships between the objects in remote sensing imagery. Yet, despite these benefits, graph neural networks are not frequently used in remote sensing. In our work, we aim to bridge this gap by presenting a novel framework for knowledge extraction in remote sensing imagery based on graph neural networks. Throughout extensive experiments, we will show an evaluation of our approach against the established benchmarks in terms of prediction performance and its intrinsic interpretability capabilities.
14	Siddharth Mehrotra	TU Delft	Siddharth Mehrotra, Folkert van Delden, Eva Bittner, Ujwal Gadiraju, Catholijn Jonker, and Myrthe Tielman	Fostering Appropriate Trust in Predictive Policing AI Systems	Law enforcement agencies worldwide are increasingly using machine learning systems for crime prevention and resource allocation. Predictive policing, a notable example, employs data analysis and algorithms to predict criminal activity and optimize resource deployment. Concerns regarding user trust levels in such systems have garnered significant attention. Under-trust may lead to inadequate reliance, while over-trust can result in over-compliance, negatively impacting tasks. Users must maintain appropriate levels of trust. Past research indicates that explanations provided by AI systems about their inner workings can enhance user understanding of when to trust or not trust the system. The role of explanations in building trust varies based on the task and user expertise. This study explores the impact of different explanation types (text, visual, and hybrid) and user expertise (retired police officers and lay users) on establishing appropriate trust in AI-based predictive policing systems. While we observed that the hybrid form of explanations significantly increased the subjective trust in AI for expert users, no form of explanation significantly helped in building appropriate trust. The findings of our study underscore the nuanced relationship between explanation types and user expertise in establishing appropriate trust, emphasizing the importance of reevaluating the use of explanations. Finally, based on our results we synthesize potential challenges along with policy recommendations to design for appropriate trust in AI-based predictive policing systems.
15	Martin Homola	Comenius University in Bratislava / Slovak.AI	Peter Anthony, Štefan Balogh, Claudia d’Amato Franco Alberto Cardillo, Franca Debole, Martin Homola, Ján Kľuka, Umberto Straccia, Alexander Šimko, Peter Švec, Daniel Trizna	Explainable Malware Detection Activity Report of TAILOR WP3.1 Coordinated Action	The coordinated action “Explainable Malware Detection” under the frame of WP3.1 aims to investigate XAI methods that can be used to explain outcomes of malware detection methods. To this end either XAI methods can be adopted and applied on top of MD tools, or MD methods can be enhanced and become more explainable. Under our CA we have so far focused on concept learning which has been applied on top of EMBER – which is a pre-classified malware dataset. One of our main results is a handcrafted ontology that can be paired with EMBER and is needed for this task. We are working with a number of concept learning toolkits (including DL learner, DL FOIL/FOCL, Fuzzy DL learner) which we hope to evaluate and compare on this use case. In the following period we would like to extend our investigation also on KB Embedding and Ontology extraction from neural classifiers. We hope to compare the applicability of these methods to state-of-the-art concept learning.
16	Piotr Skrzypczynski	Poznan University of Technology	Piotr Skrzypczynski, Tomasz Nowak	Estimating spatial uncertainty in visual perception based on machine learning	This poster presents a novel neural network architecture for estimating vehicle poses from monocular images, leveraging the HRNet backbone and task-specific heads for high accuracy. The network incorporates an uncertainty estimation method using the Unscented Transform, enhancing decision-making by providing uncertainty measures for the estimated poses. Outperforming existing methods, it offers a full pipeline for uncertainty propagation, enabling more informed decisions based on estimated car poses.
17	Nadeem Qazi	University of East London, UK	Nadeem Qazi	Enhancing Authenticity Verification with Transfer Learning and Ensemble Techniques in Facial Feature-Based Deepfake Detection	Deepfake technology, facilitated by deep learning algorithms, has emerged as a significant concern due to its potential to deceive humans with fabricated content indistinguishable from reality. The proliferation of deepfake videos presents a formidable challenge, propagating misinformation across various sectors such as social media, politics, and healthcare. Detecting and mitigating these threats is imperative for fortifying defenses and safeguarding information integrity. This paper tackles the complexities associated with deepfake detection, emphasizing the necessity for innovative approaches given the constraints of available data and the evolving nature of forgery techniques. Our proposed solution focuses on leveraging facial features and transfer learning to discern fake videos from genuine ones, aiming to identify subtle manipulations in visual content. We systematically break down videos into frames, employ the Haar cascade algorithm for facial recognition, and utilize transfer learning to extract discriminative features. We evaluate multiple pre-trained models, including VGG16, ConvNeXt- Tiny, EfficientNetB0, EfficientNetB7, DenseNet201, ResNet152V2, Xception, NASNetMobile, and MobileNetV2, for feature extraction. Subsequently, we feed these features into a Deep Artificial Neural Network (DANN) for deepfake detection and employ ensemble learning to combine the strengths of the best-performing models for enhanced accuracy. We found that the ensemble model comprising ConvNextTiny, EfficientNetB0, and EfficientNetB7 showed enhanced accuracy in detecting deep fakes compared to alternative models achieving up to 98% accuracy through ensemble learning. Index Terms—Deepfake detection, video classification, Transfer learning, EfficentNetB0, DenseNet, Ensemble learning
18	Pankaj Pandey	Norwegian University of Science and Technology	Pankaj Pandey	ENFIELD: European Lighthouse to Manifest Trustworthy and Green AI	ENFIELD is a project funded by the European Commission. ENFIELD is set to establish a distinctive European Center of Excellence focused on advancing fundamental research in Adaptive, Green, Human-Centric, and Trustworthy AI. These pillars represent novel, strategic elements crucial for developing, deploying, and accepting AI in Europe. The initiative seeks to elevate research within key sectors like healthcare, energy, manufacturing, and space by attracting top talents, technologies, and resources from leading European research and industry entities. ENFIELD aims to strengthen the EU’s competitive position in AI by conducting high-level research aligned with industry challenges, generating significant socio-economic impact for European citizens and businesses. The project envisions a dynamic European AI network comprising 30 consortium members from 18 countries, including leading educational and research institutions, large-scale businesses, SMEs, and public sector representatives. This collaborative effort will collectively tackle critical issues at the forefront of research and innovation within European AI. ENFIELD intends to deliver impactful outcomes, including over 75 unique AI solutions (algorithms, methods, simulations, services, datasets, and prototypes), 180 scientifically influential publications, and 200 peer-reviewed presentations. Additionally, four strategic documents, namely the Common Research Roadmap and Vision, the dynamic Safety and Security Risk Assessment Framework, the White Paper, and the Gender and Ethics Framework, will be produced. ENFIELD plans to support more than 76 individual researchers and 18 small-scale projects through Open Calls, facilitating exchange and innovation. The initiative will also conduct education and training activities, such as summer schools and hackathons, alongside well-designed outreach methods to enhance community engagement, expansion, and sustainability for the ENFIELD project.
19	Andres L. Marin	Universitat Politècnica de València	Andres L. Marin, M Perello-Nieto, F Martínez-Plumed, MJ Ramirez-Quintana, G Fontaras	Driving Towards Sustainability: Explainable Variational LSTM Autoencoders Energy-Efficient Driving patterns classification	Transport accounts for more than 20% of European Union (EU) Greenhouse Gas (GhG) emissions with the majority originating from road vehicles. In view of EU’s commitment to achieve climate neutrality by 2050, a series of measures at all levels needs to be taken for curbing emissions. When it comes to light road vehicles, a significant reduction can be achieved by promoting more energy efficient driving and operation. This study addresses the complex relationship between individual driving and vehicle energy consumption focusing on the drivers’ influence on vehicle energy consumption and eventual CO2 emissions. The study introduces a framework utilizing a Variational Long Short-Term Memory (LSTM) autoencoder integrated with explainable techniques to understand the impact of fuel and energy consumption events. The main focus is on the local explainability of the model’s latent space. The model ingests a multivariate input comprising speed, acceleration, friction brake power, acceleration pedal position, and jerk, aiming to distil complex, high-dimensional data into a low-dimensional latent space. This study also sheds light on the underlying factors influencing energy usage patterns through the interpretability of the latent variables. By deploying local interpretability in the latent space, the study analyzes the contribution of each input feature to each latent space, unveiling the direct impact of specific driving dynamics on energy consumption. The results of such analysis could be used to directly reproduce energy optimised naturalistic driving patterns that would not alter the drivers’ experience and at the same time offer energy and emissions savings.
20	Jiangtao Wang	Coventry University	Jiangtao Wang	Domain-Aware Machine Learning for Population Health Prediction	This talk highlights the significance of population health prediction and explores the transformative potential of machine learning (ML) and artificial intelligence (AI) in this realm. The challenge lies in the naive adoption of generic ML algorithms, leading to issues of reliability, bias, and trustworthiness. Our approach focuses on embedding domain knowledge into ML models to customize and guide predictions. The presentation includes a case study of AI-based population health monitoring in the UK, showcasing the successful integration of domain-aware machine learning. This real-world example illustrates how incorporating healthcare expertise enhances model reliability and aligns predictions with the unique characteristics of the local population. Attendees will gain insights into the methodology and benefits of domain-aware ML, offering a practical understanding of its application in population health prediction.
21	Hikaru Shindo	TU Darmstadt	Hikaru Shindo, Manuel Brack, Gopika Sudhakaran, Devendra Singh Dhami, Patrick Schramowski, Kristian Kersting	DeiSAM: Segment Anything with Deictic Prompting	Large-scale, pre-trained neural networks have demonstrated strong capabilities in various tasks, including zero-shot image segmentation. To identify concrete objects in complex scenes, humans instinctively rely on deictic descriptions in natural language, i.e., referring to something depending on the context such as “The object that is on the desk and behind the cup.”. However, deep learning approaches cannot reliably interpret such deictic representations due to their lack of reasoning capabilities in complex scenarios. To remedy this issue, we propose DeiSAM — a combination of large pre-trained neural networks with differentiable logic reasoners — for deictic promptable segmentation. Given a complex, textual segmentation description, DeiSAM leverages Large Language Models (LLMs) to generate first-order logic rules and performs differentiable forward reasoning on generated scene graphs. Subsequently, DeiSAM segments objects by matching them to the logically inferred image regions. As part of our evaluation, we propose the Deictic Visual Genome (DeiVG) dataset, containing paired visual input and complex, deictic textual prompts. Our empirical results demonstrate that DeiSAM is a substantial improvement over purely data-driven baselines for deictic promptable segmentation.
22	Jose Carlos Sola Verdú	AIJU	Jose Carlos Sola Verdú	Comprehensive and intelligent analysis for precision medicine applied to brain tumors	The following research project seeks to develop a system that incorporates and analyzes clinical data of patients affected by brain tumors in a flexible and comprehensive manner, providing answers to the individual needs of each patient in an intuitive way. This system will allow the incorporation of clinical data of patients affected by primary brain tumors, implementing algorithms based on Artificial Intelligence (AI) for their analysis. In addition, image data (histological stains, radiological images) and omics data (transcriptomics, epigenomics and oncological genomics) will be integrated retrospectively, necessary for the training of AI algorithms. The main conclusions derived from these analyses will be used in the diagnosis and classification of new cases, as well as in the evaluation of possible new markers of clinical and therapeutic utility, all through a simple and intuitive user interface. Therefore, the construction of this system will function as a test bed for experimenting with the latest AI tools on different types of data beyond imaging. This will allow us to integrate information from other sources of interest, such as gene expression variations (RNA), genetic mutations (DNA) and epigenetic changes (DNA methylation), among others. The implementation of omics data can compensate for the paucity of information in brain tumors due to their low prevalence compared to breast, lung or colorectal cancer. Although there are new techniques that allow the increase of information with little volume of initial images, the inclusion of other types of data can accelerate the correct learning of Artificial Intelligence tools.
23	Sergio Muñoz	Universidad Politécnica de Madrid	Sergio Muñoz	CF63 – Leveraging Social Agents as Mediators to Foster Trust and Comprehension of Affective Engagement with Digital Content	The overwhelming volume and diversity of digital content make it difficult for individuals to properly access and understand. This challenge is compounded by the dynamic nature of the Internet, driven by attention-seeking strategies that intricately exploit people’s unconscious emotional responses, which affects the credibility of information and increases the spread of fake news. In today’s attention society, where AI plays a crucial role, this complex landscape poses significant challenges for individuals seeking accurate and trustworthy information. Such a landscape has boosted the need for innovative approaches that promote users’ comprehension of social or emotional values present in digital content and allow them to regulate their affective responses. Advancements in sentiment analysis have provided valuable insight into identifying these values within digital content. However, there is a need to integrate these technologies into user-centred interfaces that effectively present those insights and empower users to control their affective engagement. The literature has shown the potential of social AI to offer real-time assistance in a variety of tasks, but there is limited empirical evidence of its utility in fostering comprehension, trust, and affective engagement between users and digital content. This proposal aims to contribute to the field by investigating the concept of content mediators using social agents. These agents will act as intermediaries between information sources and users, providing enriched information and perceiving users’ emotions to encourage the regulation of their affective responses. To this aim, Dr Sergio Muñoz is applying for a connectivity fund for a three-month research visit to the Group of AI for People and Society (GAIPS) at Instituto Superior Técnico. The applicant conducts teaching and research in the Intelligent Systems Group (GSI) at Universidad Politécnica de Madrid. He has experience using intelligent techniques for the promotion of well-being. Collaborating with Prof. Ana Paiva and the GAIPS team, renowned for its expertise in creating social agents, will provide valuable insights into the design and development of user-centred interfaces based on social agents with integrated emotion awareness and regulation capabilities. The results of this research will help foster comprehension and critical engagement among users as they navigate the ever-evolving digital landscape. By equipping users with insights into the emotional and social dimensions of digital content, this research is expected to also cultivate a heightened sense of trust in AI technologies.
24	Matthias König	Leiden University	Matthias König, Xiyue Zhang, Holger Hoos, Marta Kwiatkowska, Jan van Rijn	Automated Design of Linear Bounding Functions for Sigmoidal Nonlinearities in Neural Networks	The ubiquity of deep learning algorithms in various applications has amplified the need for assuring their robustness against small input perturbations such as those occurring in adversarial attacks. Existing complete verification techniques offer provable guarantees for all robustness queries but struggle to scale beyond small neural networks. To overcome this computational intractability, incomplete verification methods often rely on convex relaxation to over-approximate the nonlinearities in neural networks. Progress in tighter approximations has been achieved for piecewise linear functions. However, robustness verification of neural networks for general activation functions (e.g., Sigmoid, Tanh) remains under-explored and poses new challenges. Typically, these networks are verified using convex relaxation techniques, which involve computing linear upper and lower bounds of the nonlinear activation functions. In this work, we propose a novel parameter search method to improve the quality of these linear approximations. Specifically, we show that using a simple search method, carefully adapted to the given verification problem through state-of-the-art algorithm configuration techniques, improves the average global lower bound by 25% on average over the current state of the art on several commonly used local robustness verification benchmarks.
25	Zakaria A. Dahi	French National Institute of Research in Informatics and Automation	Zakaria A. Dahi, Francisco Chicano and Gabriel Luque	An Evolutionary Deep Learning Approach for Efficient Quantum Algorithms Transpilation	Gate-based quantum computation describes algorithms as quantum circuits. These can be seen as a set of quantum gates acting on a set of qubits. To be executable, the circuit requires complex transformations to comply with the physical constraints of the machines. This process is known as transpilation, where qubits’ layout initialisation is one of its first and most challenging steps, usually done by considering the device error properties. As the size of the quantum algorithm increases, the transpilation becomes increasingly complex and time-consuming. This constitutes a bottleneck towards agile, fast, and error-robust quantum computation. This work proposes an evolutionary deep neural network that learns the qubits’ layout initialisation of the most advanced and complex IBM heuristic used in today’s quantum machines. The aim is to progressively replace weakly scalable transpilation heuristics with machine learning models. Previous work using machine learning models for qubits’ layout initialisation suffers from some shortcomings in the proposal’s correctness and generalisation as well as benchmarks diversity, utility, and availability. The present work solves those flaws by (I) devising a complete Machine Learning pipeline including the ETL component and the evolutionary deep neural model using the linkage learning algorithm P3, (II) a modelling applicable to any quantum algorithm with a special interest to both optimisation and machine learning ones, (III) diverse and fresh benchmarks using calibration data of four real IBM quantum computers collected over 10 months (Dec. 2022 and Oct. 2023) and training dataset built using four types of quantum optimisation and machine learning algorithms, as well as random ones. The proposal has been proven to be more efficient and simple than state-of-the-art deep neural models in the literature.
26	Hao Zhou	Department of Computer Science, University of Oxford	Hao Zhou, Yongzhao Wang, Konstantinos Varsos, Nicholas Bishop, Rahul Savani, Anisoara Calinescu, Michael Wooldridge	A Strategic Analysis of Prepayments in Financial Credit Networks	In financial credit networks, prepayments enable a firm to settle its debt obligations ahead of an agreed-upon due date. Prepayments have a transformative impact on the structure of networks, influencing the financial well-being (utility) of individual firms. This study investigates prepayments from both theoretical and empirical perspectives. We first establish the computational complexity of finding prepayments that maximize welfare, assuming global coordination among firms in the financial network. Subsequently, our focus shifts to understanding the strategic behavior of individual firms in the presence of prepayments. We introduce a prepayment game where firms strategically make prepayments, delineating the existence of pure strategy Nash equilibria and analyzing the price of anarchy (stability) within this game. Recognizing the computational challenges associated with determining Nash equilibria in prepayment games, we use a simulation-based approach, known as empirical game-theoretic analysis (EGTA). Through EGTA, we are able to find Nash equilibria among a carefully-chosen set of heuristic strategies. By scrutinizing the equilibrium behavior of firms, we outline the characteristics of high-performing strategies for strategic prepayments and establish connections between our empirical and theoretical findings.
27	Anna L. Münz	RWTH Aachen University	Annelot W. Bosman, Anna L. Münz, Holger H. Hoos, and Jan N. van Rijn	A Preliminary Study to Examining Per-Class Performance Bias via Robustness Distributions	As neural networks are increasingly used in sensitive real-world applications, mitigating bias of classifiers is of crucial importance. One often-used approach to controlling quality in classification tasks is to ensure that predictive performance is balanced between different classes; however, it has been shown in previous work that even if class performance is balanced, instances of some classes are easier to perturb in such a way that they are misclassified, which indicates that per-class performance bias exists. In this preliminary study, we found that even when class performance is balanced, class robustness can vary strongly when assessing the robustness of a given neural network classifier in a more nuanced fashion. For this purpose, we use robustness distributions, i.e., empirical probability distributions of some robustness metric, such as the critical epsilon value, over a set of instances. We observed that the robustness of the same class over the same data can significantly differ from each other for different neural networks; this means that even when a neural network appears to be unbiased, it might be easier to perturb instances of a given class so that they are misclassified. Furthermore, we explored the robustness distributions when we have a predefined target class, i.e., a specific class into which an instance is misclassified after perturbation. Our empirical results indicate that in most cases, there are significant differences in robustness distributions for different classes. While our empirical results reported here are for MNIST classifiers, we are currently performing experiments using the German Traffic Sign Recognition Benchmark. Furthermore, we are running experiments with retrained networks for fairness, to see whether this has a significant effect on the per-class robustness distributions. Lastly, we aim to create a robust class fairness metric based on our findings.
28	Gregoris Mentzas	ICCS, National Technical University of Athens	Mattheos Fikardos, Katerina Lepenioti, Dimitris Apostolou, Gregoris Mentzas	A Card-based Agentic Framework for Supporting Trustworthy AI	The rapid advancements in AI have triggered the need for Trustworthy AI (TAI), which encompasses various definitions, perspectives, and technological approaches aimed at ensuring that AI is reliable and trusted by humans. Already organizations, governments, and academia have produced regulations and frameworks around TAI (e.g. the EU AI Act, the NIST Risk Management Framework), but a gap still exists between those ethical and legal guidelines and their practical applicability by companies. Despite the plethora of methods and algorithms that assess or enhance AI trustworthiness, these remain fragmented, each targeting a specific aspect of trustworthiness (e.g. accuracy, transparency, fairness, explainability). This gap increases the need for a holistic approach that would provide support to companies developing and deploying trustworthy AI systems. Our work primarily tries to address this gap with a methodological framework that unifies and entangles the AI development lifecycle with the trustworthiness requirements and integrates them within a software solution. We define the lifecycle phases of an AI system, from its design and development through deployment and monitoring, as well as the trustworthiness phases, following a risk management approach where AI risks are identified, assessed, and mitigated. In order to record and structure information related to TAI within our framework, we use and extend the “card-based” approach. Actually, we collect information through data cards, model cards and use case cards. We also propose the use of “methods cards”, which structure technical information about available algorithms, methods and software toolkits that assess and enhance AI trustworthiness. For the development of our software framework, we follow a neuro-symbolic agentic design pattern in which different roles involved in the TAI assessment process can be instantiated. This is done by enabling an LLM to be prompted and guided by navigating knowledge graphs, which have been derived from the already recorded cards. Currently, we have developed an initial prototype which, at this stage, focuses on two TAI dimensions: fairness and robustness. We have recorded more than 20 methods for each for these dimensions and developed the corresponding cards. Our aim is to further work on additional TAI dimensions and evaluate our framework in three distinct cases: disinformation and fake news detection; cancer risk identification and assessment; and management of disruptive events in ports.
29	Andrii Krutsylo	Polish Academy of Sciences	Andrii Krutsylo	(CF) Batch Sampling for Experience Replay	We address the problem of catastrophic forgetting in deep learning models trained on data streams with distribution shifts. In this setting, the model faces a series of classification tasks with new classes sequentially, without the ability to retrain on past data. This limitation is often arises due to data sensitivity or resource constraints, making it critical for the model to adapt to new information while retaining performance on previously learned tasks, a challenge known as continual learning. Experience Replay, one of the most effective methods of continual learning, involves saving and including a small number of old training samples in the current training mini-batch to maintain a distribution representative of all previously seen tasks. Typically, the effectiveness of the samples selected from memory to be replayed is evaluated based on metrics such as confidence in predictions or difference in gradients or hidden representations of the isolated samples before and after the model update on the current training mini-batch, assuming that this indicates forgetting of the sample. The problem with this approach is that if the model forgets a sample, it will also forget nearby samples in the feature space, which is very dynamic and unpredictable due to distribution shifts in the training data. And those samples would be the first to be selected by the metric, which does not take into account whether the neighbor has already been selected. That is, we would almost certainly have redundancy in the batch consisting of the highest scoring samples. The lack of diversity is usually addressed by randomly sampling before or after scoring. In either case, the resulting mini-batch would be suboptimal. As an alternative, I proposed to score the forgetting of the entire mini-batch. The new metric computes the impact caused by the model update on the hidden representations and evaluates its strength as well as its diversity, since the overall score of the batch would be the higher the greater the changes in different parts. However, this approach only provides information about the potential effectiveness of the selected batch compared to other evaluated batches. Performing an exhaustive search for optimal sample combinations after each model update remains computationally demanding, especially given the large number of possible combinations, even in a small memory buffer. The main contribution of this paper was to show that this line of research is even possible, by demonstrating that even a fast search over the limited number of randomly sampled mini-batches leads to competitive results.
30	Jakub Kwiatkowski, Krzysztof Krawiec	Poznan University of Technology	Jakub Kwiatkowski, Krzysztof Krawiec	Self-supervised Learning of Tokenized Representations for Raven Progressive Matrices	A Raven’s Progressive Matrix (RPM) problem comprises a matrix of context panels and a set of eight answer panels, containing 2D geometric objects of various properties (shape, color, size, and angle). The task is to choose the answer panel that best complements or fits the unknown pattern when placed at the missing lower-right corner of the matrix known as the query panel. We propose the Abstract Compositional Transformer (ACT), a multi-stage deep learning model comprising an image tokenizer, a transformer, and a property predictor, trained to simultaneously predict the properties of the query panel and classify (recognize) the properties of the context panels. We use the ACT to solve RPM tasks by querying it on the task with the query panel first masked-out and then replaced with each of the 8 answer panels. ACT achieves a success rate of 97.0% on the widely used RAVEN benchmark, outperforming all previously published methods.
31	Jose Carlos Sola Verdú	AIJU	Jose Carlos Sola Verdú	Development of a Digital Product Passport integration system for the toy industry	The following research project represents a significant advance towards digitization and sustainability in business and government, through the development of a Digital Product Passport, thus addressing the growing need for transparency, security and efficiency in product information management through the integration of blockchain technology. The importance of this development lies in its ability to ensure the authenticity and integrity of product data, from its origin to its final disposition. The use of blockchain provides a secure and immutable foundation, where information can be shared and verified by stakeholders without compromising confidentiality or privacy. This approach enables improved product traceability and facilitates adherence to sustainability and circular economy regulations, aligning with global initiatives such as the European Green Pact and the Circular Economy Action Plan. In addition, the Digital Passport is an essential tool for companies seeking to adapt to the demands of an increasingly sustainable market, offering a means to demonstrate the ethical and ecological provenance of their products. This project not only benefits companies in terms of compliance and efficient management, but also empowers consumers with the information they need to make informed purchasing decisions. The implementation of this research aims to facilitate the interaction of companies and governments with product data, marking an important milestone in the transition to more transparent, safe and sustainable business practices. This project demonstrates how technology can be innovatively applied to solve complex product information management challenges, reaffirming the importance of digitization in improving sustainability and efficiency in the modern era.