This work explores a simple approach to Isolated Sign Language Recognition (ISLR) by reframing the classification of pose sequences as a standard image classification task.While recent trends in Sign Language Processing (SLP) heavily favor complex temporal architectures like Transformers, we investigate the projection of spatio-temporal pose information into a static image representation.By mapping time and skeletal joints to spatial dimensions and coordinate values to color channels, we allow standard Convolutional Neural Networks (CNNs), like ResNets, to extract features effectively.Our experiments on challenging real-world ISLR datasets demonstrate that this method is not only computationally efficient, but also outperforms existing architectures like Pose-VIT and SPOTER in a simple classification setting.

The ranking task is the critical step performed during a recommendation process to predict the top-list of most-wanted products for users. Learn-to-rank algorithms have been developed to refine the ranking process. However, the underrepresentation of some demographic user categories leads to unwanted biased ranking performances that affect the fairness aspects of the recommendation. Bayesian Pairwise Ranking (BPR) is among the most popular ranking algorithms for its important ranking accuracy performance. BPR with machine learning recommendation models can unfairly perform for minority user groups. We tackle the unfairness in the recommendation by proposing the FEBPR method. Our proposal is a fair pairwise Bayesian ranking in which the data debiasing is performed by using adversarial learning fed by enriched embeddings. In our proposal, user and item embeddings are learned to obey the adversarial constraint and mislead the adversary classifier that should not be able to have a priori assumptions about user membership. Extensive experiments are performed on real-world datasets and show that the performances of the proposed debiasing method improve fairness ranking aspects, and therefore the recommendation fairness. It is also shown that our proposal outperforms state-of-the-art fairness ranking methods and presents an interesting tradeoff between the fairness aspects and ranking accuracy.

As convolutional neural networks (CNNs) have become mainstream for object recognition and image classification, the environmental impact caused by their high energy consumption (EC) is non negligible.This paper examines techniques that have the ability to reduce the EC of CNNs.It also highlights the inconsistency of metrics that are used for estimating or measuring EC, which reduces the comparability of these techniques.This review aims to shed light on the current situation and to provide a basis for future research in green machine learning.

Data imbalance can take various forms, such as uneven class distributions in the dataset.Solutions like data augmentation, sampling techniques and weighted loss functions are commonly used to address this issue.However, in segmentation tasks, an additional type of imbalance may occur at the pixel-level, with most of them belonging to the background class.This work introduces Mask-Aware Cropping (MAC), a technique to reduce pixel-level imbalance by cropping image regions containing key information about the minority class.

Sign language (SL) is a communication method used by deaf people.Static sign language recognition (SLR) is a challenging task aimed at identifying signs in images, for which acquisition of annotated data is time-consuming.To leverage unannotated data, practitioners have turned to unsupervised methods.Contrastive representation learning proved to be effective in capturing important features from unannotated data.It is known that the performance of the contrastive model depends on the data augmentation technique used during training.For various applications, a set of effective data augmentation has been identified, but it is not yet the case for SL.This paper identifies the most effective augmentation for static SLR.The results show a difference in accuracy of up to 30% between appearance-based augmentations combined with translations and augmentations based on rotations, erasing, or vertical flips.

Early lung cancer detection strongly increases survival rate. During a navigational bronchoscopy, pulmonologists perform tissue sampling for biopsies based on preoperative medical images. The bronchial carina is an airway structure that appears at each bronchus bifurcation. It is an important landmark to detect during navigations as it indicates the need to choose between multiple paths and keep track of the position in the lungs. In this paper, we assessed various deep learning pipelines including the use of semi-supervised, segmentation and recurrent methods under different setups to perform bronchial carinae detection. In contrast to most previous works that focus on phantoms, cadavers or virtual images, we exploit a large corpus of proprietary in vivo data captured during real endoscopic procedures using a mini probe. To the best of our knowledge, it is the first work that deals with this quantity of real and challenging data. After performing a comparison study, we conclude that the best performance to detect bronchial carinae is achieved by a semi-supervised pipeline that leverages the ability of nnU-Net to solve segmentation tasks, coupled with Gated Recurrent Units that extracts temporal contexts from image sequences.

As artificial intelligence (AI) systems become increasingly integrated into everyday life, the need for trustworthiness in these systems has emerged as a critical challenge.This tutorial paper addresses the complexity of building trust in AI systems by exploring recent advances in explainable AI (XAI) and related areas that go beyond mere interpretability.After reviewing recent trends in XAI, we discuss how to control AI systems, align them with societal concerns, and address the robustness, reproducibility, and evaluation concerns inherent in these systems.This review highlights the multifaceted nature of the mechanisms for building trust in AI, and we hope it will pave the way for further research in this area.1 https://digital-strategy.

Decision trees help domain experts, such as doctors and bankers, rationalize system decisions.However, existing methods lack user-friendly ways to integrate multiple constraints and identify branches for pruning.This paper introduces ChatDT, a prototype developed with a new domain-specific language and an enhanced version of the CART algorithm to address these challenges.An evaluation involving 22 participants highlights ChatDT's effectiveness, confirming its role in facilitating decision tree creation tailored to domain-specific constraints and identifying branches for pruning.

Non-linear dimensionality reduction techniques offer insights into complex datasets, yet interpreting them poses challenges.While some papers provide methods for explaining DR, and others focus on interactively exploring embeddings, there are currently no works that seamlessly combine both aspects.Our contributions, Insight-SNE, propose an interactive tool that allows exploring t-SNE embeddings and their related gradient-based explanations, as well as its evaluation with expert users.1 Supported by the Walloon region, with a Ph.D. grant from FRIA (F

In fields like biomedical imaging, it is common to manage 3D images instead of 2D ones (CT-scans, MRI, 3D-ultrasound, etc.).Although 3D-Convolutional Neural Networks (CNNs) are generally more powerful compared to their 2D counterparts for such applications, it also comes at the cost of an increase in computational resources (both in time and memory).In this work, we present a new way to build 2D representations of 3D images while minimizing the information loss by leveraging quaternions.Those quaternion CNNs are able to offer competitive performance while significantly reducing computational complexity.

This study contributes to the automatic detection of Crohn's Disease (CD), a gastrointestinal inflammatory condition.In particular, our approach deals with the challenge of data scarcity for CD by pretraining Vision Transformers (ViT) on Hyper-Kvasir and LDPolyp, two large colonscopic datasets that represent over one million images from a similar domain, using a Contrastive Loss (CL) mechanism.This approach significantly outperforms models pre-trained on ImageNet as well as models pre-trained with a Cross-Entropy Loss on the Crohn-IPI dataset.* The present research benefited from computational resources made available on Lucia, the Tier-1 supercomputer of the Walloon Region, infrastructure funded by the Walloon Region under the grant agreement n°1910247.

In the field of spatial aeronomy, atmospheric profile datasets often contain partial data.Probabilistic models, particularly Gaussian processes (GPs), offer promising solutions for filling these data gaps.However, traditional GP algorithms encounter challenges when handling multiple sequences simultaneously, both in terms of performance and computational complexity.Recently, an algorithm named MAGMA was introduced to address these issues.This paper evaluates MAGMA's performance using the SOIR Venus atmosphere dataset, marking the first application of MAGMA to atmospheric profiles.Results indicate that MAGMA represents a significant advancement towards the efficient application of GPs for extrapolating atmospheric profiles.

The k-means algorithm is used to group objects according to similarity or distance criteria. Due to its simplicity and operating principle, this algorithm has become one of the most widely used for clustering problems. However, despite its popularity, it has several limitations. One of these issues is the excessive convergence time on multidimensional datasets. To address this limitation, several fast variants of the k-means algorithm have emerged. Some optimize the process of updating centroids, while others optimize the process of assigning points to clusters and finally other methods propose to subdivide a dataset into batches and apply the algorithm to various batches to accelerate the speed of convergence of this algorithm. Despite these advancements, existing approaches are not always very optimal, and many of them optimize the algorithm while losing the efficiency of the k-means algorithm. This work address those challenges and proposes a new fast variant of the k-means algorithm capable of significantly optimizing the convergence time of the k-means algorithm while maintaining the efficiency of the naive k-means algorithm. This paper proposes a multi-level optimization of this algorithm. It proposes an efficient heuristic for determining stable and variant points from an iteration. Additionally, it optimizes the process of updating the centroid calculation process, which until now has been done using all the points in the cluster. Finally, it optimizes the assignment process using the notion of neighborhood clusters.

Low-cost gas sensors (LCS) often produce inaccurate measurements due to varying environmental conditions that are not consistent with laboratory settings, leading to inadequate productivity levels compared to high-quality sensors. To address this issue, we propose the use of Machine Learning (ML) to predict accurate concentrations of pollutant gases acquired by LCS integrated into an embedded Internet of Things platform. However, a key challenge is to optimize an accurate ML design under low memory and computation power constraints of microcontrollers (MCUs) while maintaining accurate ML scores. After data analysis and pre-processing, we assess and analyze the performance of five ML algorithms to predict the concentration of pollutants gases from multiple specifications (weather, presence of other gases, etc.). To support the experiments, datasets from three sources are used: (1) VOCSens, (2) Belgian Interregional Environment Agency cell, and (3) Visual-Crossing. Once the best model was optimized and validated, multiple hard constraints were added to the selected ML structure to satisfy material and expert requirements. Trained models were ported to be implemented locally in a MCU after comparing several porting libraries. The assembled code obtained is evaluated based on two metrics: storage memory consumption and inference time, relative to the highest attainable capacities. The improved random forest is the best ML model for the used dataset with an R2 score meeting of 0.72 and Root Means Square Error of 0.0028 ppm. The best generated Tiny-ML model needs 3% of RAM and 98% of Flash storage. The empirical results prove that the developed ML algorithm applied to LCS provides high accuracy to predict pollutant gases. This algorithm can also be used to adjust the LCS systems to provide calibrated data in real time, even if the platform being used is not particularly advanced or powerful.

Research in natural language processing has led to the creation of powerful tools for individuals, companies...However, these successes for written languages have not yet affected signed languages (SLs) to the same extent.The creation of similar tools for signed languages would benefit deaf, hard of hearing, and hearing people by making SL content, learning, and communication more accessible for everyone.SL recognition and translation are related to AI, but require collaboration with linguists and stakeholders.This paper describes related challenges from an AI researcher's point of view and summarizes the state of the art in these domains.

Ockham's razor and the curse of dimensionality are two founding principles in machine learning.First, simple models should be preferred to complex ones, in order to prevent overfitting.Second, highdimensional spaces should be avoided, whenever possible, because learning is easier in lower-dimensional spaces.These principles are often invoked to justify methodological choices or to preprocess data.However, this paper shows a counterexample where it is better to first learn a more complex model in a higher-dimensional space, and then to go back to the lowerdimensional space while dropping the additional complexity.Specifically, experiments demonstrate that Gaussian mixtures models can be learned in a higher-dimensional space and then marginalised to the target dimensionality to improve probability density estimation performances.The chosen problem is deliberately simple to facilitate the analysis, but it opens the way to similar work for more complex models and tasks. 95

Recommender systems are data-driven models that successfully provide users with personalized rankings of items (movies, books...).Meanwhile, for user minority groups, those systems can be unfair in predicting users' expectations due to biased data.Consequently, fairness remains an open challenge in the ranking prediction task.To address this issue, we propose in this paper FairBayRank, a fair Bayesian personalized ranking algorithm that deals with both fairness and ranking performance requirements.FairBayRank evaluation on real-world datasets shows that it efficiently alleviates unfairness issues while ensuring high prediction performances.

The recent advances in deep learning have been beneficial to automatic sign language recognition (SLR). However, free-to-access, usable, and accessible tools are still not widely available to the deaf community. The need for a sign language-to-text dictionary was raised by a bilingual deaf school in Belgium and linguist experts in sign languages (SL) in order to improve the autonomy of students. To meet that need, an efficient SLR system was built based on a specific transformer model. The proposed system is able to recognize 700 different signs, with a top-10 accuracy of 83%. Those results are competitive with other systems in the literature while using 10 times less parameters than existing solutions. The integration of this model into a usable and accessible web application for the dictionary is also introduced. A user-centered human-computer interaction (HCI) methodology was followed to design and implement the user interface. To the best of our knowledge, this is the first publicly released sign language-to-text dictionary using video captured by a standard camera.

When applied to critical domains, machine learning models usually need to comply with prior knowledge and domain-specific requirements. For example, one may require that a learned decision tree model should be of limited size and fair, so as to be easily interpretable, trusted, and adopted. However, most state-of-the-art models, even on decision trees , only aim to maximising expected accuracy. In this paper, we propose a framework in which a diverse family of prior and domain knowledge can be formalised and imposed as constraints on decision trees . This framework is built upon a newly introduced tree representation that leads to two generic linear programming formulations of the optimal decision tree problem. The first one targets binary features , while the second one handles continuous features without the need for discretisation . We theoretically show how a diverse family of constraints can be formalised in our framework. We validate the framework with constraints on several applications and perform extensive experiments, demonstrating empirical evidence of comparable performance w.r.t. state-of-the-art tree learners.

A plethora of dimensionality reduction techniques have emerged over the past decades, leaving researchers and analysts with a wide variety of choices for reducing their data, all the more so given some techniques come with additional hyper-parametrization (e.g., t-SNE, UMAP, etc.). Recent studies are showing that people often use dimensionality reduction as a black-box regardless of the specific properties the method itself preserves. Hence, evaluating and comparing 2D embeddings is usually qualitatively decided, by setting embeddings side-by-side and letting human judgment decide which embedding is the best. In this work, we propose a quantitative way of evaluating embeddings, that nonetheless places human perception at the center. We run a comparative study, where we ask people to select "good" and "misleading" views between scatterplots of low-dimensional embeddings of image datasets, simulating the way people usually select embeddings. We use the study data as labels for a set of quality metrics for a supervised machine learning model whose purpose is to discover and quantify what exactly people are looking for when deciding between embeddings. With the model as a proxy for human judgments, we use it to rank embeddings on new datasets, explain why they are relevant, and quantify the degree of subjectivity when people select preferred embeddings.

Recent technological advances rely on accurate decision support systems that can be perceived as black boxes due to their overwhelming complexity. This lack of transparency can lead to technical, ethical, legal, and trust issues. For example, if the control module of a self-driving car failed at detecting a pedestrian, it becomes crucial to know why the system erred. In some other cases, the decision system may reflect unacceptable biases that can generate distrust. The General Data Protection Regulation (GDPR), approved by the European Parliament in 2018, suggests that individuals should be able to obtain explanations of the decisions made from their data by automated processing, and to challenge those decisions. All these reasons have given rise to the domain of interpretable and explainable AI. AIMLAI aims at gathering researchers, experts and professionals, from inside and outside the domain of AI, interested in the topic of interpretable ML and interpretable AI. The workshop encourages interdisciplinary collaborations, with particular emphasis in knowledge management, Infovis, human computer interaction and psychology. It also welcomes applied research for use cases where interpretability matters. AIMLAI envisions to become a discussion venue for the advent of novel interpretable algorithms and explainability modules that mediate the communication between complex ML/AI systems and users.

Recurrent neural networks can be used to segment sequences such as videos, where transitions can be challenging to detect.This paper benchmarks strategies to better model the transition between states.The specific task of SL video tokenization is chosen for the evaluation, as it remains challenging.Tokenizers are the cornerstone of natural language processing pipelines.There exist powerful tokenizers for text data, but sign language (SL) video tokenizers are still under development.Benchmarked strategies prove to be useful to improve SL videos tokenization, but there is still room for improvement to better model state transitions.

Decision Trees (DTs) and Random Forests (RFs) are popular models in Machine Learning (ML) thanks to their interpretability and efficiency to solve real-world problems.However, DTs may sometimes learn rules that treat different groups of people unfairly, by paying attention to sensitive features like for example gender, age, income, language, etc.Even if several solutions have been proposed to reduce the unfairness for different ML algorithms, few of them apply to DTs.This work aims to transpose a successful method proposed by Zafar et al. [1] to reduce the unfairness in boundary based ML models to DTs.

$t$-Distributed Stochastic Neighbor Embedding ($t$-SNE) is a well-known dimensionality reduction technique used for the visualization of high-dimensional data. However, despite several improvements,$t$-SNE is not well-suited to handle large datasets. Indeed, for large datasets, the computation time required to obtain the visualizations is still too high to incorporate it in an interactive data exploration process. Since$t$-SNE can be seen as an N -body problem in physics, we present a new variant of$t$-SNE based on a popular algorithm used to solve the N -body problem in physics called Particle-Mesh (PM). The problem is solved by first computing a potential in space and deriving from it the force exerted on each body. As the potential can be computed efficiently using Fast Fourier Transforms (FFTs), this leads to a significant speed up. The mathematical correspondence between$t$-SNE and PM presented in this work could also lead to other future improvements since more advanced PM algorithms have been developed in physics for decades.

While significant progress have been made in the field of Natural Language Processing (NLP), leading the commercially available products, Sign Language Recognition (SLR) is still in its infancy. The lack of large-scale sign language datasets makes it hard to leverage new Deep Learning methods. In this paper, we introduce LSFB-CONT, a large scale dataset suited for continuous SLR along with LSFB-ISOL, a subset of LSFB-CONT for isolated SLR. Baseline SLR experiments are conducted on LSFB-ISOL and the reached accuracy measures are compared with those obtained on previous datasets. The results suggest that state-of-the-art models for action recognition still lack sufficient internal representation power to capture the high level of variations of a sign language.

Dimensionality reduction is often used for visualization without considering their understanding by users. Multidimensional scaling, for instance, provides an arbitrarily-oriented visualization. However, users can be integrated into the loop to provide clues about their understanding of the visualization. In this paper, we propose an interactive probabilistic multidimensional scaling (iPMDS) approach to compute the visualization with the lowest information loss while taking the information provided by users into account. We show that a more interpretable visualization can be obtained after interacting with the visualization while keeping a good dimensionality reduction accuracy.

Dimensionality reduction (DR) methods are useful when analyzing high dimensional data, in particular, if one wants to visualize them. t-distributed stochastic neighbor embedding (t-SNE), one of the most widely used DR methods, can preserve neighborhood information and reveal groups in embeddings. However, it may not preserve the global structure and fail to reveal the semantic information in the visualization. From a user point-of-view, a DR visualization is useful if it not only reveals hidden structures in the data but also corresponds to the user knowledge. This paper addresses these problems by proposing Hierarchical Constraint t-SNE (HC$t$-SNE), a method that allows users to integrate hierarchical constraints directly into$t$-SNE embeddings. The user constraints are encoded in an explicit tree. We transform the hierarchical information in this tree into a novel regularization term based on triplet constraints among the nodes at different levels in the tree. Our method takes advantage of semantic information provided in class labels and outperforms the original$t$-SNE and two other supervised DR methods in terms of both visual assessment and quality metrics on three classic image datasets: MNIST, Fashion-MNIST and CIFAR10.

For many applications in image analysis, learning models that are invariant to translations and rotations is paramount. This is the case, for example, in medical imaging where the objects of interest can appear at arbitrary positions, with arbitrary orientations. As of today, Convolutional Neural Networks (CNN) are one of the most powerful tools for image analysis. They achieve, thanks to convolutions, an invariance with respect to translations. In this work, we present a new type of convolutional layer that takes advantage of Bessel functions, well known in physics, to build Bessel-CNNs (B-CNNs) that are invariant to all the continuous set of possible rotation angles by design.

Black-box machine learning models can be extremely accurate. Yet, in critical applications such as in healthcare or justice, if models cannot be explained, domain experts will be reluctant to use them. A common way to explain a black-box model is to approximate it by a simpler model such as a decision tree. In this paper, we propose a co-learning framework to learn decision rules as explanations of black-box models through knowledge distillation and simultaneously constrain the black-box model by these explanations; all of this in a differentiable manner. To do so, we introduce the soft truncated Gaussian mixture analysis (STruGMA), a probabilistic model which encapsulates hyper-rectangle decision rules. With STruGMA, global explanations can be extracted by any rule learner such as decision lists, sets or trees. We provide evidences through experiments that our framework can globally explain differentiable black-box models such as neural networks. In particular, the explanation fidelity is increased, while the accuracy of the models is marginally impacted.

The Third Workshop on "Advances in Interpretable Machine Learning and Artificial Intelligence" (AIMLAI) presents contributions in the fields of (i) interpretable ML and AI, i.e., algorithms that are natively interpretable, and (ii) interpretability modules, i.e., explanation layers on top of black-box models, also called post-hoc interpretability. AIMLAI encourages interdisciplinary collaborations with particular emphasis in knowledge management, infovis, human computer interaction and psychology. It also welcomes applied research for use cases where interpretability matters.

Label noise-tolerant machine learning techniques address datasets which are affected by mislabelling of the instances. Since labelling quality is a severe issue in particular for large or streaming data sets, this setting becomes more and more relevant in the context of life-long learning, big data and crowd sourcing. In this contribution, we extend a powerful online learning method, soft robust learning vector quantisation, by a probabilistic model for noise tolerance, which is applicable for streaming data, including label-noise drift. The superiority of the technique is demonstrated in several benchmark problems.

Extreme learning machines (ELMs) are fast methods that obtain state-of-the-art results in regression. However, they are not robust to outliers and their meta-parameter (i.e., the number of neurons for standard ELMs and the regularization constant of output weights for L2 -regularized ELMs) selection is biased by such instances. This paper proposes a new robust inference algorithm for ELMs which is based on the pointwise probability reinforcement methodology. Experiments show that the proposed approach produces results which are comparable to the state of the art, while being often faster.

Feature selection and taking into account dynamic environments are two important aspects of modern data analysis and machine learning. In particular, performing feature selection on datasets where the latest instances contain more features than the initial ones is a problem that may be encountered in many application areas where new sensors are acquired. This paper proposes a method for incremental feature selection with rankings combining the information extracted before and after the introduction of new features, even when the number of instances that include these new features is small. Results on three real-world datasets show that using the ranking of features on the original, smaller-dimensional dataset improves the feature selection results performed on the new, larger-dimensional dataset.

Linear data transformations constitute essential operations in various machine learning algorithms, ranging from linear regression up to adaptive metric transformation. Often, linear scalings are not only used to improve the model accuracy, rather feature coefficients as provided by the mapping are interpreted as an indicator for the relevance of the feature for the task at hand. This principle, however, can be misleading in particular for high-dimensional or correlated features, since it easily marks irrelevant features as relevant or vice versa. In this contribution, we propose a mathematical formalisation of the minimum and maximum feature relevance for a given linear transformation which can efficiently be solved by means of linear programming. We evaluate the method in several benchmarks, where it becomes apparent that the minimum and maximum relevance closely resembles what is often referred to as weak and strong relevance of the features; hence unlike the mere scaling provided by the linear mapping, it ensures valid interpretability.

Label noise is an important issue in classification, with many potential negative consequences. For example, the accuracy of predictions may decrease, whereas the complexity of inferred models and the number of necessary training samples may increase. Many works in the literature have been devoted to the study of label noise and the development of techniques to deal with label noise. However, the field lacks a comprehensive survey on the different types of label noise, their consequences and the algorithms that consider label noise. This paper proposes to fill this gap. First, the definitions and sources of label noise are considered and a taxonomy of the types of label noise is proposed. Second, the potential consequences of label noise are discussed. Third, label noise-robust, label noise cleansing, and label noise-tolerant algorithms are reviewed. For each category of approaches, a short discussion is proposed to help the practitioner to choose the most suitable technique in its own particular field of application. Eventually, the design of experiments is also discussed, what may interest the researchers who would like to test their own algorithms. In this paper, label noise consists of mislabeled instances: no additional information is assumed to be available like e.g., confidences on labels.