Thomas George

Thomas George, Guillaume Lajoie, Aristide Baratin

Among attempts at giving a theoretical account of the success of deep neural networks, a recent line of work has identified a so-called lazy training regime in which the network can be well approximated by its linearization around initialization. Here we investigate the comparative effect of the lazy (linear) and feature learning (non-linear) regimes on subgroups of examples based on their difficulty. Specifically, we show that easier examples are given more weight in feature learning mode, resulting in faster training compared to more difficult ones. In other words, the non-linear dynamics tends to sequentialize the learning of examples of increasing difficulty. We illustrate this phenomenon across different ways to quantify example difficulty, including c-score, label noise, and in the presence of easy-to-learn spurious correlations. Our results reveal a new understanding of how deep networks prioritize resources across example difficulty.

Ilies Chibane, Thomas George, Pierre Nodet et al.

Mislabeled data is a pervasive issue that undermines the performance of machine learning systems in real-world applications. An effective approach to mitigate this problem is to detect mislabeled instances and subject them to special treatment, such as filtering or relabeling. Automatic mislabeling detection methods typically rely on training a base machine learning model and then probing it for each instance to obtain a trust score that each provided label is genuine or incorrect. The properties of this base model are thus of paramount importance. In this paper, we investigate the impact of calibrating this model. Our empirical results show that using calibration methods improves the accuracy and robustness of mislabeled instance detection, providing a practical and effective solution for industrial applications.

Jeanne Monnier, Thomas George, Frédéric Guyard et al.

Fairness in machine learning remains challenging due to its ethical complexity, the absence of a universal definition, and the need for context-specific bias metrics. Existing methods still struggle with intersectionality, multiclass settings, and limited flexibility and generality. To address these gaps, we introduce MIFair, a unified framework for bias assessment and mitigation based on mutual information. MIFair provides a flexible metric template and an in-processing mitigation method inspired by the Prejudice Remover, defining group fairness as statistical independence between prediction-derived variables and sensitive attributes. We further strengthen its information-theoretic foundation by establishing equivalences with widely used fairness notions such as independence and separation. MIFair naturally supports intersectionality, complex subgroup structures, and multiclass classification and employs regularization-based training to reduce bias according to the selected metric. Its key advantage is its versatility: it consolidates diverse fairness requirements into a single coherent framework, enabling consistent benchmarking and simplifying practical use. Experiments on real-world tabular and image datasets show that MIFair effectively reduces bias, including previously unaddressed multi-attribute scenarios, while maintaining strong predictive performance across the evaluated settings.

Thomas George, Pierre Nodet, Alexis Bondu et al.

Mislabeled examples are ubiquitous in real-world machine learning datasets, advocating the development of techniques for automatic detection. We show that most mislabeled detection methods can be viewed as probing trained machine learning models using a few core principles. We formalize a modular framework that encompasses these methods, parameterized by only 4 building blocks, as well as a Python library that demonstrates that these principles can actually be implemented. The focus is on classifier-agnostic concepts, with an emphasis on adapting methods developed for deep learning models to non-deep classifiers for tabular data. We benchmark existing methods on (artificial) Completely At Random (NCAR) as well as (realistic) Not At Random (NNAR) labeling noise from a variety of tasks with imperfect labeling rules. This benchmark provides new insights as well as limitations of existing methods in this setup.

Aziz Bacha, Thomas George

Deep neural networks are often considered opaque systems, prompting the need for explainability methods to improve trust and accountability. Existing approaches typically attribute test-time predictions either to input features (e.g., pixels in an image) or to influential training examples. We argue that both perspectives should be studied jointly. This work explores *training feature attribution*, which links test predictions to specific regions of specific training images and thereby provides new insights into the inner workings of deep models. Our experiments on vision datasets show that training feature attribution yields fine-grained, test-specific explanations: it identifies harmful examples that drive misclassifications and reveals spurious correlations, such as patch-based shortcuts, that conventional attribution methods fail to expose.

Irene Li, Thomas George, Alexander Fabbri et al.

Effective human learning depends on a wide selection of educational materials that align with the learner's current understanding of the topic. While the Internet has revolutionized human learning or education, a substantial resource accessibility barrier still exists. Namely, the excess of online information can make it challenging to navigate and discover high-quality learning materials. In this paper, we propose the educational resource discovery (ERD) pipeline that automates web resource discovery for novel domains. The pipeline consists of three main steps: data collection, feature extraction, and resource classification. We start with a known source domain and conduct resource discovery on two unseen target domains via transfer learning. We first collect frequent queries from a set of seed documents and search on the web to obtain candidate resources, such as lecture slides and introductory blog posts. Then we introduce a novel pretrained information retrieval deep neural network model, query-document masked language modeling (QD-MLM), to extract deep features of these candidate resources. We apply a tree-based classifier to decide whether the candidate is a positive learning resource. The pipeline achieves F1 scores of 0.94 and 0.82 when evaluated on two similar but novel target domains. Finally, we demonstrate how this pipeline can benefit an application: leading paragraph generation for surveys. This is the first study that considers various web resources for survey generation, to the best of our knowledge. We also release a corpus of 39,728 manually labeled web resources and 659 queries from NLP, Computer Vision (CV), and Statistics (STATS).

Swapnil Hingmire, Irene Li, Rena Kawamura et al.

A classification scheme of a scientific subject gives an overview of its body of knowledge. It can also be used to facilitate access to research articles and other materials related to the subject. For example, the ACM Computing Classification System (CCS) is used in the ACM Digital Library search interface and also for indexing computer science papers. We observed that a comprehensive classification system like CCS or Mathematics Subject Classification (MSC) does not exist for Computational Linguistics (CL) and Natural Language Processing (NLP). We propose a classification scheme -- CLICKER for CL/NLP based on the analysis of online lectures from 77 university courses on this subject. The currently proposed taxonomy includes 334 topics and focuses on educational aspects of CL/NLP; it is based primarily, but not exclusively, on lecture notes from NLP courses. We discuss how such a taxonomy can help in various real-world applications, including tutoring platforms, resource retrieval, resource recommendation, prerequisite chain learning, and survey generation.

Timothée Lesort, Thomas George, Irina Rish

We study how different output layer parameterizations of a deep neural network affects learning and forgetting in continual learning settings. The following three effects can cause catastrophic forgetting in the output layer: (1) weights modifications, (2) interference, and (3) projection drift. In this paper, our goal is to provide more insights into how changing the output layer parameterization may address (1) and (2). Some potential solutions to those issues are proposed and evaluated here in several continual learning scenarios. We show that the best-performing type of output layer depends on the data distribution drifts and/or the amount of data available. In particular, in some cases where a standard linear layer would fail, changing parameterization is sufficient to achieve a significantly better performance, without introducing any continual-learning algorithm but instead by using standard SGD to train a model. Our analysis and results shed light on the dynamics of the output layer in continual learning scenarios and suggest a way of selecting the best type of output layer for a given scenario.

Aristide Baratin, Thomas George, César Laurent et al.

We approach the problem of implicit regularization in deep learning from a geometrical viewpoint. We highlight a regularization effect induced by a dynamical alignment of the neural tangent features introduced by Jacot et al, along a small number of task-relevant directions. This can be interpreted as a combined mechanism of feature selection and compression. By extrapolating a new analysis of Rademacher complexity bounds for linear models, we motivate and study a heuristic complexity measure that captures this phenomenon, in terms of sequences of tangent kernel classes along optimization paths.

César Laurent, Camille Ballas, Thomas George et al.

By removing parameters from deep neural networks, unstructured pruning methods aim at cutting down memory footprint and computational cost, while maintaining prediction accuracy. In order to tackle this otherwise intractable problem, many of these methods model the loss landscape using first or second order Taylor expansions to identify which parameters can be discarded. We revisit loss modelling for unstructured pruning: we show the importance of ensuring locality of the pruning steps. We systematically compare first and second order Taylor expansions and empirically show that both can reach similar levels of performance. Finally, we show that better preserving the original network function does not necessarily transfer to better performing networks after fine-tuning, suggesting that only considering the impact of pruning on the loss might not be a sufficient objective to design good pruning criteria.

Thomas George, César Laurent, Xavier Bouthillier et al.

Optimization algorithms that leverage gradient covariance information, such as variants of natural gradient descent (Amari, 1998), offer the prospect of yielding more effective descent directions. For models with many parameters, the covariance matrix they are based on becomes gigantic, making them inapplicable in their original form. This has motivated research into both simple diagonal approximations and more sophisticated factored approximations such as KFAC (Heskes, 2000; Martens & Grosse, 2015; Grosse & Martens, 2016). In the present work we draw inspiration from both to propose a novel approximation that is provably better than KFAC and amendable to cheap partial updates. It consists in tracking a diagonal variance, not in parameter coordinates, but in a Kronecker-factored eigenbasis, in which the diagonal approximation is likely to be more effective. Experiments show improvements over KFAC in optimization speed for several deep network architectures.