Céline Hudelot

Manuel Faysse, Gautier Viaud, Céline Hudelot et al. · meta-ai

Instruction Fine-Tuning (IFT) is a powerful paradigm that strengthens the zero-shot capabilities of Large Language Models (LLMs), but in doing so induces new evaluation metric requirements. We show LLM-based metrics to be well adapted to these requirements, and leverage them to conduct an investigation of task-specialization strategies, quantifying the trade-offs that emerge in practical industrial settings. Our findings offer practitioners actionable insights for real-world IFT model deployment.

Malik Boudiaf, Etienne Bennequin, Myriam Tami et al.

We tackle the Few-Shot Open-Set Recognition (FSOSR) problem, i.e. classifying instances among a set of classes for which we only have a few labeled samples, while simultaneously detecting instances that do not belong to any known class. We explore the popular transductive setting, which leverages the unlabelled query instances at inference. Motivated by the observation that existing transductive methods perform poorly in open-set scenarios, we propose a generalization of the maximum likelihood principle, in which latent scores down-weighing the influence of potential outliers are introduced alongside the usual parametric model. Our formulation embeds supervision constraints from the support set and additional penalties discouraging overconfident predictions on the query set. We proceed with a block-coordinate descent, with the latent scores and parametric model co-optimized alternately, thereby benefiting from each other. We call our resulting formulation \textit{Open-Set Likelihood Optimization} (OSLO). OSLO is interpretable and fully modular; it can be applied on top of any pre-trained model seamlessly. Through extensive experiments, we show that our method surpasses existing inductive and transductive methods on both aspects of open-set recognition, namely inlier classification and outlier detection.

Grégoire Petit, Michael Soumm, Eva Feillet et al.

Class-Incremental Learning (CIL) aims to build classification models from data streams. At each step of the CIL process, new classes must be integrated into the model. Due to catastrophic forgetting, CIL is particularly challenging when examples from past classes cannot be stored, the case on which we focus here. To date, most approaches are based exclusively on the target dataset of the CIL process. However, the use of models pre-trained in a self-supervised way on large amounts of data has recently gained momentum. The initial model of the CIL process may only use the first batch of the target dataset, or also use pre-trained weights obtained on an auxiliary dataset. The choice between these two initial learning strategies can significantly influence the performance of the incremental learning model, but has not yet been studied in depth. Performance is also influenced by the choice of the CIL algorithm, the neural architecture, the nature of the target task, the distribution of classes in the stream and the number of examples available for learning. We conduct a comprehensive experimental study to assess the roles of these factors. We present a statistical analysis framework that quantifies the relative contribution of each factor to incremental performance. Our main finding is that the initial training strategy is the dominant factor influencing the average incremental accuracy, but that the choice of CIL algorithm is more important in preventing forgetting. Based on this analysis, we propose practical recommendations for choosing the right initial training strategy for a given incremental learning use case. These recommendations are intended to facilitate the practical deployment of incremental learning.

Jun Zhu, Céline Hudelot

Works on learning job title representation are mainly based on \textit{Job-Transition Graph}, built from the working history of talents. However, since these records are usually messy, this graph is very sparse, which affects the quality of the learned representation and hinders further analysis. To address this specific issue, we propose to enrich the graph with additional nodes that improve the quality of job title representation. Specifically, we construct \textit{Job-Transition-Tag Graph}, a heterogeneous graph containing two types of nodes, i.e., job titles and tags (i.e., words related to job responsibilities or functionalities). Along this line, we reformulate job title representation learning as the task of learning node embedding on the \textit{Job-Transition-Tag Graph}. Experiments on two datasets show the interest of our approach.

Hippolyte Gisserot-Boukhlef, Ricardo Rei, Emmanuel Malherbe et al.

Neural metrics for machine translation (MT) evaluation have become increasingly prominent due to their superior correlation with human judgments compared to traditional lexical metrics. Researchers have therefore utilized neural metrics through quality-informed decoding strategies, achieving better results than likelihood-based methods. With the rise of Large Language Models (LLMs), preference-based alignment techniques have gained attention for their potential to enhance translation quality by optimizing model weights directly on preferences induced by quality estimators. This study focuses on Contrastive Preference Optimization (CPO) and conducts extensive experiments to evaluate the impact of preference-based alignment on translation quality. Our findings indicate that while CPO consistently outperforms Supervised Fine-Tuning (SFT) on high-quality data with regard to the alignment metric, it may lead to instability across downstream evaluation metrics, particularly between neural and lexical ones. Additionally, we demonstrate that relying solely on the base model for generating candidate translations achieves performance comparable to using multiple external systems, while ensuring better consistency across downstream metrics.

Thomas Cordier, Victor Bouvier, Gilles Hénaff et al.

Machine Learning models are prone to fail when test data are different from training data, a situation often encountered in real applications known as distribution shift. While still valid, the training-time knowledge becomes less effective, requiring a test-time adaptation to maintain high performance. Following approaches that assume batch-norm layer and use their statistics for adaptation, we propose a Test-Time Adaptation with Principal Component Analysis (TTAwPCA), which presumes a fitted PCA and adapts at test time a spectral filter based on the singular values of the PCA for robustness to corruptions. TTAwPCA combines three components: the output of a given layer is decomposed using a Principal Component Analysis (PCA), filtered by a penalization of its singular values, and reconstructed with the PCA inverse transform. This generic enhancement adds fewer parameters than current methods. Experiments on CIFAR-10-C and CIFAR- 100-C demonstrate the effectiveness and limits of our method using a unique filter of 2000 parameters.

Antonin Poché, Thomas Mullor, Gabriele Sarti et al.

Interpreto is a Python library for post-hoc explainability of text HuggingFace models, from early BERT variants to LLMs. It provides two complementary families of methods: attributions and concept-based explanations. The library connects recent research to practical tooling for data scientists, aiming to make explanations accessible to end users. It includes documentation, examples, and tutorials. Interpreto supports both classification and generation models through a unified API. A key differentiator is its concept-based functionality, which goes beyond feature-level attributions and is uncommon in existing libraries. The library is open source; install via pip install interpreto. Code and documentation are available at https://github.com/FOR-sight-ai/interpreto.

Manuel Faysse, Patrick Fernandes, Nuno M. Guerreiro et al. · meta-ai

We introduce CroissantLLM, a 1.3B language model pretrained on a set of 3T English and French tokens, to bring to the research and industrial community a high-performance, fully open-sourced bilingual model that runs swiftly on consumer-grade local hardware. To that end, we pioneer the approach of training an intrinsically bilingual model with a 1:1 English-to-French pretraining data ratio, a custom tokenizer, and bilingual finetuning datasets. We release the training dataset, notably containing a French split with manually curated, high-quality, and varied data sources. To assess performance outside of English, we craft a novel benchmark, FrenchBench, consisting of an array of classification and generation tasks, covering various orthogonal aspects of model performance in the French Language. Additionally, rooted in transparency and to foster further Large Language Model research, we release codebases, and dozens of checkpoints across various model sizes, training data distributions, and training steps, as well as fine-tuned Chat models, and strong translation models. We evaluate our model through the FMTI framework, and validate 81 % of the transparency criteria, far beyond the scores of even most open initiatives. This work enriches the NLP landscape, breaking away from previous English-centric work in order to strengthen our understanding of multilinguality in language models.

António Loison, Quentin Macé, Antoine Edy et al.

Retrieval-Augmented Generation (RAG) pipelines must address challenges beyond simple single-document retrieval, such as interpreting visual elements (tables, charts, images), synthesizing information across documents, and providing accurate source grounding. Existing benchmarks fail to capture this complexity, often focusing on textual data, single-document comprehension, or evaluating retrieval and generation in isolation. We introduce ViDoRe v3, a comprehensive multimodal RAG benchmark featuring multi-type queries over visually rich document corpora. It covers 10 datasets across diverse professional domains, comprising ~26,000 document pages paired with 3,099 human-verified queries, each available in 6 languages. Through 12,000 hours of human annotation effort, we provide high-quality annotations for retrieval relevance, bounding box localization, and verified reference answers. Our evaluation of state-of-the-art RAG pipelines reveals that visual retrievers outperform textual ones, late-interaction models and textual reranking substantially improve performance, and hybrid or purely visual contexts enhance answer generation quality. However, current models still struggle with non-textual elements, open-ended queries, and fine-grained visual grounding. To encourage progress in addressing these challenges, the benchmark is released under a commercially permissive license at https://hf.co/vidore.

Géraud Faye, Wassila Ouerdane, Guillaume Gadek et al.

Misinformation detection is a critical task that can benefit significantly from the integration of external knowledge, much like manual fact-checking. In this work, we propose a novel method for representing textual documents that facilitates the incorporation of information from a knowledge base. Our approach, Text Encoding with Graph (TEG), processes documents by extracting structured information in the form of a graph and encoding both the text and the graph for classification purposes. Through extensive experiments, we demonstrate that this hybrid representation enhances misinformation detection performance compared to using language models alone. Furthermore, we introduce TEGRA, an extension of our framework that integrates domain-specific knowledge, further enhancing classification accuracy in most cases.

Hippolyte Gisserot-Boukhlef, Manuel Faysse, Emmanuel Malherbe et al. · meta-ai

Neural Information Retrieval (NIR) has significantly improved upon heuristic-based Information Retrieval (IR) systems. Yet, failures remain frequent, the models used often being unable to retrieve documents relevant to the user's query. We address this challenge by proposing a lightweight abstention mechanism tailored for real-world constraints, with particular emphasis placed on the reranking phase. We introduce a protocol for evaluating abstention strategies in black-box scenarios (typically encountered when relying on API services), demonstrating their efficacy, and propose a simple yet effective data-driven mechanism. We provide open-source code for experiment replication and abstention implementation, fostering wider adoption and application in diverse contexts.

Arthur Ledaguenel, Céline Hudelot, Mostepha Khouadjia

The last decades have seen a drastic improvement of Machine Learning (ML), mainly driven by Deep Learning (DL). However, despite the resounding successes of ML in many domains, the impossibility to provide guarantees of conformity and the fragility of ML systems (faced with distribution shifts, adversarial attacks, etc.) have prevented the design of trustworthy AI systems. Several research paths have been investigated to mitigate this fragility and provide some guarantees regarding the behavior of ML systems, among which are neurosymbolic AI and conformal prediction. Neurosymbolic artificial intelligence is a growing field of research aiming to combine neural network learning capabilities with the reasoning abilities of symbolic systems. One of the objective of this hybridization can be to provide theoritical guarantees that the output of the system will comply with some prior knowledge. Conformal prediction is a set of techniques that enable to take into account the uncertainty of ML systems by transforming the unique prediction into a set of predictions, called a confidence set. Interestingly, this comes with statistical guarantees regarding the presence of the true label inside the confidence set. Both approaches are distribution-free and model-agnostic. In this paper, we see how these two approaches can complement one another. We introduce several neurosymbolic conformal prediction techniques and explore their different characteristics (size of confidence sets, computational complexity, etc.).

Théo Sourget, David Restrepo, Céline Hudelot et al.

Motivated by the strong performance of CLIP-based models in natural image-text domains, recent efforts have adapted these architectures to medical tasks, particularly in radiology, where large paired datasets of images and reports, such as chest X-rays, are available. While these models have shown encouraging results in terms of accuracy and discriminative performance, their fairness and robustness in the different clinical tasks remain largely underexplored. In this study, we extensively evaluate six widely used CLIP-based models on chest X-ray classification using three publicly available datasets: MIMIC-CXR, NIH-CXR14, and NEATX. We assess the models fairness across six conditions and patient subgroups based on age, sex, and race. Additionally, we assess the robustness to shortcut learning by evaluating performance on pneumothorax cases with and without chest drains. Our results indicate performance gaps between patients of different ages, but more equitable results for the other attributes. Moreover, all models exhibit lower performance on images without chest drains, suggesting reliance on spurious correlations. We further complement the performance analysis with a study of the embeddings generated by the models. While the sensitive attributes could be classified from the embeddings, we do not see such patterns using PCA, showing the limitations of these visualisation techniques when assessing models. Our code is available at https://github.com/TheoSourget/clip_cxr_fairness

Etienne Bennequin, Victor Bouvier, Myriam Tami et al.

Few-Shot Learning (FSL) algorithms have made substantial progress in learning novel concepts with just a handful of labelled data. To classify query instances from novel classes encountered at test-time, they only require a support set composed of a few labelled samples. FSL benchmarks commonly assume that those queries come from the same distribution as instances in the support set. However, in a realistic set-ting, data distribution is plausibly subject to change, a situation referred to as Distribution Shift (DS). The present work addresses the new and challenging problem of Few-Shot Learning under Support/Query Shift (FSQS) i.e., when support and query instances are sampled from related but different distributions. Our contributions are the following. First, we release a testbed for FSQS, including datasets, relevant baselines and a protocol for a rigorous and reproducible evaluation. Second, we observe that well-established FSL algorithms unsurprisingly suffer from a considerable drop in accuracy when facing FSQS, stressing the significance of our study. Finally, we show that transductive algorithms can limit the inopportune effect of DS. In particular, we study both the role of Batch-Normalization and Optimal Transport (OT) in aligning distributions, bridging Unsupervised Domain Adaptation with FSL. This results in a new method that efficiently combines OT with the celebrated Prototypical Networks. We bring compelling experiments demonstrating the advantage of our method. Our work opens an exciting line of research by providing a testbed and strong baselines. Our code is available at https://github.com/ebennequin/meta-domain-shift.

Nicolas Boizard, Kevin El Haddad, Céline Hudelot et al.

Deploying large language models (LLMs) of several billion parameters can be impractical in most industrial use cases due to constraints such as cost, latency limitations, and hardware accessibility. Knowledge distillation (KD) offers a solution by compressing knowledge from resource-intensive large models to smaller ones. Various strategies exist, some relying on the text generated by the teacher model and optionally utilizing his logits to enhance learning. However, these methods based on logits often require both teacher and student models to share the same tokenizer, limiting their applicability across different LLM families. In this paper, we introduce Universal Logit Distillation (ULD) loss, grounded in optimal transport, to address this limitation. Our experimental results demonstrate the effectiveness of ULD loss in enabling distillation across models with different architectures and tokenizers, paving the way to a more widespread use of distillation techniques.

Théo Moutakanni, Piotr Bojanowski, Guillaume Chassagnon et al.

AI Foundation models are gaining traction in various applications, including medical fields like radiology. However, medical foundation models are often tested on limited tasks, leaving their generalisability and biases unexplored. We present RayDINO, a large visual encoder trained by self-supervision on 873k chest X-rays. We compare RayDINO to previous state-of-the-art models across nine radiology tasks, from classification and dense segmentation to text generation, and provide an in depth analysis of population, age and sex biases of our model. Our findings suggest that self-supervision allows patient-centric AI proving useful in clinical workflows and interpreting X-rays holistically. With RayDINO and small task-specific adapters, we reach state-of-the-art results and improve generalization to unseen populations while mitigating bias, illustrating the true promise of foundation models: versatility and robustness.

Nicolas Boizard, Hippolyte Gisserot-Boukhlef, Duarte M. Alves et al. · meta-ai

General-purpose multilingual vector representations, used in retrieval, regression and classification, are traditionally obtained from bidirectional encoder models. Despite their wide applicability, encoders have been recently overshadowed by advances in generative decoder-only models. However, many innovations driving this progress are not inherently tied to decoders. In this paper, we revisit the development of multilingual encoders through the lens of these advances, and introduce EuroBERT, a family of multilingual encoders covering European and widely spoken global languages. Our models outperform existing alternatives across a diverse range of tasks, spanning multilingual capabilities, mathematics, and coding, and natively supporting sequences of up to 8,192 tokens. We also examine the design decisions behind EuroBERT, offering insights into our dataset composition and training pipeline. We publicly release the EuroBERT models, including intermediate training checkpoints, together with our training framework.

Hippolyte Gisserot-Boukhlef, Nicolas Boizard, Manuel Faysse et al. · meta-ai

Learning high-quality text representations is fundamental to a wide range of NLP tasks. While encoder pretraining has traditionally relied on Masked Language Modeling (MLM), recent evidence suggests that decoder models pretrained with Causal Language Modeling (CLM) can be effectively repurposed as encoders, often surpassing traditional encoders on text representation benchmarks. However, it remains unclear whether these gains reflect an inherent advantage of the CLM objective or arise from confounding factors such as model and data scale. In this paper, we address this question through a series of large-scale, carefully controlled pretraining ablations, training a total of 38 models ranging from 210 million to 1 billion parameters, and conducting over 15,000 fine-tuning and evaluation runs. We find that while training with MLM generally yields better performance across text representation tasks, CLM-trained models are more data-efficient and demonstrate improved fine-tuning stability. Building on these findings, we experimentally show that a biphasic training strategy that sequentially applies CLM and then MLM, achieves optimal performance under a fixed computational training budget. Moreover, we demonstrate that this strategy becomes more appealing when initializing from readily available pretrained CLM models, reducing the computational burden needed to train best-in-class encoder models. We release all project artifacts at https://hf.co/MLMvsCLM to foster further research.

Arthur Ledaguenel, Céline Hudelot, Mostepha Khouadjia

Neurosymbolic AI is a growing field of research aiming to combine neural networks learning capabilities with the reasoning abilities of symbolic systems. This hybridization can take many shapes. In this paper, we propose a new formalism for supervised multi-label classification with propositional background knowledge. We introduce a new neurosymbolic technique called semantic conditioning at inference, which only constrains the system during inference while leaving the training unaffected. We discuss its theoritical and practical advantages over two other popular neurosymbolic techniques: semantic conditioning and semantic regularization. We develop a new multi-scale methodology to evaluate how the benefits of a neurosymbolic technique evolve with the scale of the network. We then evaluate experimentally and compare the benefits of all three techniques across model scales on several datasets. Our results demonstrate that semantic conditioning at inference can be used to build more accurate neural-based systems with fewer resources while guaranteeing the semantic consistency of outputs.

Nicolas Boizard, Hippolyte Gisserot-Boukhlef, Kevin El-Haddad et al.

Large Language Models (LLMs) with reasoning capabilities have achieved state-of-the-art performance on a wide range of tasks. Despite its empirical success, the tasks and model scales at which reasoning becomes effective, as well as its training and inference costs, remain underexplored. In this work, we rely on a synthetic data distillation framework to conduct a large-scale supervised study. We compare Instruction Fine-Tuning (IFT) and reasoning models of varying sizes, on a wide range of math-centric and general-purpose tasks, evaluating both multiple-choice and open-ended formats. Our analysis reveals that reasoning consistently improves model performance, often matching or surpassing significantly larger IFT systems. Notably, while IFT remains Pareto-optimal in training and inference costs, reasoning models become increasingly valuable as model size scales, overcoming IFT performance limits on reasoning-intensive and open-ended tasks.

Alexis-Raja Brachet, Pierre-Yves Richard, Céline Hudelot

In the trend of hybrid Artificial Intelligence techniques, Physical-Informed Machine Learning has seen a growing interest. It operates mainly by imposing data, learning, or architecture bias with simulation data, Partial Differential Equations, or equivariance and invariance properties. While it has shown great success on tasks involving one physical domain, such as fluid dynamics, existing methods are not adapted to tasks with complex multi-physical and multi-domain phenomena. In addition, it is mainly formulated as an end-to-end learning scheme. To address these challenges, we propose to leverage Bond Graphs, a multi-physics modeling approach, together with Message Passing Graph Neural Networks. We propose a Neural Bond graph Encoder (NBgE) producing multi-physics-informed representations that can be fed into any task-specific model. It provides a unified way to integrate both data and architecture biases in deep learning. Our experiments on two challenging multi-domain physical systems - a Direct Current Motor and the Respiratory System - demonstrate the effectiveness of our approach on a multivariate time-series forecasting task.

Charlotte Claye, Pierre Marschall, Wassila Ouerdane et al.

Single-cell RNA-seq foundation models achieve strong performance on downstream tasks but remain black boxes, limiting their utility for biological discovery. Recent work has shown that sparse dictionary learning can extract concepts from deep learning models, with promising applications in biomedical imaging and protein models. However, interpreting biological concepts remains challenging, as biological sequences are not inherently human-interpretable. We introduce a novel concept-based interpretability framework for single-cell RNA-seq models with a focus on concept interpretation and evaluation. We propose an attribution method with counterfactual perturbations that identifies genes that influence concept activation, moving beyond correlational approaches like differential expression analysis. We then provide two complementary interpretation approaches: an expert-driven analysis facilitated by an interactive interface and an ontology-driven method with attribution-based biological pathway enrichment. Applying our framework to two well-known single-cell RNA-seq models from the literature, we interpret concepts extracted by Top-K Sparse Auto-Encoders trained on two immune cell datasets. With a domain expert in immunology, we show that concepts improve interpretability compared to individual neurons while preserving the richness and informativeness of the latent representations. This work provides a principled framework for interpreting what biological knowledge foundation models have encoded, paving the way for their use for hypothesis generation and discovery.

Darpan Aswal, Céline Hudelot

Large Language Models have found success in a variety of applications; however, their safety remains a matter of concern due to the existence of various types of jailbreaking methods. Despite significant efforts, alignment and safety fine-tuning only provide a certain degree of robustness against jailbreak attacks that covertly mislead LLMs towards the generation of harmful content. This leaves them prone to a number of vulnerabilities, ranging from targeted misuse to accidental profiling of users. This work introduces \textbf{LLMSymGuard}, a novel framework that leverages Sparse Autoencoders (SAEs) to identify interpretable concepts within LLM internals associated with different jailbreak themes. By extracting semantically meaningful internal representations, LLMSymGuard enables building symbolic, logical safety guardrails -- offering transparent and robust defenses without sacrificing model capabilities or requiring further fine-tuning. Leveraging advances in mechanistic interpretability of LLMs, our approach demonstrates that LLMs learn human-interpretable concepts from jailbreaks, and provides a foundation for designing more interpretable and logical safeguard measures against attackers. Code will be released upon publication.

Alexis-Raja Brachet, Pierre-Yves Richard, Céline Hudelot

While boundaries between data modalities are vanishing, the usual successful deep models are still challenged by simple ones in the time-series forecasting task. Our hypothesis is that this task needs models that are able to learn the data underlying dynamics. We propose to validate it through both systemic and empirical studies. We develop an original $\texttt{PRO-DYN}$ nomenclature to analyze existing models through the lens of dynamics. Two observations thus emerged: $\textbf{1}$. under-performing architectures learn dynamics at most partially, $\textbf{2}$. the location of the dynamics block at the model end is of prime importance. We conduct extensive experiments to confirm our observations on a set of performance-varying models with diverse backbones. Results support the need to incorporate a learnable dynamics block and its use as the final predictor.

Walid Yassine, Martin Charachon, Céline Hudelot et al.

Assessing cancer progression in liver CT scans is a clinical challenge, requiring a comparison of scans at different times for the same patient. Practitioners must identify existing tumors, compare them with prior exams, identify new tumors, and evaluate overall disease evolution. This process is particularly complex in liver examinations due to misalignment between exams caused by several factors. Indeed, longitudinal liver examinations can undergo different non-pathological and pathological changes due to non-rigid deformations, the appearance or disappearance of pathologies, and other variations. In such cases, existing registration approaches, mainly based on intrinsic features may distort tumor regions, biasing the tumor progress evaluation step and the corresponding diagnosis. This work proposes a registration method based only on geometrical and anatomical information from liver segmentation, aimed at aligning longitudinal liver images for aided diagnosis. The proposed method is trained and tested on longitudinal liver CT scans, with 317 patients for training and 53 for testing. Our experimental results support our claims by showing that our method is better than other registration techniques by providing a smoother deformation while preserving the tumor burden (total volume of tissues considered as tumor) within the volume. Qualitative results emphasize the importance of smooth deformations in preserving tumor appearance.

Manuel Faysse, Hugues Sibille, Tony Wu et al.

Documents are visually rich structures that convey information through text, but also figures, page layouts, tables, or even fonts. Since modern retrieval systems mainly rely on the textual information they extract from document pages to index documents -often through lengthy and brittle processes-, they struggle to exploit key visual cues efficiently. This limits their capabilities in many practical document retrieval applications such as Retrieval Augmented Generation (RAG). To benchmark current systems on visually rich document retrieval, we introduce the Visual Document Retrieval Benchmark ViDoRe, composed of various page-level retrieval tasks spanning multiple domains, languages, and practical settings. The inherent complexity and performance shortcomings of modern systems motivate a new concept; doing document retrieval by directly embedding the images of the document pages. We release ColPali, a Vision Language Model trained to produce high-quality multi-vector embeddings from images of document pages. Combined with a late interaction matching mechanism, ColPali largely outperforms modern document retrieval pipelines while being drastically simpler, faster and end-to-end trainable. We release models, data, code and benchmarks under open licenses at https://hf.co/vidore.

Arthur Ledaguenel, Céline Hudelot, Mostepha Khouadjia

Neurosymbolic artificial intelligence is a growing field of research aiming to combine neural network learning capabilities with the reasoning abilities of symbolic systems. Informed multi-label classification is a sub-field of neurosymbolic AI which studies how to leverage prior knowledge to improve neural classification systems. Recently, a family of neurosymbolic techniques for informed classification based on probabilistic reasoning has gained significant traction. Unfortunately, depending on the language used to represent prior knowledge, solving certain probabilistic reasoning problems can become prohibitively hard when the number of classes increases. Therefore, the asymptotic complexity of probabilistic reasoning is of cardinal importance to assess the scalability of such techniques. In this paper, we develop a unified formalism for four probabilistic reasoning problems. Then, we compile several known and new tractability results into a single complexity map of probabilistic reasoning. We build on top of this complexity map to characterize the domains of scalability of several techniques. We hope this work will help neurosymbolic AI practitioners navigate the scalability landscape of probabilistic neurosymbolic techniques.

Eva Feillet, Adrian Popescu, Céline Hudelot

Class-incremental learning deals with sequential data streams composed of batches of classes. Various algorithms have been proposed to address the challenging case where samples from past classes cannot be stored. However, selecting an appropriate algorithm for a user-defined setting is an open problem, as the relative performance of these algorithms depends on the incremental settings. To solve this problem, we introduce an algorithm recommendation method that simulates the future data stream. Given an initial set of classes, it leverages generative models to simulate future classes from the same visual domain. We evaluate recent algorithms on the simulated stream and recommend the one which performs best in the user-defined incremental setting. We illustrate the effectiveness of our method on three large datasets using six algorithms and six incremental settings. Our method outperforms competitive baselines, and performance is close to that of an oracle choosing the best algorithm in each setting. This work contributes to facilitate the practical deployment of incremental learning.

Umang Aggarwal, Adrian Popescu, Céline Hudelot

Active learning aims to optimize the dataset annotation process when resources are constrained. Most existing methods are designed for balanced datasets. Their practical applicability is limited by the fact that a majority of real-life datasets are actually imbalanced. Here, we introduce a new active learning method which is designed for imbalanced datasets. It favors samples likely to be in minority classes so as to reduce the imbalance of the labeled subset and create a better representation for these classes. We also compare two training schemes for active learning: (1) the one commonly deployed in deep active learning using model fine tuning for each iteration and (2) a scheme which is inspired by transfer learning and exploits generic pre-trained models and train shallow classifiers for each iteration. Evaluation is run with three imbalanced datasets. Results show that the proposed active learning method outperforms competitive baselines. Equally interesting, they also indicate that the transfer learning training scheme outperforms model fine tuning if features are transferable from the generic dataset to the unlabeled one. This last result is surprising and should encourage the community to explore the design of deep active learning methods.

Umang Aggarwal, Adrian Popescu, Eden Belouadah et al.

Deep learning approaches are successful in a wide range of AI problems and in particular for visual recognition tasks. However, there are still open problems among which is the capacity to handle streams of visual information and the management of class imbalance in datasets. Existing research approaches these two problems separately while they co-occur in real world applications. Here, we study the problem of learning incrementally from imbalanced datasets. We focus on algorithms which have a constant deep model complexity and use a bounded memory to store exemplars of old classes across incremental states. Since memory is bounded, old classes are learned with fewer images than new classes and an imbalance due to incremental learning is added to the initial dataset imbalance. A score prediction bias in favor of new classes appears and we evaluate a comprehensive set of score calibration methods to reduce it. Evaluation is carried with three datasets, using two dataset imbalance configurations and three bounded memory sizes. Results show that most calibration methods have beneficial effect and that they are most useful for lower bounded memory sizes, which are most interesting in practice. As a secondary contribution, we remove the usual distillation component from the loss function of incremental learning algorithms. We show that simpler vanilla fine tuning is a stronger backbone for imbalanced incremental learning algorithms.

Umang Aggarwal, Adrian Popescu, Céline Hudelot

When we can not assume a large amount of annotated data , active learning is a good strategy. It consists in learning a model on a small amount of annotated data (annotation budget) and in choosing the best set of points to annotate in order to improve the previous model and gain in generalization. In deep learning, active learning is usually implemented as an iterative process in which successive deep models are updated via fine tuning, but it still poses some issues. First, the initial batch of annotated images has to be sufficiently large to train a deep model. Such an assumption is strong, especially when the total annotation budget is reduced. We tackle this issue by using an approach inspired by transfer learning. A pre-trained model is used as a feature extractor and only shallow classifiers are learned during the active iterations. The second issue is the effectiveness of probability or feature estimates of early models for AL task. Samples are generally selected for annotation using acquisition functions based only on the last learned model. We introduce a novel acquisition function which exploits the iterative nature of AL process to select samples in a more robust fashion. Samples for which there is a maximum shift towards uncertainty between the last two learned models predictions are favored. A diversification step is added to select samples from different regions of the classification space and thus introduces a representativeness component in our approach. Evaluation is done against competitive methods with three balanced and imbalanced datasets and outperforms them.

Jun Zhu, Gautier Viaud, Céline Hudelot

Recently, due to the ubiquity and supremacy of E-recruitment platforms, job recommender systems have been largely studied. In this paper, we tackle the next job application problem, which has many practical applications. In particular, we propose to leverage next-item recommendation approaches to consider better the job seeker's career preference to discover the next relevant job postings (referred to jobs for short) they might apply for. Our proposed model, named Personalized-Attention Next-Application Prediction (PANAP), is composed of three modules. The first module learns job representations from textual content and metadata attributes in an unsupervised way. The second module learns job seeker representations. It includes a personalized-attention mechanism that can adapt the importance of each job in the learned career preference representation to the specific job seeker's profile. The attention mechanism also brings some interpretability to learned representations. Then, the third module models the Next-Application Prediction task as a top-K search process based on the similarity of representations. In addition, the geographic location is an essential factor that affects the preferences of job seekers in the recruitment domain. Therefore, we explore the influence of geographic location on the model performance from the perspective of negative sampling strategies. Experiments on the public CareerBuilder12 dataset show the interest in our approach.

Victor Pellegrain, Myriam Tami, Michel Batteux et al.

The increasing complexity of Industry 4.0 systems brings new challenges regarding predictive maintenance tasks such as fault detection and diagnosis. A corresponding and realistic setting includes multi-source data streams from different modalities, such as sensors measurements time series, machine images, textual maintenance reports, etc. These heterogeneous multimodal streams also differ in their acquisition frequency, may embed temporally unaligned information and can be arbitrarily long, depending on the considered system and task. Whereas multimodal fusion has been largely studied in a static setting, to the best of our knowledge, there exists no previous work considering arbitrarily long multimodal streams alongside with related tasks such as prediction across time. Thus, in this paper, we first formalize this paradigm of heterogeneous multimodal learning in a streaming setting as a new one. To tackle this challenge, we propose StreaMulT, a Streaming Multimodal Transformer relying on cross-modal attention and on a memory bank to process arbitrarily long input sequences at training time and run in a streaming way at inference. StreaMulT improves the state-of-the-art metrics on CMU-MOSEI dataset for Multimodal Sentiment Analysis task, while being able to deal with much longer inputs than other multimodal models. The conducted experiments eventually highlight the importance of the textual embedding layer, questioning recent improvements in Multimodal Sentiment Analysis benchmarks.

Edoardo Ramalli, Alberto Parravicini, Guido Walter Di Donato et al.

Drug repurposing is more relevant than ever due to drug development's rising costs and the need to respond to emerging diseases quickly. Knowledge graph embedding enables drug repurposing using heterogeneous data sources combined with state-of-the-art machine learning models to predict new drug-disease links in the knowledge graph. As in many machine learning applications, significant work is still required to understand the predictive models' behavior. We propose a structured methodology to understand better machine learning models' results for drug repurposing, suggesting key elements of the knowledge graph to improve predictions while saving computational resources. We reduce the training set of 11.05% and the embedding space by 31.87%, with only a 2% accuracy reduction, and increase accuracy by 60% on the open ogbl-biokg graph adding only 1.53% new triples.

Martin Charachon, Paul-Henry Cournède, Céline Hudelot et al.

Providing a human-understandable explanation of classifiers' decisions has become imperative to generate trust in their use for day-to-day tasks. Although many works have addressed this problem by generating visual explanation maps, they often provide noisy and inaccurate results forcing the use of heuristic regularization unrelated to the classifier in question. In this paper, we propose a new general perspective of the visual explanation problem overcoming these limitations. We show that visual explanation can be produced as the difference between two generated images obtained via two specific conditional generative models. Both generative models are trained using the classifier to explain and a database to enforce the following properties: (i) All images generated by the first generator are classified similarly to the input image, whereas the second generator's outputs are classified oppositely. (ii) Generated images belong to the distribution of real images. (iii) The distances between the input image and the corresponding generated images are minimal so that the difference between the generated elements only reveals relevant information for the studied classifier. Using symmetrical and cyclic constraints, we present two different approximations and implementations of the general formulation. Experimentally, we demonstrate significant improvements w.r.t the state-of-the-art on three different public data sets. In particular, the localization of regions influencing the classifier is consistent with human annotations.

Yassine Ouali, Céline Hudelot, Myriam Tami

In this paper, we explore contrastive learning for few-shot classification, in which we propose to use it as an additional auxiliary training objective acting as a data-dependent regularizer to promote more general and transferable features. In particular, we present a novel attention-based spatial contrastive objective to learn locally discriminative and class-agnostic features. As a result, our approach overcomes some of the limitations of the cross-entropy loss, such as its excessive discrimination towards seen classes, which reduces the transferability of features to unseen classes. With extensive experiments, we show that the proposed method outperforms state-of-the-art approaches, confirming the importance of learning good and transferable embeddings for few-shot learning.

Antoine Plumerault, Hervé Le Borgne, Céline Hudelot

Among the wide variety of image generative models, two models stand out: Variational Auto Encoders (VAE) and Generative Adversarial Networks (GAN). GANs can produce realistic images, but they suffer from mode collapse and do not provide simple ways to get the latent representation of an image. On the other hand, VAEs do not have these problems, but they often generate images less realistic than GANs. In this article, we explain that this lack of realism is partially due to a common underestimation of the natural image manifold dimensionality. To solve this issue we introduce a new framework that combines VAE and GAN in a novel and complementary way to produce an auto-encoding model that keeps VAEs properties while generating images of GAN-quality. We evaluate our approach both qualitatively and quantitatively on five image datasets.

Martin Charachon, Céline Hudelot, Paul-Henry Cournède et al.

Explaining decisions of black-box classifiers is paramount in sensitive domains such as medical imaging since clinicians confidence is necessary for adoption. Various explanation approaches have been proposed, among which perturbation based approaches are very promising. Within this class of methods, we leverage a learning framework to produce our visual explanations method. From a given classifier, we train two generators to produce from an input image the so called similar and adversarial images. The similar image shall be classified as the input image whereas the adversarial shall not. Visual explanation is built as the difference between these two generated images. Using metrics from the literature, our method outperforms state-of-the-art approaches. The proposed approach is model-agnostic and has a low computation burden at prediction time. Thus, it is adapted for real-time systems. Finally, we show that random geometric augmentations applied to the original image play a regularization role that improves several previously proposed explanation methods. We validate our approach on a large chest X-ray database.

Victor Bouvier, Philippe Very, Clément Chastagnol et al.

Unsupervised Domain Adaptation (UDA) aims to bridge the gap between a source domain, where labelled data are available, and a target domain only represented with unlabelled data. If domain invariant representations have dramatically improved the adaptability of models, to guarantee their good transferability remains a challenging problem. This paper addresses this problem by using active learning to annotate a small budget of target data. Although this setup, called Active Domain Adaptation (ADA), deviates from UDA's standard setup, a wide range of practical applications are faced with this situation. To this purpose, we introduce \textit{Stochastic Adversarial Gradient Embedding} (SAGE), a framework that makes a triple contribution to ADA. First, we select for annotation target samples that are likely to improve the representations' transferability by measuring the variation, before and after annotation, of the transferability loss gradient. Second, we increase sampling diversity by promoting different gradient directions. Third, we introduce a novel training procedure for actively incorporating target samples when learning invariant representations. SAGE is based on solid theoretical ground and validated on various UDA benchmarks against several baselines. Our empirical investigation demonstrates that SAGE takes the best of uncertainty \textit{vs} diversity samplings and improves representations transferability substantially.

Yassine Ouali, Céline Hudelot, Myriam Tami

In this work, we propose a new unsupervised image segmentation approach based on mutual information maximization between different constructed views of the inputs. Taking inspiration from autoregressive generative models that predict the current pixel from past pixels in a raster-scan ordering created with masked convolutions, we propose to use different orderings over the inputs using various forms of masked convolutions to construct different views of the data. For a given input, the model produces a pair of predictions with two valid orderings, and is then trained to maximize the mutual information between the two outputs. These outputs can either be low-dimensional features for representation learning or output clusters corresponding to semantic labels for clustering. While masked convolutions are used during training, in inference, no masking is applied and we fall back to the standard convolution where the model has access to the full input. The proposed method outperforms current state-of-the-art on unsupervised image segmentation. It is simple and easy to implement, and can be extended to other visual tasks and integrated seamlessly into existing unsupervised learning methods requiring different views of the data.

Yassine Ouali, Victor Bouvier, Myriam Tami et al.

Learning Invariant Representations has been successfully applied for reconciling a source and a target domain for Unsupervised Domain Adaptation. By investigating the robustness of such methods under the prism of the cluster assumption, we bring new evidence that invariance with a low source risk does not guarantee a well-performing target classifier. More precisely, we show that the cluster assumption is violated in the target domain despite being maintained in the source domain, indicating a lack of robustness of the target classifier. To address this problem, we demonstrate the importance of enforcing the cluster assumption in the target domain, named Target Consistency (TC), especially when paired with Class-Level InVariance (CLIV). Our new approach results in a significant improvement, on both image classification and segmentation benchmarks, over state-of-the-art methods based on invariant representations. Importantly, our method is flexible and easy to implement, making it a complementary technique to existing approaches for improving transferability of representations.

Victor Bouvier, Philippe Very, Clément Chastagnol et al.

Unsupervised Domain Adaptation (UDA) has attracted a lot of attention in the last ten years. The emergence of Domain Invariant Representations (IR) has improved drastically the transferability of representations from a labelled source domain to a new and unlabelled target domain. However, a potential pitfall of this approach, namely the presence of \textit{label shift}, has been brought to light. Some works address this issue with a relaxed version of domain invariance obtained by weighting samples, a strategy often referred to as Importance Sampling. From our point of view, the theoretical aspects of how Importance Sampling and Invariant Representations interact in UDA have not been studied in depth. In the present work, we present a bound of the target risk which incorporates both weights and invariant representations. Our theoretical analysis highlights the role of inductive bias in aligning distributions across domains. We illustrate it on standard benchmarks by proposing a new learning procedure for UDA. We observed empirically that weak inductive bias makes adaptation more robust. The elaboration of stronger inductive bias is a promising direction for new UDA algorithms.

Yassine Ouali, Céline Hudelot, Myriam Tami

Deep neural networks demonstrated their ability to provide remarkable performances on a wide range of supervised learning tasks (e.g., image classification) when trained on extensive collections of labeled data (e.g., ImageNet). However, creating such large datasets requires a considerable amount of resources, time, and effort. Such resources may not be available in many practical cases, limiting the adoption and the application of many deep learning methods. In a search for more data-efficient deep learning methods to overcome the need for large annotated datasets, there is a rising research interest in semi-supervised learning and its applications to deep neural networks to reduce the amount of labeled data required, by either developing novel methods or adopting existing semi-supervised learning frameworks for a deep learning setting. In this paper, we provide a comprehensive overview of deep semi-supervised learning, starting with an introduction to the field, followed by a summarization of the dominant semi-supervised approaches in deep learning.

Yassine Ouali, Céline Hudelot, Myriam Tami

In this paper, we present a novel cross-consistency based semi-supervised approach for semantic segmentation. Consistency training has proven to be a powerful semi-supervised learning framework for leveraging unlabeled data under the cluster assumption, in which the decision boundary should lie in low-density regions. In this work, we first observe that for semantic segmentation, the low-density regions are more apparent within the hidden representations than within the inputs. We thus propose cross-consistency training, where an invariance of the predictions is enforced over different perturbations applied to the outputs of the encoder. Concretely, a shared encoder and a main decoder are trained in a supervised manner using the available labeled examples. To leverage the unlabeled examples, we enforce a consistency between the main decoder predictions and those of the auxiliary decoders, taking as inputs different perturbed versions of the encoder's output, and consequently, improving the encoder's representations. The proposed method is simple and can easily be extended to use additional training signal, such as image-level labels or pixel-level labels across different domains. We perform an ablation study to tease apart the effectiveness of each component, and conduct extensive experiments to demonstrate that our method achieves state-of-the-art results in several datasets.

Antoine Plumerault, Hervé Le Borgne, Céline Hudelot

Recent deep generative models are able to provide photo-realistic images as well as visual or textual content embeddings useful to address various tasks of computer vision and natural language processing. Their usefulness is nevertheless often limited by the lack of control over the generative process or the poor understanding of the learned representation. To overcome these major issues, very recent work has shown the interest of studying the semantics of the latent space of generative models. In this paper, we propose to advance on the interpretability of the latent space of generative models by introducing a new method to find meaningful directions in the latent space of any generative model along which we can move to control precisely specific properties of the generated image like the position or scale of the object in the image. Our method does not require human annotations and is particularly well suited for the search of directions encoding simple transformations of the generated image, such as translation, zoom or color variations. We demonstrate the effectiveness of our method qualitatively and quantitatively, both for GANs and variational auto-encoders.

Béatrice Mazoyer, Nicolas Hervé, Céline Hudelot et al.

In this work, we evaluate the performance of recent text embeddings for the automatic detection of events in a stream of tweets. We model this task as a dynamic clustering problem.Our experiments are conducted on a publicly available corpus of tweets in English and on a similar dataset in French annotated by our team. We show that recent techniques based on deep neural networks (ELMo, Universal Sentence Encoder, BERT, SBERT), although promising on many applications, are not very suitable for this task. We also experiment with different types of fine-tuning to improve these results on French data. Finally, we propose a detailed analysis of the results obtained, showing the superiority of tf-idf approaches for this task.

Régis Pierrard, Jean-Philippe Poli, Céline Hudelot

Despite the recent successes of deep learning, such models are still far from some human abilities like learning from few examples, reasoning and explaining decisions. In this paper, we focus on organ annotation in medical images and we introduce a reasoning framework that is based on learning fuzzy relations on a small dataset for generating explanations. Given a catalogue of relations, it efficiently induces the most relevant relations and combines them for building constraints in order to both solve the organ annotation task and generate explanations. We test our approach on a publicly available dataset of medical images where several organs are already segmented. A demonstration of our model is proposed with an example of explained annotations. It was trained on a small training set containing as few as a couple of examples.

Victor Bouvier, Philippe Very, Céline Hudelot et al.

Learning representations which remain invariant to a nuisance factor has a great interest in Domain Adaptation, Transfer Learning, and Fair Machine Learning. Finding such representations becomes highly challenging in NLP tasks since the nuisance factor is entangled in a raw text. To our knowledge, a major issue is also that only few NLP datasets allow assessing the impact of such factor. In this paper, we introduce two generalization metrics to assess model robustness to a nuisance factor: \textit{generalization under target bias} and \textit{generalization onto unknown}. We combine those metrics with a simple data filtering approach to control the impact of the nuisance factor on the data and thus to build experimental biased datasets. We apply our method to standard datasets of the literature (\textit{Amazon} and \textit{Yelp}). Our work shows that a simple text classification baseline (i.e., sentiment analysis on reviews) may be badly affected by the \textit{product ID} (considered as a nuisance factor) when learning the polarity of a review. The method proposed is generic and applicable as soon as the nuisance variable is annotated in the dataset.

Victor Bouvier, Philippe Very, Céline Hudelot et al.

Unsupervised Domain Adaptation aims to learn a model on a source domain with labeled data in order to perform well on unlabeled data of a target domain. Current approaches focus on learning \textit{Domain Invariant Representations}. It relies on the assumption that such representations are well-suited for learning the supervised task in the target domain. We rather believe that a better and minimal assumption for performing Domain Adaptation is the \textit{Hidden Covariate Shift} hypothesis. Such approach consists in learning a representation of the data such that the label distribution conditioned on this representation is domain invariant. From the Hidden Covariate Shift assumption, we derive an optimization procedure which learns to match an estimated joint distribution on the target domain and a re-weighted joint distribution on the source domain. The re-weighting is done in the representation space and is learned during the optimization procedure. We show on synthetic data and real world data that our approach deals with both \textit{Target Shift} and \textit{Concept Drift}. We report state-of-the-art performances on Amazon Reviews dataset \cite{blitzer2007biographies} demonstrating the viability of this approach.

Julien Girard, Youssef Tamaazousti, Hervé Le Borgne et al.

Many real-world visual recognition use-cases can not directly benefit from state-of-the-art CNN-based approaches because of the lack of many annotated data. The usual approach to deal with this is to transfer a representation pre-learned on a large annotated source-task onto a target-task of interest. This raises the question of how well the original representation is "universal", that is to say directly adapted to many different target-tasks. To improve such universality, the state-of-the-art consists in training networks on a diversified source problem, that is modified either by adding generic or specific categories to the initial set of categories. In this vein, we proposed a method that exploits finer-classes than the most specific ones existing, for which no annotation is available. We rely on unsupervised learning and a bottom-up split and merge strategy. We show that our method learns more universal representations than state-of-the-art, leading to significantly better results on 10 target-tasks from multiple domains, using several network architectures, either alone or combined with networks learned at a coarser semantic level.