Graph Neural Networks (GNNs) have proven highly effective across diverse applications, including semi-supervised learning, point cloud semantic segmentation, materials modeling, and drug discovery. Nonetheless, GNNs primarily excel at capturing pairwise interactions, making them less capable of representing complex multi-way relationships inherent in interconnected systems. Graphs are indeed a subset of a broader mathematical construct known as simplicial complexes, which represent geometrical objects through combinatorial data. Particularly, graphs correspond to simplicial 1-complexes, encompassing nodes and edges, with nodes alone considered simplicial 0-complexes. Simplicial complexes provide an elegant framework for geometry processing, particularly in the context of triangular meshes, which are naturally represented with simplicial 2-complexes. Recent advances in topological machine learning emphasize the advantages of working with data defined on simplicial complexes, sparking a growing interest in the design of simplicial neural network architectures. Our project, Deep Simplicial Neural Networks for Advanced Geometry Processing (DeSNAP), is committed to creating robust and mathematically sound methodologies for designing deep learning models tailored for simplicial complexes. This endeavor presents its own set of challenges and open research questions, including computational complexity, over-smoothing, and over-squashing. We intend to leverage the innovations from DeSNAP to process 3D meshes represented as triangle meshes (simplices of dimension 2). Our primary focus centers on important applications in geometry processing, including the domains of mesh compression and generation.

The XAI4AML (Explainable AI for Anti-Money Laundering) chair will explore how AI and explainability affect the optimal level of financial regulation, including how different levels of explainability and regulation may affect the costs and benefits associated with deploying AI-based solutions for anti-money laundering (AML) enforcement. Traditional approaches used by banks to fight money laundering are both costly (€20 billion per year in Europe), and relatively ineffective, being based on deterministic rule-based models. Current AML systems generate many false positives, while at the same time missing large amounts of truly suspicious transactions. Professional criminals use sophisticated techniques to disguise transfers as normal-looking transactions. AI can reduce false positives and bring about greater effectiveness by identifying otherwise invisible trends across large data sets. However, problems of explainability, together with regulatory uncertainty, are the main barriers to implementing AI in AML systems. My interdisciplinary chair, combining economics, law (with Winston Maxwell, Director Law & Technology) and AI/data science (Stéphan Clémençon, Professor Applied Mathematics), will contribute to the economic literature on the economics of financial regulation and financial crime, while at the same time contributing to an operational need for clarity on what constitutes and “explainable” AI system for AML. The results will have a positive impact on designers of AI-based AML systems (such as the French fintech Bleckwen, partner of the chair), on users of the systems (such as banks and consulting firms, in particular PWC specialized in banking operations and compliance, partner of the chair), and on the financial regulator (ACPR, partner of the chair).

Artificial neural networks are nowadays one of the most studied and used algorithms for Artificial Intelligence to solve a huge variety of tasks. However, many deep models share a common drawback: their growing complexity challenges the computational capability of embedded devices and poses questions around the power consumption when deployed on-the-field. Neural Architecture Search targets automatic research of neural network architectures, optimizing metrics like performance and latency. However, typical NAS approaches do not keep into account the target hardware architecture but they optimize over generic metrics like FLOPs and memory footprint, and do not take into account the possible presence of algorithmic biases. Indeed, after the recent AI Act, neural AI will be regulated also against the presence of these biases, which find their best ground to prosper in more compact architectures. BANERA proposes to: a) study the bias propagation inside neural networks with the employment of tools from information theory, and to propose a differentiable term to minimize the bias propagation; b) design a penalty term which self-adapts to specific hardware architectures (hence, keeping also into account more metrics like for example caching and memory transfers), in order to obtain a "green" and optimized deep model; c) research biasing/debiasing sub-structures in the generated models to drive the fast design of efficient AI models which are also algorithmic bias-free.

The speaker diarization (SD) aims to answer the question: “who speaks and when ?”. It still remains a challenging problem due to its various complex real scenario configurations (propagation environment, large number and moving speaker ...). In presence of at least two speakers (meeting, phone conversation, TV show ...), the SD is essential for good performance of automatic transcription or translation algorithms. For the last decade, SD has been focusing on many deep neural networks (DNN) architectures (end-to-end, autoencoder, recurrent neural network, transformer ...) in order to alleviate its high nonlinear complexity. One popular DNN architecture used for SD is the autoencoder which takes two neural networks into account: the encoder mapping the input in a so-called latent space and the decoder which transforms the latent variable to some output data supposed to be identical to the input ones. In parallel, recent papers make use of a multichannel audio dataset as an input of SD using DNN and considerably increase the performances. Most of the aforementioned SD with DNN lacks interpretability although a good average performance has been shown. It consequently makes a DNN hard to train and a low-level adaptability in case of uncommon scenarios of SD not included in the training dataset. Scientific challenges arise from what was stated before as 1) proposing a robust and interpretable DNN architecture that considers 2) multichannel audio input and 3) other multimodal information. One architecture that was proved to reinforce the interpretability and performance of autoencoder is the variational autoencoder (VAE). The VAE assumes a probabilistic model on input data that leads to variational techniques for autoencoder parameter estimation. SAROUMANE project aims to develop new methodologies for MSD combining unified heavy-tailed probabilistic models on multichannel audio signal and multimodal data with a VAE architecture.

APEX-technologies is a SME specialized in instrumentation for tests and characterisation of devices and high bit-rate/high frequency optical systems. Measurement equipments proposed concern the spectral and time analysis of Optical signal such as those appearing in Optical communication systems and sensors. On the other side, Télécom ParisTech’s Optical communication research group is well recognized for his theoretical and experimental activities in the field and, being permanently confronted to state-of-the-art realisation is led to develop its own characterisation concepts and tools and experimental platforms. Several original measurement techniques have been proposed in order to assess the functional performances of devices and systems when no commercial equipment existed. Exhibiting complementary objectives, APEX and Télécom ParisTEch participated in an ANR project OCELOT started in 2011. Convergence between both entities research and innovation strategies was confirmed when APEX employed recently two former PhD and postdoc engineer from Télécom ParisTech. A large number of common technical interests have been indentified through the interaction for which the expertise and knowledge of the academic partner complements the industrial know-how and problematic for the SME. In addition, APEX can propose some new scientific challenges for Télécom ParisTech to meet. The Common Laboratory programme is structured around Research and Innovation Actions (ARI), whose duration will be defined by the Steering Committee. These actions are quite different from a well-defined project in the sense that several of them run in parallel, they are self-consistent, they can be reoriented any time when required. The can range from initial literature studies to the experimental tests of a new ideas. The are not aiming at immediate success but always towards innovations, either for improving existing commercial equipments form APEX or proposing a new range of equipments. Presently three ARI have been identified and can start immediately. ARI are evaluated on a regular basis every 6 months. Then they can be extended or stopped and/or replaced by a new one referring of not to the result obtained. The three initial year of the Common Laboratory within the scope of ANR are seen as a starting activity that would continue in a longer interaction following existing example previously experimented in Télécom ParisTech.