Artificial intelligence (AI), and specifically deep learning (DL) models, are rapidly gaining traction in healthcare to analyze complex medical images and support clinical decision-making. However, DL models are often considered black boxes due to the lack of a clear explanation when providing predictions. Explainable artificial intelligence (XAI) methods are emerging as an effective way to make models explainable for developers and provide interpretable outputs for clinicians. This review presents a taxonomy of the most widely used XAI methods for image classification, with related benefits and drawbacks. Furthermore, it examines whether the type of classifier affects the choice of an explainability technique and investigates the impact of black boxes on the healthcare environment. The analysis considered papers published between January 2020 and July 2025 in Scopus and Google Scholar, utilizing the PRISMA guidelines to enhance reporting. Sixty-nine papers were identified as suitable for classifying XAI methods in four categories based on backpropagation, perturbation, attention, and concept. The results show increased use of backpropagation-based techniques, which offer simple and intuitive heatmaps. Perturbation-based methods are frequently employed to validate model robustness, but they are computationally expensive. Finally, concept-based and attention-based approaches are less widespread but represent a promising solution towards explanations that align with human semantics and reflect the intrinsic model behavior. Future research should focus on combined approaches and concept methods that generate explanations in the same semantic field as clinicians and are computationally suitable for healthcare environments, paving the way for transparent and clinically reliable DL systems.

In collaborative filtering, recommendations are made using user feedback on a few products.In this paper, we show that even if sensitive attributes are not used to fit the models, a disparate impact may nevertheless affect recommendations.We propose a definition of fairness for the recommender system that expresses that the ranking of items should be independent of sensitive attribute.We design a coclustering of users and items that processes exogenous sensitive attributes to remove their influence to return fair recommendations.We prove that our model ensures approximately fair recommendations provided that the classification of users approximately respects statistical parity.

Camera-based end-to-end driving neural networks bring the promise of a low-cost system that maps camera images to driving control commands. These networks are appealing because they replace laborious hand engineered building blocks but their black-box nature makes them difficult to delve in case of failure. Recent works have shown the importance of using an explicit intermediate representation that has the benefits of increasing both the interpretability and the accuracy of networks' decisions. Nonetheless, these camera-based networks reason in camera view where scale is not homogeneous and hence not directly suitable for motion forecasting. In this paper, we introduce a novel monocular camera-only holistic end-to-end trajectory planning network with a Bird-Eye-View (BEV) intermediate representation that comes in the form of binary Occupancy Grid Maps (OGMs). To ease the prediction of OGMs in BEV from camera images, we introduce a novel scheme where the OGMs are first predicted as semantic masks in camera view and then warped in BEV using the homography between the two planes. The key element allowing this transformation to be applied to 3D objects such as vehicles, consists in predicting solely their footprint in camera-view, hence respecting the flat world hypothesis implied by the homography.

This paper presents three different approaches to inject a location information in semantic segmentation Convolutional Neural Networks (CNN) applied to urban scenes. The assumption that a location information would improve semantic segmentation performance emerges from the idea that some elements of urban scenes are located in a predictable manner. This assumption is confronted to realistic data on the CARLA autonomous driving simulator, which is used to create our own synthetic dataset with images, depth maps and bird-eye-view cartographic images. Simulators circumvent the difficulties due to the scarcity of publicly available synchronous labeled images and location information. We encode the location information as a cartographic image to process it as a camera image. We assess the relevance of injecting the cartographic information in three different manners: as a Conditional Random Field potential, as an additional task and as an additional encoder input of a CNN. The three methods are evaluated and compared with a state of the art CNN with regards to the pixel-wise accuracy, mean intersection over union and intersection over union of some important classes. The multi-encoder approach improves the intersection over union of the pedestrians, vehicles and traffic signs classes by respectively 4%, 1.6% and 9%.

In inductive transfer learning, fine-tuning pre-trained convolutional networks substantially outperforms training from scratch. When using fine-tuning, the underlying assumption is that the pre-trained model extracts generic features, which are at least partially relevant for solving the target task, but would be difficult to extract from the limited amount of data available on the target task. However, besides the initialization with the pre-trained model and the early stopping, there is no mechanism in fine-tuning for retaining the features learned on the source task. In this paper, we investigate several regularization schemes that explicitly promote the similarity of the final solution with the initial model. We show the benefit of having an explicit inductive bias towards the initial model, and we eventually recommend a simple $L^2$ penalty with the pre-trained model being a reference as the baseline of penalty for transfer learning tasks.

In many learning tasks, including semantic image segmentation, performance can be effectively improved through the fine-tuning of a pre-trained convolutional network, instead of training from scratch. With fine-tuning, the underlying assumption is that the pre-trained model extracts generic features, which are at least partially relevant for solving a segmentation task, but that would be difficult to extract from the smaller amount of data that is available for training in urban driving scenes segmentation. However, besides the initialization with the pre-trained model and the early stopping, there is no mechanism in classical fine-tuning approaches for keeping the generic features. Even worse, the standard weight decay drives the parameters towards the origin and affects the learned features. In this paper, we show that a simple regularization that uses the pre-trained model as a reference consistently improves the performance when applied to semantic urban driving scene segmentation. Experiments are done on the Cityscapes dataset, with four different architectures of convolutional networks.

We consider the problem of generating diverse, personalized recommendations such that a small set of recommended items covers a broad range of the user's interests. We represent items in a similarity graph, and we formulate the relevance/diversity trade-off as finding a small set of unrated items that best covers a subset of items positively rated by the user. In contrast to previous approaches, our method does not rely on an explicit trade-off between a relevance objective and a diversity objective, as the estimations of relevance and diversity are implicit in the coverage criterion. We show on several benchmark datasets that our approach compares favorably to the state-of-the-art diversification methods according to various relevance and diversity measures.

This paper tackles the problem of endogenous link prediction for knowledge base completion. Knowledge bases can be represented as directed graphs whose nodes correspond to entities and edges to relationships. Previous attempts either consist of powerful systems with high capacity to model complex connectivity patterns, which unfortunately usually end up overfitting on rare relationships, or in approaches that trade capacity for simplicity in order to fairly model all relationships, frequent or not. In this paper, we propose Tatec, a happy medium obtained by complementing a high-capacity model with a simpler one, both pre-trained separately and then combined. We present several variants of this model with different kinds of regularization and combination strategies and show that this approach outperforms existing methods on different types of relationships by achieving state-of-the-art results on four benchmarks of the literature.

Assessing the uncertainty pertaining to the conclusions derived from experimental data is challenging when there is a high number of possible explanations compared to the number of experiments. We propose a new two-stage “screen and clean” procedure for assessing the uncertainties pertaining to the selection of relevant variables in high-dimensional regression problems. In this two-stage method, screening consists in selecting a subset of candidate variables by a sparsity-inducing penalized regression, while cleaning consists in discarding all variables that do not pass a significance test. This test was originally based on ordinary least squares regression. We propose to improve the procedure by conveying more information from the screening stage to the cleaning stage. Our cleaning stage is based on an adaptively penalized regression whose weights are adjusted in the screening stage. Our procedure is amenable to the computation of p-values, allowing to control the False Discovery Rate. Our experiments show the benefits of our procedure, as we observe a systematic improvement of sensitivity compared to the original procedure.

This paper considers the problem of estimation and variable selection for large high-dimensional data (high number of predictors p and large sample size N, without excluding the possibility that N < p) resulting from an individually matched case-control study. We develop a simple algorithm for the adaptation of the Lasso and related methods to the conditional logistic regression model. Our proposal relies on the simplification of the calculations involved in the likelihood function. Then, the proposed algorithm iteratively solves reweighted Lasso problems using cyclical coordinate descent, computed along a regularization path. This method can handle large problems and deal with sparse features efficiently. We discuss benefits and drawbacks with respect to the existing available implementations. We also illustrate the interest and use of these techniques on a pharmacoepidemiological study of medication use and traffic safety.

BACKGROUND: Huge amounts of electronic biomedical documents, such as molecular biology reports or genomic papers are generated daily. Nowadays, these documents are mainly available in the form of unstructured free texts, which require heavy processing for their registration into organized databases. This organization is instrumental for information retrieval, enabling to answer the advanced queries of researchers and practitioners in biology, medicine, and related fields. Hence, the massive data flow calls for efficient automatic methods of text-mining that extract high-level information, such as biomedical events, from biomedical text. The usual computational tools of Natural Language Processing cannot be readily applied to extract these biomedical events, due to the peculiarities of the domain. Indeed, biomedical documents contain highly domain-specific jargon and syntax. These documents also describe distinctive dependencies, making text-mining in molecular biology a specific discipline. RESULTS: We address biomedical event extraction as the classification of pairs of text entities into the classes corresponding to event types. The candidate pairs of text entities are recursively provided to a multiclass classifier relying on Support Vector Machines. This recursive process extracts events involving other events as arguments. Compared to joint models based on Markov Random Fields, our model simplifies inference and hence requires shorter training and prediction times along with lower memory capacity. Compared to usual pipeline approaches, our model passes over a complex intermediate problem, while making a more extensive usage of sophisticated joint features between text entities. Our method focuses on the core event extraction of the Genia task of BioNLP challenges yielding the best result reported so far on the 2013 edition.

Extracting biomedical events from scientific articles to automatically update dedicated knowledge bases has become a popular research topic with important applications. Most existing approaches are either pipeline models of specific classifiers, usually subject to cascading errors, or joint structured models, more efficient but also more costly and complicated to train. This paper proposes a system based on a pairwise model that transforms event extraction into a simple multi-class problem of classifying pairs of text entities. Such pairs are recursively provided to the classifier, allowing to extract events involving other events as arguments. This model facilitates inference compared to joint models while % relying on a single main classifier compared to being more direct and efficient than usual pipeline approaches. This method yields the best results reported so far on the BioNLP 2011 and 2013 Genia tasks.

Data representation is a crucial issue in signal processing and machine learning. In this work, we propose to guide the learning process with a prior knowledge describing how similarities between examples are organized. This knowledge is encoded in a tree structure that represents nested groups of similarities that are the pyramids of kernels. We propose a framework that learns a Support Vector Machine (SVM) on pyramids of arbitrary heights and identifies the relevant groups of similarities groups are relevant for classifying the examples. A weighted combination of (groups of) similarities is learned jointly with the SVM parameters, by optimizing a criterion that is shown to be an equivalent formulation regularized with a mixed norm of the original fitting problem. Our approach is illustrated on a Brain Computer Interfaces classification problem.

Modelling long-term dependencies in time series has proved very difficult to achieve with traditional machine-learning methods. This problem occurs when considering music data. In this paper, we introduce predictive models for melodies. We decompose melodic modelling into two subtasks. We first propose a rhythm model based on the distributions of distances between subsequences. Then, we define a generative model for melodies given chords and rhythms based on modelling sequences of Narmour features. The rhythm model consistently outperforms a standard hidden markov model (HMM) in terms of conditional prediction accuracy on two different music databases. Using a similar evaluation procedure, the proposed melodic model consistently outperforms an input/output HMM. Furthermore, these models are able to generate realistic melodies given appropriate musical contexts.

Modeling long-term dependencies in time series has proved very difficult to achieve with traditional machine learning methods. This problem occurs when considering music data. In this paper, we introduce a model for rhythms based on the distributions of distances between subsequences. A specific implementation of the model when considering Hamming distances over a simple rhythm representation is described. The proposed model consistently outperforms a standard Hidden Markov Model in terms of conditional prediction accuracy on two different music databases. 1.

The Support Vector Machine (SVM) is an acknowledged powerful tool for building classifiers, but it lacks flexibility, in the sense that the kernel is chosen prior to learning. Multiple Kernel Learning (MKL) enables to learn the kernel, from an ensemble of basis kernels, whose combination is optimized in the learning process. Here, we propose Composite Kernel Learning to address the situation where distinct components give rise to a group structure among kernels. Our formulation of the learning problem encompasses several setups, putting more or less emphasis on the group structure. We characterize the convexity of the learning problem, and provide a general wrapper algorithm for computing solutions. Finally, we illustrate the behavior of our method on multi-channel data where groups correpond to channels.

We consider the problem of binary classification where the classifier may abstain instead of classifying each observation. The Bayes decision rule for this setup, known as Chow's rule, is defined by two thresholds on posterior probabilities. From simple desiderata, namely the consistency and the sparsity of the classifier, we derive the double hinge loss function that focuses on estimating conditional probabilities only in the vicinity of the threshold points of the optimal decision rule. We show that, for suitable kernel machines, our approach is universally consistent. We cast the problem of minimizing the double hinge loss as a quadratic program akin to the standard SVM optimization problem and propose an active set method to solve it efficiently. We finally provide preliminary experimental results illustrating the interest of our constructive approach to devising loss functions.

The scores returned by support vector machines are often used as a confidence measures in the classification of new examples. However, there is no theoretical argument sustaining this practice. Thus, when classification uncertainty has to be assessed, it is safer to resort to classifiers estimating conditional probabilities of class labels. Here, we focus on the ambiguity in the vicinity of the boundary decision. We propose an adaptation of maximum likelihood estimation, instantiated on logistic regression. The model outputs proper conditional probabilities into a user-defined interval and is less precise elsewhere. The model is also sparse, in the sense that few examples contribute to the solution. The computational efficiency is thus improved compared to logistic regression. Furthermore, preliminary experiments show improvements over standard logistic regression and performances similar to support vector machines.

An efficient and general multiple kernel learning (MKL) algorithm has been recently proposed by Sonnenburg et al. (2006). This approach has opened new perspectives since it makes the MKL approach tractable for large-scale problems, by iteratively using existing support vector machine code. However, it turns out that this iterative algorithm needs several iterations before converging towards a reasonable solution. In this paper, we address the MKL problem through an adaptive 2-norm regularization formulation. Weights on each kernel matrix are included in the standard SVM empirical risk minimization problem with a l1 constraint to encourage sparsity. We propose an algorithm for solving this problem and provide an new insight on MKL algorithms based on block 1-norm regularization by showing that the two approaches are equivalent. Experimental results show that the resulting algorithm converges rapidly and its efficiency compares favorably to other MKL algorithms.

Hierarchical penalization is a generic framework for incorporating prior informa- tion in the fitting of statistical models, when the explicative variables are organized in a hierarchical structure. The penalizer is a convex functional that performs soft selection at the group level, and shrinks variables within each group. This favors solutions with few leading terms in the final combination. The framework, orig- inally derived for taking prior knowledge into account, is shown to be useful in linear regression, when several parameters are used to model the influence of one feature, or in kernel regression, for learning multiple kernels. Keywords – Optimization: constrained and convex optimization. Supervised learning: regression, kernel methods, sparsity and feature selection.

In this paper, we show that the hinge loss can be interpreted as the neg-log-likelihood of a semi-parametric model of posterior probabilities. From this point of view, SVMs represent the parametric component of a semi-parametric model fitted by a maximum a posteriori estimation pro- cedure. This connection enables to derive a mapping from SVM scores to estimated posterior probabilities. Unlike previous proposals, the sug- gested mapping is interval-valued, providing a set of posterior probabil- ities compatible with each SVM score. This framework offers a new way to adapt the SVM optimization problem to unbalanced classifica- tion, when decisions result in unequal (asymmetric) losses. Experiments show improvements over state-of-the-art procedures.

We consider the semi-supervised learning problem, where a decision rule is to be learned from labeled and unlabeled data. In this framework, we motivate minimum entropy regularization, which enables to incorporate unlabeled data in the standard supervised learning. Our approach in- cludes other approaches to the semi-supervised problem as particular or limiting cases. A series of experiments illustrates that the proposed solu- tion benefits from unlabeled data. The method challenges mixture mod- els when the data are sampled from the distribution class spanned by the generative model. The performances are definitely in favor of minimum entropy regularization when generative models are misspecified, and the weighting of unlabeled data provides robustness to the violation of the "cluster assumption". Finally, we also illustrate that the method can also be far superior to manifold learning in high dimension spaces. 1 Introduction In the classical supervised learning classification framework, a decision rule is to be learned from a learning set Ln = {xi, yi}ni , where each example is described by a pattern =1 xi X and by the supervisor's response yi = {1, . . . , K }. We consider semi-supervised learning, where the supervisor's responses are limited to a subset of Ln. In the terminology used here, semi-supervised learning refers to learning a decision rule on X from labeled and unlabeled data. However, the related problem of transductive learning, i.e. of predicting labels on a set of predefined patterns, is addressed as a side issue. Semi- supervised problems occur in many applications where labeling is performed by human experts. They have been receiving much attention during the last few years, but some important issues are unresolved [10]. In the probabilistic framework, semi-supervised learning can be modeled as a missing data problem, which can be addressed by generative models such as mixture models thanks to the EM algorithm and extensions thereof [6].Generative models apply to the joint den- sity of patterns and class (X, Y ). They have appealing features, but they also have major drawbacks. Their estimation is much more demanding than discriminative models, since the model of P (X, Y ) is exhaustive, hence necessarily more complex than the model of This work was supported in part by the IST Programme of the European Community, under the PASCAL Network of Excellence IST-2002-506778. This publication only reflects the authors' views. P (Y |X). More parameters are to be estimated, resulting in more uncertainty in the es- timation process. The generative model being more precise, it is also more likely to be misspecified. Finally, the fitness measure is not discriminative, so that better models are not necessarily better predictors of class labels. These difficulties have lead to proposals aiming at processing unlabeled data in the framework of supervised classification [1, 5, 11]. Here, we propose an estimation principle applicable to any probabilistic classifier, aiming at making the most of unlabeled data when they are beneficial, while providing a control on their contribution to provide robustness to the learning scheme. 2 Derivation of the Criterion

Most machine learning researchers perform quantitative experiments to estimate generalization error and compare algorithm performances. In order to draw statistically convincing conclusions, it is important to esti- mate the uncertainty of such estimates. This paper studies the estimation of uncertainty around the K-fold cross-validation estimator. The main theorem shows that there exists no universal unbiased estimator of the variance of K-fold cross-validation. An analysis based on the eigende- composition of the covariance matrix of errors helps to better understand the nature of the problem and shows that naive estimators may grossly underestimate variance, as con£rmed by numerical experiments.

This paper introduces an algorithm for the automatic relevance determi- nation of input variables in kernelized Support Vector Machines. Rele- vance is measured by scale factors defining the input space metric, and feature selection is performed by assigning zero weights to irrelevant variables. The metric is automatically tuned by the minimization of the standard SVM empirical risk, where scale factors are added to the usual set of parameters defining the classifier. Feature selection is achieved by constraints encouraging the sparsity of scale factors. The resulting algorithm compares favorably to state-of-the-art feature selection proce- dures and demonstrates its effectiveness on a demanding facial expres- sion recognition problem.

In many discrimination problems a large amount of data is available but only a few of them are labeled. This provides a strong motivation to improve or develop methods for semi-supervised learning. In this paper, boosting is generalized to this task within the optimization framework of MarginBoost . We extend the margin definition to unlabeled data and develop the gradient descent algorithm that corresponds to the resulting margin cost function. This meta-learning scheme can be applied to any base classifier able to benefit from unlabeled data. We propose here to apply it to mixture models trained with an Expectation-Maximization algorithm. Promising results are presented on benchmarks with different rates of labeled data.

Bagging is a procedure averaging estimators trained on bootstrap samples. Numerous experiments have shown that bagged estimates often yield better results than the original predictor, and several explanations have been given to account for this gain. However, six years from its introduction, bagging is still not fully understood. Most explanations given until now are based on global properties of the estimates. Here, we focus on the local effects on leverage points, i.e., on observations whose fitted values are largely determined by the corresponding response values. These points are shown experimentally to be down-weighted by bagging. The performance of the bagged estimate depends on the goodness of these points for the original estimator. Illustrative examples findings are supported by the study of smoothing matrix, and their consequences are discussed.

There are two archetypal ways to control the complexity of a flexible regressor: subset selection and ridge regression. In neural-networks jargon, they are, respectively, known as pruning and weight decay. These techniques may also be adapted to estimate which features of the input space are relevant for predicting the output variables. Relevance is given by a binary indicator for subset selection, and by a continuous rating for ridge regression. This paper shows how to achieve such a rating for a multilayer perceptron trained with noise (or jitter). Noise injection (NI) is modified in order to penalize heavily irrelevant features. The proposed algorithm is attractive as it requires the tuning of a single parameter. This parameter controls the complexity of the model (effective number of parameters) together with the rating of feature relevances (effective input space dimension). Bounds on the effective number of parameters support that the stability of this adaptive scheme is enforced by the constraints applied to the admissible set of relevance indices. The good properties of the algorithm are confirmed by satisfactory experimental results on simulated data sets.

Noise injection consists of adding noise to the inputs during neural network training. Experimental results suggest that it might improve the generalization ability of the resulting neural network. A justification of this improvement remains elusive: describing analytically the average perturbed cost function is difficult, and controlling the fluctuations of the random perturbed cost function is hard. Hence, recent papers suggest replacing the random perturbed cost by a (deterministic) Taylor approximation of the average perturbed cost function. This article takes a different stance: when the injected noise is gaussian, noise injection is naturally connected to the action of the heat kernel. This provides indications on the relevance domain of traditional Taylor expansions and shows the dependence of the quality of Taylor approximations on global smoothness properties of neural networks under consideration. The connection between noise injection and heat kernel also enables controlling the fluctuations of the random perturbed cost function. Under the global smoothness assumption, tools from gaussian analysis provide bounds on the tail behavior of the perturbed cost. This finally suggests that mixing input perturbation with smoothness-based penalization might be profitable.

The generalization capacity of neural networks learning from examples is important. Several authors showed experimentally that training a neural network with noise injected inputs could improve its generalization abilities. In the original paper (ibid., vol. 22, no. 3. p. 436-40, 1992), Matsuoka explained this fact in a formal way, claiming that using noise injected inputs is equivalent to reduce the sensitivity of the network. However, the author states that an error in Matsuoka's calculations lead him to inadequate conclusions. This paper corrects these calculations and conclusions.>