Our project has two goals: a) make machines recognise speech more like humans do, and b) validate our understanding about human speech perception through the use of data-driven techniques. MIM aims at proposing computational models that predict human speech recognition at a fine resolution. Current approaches to intelligibility prediction provide macroscopic estimates consisting of aggregates over many stimuli and listeners. By leveraging recent developments in the field of Artificial Intelligence, models could predict recognition at a sub-lexical level. Deep learning (DL) has improved automatic speech recognition performance significantly, achieving super-human transcription in conversational tasks. We plan to build DL models to predict human listening tests responses aiming at improving individualization of hearing solutions. Scarcity and variability of human listening data, and the interpretation problem in DL are two of the main issues that we will tackle.

The objective of the SargAlert project is to significantly improve the forecasts of the strandings of the invasive algal species Sargassum in the tropical Atlantic Ocean, in the Caribbean Sea and on the Brazilian coast. The synergy between satellite data / ocean transport modeling / in-situ measurements will be used for that purpose. SargAlert will provide alert bulletins to end-users such as territorial authority, tourism, fishers. The challenges that will be addressed by SargAlert are as follows: - detection and monitoring of at different time (hour to daily) and spatial (20 m to 5 km) scales using a multi-sensor satellite data analysis (Low Earth and GEOstationary orbits), - improvement of Sargassum stranding forecasts by combining physical transport models with artificial intelligence approaches, - validation of satellite data products and forecast models using in-situ measurements, - production of alert bulletins to address societal issues. The innovative developments of the project will enable an integrative approach of the Sargassum stranding issues: synergy between satellite data, understanding of Sargassum spatio-temporal distribution, transport forecast. Improvements of ocean modeling of dynamics will benefit societal authorities to better respond to the risks induced by the more frequent and intense Sargassum blooms in the Atlantic Ocean. The operational Sargassum forecast center will thus have all required inputs to provide reliable forecasts in near real time. This federative and interdisciplinary project includes complementary partners from academic laboratories, including a human science team (AEM, IRISA, LATMOS, LC2S, LIS, Marbec, MIO, UFPE/UFRPE), from an operational forecast center (Météo-France) and from a national satellite data center (AERIS/ICARE).

Despite great enthusiasm for deep learning in NLP, concern is rising about its limitations. First, neural models are often blackboxes, and their behavior is hard to interpret. Second, benchmark-based evaluation overlooks biases, questioning the robustness and coverage of the resulting generalisations, yielding a landscape of overall diversity. The goal of the SELEXINI project is to address these issues by developing **weakly supervised methods to induce semantic lexicons** from raw corpora, which will then be **seamlessly integrated with semantic text processing models**. Lexical units are seen as useful abstractions that allow representing complex phenomena (e.g. polysemy, similarity, multiword units) associated with interpretable labels, avoiding the overhead and opaqueness of contextualized embeddings (one vector per occurrence). Moreover, our lexicon will combine continuous data (embeddings, clusters) and symbolic data (labels). We will model single and multiword units, their senses, and their semantic frames (arguments, roles). Hence, we propose a new "by-construction" view on interpretability, which can be seen as an alternative to methods trying to dissect complex neural models. For extrinsic evaluation of interpretability and diversity, the induced lexicon will be integrated into standard deep learning models in downstream tasks requiring semantic information: machine reading comprehension and multiword expressions identification. We will develop an experimental protocol to assess the lexicon-corpus complementarity on diverse linguistic phenomena, and to assess the lexicon's usefulness for non-expert end users requiring interpretable results. We expect that this original approach will increase both the interpretability of models and the coverage of diverse phenomena (e.g. rare/unseen items in training data).

OMRES is a project at the forefront of Europe's green energy transition, targeting the seamless integration of zero-emission power technologies into the existing power system. The project's core objective is the development of digitalized operating and maintenance solutions for hybrid renewable energy sources (RES) systems, including storage, with a focus on enhancing technology performance and contributing to zero-emission power production. Scientifically, OMRES aims to achieve several key objectives: (a) advancing RES power forecasting models through atmospheric modeling for improved energy yield prediction accuracy, (b) designing fault diagnosis and lifetime extension tools for power electronics inverters to enhance the overall reliability of RES, (c) developing bidding and control strategies to optimize the operation of hybrid RES sites, maximizing profitability and offering ancillary services in both Grid Forming and Grid Following modes, and (d) implementing a real-time digital twin framework with enhanced data analytics for the validation of OMRES solutions in operational environments. By tackling challenges arising from the ongoing clean energy transition, including RES integration issues, cost reduction, and improving the reliability and lifetime of RES-storage systems, OMRES leverages digitalization to develop advanced solutions aligned with Transition Initiative 2 goals. This project is poised to make significant contributions to overcoming challenges associated with the integration,

Most of our cognitive activities require processing and memorizing sequences of information. Whether sequences of motor gestures, sounds, letters, or words, we memorize them by associating the elements that make up these sequences. The field of implicit statistical learning aims at understanding the associative mechanisms that allow us to memorize these sequences of information. The HEBBIAN project aims to better understand these fundamental associative mechanisms by relying on recent theoretical models whose general framework is Hebbian learning. This project aims to experimentally study three main questions concerning 1) the role of the spacing between two repetitions of the same sequence in the memorization of this sequence ; 2) the dynamic of sequence encoding as a function of sequence size, number, and learning context ; 3) the problem of parts/whole relations between sequences of different sizes. This experimental work will be carried out in a comparative perspective with humans and non-human primates (Guinea baboons, Papio papio) using serial pointing tasks and classical psycholinguistic tasks, such as the naming or lexical decision tasks, in which a sequence is systematically repeated without the subjects being informed. This experimental work will be done in conjunction with the development and evaluation of two types of models based on the principles of Hebbian learning (psychological models and others that are more plausible on the neurobiological level). All of this work should allow for significant progress in our understanding and conception of these fundamental associative mechanisms.