INRIA a CCSD electronic archive server
Not a member yet
122212 research outputs found
Sort by
Evaluating Multichannel Speech Enhancement Algorithms at the Phoneme Scale Across Genders
No full textInternational audienceMultichannel speech enhancement algorithms are essential for improving the intelligibility of speech signals in noisy environments. These algorithms are usually evaluated at the utterance level, but this approach overlooks the disparities in acoustic characteristics that are observed in different phoneme categories and between male and female speakers. In this paper, we investigate the impact of gender and phonetic content on speech enhancement algorithms. We motivate this approach by outlining phoneme-and gender-specific spectral features. Our experiments reveal that while utterance-level differences between genders are minimal, significant variations emerge at the phoneme level. Results show that the tested algorithms better reduce interference with fewer artifacts on female speech, particularly in plosives, fricatives, and vowels. Additionally, they demonstrate greater performance for female speech in terms of perceptual and speech recognition metrics
Checking Linear Integer Arithmetic Proofs in Lambdapi
No full textInternational audienceModern SMT solvers can generate proofs of unsatisfiability so that the result can be checked independently. A dependable approach to verify these proofs is to reconstruct them within a proof assistant. In previous work, the SMT checker Carcara was extended to reconstruct SMT proofs in Lambdapi – a proof assistant designed for interoperability, supporting the import and export of proofs for integration with other proof assistants such as Rocq, Lean, or HOL-Light. Whereas that work was limited to SMT theories without arithmetic, we here present an extension that enables the reconstruction of SMT proofs involving linear integer arithmetic.</div
A Generalized Stochastic Formulation of the Ekman–Stokes Model with Statistical Analyses
No full textInternational audienceAbstract To describe the upper-ocean Ekman boundary layer, a novel stochastic model is formulated to better capture the interplay between random components associated with the numerically unresolved physical process induced by wind, waves, and currents. Accounting for the uncertainty of unresolved velocity fluctuations, resulting Ekman velocities display increased variability, higher kinetic energy, and more frequent extreme events than in an Ekman–Stokes model with time-correlated additive Gaussian forcing. Near the surface, stronger correlations and skewed probability distributions are also revealed. Sensitivity analyses highlight impacts of transient winds and surface waves, which deepen circulation and increase the dispersion of realizations in a balanced way according to error representations. Energy and transport are maximized when the mean directions of wind velocity and surface waves coincide. This model offers a new approach to describing upper-ocean dynamics, providing improved insights for anticipating vertical fluxes, and enhancing predictive capabilities. Significance Statement The upper-ocean boundary layer plays a critical role in regulating air–sea interactions, vertical mixing, and climate variability. Many existing boundary layer models employ deterministic turbulence closures that do not account for the inherently random nature of unresolved small-scale processes. This study presents a novel stochastic framework that explicitly incorporates the random effects of wind forcing, surface waves, and turbulent transport, while maintaining key physical balances. The results demonstrate enhanced variability and a higher frequency of extreme events, emphasizing the relevance of stochasticity in upper-ocean dynamics. This approach provides a physically consistent representation of boundary layer processes, with potential benefits for improving ocean modeling and forecasting of an ensemble of realizations at coarse resolution
Control of blow-up profiles for the mass-critical focusing nonlinear Schrödinger equation on bounded domains
No full textIn this paper, we consider the mass-critical focusing nonlinear Schrödinger on bounded two-dimensional domains with Dirichlet boundary conditions. In the absence of control, it is well-known that free solutions starting from initial data sufficiently large can blow-up. More precisely, given a finite number of points, there exists particular profiles blowing up exactly at these points at the blow-up time. For pertubations of these profiles, we show that, with the help of an appropriate nonlinear feedback law located in an open set containing the blow-up points, the blow-up can be prevented from happening. More specifically, we construct a small-time control acting just before the blow-up time. The solution may then be extended globally in time. This is the first result of control for blow-up profiles for nonlinear Schrödinger type equations. Assuming further a geometrical control condition on the support of the control, we are able to prove a null-controllability result for such blow-up profiles. Finally, we discuss possible extensions to three-dimensional domains
How Hard is it to Confuse a World Model?
No full textIn reinforcement learning (RL) theory, the concept of most confusing instances is central to establishing regret lower bounds, that is, the minimal exploration needed to solve a problem. Given a reference model and its optimal policy, a most confusing instance is the statistically closest alternative model that makes a suboptimal policy optimal. While this concept is well-studied in multi-armed bandits and ergodic tabular Markov decision processes, constructing such instances remains an open question in the general case. In this paper, we formalize this problem for neural network world models as a constrained optimization: finding a modified model that is statistically close to the reference one, while producing divergent performance between optimal and suboptimal policies. We propose an adversarial training procedure to solve this problem and conduct an empirical study across world models of varying quality. Our results suggest that the degree of achievable confusion correlates with uncertainty in the approximate model, which may inform theoretically-grounded exploration strategies for deep model-based RL.</div
Machine Learning Guided Equality Saturation
No full textEquality saturation has successfully been applied in many domains. Yet, scaling issues hold back its success in even more applications. The underlying e-graph data structure can grow rapidly quickly consuming all available resources. Guided Equality Saturation proposed a solution by breaking challenging rewrite problems into a sequence of equality saturations. This enables the technique to scale further and solve complex rewrite problems far out of reach of standard equality saturation. However, this technique relies on the human experts to provide insights in the form of guides that describe when to stop one equality saturation and start the next. In this talk, we are going to present our ongoing efforts to reduce the reliance on human experts. In our Machine Learning Guided Equality Saturation, the ambition is to automatically generate guides using a machine learning model to enable the scaling of Equality Saturation to more complex applications. We report on the current state of our research and the machine learning model we are developing
Symmetry reduction for testing -block-positivity via extendibility
No full textInternational audienceWe study the problem of testing -block-positivity via symmetric -extendibility by taking the tensor product with a -dimensional maximally entangled state. We exploit the unitary symmetry of the maximally entangled state to reduce the size of the corresponding semidefinite program (SDP). For example, for , the SDP is reduced from one block of size to blocks of size
A typology of quantum algorithms
No full textWe draw the current landscape of quantum algorithms, by classifying about 130 quantum algorithms, according to the fundamental mathematical problems they solve, their real-world applications, the main subroutines they employ, and several other relevant criteria. The primary objectives include revealing trends of algorithms, identifying promising fields for implementations in the NISQ era, and identifying the key algorithmic primitives that power quantum advantage
Unsupervised anomaly detection using Bayesian flow networks: application to brain FDG pet in the context of Alzheimer’s disease
No full textInternational audienceUnsupervised anomaly detection (UAD) plays a crucial role in neuroimaging for identifying deviations from healthy subject data and thus facilitating the diagnosis of neurological disorders. In this work, we focus on Bayesian flow networks (BFNs), a novel class of generative models, which have not yet been applied to medical imaging or anomaly detection. BFNs combine the strength of diffusion frameworks and Bayesian inference. We introduce AnoBFN, an extension of BFNs for UAD, designed to: i) perform conditional image generation under high levels of spatially correlated noise, and ii) preserve subject specificity by incorporating a recursive feedback from the input image throughout the generative process. We evaluate AnoBFN on the challenging task of Alzheimer's disease-related anomaly detection in FDG PET images. Our approach outperforms other state-of-the-art methods based on VAEs (β-VAE), GANs (f-AnoGAN), and diffusion models (AnoDDPM), demonstrating its effectiveness at detecting anomalies while reducing false positive rates
