What is the shortest path between two data points lying in a high-dimensional space? While the answer is trivial in Euclidean geometry, it becomes significantly more complex when the data lies on a curved manifold -- requiring a Riemannian metric to describe the space's local curvature. Estimating such a metric, however, remains a major challenge in high dimensions. In this work, we propose a method for deriving Riemannian metrics directly from pretrained Energy-Based Models (EBMs) -- a class of generative models that assign low energy to high-density regions. These metrics define spatially varying distances, enabling the computation of geodesics -- shortest paths that follow the data manifold's intrinsic geometry. We introduce two novel metrics derived from EBMs and show that they produce geodesics that remain closer to the data manifold and exhibit lower curvature distortion, as measured by alignment with ground-truth trajectories. We evaluate our approach on increasingly complex datasets: synthetic datasets with known data density, rotated character images with interpretable geometry, and high-resolution natural images embedded in a pretrained VAE latent space. Our results show that EBM-derived metrics consistently outperform established baselines, especially in high-dimensional settings. Our work is the first to derive Riemannian metrics from EBMs, enabling data-aware geodesics and unlocking scalable, geometry-driven learning for generative modeling and simulation.

Multimodal magnetic resonance imaging (MRI) constitutes the first line of investigation for clinicians in the care of brain tumors, providing crucial insights for surgery planning, treatment monitoring, and biomarker identification. Pre-training on large datasets have been shown to help models learn transferable representations and adapt with minimal labeled data. This behavior is especially valuable in medical imaging, where annotations are often scarce. However, applying this paradigm to multimodal medical data introduces a challenge: most existing approaches assume that all imaging modalities are available during both pre-training and fine-tuning. In practice, missing modalities often occur due to acquisition issues, specialist unavailability, or specific experimental designs on small in-house datasets. Consequently, a common approach involves training a separate model for each desired modality combination, making the process both resource-intensive and impractical for clinical use. Therefore, we introduce BM-MAE, a masked image modeling pre-training strategy tailored for multimodal MRI data. The same pre-trained model seamlessly adapts to any combination of available modalities, extracting rich representations that capture both intra- and inter-modal information. This allows fine-tuning on any subset of modalities without requiring architectural changes, while still benefiting from a model pre-trained on the full set of modalities. Extensive experiments show that the proposed pre-training strategy outperforms or remains competitive with baselines that require separate pre-training for each modality subset, while substantially surpassing training from scratch on several downstream tasks. Additionally, it can quickly and efficiently reconstruct missing modalities, highlighting its practical value. Code and trained models are available at: this https URL

Image segmentation is a challenging task influenced by multiple sources of uncertainty, such as the data labeling process or the sampling of training data. In this paper we focus on binary segmentation and address these challenges using conformal prediction, a family of model- and data-agnostic methods for uncertainty quantification that provide finite-sample theoretical guarantees and applicable to any pretrained predictor. Our approach involves computing nonconformity scores, a type of prediction residual, on held-out calibration data not used during training. We use dilation, one of the fundamental operations in mathematical morphology, to construct a margin added to the borders of predicted segmentation masks. At inference, the predicted set formed by the mask and its margin contains the ground-truth mask with high probability, at a confidence level specified by the user. The size of the margin serves as an indicator of predictive uncertainty for a given model and dataset. We work in a regime of minimal information as we do not require any feedback from the predictor: only the predicted masks are needed for computing the prediction sets. Hence, our method is applicable to any segmentation model, including those based on deep learning; we evaluate our approach on several medical imaging applications.

Recent advances in object detectors have led to their adoption for industrial uses. However, their deployment in safety-critical applications is hindered by the inherent lack of reliability of neural networks and the complex structure of object detection models. To address these challenges, we turn to Conformal Prediction, a post-hoc predictive uncertainty quantification procedure with statistical guarantees that are valid for any dataset size, without requiring prior knowledge on the model or data distribution. Our contribution is manifold. First, we formally define the problem of Conformal Object Detection (COD). We introduce a novel method, Sequential Conformal Risk Control (SeqCRC), that extends the statistical guarantees of Conformal Risk Control to two sequential tasks with two parameters, as required in the COD setting. Then, we present old and new loss functions and prediction sets suited to applying SeqCRC to different cases and certification requirements. Finally, we present a conformal toolkit for replication and further exploration of our method. Using this toolkit, we perform extensive experiments that validate our approach and emphasize trade-offs and other practical consequences.

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.

Over the past few years, the acceleration of computing resources and research in deep learning has led to significant practical successes in a range of tasks, including in particular in computer vision. Building on these advances, reinforcement learning has also seen a leap forward with the emergence of agents capable of making decisions directly from visual observations. Despite these successes, the over-parametrization of neural architectures leads to memorization of the data used during training and thus to a lack of generalization. Reinforcement learning agents based on visual inputs also suffer from this phenomenon by erroneously correlating rewards with unrelated visual features such as background elements. To alleviate this problem, we introduce a new regularization technique consisting of channel-consistent local permutations (CLOP) of the feature maps. The proposed permutations induce robustness to spatial correlations and help prevent overfitting behaviors in RL. We demonstrate, on the OpenAI Procgen Benchmark, that RL agents trained with the CLOP method exhibit robustness to visual changes and better generalization properties than agents trained using other state-of-the-art regularization techniques. We also demonstrate the effectiveness of CLOP as a general regularization technique in supervised learning.

As the interest in autonomous systems continues to grow, one of the major challenges is collecting sufficient and representative real-world data. Despite the strong practical and commercial interest in autonomous landing systems in the aerospace field, there is a lack of open-source datasets of aerial images. To address this issue, we present a dataset-lard-of high-quality aerial images for the task of runway detection during approach and landing phases. Most of the dataset is composed of synthetic images but we also provide manually labelled images from real landing footages, to extend the detection task to a more realistic setting. In addition, we offer the generator which can produce such synthetic front-view images and enables automatic annotation of the runway corners through geometric transformations. This dataset paves the way for further research such as the analysis of dataset quality or the development of models to cope with the detection tasks. Find data, code and more up-to-date information at this https URL

We propose a post-hoc, computationally lightweight method to quantify predictive uncertainty in semantic image segmentation. Our approach uses conformal prediction to generate statistically valid prediction sets that are guaranteed to include the ground-truth segmentation mask at a predefined confidence level. We introduce a novel visualization technique of conformalized predictions based on heatmaps, and provide metrics to assess their empirical validity. We demonstrate the effectiveness of our approach on well-known benchmark datasets and image segmentation prediction models, and conclude with practical insights.

In theory, the choice of ReLU(0) in [0, 1] for a neural network has a negligible influence both on backpropagation and training. Yet, in the real world, 32 bits default precision combined with the size of deep learning problems makes it a hyperparameter of training methods. We investigate the importance of the value of ReLU'(0) for several precision levels (16, 32, 64 bits), on various networks (fully connected, VGG, ResNet) and datasets (MNIST, CIFAR10, SVHN, ImageNet). We observe considerable variations of backpropagation outputs which occur around half of the time in 32 bits precision. The effect disappears with double precision, while it is systematic at 16 bits. For vanilla SGD training, the choice ReLU'(0) = 0 seems to be the most efficient. For our experiments on ImageNet the gain in test accuracy over ReLU'(0) = 1 was more than 10 points (two runs). We also evidence that reconditioning approaches as batch-norm or ADAM tend to buffer the influence of ReLU'(0)'s value. Overall, the message we convey is that algorithmic differentiation of nonsmooth problems potentially hides parameters that could be tuned advantageously.

With increasing processing power, deploying AI models for remote sensing directly onboard satellites is becoming feasible. However, new constraints arise, mainly when using raw, unprocessed sensor data instead of preprocessed ground-based products. While current solutions primarily rely on preprocessed sensor images, few approaches directly leverage raw data. This study investigates the effects of utilising raw data on deep learning models for object detection and classification tasks. We introduce a simulation workflow to generate raw-like products from high-resolution L1 imagery, enabling systemic evaluation. Two object detection models (YOLOv11n and YOLOX-S) are trained on both raw and L1 datasets, and their performance is compared using standard detection metrics and explainability tools. Results indicate that while both models perform similarly at low to medium confidence thresholds, the model trained on raw data struggles with object boundary identification at high confidence levels. It suggests that adapting AI architectures with improved contouring methods can enhance object detection on raw images, improving onboard AI for remote sensing.

A variety of methods have been proposed to try to explain how deep neural networks make their decisions. Key to those approaches is the need to sample the pixel space efficiently in order to derive importance maps. However, it has been shown that the sampling methods used to date introduce biases and other artifacts, leading to inaccurate estimates of the importance of individual pixels and severely limit the reliability of current explainability methods. Unfortunately, the alternative -- to exhaustively sample the image space is computationally prohibitive. In this paper, we introduce EVA (Explaining using Verified perturbation Analysis) -- the first explainability method guarantee to have an exhaustive exploration of a perturbation space. Specifically, we leverage the beneficial properties of verified perturbation analysis -- time efficiency, tractability and guaranteed complete coverage of a manifold -- to efficiently characterize the input variables that are most likely to drive the model decision. We evaluate the approach systematically and demonstrate state-of-the-art results on multiple benchmarks.

Explainable Artificial Intelligence (XAI) has become increasingly significant for improving the interpretability and trustworthiness of machine learning models. While saliency maps have stolen the show for the last few years in the XAI field, their ability to reflect models' internal processes has been questioned. Although less in the spotlight, example-based XAI methods have continued to improve. It encompasses methods that use examples as explanations for a machine learning model's predictions. This aligns with the psychological mechanisms of human reasoning and makes example-based explanations natural and intuitive for users to understand. Indeed, humans learn and reason by forming mental representations of concepts based on examples. This paper provides an overview of the state-of-the-art in natural example-based XAI, describing the pros and cons of each approach. A "natural" example simply means that it is directly drawn from the training data without involving any generative process. The exclusion of methods that require generating examples is justified by the need for plausibility which is in some regards required to gain a user's trust. Consequently, this paper will explore the following family of methods: similar examples, counterfactual and semi-factual, influential instances, prototypes, and concepts. In particular, it will compare their semantic definition, their cognitive impact, and added values. We hope it will encourage and facilitate future work on natural example-based XAI.

This study introduces a lightweight U-Net model optimized for real-time semantic segmentation of aerial images, targeting the efficient utilization of Commercial Off-The-Shelf (COTS) embedded computing platforms. We maintain the accuracy of the U-Net on a real-world dataset while significantly reducing the model's parameters and Multiply-Accumulate (MAC) operations by a factor of 16. Our comprehensive analysis covers three hardware platforms (CPU, GPU, and FPGA) and five different toolchains (TVM, FINN, Vitis AI, TensorFlow GPU, and cuDNN), assessing each on metrics such as latency, power consumption, memory footprint, energy efficiency, and FPGA resource usage. The results highlight the trade-offs between these platforms and toolchains, with a particular focus on the practical deployment challenges in real-world applications. Our findings demonstrate that while the FPGA with Vitis AI emerges as the superior choice due to its performance, energy efficiency, and maturity, it requires specialized hardware knowledge, emphasizing the need for a balanced approach in selecting embedded computing solutions for semantic segmentation tasks

Real-life medical data is often multimodal and incomplete, fueling the growing need for advanced deep learning models capable of integrating them efficiently. The use of diverse modalities, including histopathology slides, MRI, and genetic data, offers unprecedented opportunities to improve prognosis prediction and to unveil new treatment pathways. Contrastive learning, widely used for deriving representations from paired data in multimodal tasks, assumes that different views contain the same task-relevant information and leverages only shared information. This assumption becomes restrictive when handling medical data since each modality also harbors specific knowledge relevant to downstream tasks. We introduce DRIM, a new multimodal method for capturing these shared and unique representations, despite data sparsity. More specifically, given a set of modalities, we aim to encode a representation for each one that can be divided into two components: one encapsulating patient-related information common across modalities and the other, encapsulating modality-specific details. This is achieved by increasing the shared information among different patient modalities while minimizing the overlap between shared and unique components within each modality. Our method outperforms state-of-the-art algorithms on glioma patients survival prediction tasks, while being robust to missing modalities. To promote reproducibility, the code is made publicly available at this https URL

We study the problem of approximating the level set of an unknown function by sequentially querying its values. We introduce a family of algorithms called Bisect and Approximate through which we reduce the level set approximation problem to a local function approximation problem. We then show how this approach leads to rate-optimal sample complexity guarantees for H{ö}lder functions, and we investigate how such rates improve when additional smoothness or other structural assumptions hold true.

Various methods have been proposed to detect objects while reducing the cost of data annotation. For instance, weakly supervised object detection (WSOD) methods rely only on image-level annotations during training. Unfortunately, data annotation remains expensive since annotators must provide the categories describing the content of each image and labeling is restricted to a fixed set of categories. In this paper, we propose a method to locate and label objects in an image by using a form of weaker supervision: image-caption pairs. By leveraging recent advances in vision-language (VL) models and self-supervised vision transformers (ViTs), our method is able to perform phrase grounding and object detection in a weakly supervised manner. Our experiments demonstrate the effectiveness of our approach by achieving a 47.51% recall@1 score in phrase grounding on Flickr30k Entities and establishing a new state-of-the-art in object detection by achieving 21.1 mAP 50 and 10.5 mAP 50:95 on MS COCO when exclusively relying on image-caption pairs.

Out-of-distribution (OOD) detection is crucial for ensuring the reliability of deep learning models in real-world applications. Existing methods typically focus on feature representations or output-space analysis, often assuming a distribution over these spaces or leveraging gradient norms with respect to model parameters. However, these approaches struggle to distinguish near-OOD samples and often require extensive hyper-parameter tuning, limiting their this http URL this work, we propose GRadient-aware Out-Of-Distribution detection (GROOD), a method that derives an OOD prototype from synthetic samples and computes class prototypes directly from In-distribution (ID) training data. By analyzing the gradients of a nearest-class-prototype loss function concerning an artificial OOD prototype, our approach achieves a clear separation between in-distribution and OOD samples. Experimental evaluations demonstrate that gradients computed from the OOD prototype enhance the distinction between ID and OOD data, surpassing established baselines in robustness, particularly on ImageNet-1k. These findings highlight the potential of gradient-based methods and prototype-driven approaches in advancing OOD detection within deep neural networks.

We propose a new method, dubbed One Class Signed Distance Function (OCSDF), to perform One Class Classification (OCC) by provably learning the Signed Distance Function (SDF) to the boundary of the support of any distribution. The distance to the support can be interpreted as a normality score, and its approximation using 1-Lipschitz neural networks provides robustness bounds against $l2$ adversarial attacks, an under-explored weakness of deep learning-based OCC algorithms. As a result, OCSDF comes with a new metric, certified AUROC, that can be computed at the same cost as any classical AUROC. We show that OCSDF is competitive against concurrent methods on tabular and image data while being way more robust to adversarial attacks, illustrating its theoretical properties. Finally, as exploratory research perspectives, we theoretically and empirically show how OCSDF connects OCC with image generation and implicit neural surface parametrization. Our code is available at this https URL

Ensuring fairness in NLP models is crucial, as they often encode sensitive attributes like gender and ethnicity, leading to biased outcomes. Current concept erasure methods attempt to mitigate this by modifying final latent representations to remove sensitive information without retraining the entire model. However, these methods typically rely on linear classifiers, which leave models vulnerable to non-linear adversaries capable of recovering sensitive information. We introduce Targeted Concept Erasure (TaCo), a novel approach that removes sensitive information from final latent representations, ensuring fairness even against non-linear classifiers. Our experiments show that TaCo outperforms state-of-the-art methods, achieving greater reductions in the prediction accuracy of sensitive attributes by non-linear classifier while preserving overall task performance. Code is available on this https URL