Institution profile

Universite Grenoble Alpes

Academic institutioneurope · fr
Official website
Research library205linked papers
Opportunities0open roles
Selected work

Representative Papers

Non-Stationary Functional Bilevel Optimization

Jan 21, 2026

Existing function-space bilevel optimization methods struggle to handle online non-stationary environments. This work proposes SmoothFBO, the first algorithm to extend function-space bilevel optimization to such settings. By employing a time-smoothed stochastic hypergradient estimator, SmoothFBO reduces variance in gradient estimates, enabling stable outer-loop updates and achieving sublinear regret. The method offers strong theoretical guarantees and scalability, while naturally encompassing classical parametric bilevel optimization as a special case. Empirical evaluations on non-stationary hyperparameter optimization and model-based reinforcement learning tasks demonstrate that SmoothFBO significantly outperforms existing approaches, confirming its effectiveness and broad applicability.

2 citationsRead paper

A 1Mb Mixed-Precision Quantized Encoder for Image Classification and Patch-Based Compression

Aug 01, 2022IEEE transactions on circuits and systems for video technology (Print)

To address the demand for high-energy-efficiency, low-resource hardware accelerators for edge-based image classification and compression, this work proposes a reconfigurable ASIC neural network encoder consuming only 1 MB of on-chip resources and supporting mixed-precision computation (3-bit/2-bit/1-bit). Methodologically, it introduces a novel linear symmetric quantization with adaptive scaling factors to ensure training stability under ultra-low-bit weight quantization, and replaces batch normalization with layer-shared shift-based normalization to drastically reduce hardware overhead. Furthermore, it unifies classification and compression acceleration via structural pruning, block-wise encoding, and remote full-frame decoding. Evaluated on CIFAR-10, the encoder achieves 87.5% classification accuracy and—uniquely among learned image codecs—enables end-to-end, block-artifact-free image compression. Its fixed-bitrate performance significantly surpasses state-of-the-art block-based compression methods.

2 citationsRead paper

Comparative Analysis of Ray Tracing and Rayleigh Fading Models for Distributed MIMO Systems in Industrial Environments

Mar 03, 2025

This work addresses channel modeling and performance evaluation of distributed MIMO (D-MIMO) in industrial environments. We systematically compare, for the first time, deterministic ray-tracing models against stochastic Rayleigh fading models in predicting downlink/uplink single-user capacity. Leveraging a real-world 3D factory map, we construct multiple deployment scenarios to quantify how network densification affects user equipment (UE) multi-access point (AP) connectivity and coverage gain. Results show that densification significantly enhances D-MIMO capacity. Ray tracing more accurately captures spatial correlation and realistic propagation characteristics, whereas the Rayleigh model offers superior computational efficiency and maintains acceptable prediction error (<15%) in typical factory settings. The study establishes fundamental trade-offs among modeling accuracy, spatial correlation fidelity, and computational overhead, providing both theoretical guidance and empirical evidence for selecting appropriate D-MIMO channel models in industrial wireless systems.

1 citations1 influentialRead paper

Diffusion-based Frameworks for Unsupervised Speech Enhancement

Jan 14, 2026

This work addresses the challenge of unsupervised single-channel speech enhancement in the absence of paired noisy-clean speech data by proposing a novel diffusion-based framework. The method explicitly models noise as a latent variable under an unsupervised setting and integrates an expectation-maximization (EM) algorithm, where the E-step jointly samples latent representations of both speech and noise. Crucially, it replaces conventional NMF-based noise models with a diffusion prior to construct a conditional diffusion generative mechanism. Experimental results demonstrate significant improvements in speech quality and intelligibility on the WSJ0-QUT and VoiceBank-DEMAND datasets: under matched conditions, the diffusion-based noise model achieves optimal performance, while under mismatched conditions, its explicit NMF modeling variant even surpasses several supervised baselines.

1 citationsRead paper
Recent publications

Latest Papers