INRIA a CCSD electronic archive server
Not a member yet
122212 research outputs found
Sort by
Decentralized Online Convex Optimization with Unknown Feedback Delays
No full textInternational audienceDecentralized online convex optimization (D-OCO), where multiple agents within a network collaboratively learn optimal decisions in real-time, arises naturally in applications such as federated learning, sensor networks, and multi-agent control.In this paper, we study D-OCO under unknown, time-and agent-varying feedback delays. While recent work has addressed this problem (Nguyen et al., 2024), existing algorithms assume prior knowledge of the total delay over agents and still suffer from suboptimal dependence on both the delay and network parameters. To overcome these limitations, we propose a novel algorithm that achieves an improved regret bound of O N √ d tot + N √ T(1-σ2) 1/4 , where T is the total horizon, d tot denotes the average total delay across agents, N is the number of agents, and 1 -σ 2 is the spectral gap of the network. Our approach builds upon recent advances in D-OCO (Wan et al., 2024a), but crucially incorporates an adaptive learning rate mechanism via a decentralized communication protocol. This enables each agent to estimate delays locally using a gossip-based strategy without the prior knowledge of the total delay. We further extend our framework to the strongly convex setting and derive a sharper regret bound of O N δmax ln T α, where α is the strong convexity parameter and δ max is the maximum number of missing observations averaged over agents. We also show that our upper bounds for both settings are tight up to logarithmic factors. Experimental results validate the effectiveness of our approach, showing improvements over existing benchmark algorithms.coordinator. Early foundational work in decentralized optimization focused on offline settings, leveraging techniques from gossip algorithms -originally used to achieve consensus to enable distributed optimization Boyd et al. (2011); Nedic & Ozdaglar (2009). The first formal treatment of the online counterpart was given by Hosseini et al. ( 2013), who analyzed a dual averaging algorithm and established sublinear regret guarantees. Specifically, they showed that a regret bound of O(N 5/4 √ T /(1 -σ 2 ) 1/2 ) is achievable, where σ 2 is the second highest singular value of the communication matrix W , whose definition is shown in later sections. Since then, various algorithmic approaches have been developed, including decentralized mirror descent Shahrampour & Jadbabaie (2018), for which a similar regret rate is provable and accelerated gossiping for D-OCO Wan et al. (2024a). The method from (Wan et al., 2024a) notably improves the previous regret bound by a factor of (1 -σ 2 (W )) -1/4 N 1/4 / log(N ). The D-OCO framework has seen various extensions, including work on settings with dynamic networks Hosseini et al. (2016); Lei et al. (2020). For a comprehensive overview of such developments, we refer the reader to the recent monograph by Yuan et al. (2024).Online learning with delayed feedbacks Our work is closely related to the literature on online learning with delayed feedback, initiated by Weinberger & Ordentlich (2002). They considered the setting with uniform, known per-round delays and proposed a general reduction to non-delayed online learning. Subsequent studies extended these results to handle non-uniform delays (Joulani et al., 2013).</div
Encode the Cake and Eat it Too: Controlling computation in type theory, locally
No full textInternational audienceProof assistants based on dependent type theory such as Agda, Lean and Rocq identify objects up to computation during proof checking. This takes away some of the proof burden from the user and even provides a way to get very efficient automation. Recently, Agda and Rocq have been extended to support user-defined computation. While they already prove very useful, user-defined computation rules are global: once they are added, they are here to stay. Importing a development that makes use of those rules then means relying on them, whether we want it or not, which can lead to unwanted incompatibilities. We design LRTT, a type theory with support for local abstraction over user-defined computation rules. This takes the form of a prenex quantification at the definition level. This quantification is supplemented with the possibility to provide one or several instantiations that verify the equations definitionally. We show that a procedure inlining definitions abstracting over definitional equality is possible, in the style of monomorphisation or of C++ templates. In the process we get a conservativity result over more conventional Martin-Löf type theories. There are several benefits to such a system. First, it provides encapsulation for user-defined computation rules, which is important to avoid unwanted bad interactions and limits the scope in which invariants of type theory (such as termination, confluence, type preservation and consistency) are broken. Second, abstraction lets users factorise code that crucially relies on definitional equality, as well as hide implementation details that are irrelevant in some settings. Finally, it gives a way to encode certain features without paying the price of the encoding. We showcase such examples in a prototype implementation as an extension of the Rocq Prover. Additionally, all the results in this have been formalised in Rocq
Impact de l'écho partiel en IRM de flux 4D: un éclairage par IRM synthétique
No full textPurpose: The aim of this study is to investigate the impact of the partial echo on 4D flow MRI sequences thanks to in silico coupled MRI-CFD (Computational Fluid Dynamics) simulations.Methods: Two sequences are studied: one with a full echo (FE) and another using partial echo (PE) with an echo symmetry fraction of 0.75. MRI-CFD simulations are conducted on an in silico pulsatile flow phantom for a sinusoidal inflow and a physiological inflow typical of the ascending aorta.Results: For both inflow signals, PE-based simulations exhibited better compliance with their matching CFD simulations compared to the FE ones. Conclusion:The reduction of flow misregistration artifacts achieved through the use of PE appears to be more beneficial than the drawback of incomplete k-space filling. The MRI-CFD framework presented in this study appears as a useful tool to investigate the design of MRI sequences and to stratify its different sources of errors.</div
Eye2Heart: A reduced mathematical model bridging cardiovascular and ocular hemodynamics
No full textInternational audienceThe cardiovascular and ocular systems are intricately connected, with hemodynamic interactions playing a crucial role in both physiological regulation and pathological conditions. However, existing models often treat these systems separately, thus limiting the understanding of their interdependence. In this study, we present the Eye2Heart model, which is a novel closed-loop mathematical framework that integrates cardiovascular and ocular dynamics. Using an electrical-hydraulic analogy, the model describes the interactions between the heart and retinal circulation through a nonlinear system of ordinary differential equations. The model is tested against clinical and experimental data, thus demonstrating its ability to reproduce key cardiovascular parameters (e.g., stroke volume, cardiac output) and ocular hemodynamics (e.g., retinal blood flow). Additionally, we explore in silico the effects of intraocular pressure and left ventricular compliance on both local ocular and global systemic circulation, thus revealing critical dependencies between cardiovascular and ocular health. The results highlight the model's potential for studying cardiovascular diseases with ocular manifestations and support emerging research in oculomics by providing a mechanistic basis to interpret ocular biomarkers within a systemic context. This paves the way for patient-specific data integration and broader applications in personalized medicine
Compressing image encoders via latent distillation
No full textDeep learning models for image compression often face practical limitations in hardware-constrained applications. Although these models achieve high-quality reconstructions, they are typically complex, heavyweight, and require substantial training data and computational resources. We propose a methodology to partially compress these networks by reducing the size of their encoders. Our approach uses a simplified knowledge distillation strategy to approximate the latent space of the original models with less data and shorter training, yielding lightweight encoders from heavyweight ones. We evaluate the resulting lightweight encoders across two different architectures on the image compression task. Experiments show that our method preserves reconstruction quality and statistical fidelity better than training lightweight encoders with the original loss, making it practical for resource-limited environments
Classification par tomodensitométrie des plaques carotidiennes symptomatiques et asymptomatiques à l'aide des caractéristiques du spectre de Schrödinger
No full textInternational audienceThis paper presents a novel methodology for classifying symptomatic and asymptomatic carotid artery plaques from CT images using quantum-inspired features. The proposed approach applies two-dimensional semi-classical signal analysis (2D-SCSA) to extract spectral features from each slice and subsequently constructs spatial sequences that track the evolution of these features across the entire volume. Statistical and frequency-domain descriptors computed from these spatial sequences capture three-dimensional morphological and textural characteristics of the plaques. Using stratified group K-fold cross-validation with hyperparameter tuning, multiple machine-learning models are trained on the extracted features. Experimental results on real clinical data confirm the effectiveness of this hierarchical feature-extraction strategy, demonstrating its strong potential to improve clinical diagnosis and risk assessment of carotid artery disease.Cet article présente une méthodologie novatrice pour la classification des plaques carotidiennes symptomatiques et asymptomatiques à partir d'images tomodensitométriques, grâce à l'utilisation de caractéristiques inspirées de la mécanique quantique. L'approche proposée applique une analyse de signal semi-classique bidimensionnelle (2D-SCSA) pour extraire les caractéristiques spectrales de chaque coupe, puis construit des séquences spatiales qui suivent l'évolution de ces caractéristiques dans l'ensemble du volume. Les descripteurs statistiques et fréquentiels calculés à partir de ces séquences spatiales capturent les caractéristiques morphologiques et texturales tridimensionnelles des plaques. À l'aide d'une validation croisée stratifiée à K plis avec optimisation des hyperparamètres, plusieurs modèles d'apprentissage automatique sont entraînés sur les caractéristiques extraites. Les résultats expérimentaux obtenus sur des données cliniques réelles confirment l'efficacité de cette stratégie hiérarchique d'extraction de caractéristiques, démontrant son fort potentiel pour améliorer le diagnostic clinique et l'évaluation du risque de maladie carotidienne
Assessing the Efficacy of Artefact Synthesis and Transfer Learning for Quality Control in Clinical 3D FLAIR Brain MRI
No full textInternational audienceThe recent advent of clinical data warehouses (CDWs) has greatly simplified the sharing of large medical datasets, including imaging, for research purposes. MRI scans can be affected by various artefacts such as motion, noise and low grey/white matter contrast, which can degrade image quality. Given the abundance of MRI scans available in CDWs, it is imperative to develop tools to automatically detect images corrupted by these artefacts, in order to guarantee the reliability of the data used for research. We have previously developed a method for detecting moderate artefacts in 3D T1-weighted brain MRI. We propose to test and extend our approach to the FLAIR sequence. Our approach involves pretraining models on research data augmented with synthetic artefacts, followed by a fine-tuning phase designed to adapt these models to routine clinical data, either keeping the clinical data unchanged, or balancing the classes by simulating artefacts on some of the artefact-free MRIs. This adaptation relies on the manual annotation of 637 FLAIR images. We evaluate whether this strategy can generalise to new MRI sequences and clinical settings where manual labels are limited. While leveraging synthetic artefacts and research datasets shows some benefit, the performance gains remain modest compared to models trained from scratch on clinical data, and the results fall short of human-level accuracy. A key limitation remains the scarcity of representative, annotated clinical images, which constrains overall performance
Guaranteed stability bounds for second-order PDE problems satisfying a Garding inequality
No full textWe propose an algorithm to numerically determined whether a second-order linear PDE problem satisfying a Gårding inequality is well-posed. This algorithm further provides a lower bound to the inf-sup constant of the weak formulation, which may in turn be used for a posteriori error estimation purposes. Our numerical lower bound is based on two discrete singular value problems involving a Lagrange finite element discretization coupled with an a posteriori error estimator based on flux reconstruction techniques. We show that if the finite element discretization is sufficiently rich, our lower bound underestimates the optimal constant only by a factor roughly equal to two
Redatuming via an iterative primal-dual TRAC approach
No full textIn inverse problems, redatuming data consists in virtually moving the sensors from the original acquisition location to an arbitrary position. This is an essential tool for target oriented inversion. This paper presents an exact redatuming method based on the Time Reversal Absorbing Conditions (TRAC) approach. Unlike conventional methods, our approach is built on a convergent iterative process instead of a least-squares formulation. Numerical results and comparisons with other approaches illustrate the efficiency and convergence of our method applied to the stationary Helmholtz equation
CP4SBI: Local Conformal Calibration of Credible Sets in Simulation-Based Inference
No full textInternational audienceCurrent experimental scientists have been increasingly relying on simulation-based inference (SBI) to invert complex non-linear models with intractable likelihoods. However, posterior approximations obtained with SBI are often miscalibrated, causing credible regions to undercover true parameters. We develop CP4SBI, a model-agnostic conformal calibration framework that constructs credible sets with local Bayesian coverage. Our two proposed variants, namely local calibration via regression trees and CDF-based calibration, enable finite-sample local coverage guarantees for any scoring function, including HPD, symmetric, and quantile-based regions. Experiments on widely used SBI benchmarks demonstrate that our approach improves the quality of uncertainty quantification for neural posterior estimators using both normalizing flows and score-diffusion modeling