HAL-CentraleSupelec
Not a member yet
77624 research outputs found
Sort by
Assessing the Difficulty of Inference Types in Natural Language Inference for Clinical Trials
No full textLarge Language Models (LLMs) achieve competitive results on Natural Language Inference when applied to clinical trials; however, it is not yet clear on which type of inference LLMs perform well or poorly. We address this by proposing new supplementary annotations to the existing NLI4CT dataset on the types of inferences observed in clinical trials. Our dataset supplements NLI4CT with a total of 1,851 new annotations using our carefully crafted guidelines for 17 types of inferences. To investigate how the inference types impact the performance of LLMs, we prompt Flan-T5, Llama, Mistral, and Qwen and investigate their performance using our newly annotated dataset. We found that logical inferences have a negative impact on Mixtral, Qwen-7B, and Qwen-14B's overall performance, while numerical inferences have a negative impact on Flan-T5-XL and Mixtral. Further analysis shows that understanding the CTR's structure by itself remains challenging for MMed-Llama-3. Other parameters, such as the number of inference types involved or the type of section used in the premise, also impact models' performance. Our code and dataset are publicly available on GitHub
Apprentissage basé sur les modèles pour les systèmes sans fil de grande dimension
No full textThe field of wireless communications has traditionally relied on signal processing techniques, based on mathematical models, to describe and optimize wireless transmission of information. When the assumptions underlying these models are not fully met, conventional methods often yield suboptimal performance. Recently, machine learning methods have been used to address this issue, owing to their strong adaptation capabilities. However, these methods also introduce significant challenges in terms of interpretability and complexity.This thesis focuses on the use of the model-based machine learning paradigm in wireless communication or sensing systems. This paradigm leverages mathematical models used in classical signal processing approaches to structure, initialize, or optimize learning methods, thereby combining the structure and interpretability of classical methods with the adaptability of machine learning. This paradigm is studied through three specific use cases: location-to-channel mapping learning, hardware impairments compensation, and channel compression.Le domaine des communications sans fil s'est traditionnellement appuyé sur des techniques de traitement du signal, basées sur des modèles mathématiques, pour décrire et optimiser la transmission sans fil de l'information. Lorsque les hypothèses sous-jacentes à ces modèles ne sont pas pleinement satisfaites, les méthodes conventionnelles présentent souvent des performances sous-optimales. Récemment, des approches basées sur l'apprentissage automatique ont été utilisées pour traiter ce problème en raison de leurs fortes capacités d'adaptation. Néanmoins, ces méthodes introduisent également des problématiques importantes, notamment en termes d'interprétabilité et de complexité.Cette thèse se concentre sur l'utilisation du paradigme de l'apprentissage automatique basé sur les modèles, dans les systèmes sans fil de communication ou de détection. Ce paradigme s'appuie sur les modèles mathématiques utilisés dans les approches conventionnelles de traitement du signal pour structurer, initialiser ou optimiser des méthodes d'apprentissage, combinant ainsi la structure et l'interprétabilité des méthodes conventionnelles, à l'adaptabilité de l'apprentissage automatique. Ce paradigme est étudié à travers trois cas d'usage spécifiques : l'apprentissage de la fonction position-canal, la compensation des imperfections matérielles et la compression de canal
Grands modèles de langue pour la détection de pathologies psychiatriques : promesses, réalité, et enjeux
No full textInternational audienceLes troubles de la santé mentale concernent une part croissante de la population. Face aux difficultés rencontrées par nos systèmes de santé, les technologies de TAL ont pu être présentées comme des alternatives prometteuses, en particulier pour l'aide au diagnostic. Pour comprendre la réalité des contributions apportées, nous avons effectué une revue de la littérature sur l'utilisation de LLM pour la détection de pathologies psychiatriques. Les résultats obtenus permettent d'analyser l'adéquation entre promesses mobilisées et état actuel de la recherche. Ceci nous amène finalement à évoquer les enjeux soulevés par l'intégration de LLM dans le domaine de la psychiatrie.</div
Scalable Reliability Assessment of DNNs through Simultaneous Fault Injection
No full textInternational audienceDeep Neural Networks (DNNs) are being deployed in safety-critical applications, where resilience to transient faults is essential. Traditional fault injection methods often face challenges in scaling efficiently to larger models, whereas the majority of existing speedup techniques is closely linked to specific hardware or architectural configurations. To speed up the assessment of single-fault effects, an approach based on simultaneous injection of faults has demonstrated promising results, where multiple non-interacting faults are injected concurrently during a single workload execution. Nevertheless, the applicability of this method to DNNs has not been explored. In this study, we investigate the use of simultaneous injection of faults in DNNs and observe that faults can easily interact with one another due to DNNs' densely connected structure. These fault interactions can create "artificial" masking effects, leading to the misclassification of faults as non-critical (called false negatives), ultimately compromising the accuracy of the reliability assessment. To overcome this phenomenon, we propose an approach to mitigate the effects of such fault interaction during simultaneous injection of faults in DNNs, ensuring accurate assessment. Furthermore, we propose a strategy to further accelerate the assessment by pruning non-critical inputs from the DNN input batch during fault injection, further improving the speedup with negligible accuracy loss. To our knowledge, this is the first approach to enable accurate and efficient simultaneous injection of faults into DNNs, supporting fast reliability assessment applicable to different abstraction levels. We experiment with nearly 42 million injections at both software (SW) and RTL, achieving very low false negatives (as low as 0%, avg 0.2%) and an average injection time gain of 3.82× (RTL) and 5.29× (SW) over existing DNN fault injection approaches
Compressed Consecutive Pattern Matching
No full textInternational audienceOriginating from the work of Navarro and Thankanchan [TCS 2016], the problem of consecutive pattern matching is a variant of the fundamental pattern matching problem, where one is given a text and a pair of patterns , and must compute consecutive occurrences of in the text. Assuming that the text is given as a straight-line program of size , we develop an algorithm that computes all consecutive occurrences of in optimal time. As a corollary, we also derive an algorithm that reports all co-occurrences separated by a distance in O(g+|p_1|+|p_2|+\occ) time and an algorithm that reports the top- closest co-occurrences in time
Predicting large scale cosmological structure evolution with GAN-based autoencoders
No full textInternational audienceCosmological simulations play a key role in the prediction and understanding of large scale structure formation from initial conditions. We make use of GAN-based Autoencoders (AEs) in an attempt to predict structure evolution within simulations. The AEs are trained on images and cubes issued from respectively 2D and 3D N-body simulations describing the evolution of the dark matter (DM) field. We find that while the AEs can predict structure evolution for 2D simulations of DM fields well, using only the density fields as input, they perform significantly more poorly in similar conditions for 3D simulations. However, additionally providing velocity fields as inputs greatly improves results, with similar predictions regardless of time-difference between input and target
A solution to the mystery of the sub-harmonic series via a linear model of the cochlea
No full textIn this paper, we study a simple linear model of the cochlea as a set of vibrating strings. We make hypothesis that the information sent to the auditory cortex is the energy stored in the strings and consider all oscillation modes of the strings. We show the emergence of the sub-harmonic series whose existence was hypothesized in the XVI century to explain the consonance of the minor chord. We additionally show how the nonlinearity of the energy can be used to study the emergence of the combination tone (Tartini’s third sound) shedding new light on this long debated subject
Fast and General Automatic Differentiation for Finite-State Methods
No full textWe propose a new method, that we coined the “morphism-trick”, to integrate custom implementations of vector-Jacobian products in automatic differentiation softwares, applicable to a wide range of semiring-based computations. Our approach leads to efficient and semiring-agnostic implementations of the backward pass of dynamic programming algorithms. For the particular case of finite-state methods, we introduce an algorithm that computes and differentiates the ⊕-sum of all paths’ weight of a finite-state automaton. Results show that, with minimal effort from the user, our novel library allows computing the gradient of a function w.r.t. to the weights of a finite state automaton orders of magnitude faster than state-of-the-art automatic differentiation systems. Implementations are made available via an open-source library distributed under a permissive license
QoS-Aware Approximate Task Mapping on Heterogeneous Multicore Platforms with DVFS and Task Migration
No full textInternational audienceHeterogeneous Multicore Platforms (HMPs) have been widely adopted to execute tasks across a range of applications. Under limited system resources and diverse application requirements, allocating and executing dependent Approximate Computing (AC) tasks on these platforms to achieve high Quality-of-Service (QoS) is challenging. Dynamic Voltage and Frequency Scaling (DVFS) and task migration have proven effective for improving QoS while balancing time and energy consumption. However, existing approaches often overlook the migration overhead and the resulting dynamic changes in task dependencies, which can adversely affect mapping outcomes. To address these issues, this paper presents a novel AC task mapping method that maximizes system QoS under multiple constraints on HMPs, accounting for task migration overhead, DVFS, and changes in Directed Acyclic Graph (DAG) topology. We first formulate this joint design problem as a complex nonlinear programming problem. Next, we linearize the nonlinear terms without performance loss by introducing auxiliary variables and additional constraints. Building on this formulation, we propose an optimal (OPT) and a low-complexity Heuristic Algorithm (HEU), derived from problem decomposition and a greedy strategy, which divides the Mixed-Integer Non-Linear Programming (MINLP) problem into two smaller subproblems with fewer variables and constraints, solving them sequentially. The simulation results show that the proposed OPT method achieves higher QoS performance, measured at about 2.389 times on average and up to 4.115 times, while its feasibility is increased to about 3.263 times on average and up to 9.667 times, compared to other state-of-the-art methods. In addition, the average QoS of the proposed HEU method is about 0.577 times that of the proposed method, but its computation time is over a thousand times shorter
Multi-Taxonomy Vulnerability Classification with Hierarchically Finetuned Language Models
No full textInternational audienceThe number of newly disclosed vulnerabilities continues to increase rapidly each year. When disclosed, vulnerability entries typically include an initial free-form textual description that provides a basic overview of the vulnerability. Mapping these descriptions to structured taxonomies such as MITRE ATT&CK, CWE, and CAPEC enhances contextual understanding and supports prioritization from the moment of disclosure, but manual mapping is complex and time-consuming. We present a unified framework for automated multi-label vulnerability classification with Language Models (LMs), implemented in our open-source library CVE-LMTune. It combines: (1) a multi-stage pipeline for building an up-to-date, multi-taxonomy labeled dataset; (2) a standardized finetuning and evaluation protocol addressing extreme multi-label imbalance; and (3) a hierarchical cascade that decomposes large class spaces of these frameworks into smaller, tractable sub-problems. Our experiments indicate that fine-tuned encoder-only LMs outperform text-generation models, which struggle in handling this high-cardinality, imbalanced task in the token space. Applying our hierarchical approach to the best model, SecureBERT, consistently improves weighted-F1 over the flat version: +12% on CWE (90%), +8% on CAPEC (92%), and +12% on MITRE ATT&CK (93%). We further propose a shared-embedding strategy that cuts hierarchical inference costs toward flat-model efficiency and demonstrate its robustness on newly disclosed vulnerabilities