Eva Feillet · Assistant Professor · Université Paris-Saclay · LISN

Adaptive and resource-efficient machine learning

I study learning systems that must adapt to evolving data while operating under constraints on memory, data, and computation. My work connects continual learning, representation learning, computer vision, and deep learning architectures. I am an Assistant Professor (Maîtresse de conférences en informatique) at Université Paris-Saclay, a member of the LISN laboratory, and I teach at the UFR des Sciences d'Orsay.

Research programme Selected publications Collaboration

Google Scholar GitHub ORCID CV

Research programme

Adaptive and resource-efficient machine learning

My research combines continual learning and efficient representation architectures, with applications across vision, speech, language, and scientific machine learning.

Continual and class-incremental learning

Approach

Analyse and recommend class-incremental learning methods, taking into account model initialization, representation transfer, and memory constraints.

Research question

How should an incremental learner update its representation when earlier examples cannot be retained and the future data stream not known yet? How to evaluate models when the relation between classes, and visual appearance evolves over time?

Selected publications

Reality Check on Pre-training
Temporal Dynamics in Visual Data
AdvisIL

Efficient representation architectures

Approach

Study alternatives to standard attention and mixing mechanisms, including polynomial token mixing, multipole attention, and explicit mode mixing in Fourier neural operators.

Research question

Which architectural mechanisms preserve useful global interactions while reducing memory or computational costs in speech, vision, and operator learning?

Selected publications

Polynomial Mixing for Speech Encoders
Multipole Attention

Adaptive learning across modalities

Perspectives

I am also interested in efficient vision-language architectures, and in connecting continual learning with multimodal systems, particularly vision–language models and speech-to-text models.

Ongoing work

Current projects

News! Check-out our pre-print

Preprint, 2026 · Under submission

Higher-Order Fourier Neural Operator: Explicit Mode Mixer for Nonlinear PDEs

Colagrande, A., Caillon, P., Feillet, E., Allauzen, A.

arXiv · Code

Current supervision

Small and adaptive language models

Luc Pommeret is starting a PhD in autumn 2026 on efficient adaptation and continual evaluation of small language models.

Efficient methods for scientific machine learning

How to design deep learning architectures for accelerating and reducing the cost of numerical simulations? See the great work of Alex Colagrande on Neural Operators for approximating PDE solutions!

Parameter-efficient spoken-language understanding

Ongoing project with Sahar Ghannay on efficient adaptation of spoken-language understanding models, with a focus on low-resource languages.

Selected work, by topic

Selected publications

Efficient architectures

ICASSP 2026

Polynomial Mixing for Efficient Self-supervised Speech Encoders

We propose a polynomial token mixer as a drop-in alternative to multi-head self-attention in a self-supervised speech encoder. The proposed mixer offers a trade-off between recognition performance, runtime, and memory use.

Paper · Code

ECLR Workshop at ICCV 2025

Linear Attention with Global Context: A Multipole Attention Mechanism for Vision and Physics

We propose MANO, a linear-attention mechanism intended to retain global contextual interactions while limiting the cost associated with standard attention. MANO shows competitive performance in image classification and physics-oriented benchmarks.

Paper · Code

Continual learning

WACV 2025

A Reality Check on Pre-training for Exemplar-free Class-Incremental Learning

This study compares several initialization and pre-training strategies across exemplar-free class-incremental learning algorithms and datasets. We analyse how the usefulness of pre-training depends on domain alignment and provide practical recommendations for designing continual learning systems.

Paper · Supplementary material

WACV 2023

AdvisIL — A Class-Incremental Learning Advisor

AdvisIL addresses algorithm recommendation for class-incremental learning. AdvisIL recommendation is based on the characteristics of an incremental scenario to support the choice of a suitable learning method and deep architecture.

Paper · Supplementary material · Code

Robustness to changing data

WACV 2025

Temporal Dynamics in Visual Data: Analyzing the Impact of Time on Classification Accuracy

We introduce VCT-107, a visual dataset organised by collection period, and analyse how classification accuracy changes when training and test data come from different periods. The results motivate explicit treatment of temporal shift and regular model updating.

Paper · Supplementary material · Data

ICPR 2024

Recommendation of Data-free Class-Incremental Learning Algorithms by Simulating Future Data

This work studies the recommendation of data-free class-incremental algorithms by constructing simulations of future data. It allows selecting a learning strategy based on assumptions about the semantic content of the future data.

Paper · Supplementary material

Background

Biography

Eva Feillet is an Assistant Professor at Université Paris-Saclay, a member of the LISN laboratory, and teaches at the UFR des Sciences d’Orsay. Previously, she was a postdoctoral researcher in frugal deep learning at Université Paris-Dauphine, where she worked with Alexandre Allauzen in the MILES team at the LAMSADE laboratory. She participated in the SHARP project, part of the PEPR IA programme.

She holds a PhD from Université Paris-Saclay on deep continual learning for image classification. Her doctoral work was supervised by Céline Hudelot at CentraleSupélec MICS, and by Adrian Popescu and Marina Reyboz at CEA List. She also holds an engineering degree from CentraleSupélec (equivalent to an M.Sc. in AI).

Contact

Research collaborations and student projects

I am open to discussions related to advancing research on adaptive learning, efficient representation architectures, or multimodal learning. I am also interested in applying my expertise to practical challenges. Please contact me to discuss potential collaborations.

Email: name [dot] surname [at] universite-p#ris-s#clay.fr