Adaptive AI

Project: Pruning-Aware Adapters for Bitrate and Complexity Scalable LIC Models

Third parties involved: Gabriele Spadaro – The University of Turin
The project reduces the complexity of deep neural networks for learned image compression by combining transfer learning, Low-Rank Adapters, and learnable pruning. It develops a lightweight framework that adapts a single pre-trained compression model to different bitrates, domains, and computational constraints without full retraining or multiple large models. By using pruning masks as dynamic adapters, the model adjusts its size, bitrate, and computational load according to available resources. The project supports efficient, scalable, and adaptable codecs for real-world deployment, contributing to sustainable and energy-efficient AI technologies.

Project: Improving Edge AI Performance by Federation and Adaptive Model Selection

Third parties involved: Ismail Ari, Habtamu Abie, Sandeep Pirbhulal – Ozyegin University
Running DNN models at the edge enables critical field apps including object classification, sensing & control. However, embedded & mobile devices are constrained in computation, energy, and communication. To improve operational efficiency, the prosejct proposes to demonstrate a federated learning (FL) scenario with adaptive & online model selection by starting with a set of pretrained DNN models creating a heterogeneous setting (Pi4/5s, Jetsons, Arduino with different sensors or motors), monitoring model & device performance & costs in accuracy, training/inference time, energy during FL, and switching models online among devices. We expect to reach optimal HW/SW configurations.

Project: SHACKLE: SHape-based pAtterns for Constraining KnowLedge graph Embeddings

Third parties involved: Pierre Monnin – Université Côte d’Azur
The project SHACKLE advances trustworthy AI by combining sub-symbolic methods with the validation schemata of knowledge graphs to reduce errors, increase explainability, and enforce trust. Building on ontologies and shape constraints for soundness and completeness, it explores a currently under-studied path toward neurosymbolic integration.

Project: ADAPT-BioEL: Zero Shot Biomedical Entity Matching and Linking

Third parties involved: Aikaterini Gkirtzou – National Technical University of Athens
The project develops adaptive experience replay strategies for LoRA-based continual learning in foundation models. By exploiting LoRA’s low-rank update structure, it proposes rank-aware and subspace-driven sampling methods based on update geometry, including alignment, coverage, and orthogonality. These methods will be validated against existing replay techniques and assessed for memory use, computational efficiency, and model stability. The project contributes to more sustainable and scalable foundation model deployment, with expected outcomes including improved learning performance, open-source tools, and practical benefits for research and industry.

Project: Semantics-Aware Image Generation From a Scene Sketch

Third parties involved: Ahmed Bourouis – University of Surret
The project explores how freehand sketches can be better integrated into diffusion models through noise modulation to improve controllability, adaptability, and sketch fidelity in image generation. By using sketch information to guide the diffusion process, it aims to produce outputs that better preserve the structure and intent of the input sketch while maintaining high image quality. The work includes benchmarking existing methods, building sketch image datasets, developing noise modulation strategies, and defining evaluation metrics. The project contributes to more adaptive and user-controllable generative AI systems for creative applications.

Project: CHARM: CH4 and CO High-Resolution Mapper

Third parties involved: Oualid Yahia – University of Stirling
The project proposes CHARM, an adaptive AI framework for downscaling Sentinel-5P methane and carbon monoxide products to support scalable environmental monitoring of generative AI systems. By combining multimodal Earth Observation data with a Mixture-of-Experts architecture, CHARM enables context-aware modelling while reducing computational costs through sparse expert activation. The framework integrates data from MODIS, Landsat8/9, land cover, elevation, in-situ measurements, and ERA5 reanalysis, with outputs validated against ground-based monitoring data. The project contributes to energy-efficient, high-resolution emissions monitoring and provides a practical tool for
assessing the carbon footprint of generative AI deployment.

Project: Adaptive Intelligence in Multi-Agent Systems: When Collective meets DRL

Third parties involved: Chuhao Qin – University of Leeds

This project aims to advance AI by enhancing the robustness and adaptability of multi-agent systems. By integrating collective learning and multi-agent deep reinforcement learning (MADRL), we address challenges such as biased information propagation and lack of flexibility. Our approach involves developing an adaptive model for real-time learning, with activities spanning data collection, algorithm development, and evaluation across domains like voice conversation. Expected outcomes include novel AI approaches, opensource contributions, publications, and collaborative partnerships.

Project: CXAI: Cautious explainable artificial intelligence

Third parties involved: Sébastien Destercke – CNRS (Centre National de la Recherche)

This project aims to ensure robustness and trustworthiness by developing classifiers that return set-valued predictions. It addresses two challenges: (1) how to evaluate set-valued predictions and calibrate them to a user’s attitude toward imprecision to build a calibrated, optimal robust imprecise classifier; and (2) how to explain set-valued predictions—both the need for robustness and how robust set-valued models can help test the robustness of classical XAI techniques.

Project: Robust Multimodal Continual Learning for Robotics

Third parties involved: Nicolas Kuske – Artificial and Natural Intelligence Toulouse Institute
This project advances multimodal continual learning (MMCL) by integrating audio-visual cues into reinforcement learning (RL) for robotic manipulation. Objectives include (1) developing a VR-based RL environment for testing and (2) optimizing Global Latent Workspace (GLW) and Semantic-Aware Multimodal (SAMM) models with attentional mechanisms. Activities involve VR setup, model evaluation, and hybrid model fusion. Expected outcomes are a robust MMCL framework handling noisy sensory inputs, adaptable to tasks like robotic pick-and-place. Added European value arises from TU/e (Netherlands) and ANITI (France) collaboration, combining expertise in continual and multimodal learning to advance AI innovation and support adaptive robotics.

Project: LoRA-Specific Adaptive Replay for Continual Learning

Third parties involved: Andrii Krutsylo – Institute of Computer Science Polish Academy of Sciences
The project develops adaptive experience replay strategies for LoRA-based continual learning in foundation models. By exploiting LoRA’s low-rank update structure, it proposes rank-aware and subspace-driven sampling methods based on update geometry, including alignment, coverage, and orthogonality. These methods will be validated against existing replay techniques and assessed for memory use, computational efficiency, and model stability. The project contributes to more sustainable and scalable foundation model deployment, with expected outcomes including improved learning performance, open-source tools, and practical benefits for research and industry.

Project: MoSE: Mixture-of-Specialized-Experts for efficient continual adaptation

Third parties involved: Van-Tuan Tran – Trinity College Dublin
The project proposes MoSE, a Mixture-of-SpecializedExperts framework to improve parameter-efficient fine-tuning of foundation models across multiple tasks. It addresses task interference and catastrophic forgetting through threshold-based routing, activating only the experts most relevant to each task while reducing redundancy and computational overhead. The framework also introduces regularisation for expert specialisation and a weights correction algorithm to manage conflicts during sequential adaptation. Evaluated on language and vision benchmarks, MoSE supports more efficient, stable, and scalable foundation model adaptation, contributing to sustainable and reusable AI technologies.