Program

This is an evolving program. Schedule only contains title, abstract are at the bottom.

List of confirmed participants so far in no particular order (more to come):

• Julian Lienen (Paderborn University)

• Willem Waegeman (Ghent University)

• Sébastien Destercke (Compiègne University)

• Thomas Mortier (Ghent University)

• Henrik Bostrom (KTH royal institute of technology) - online

• Cassio de Campos (Eindhoven University)

• Benjamin Quost (Compiègne University)

• Soundouss Messoudi (Compiègne University)

• Titouan Lorieul (INRIA - Montpellier University)

• Loic Adam (Compiègne University)

• Nicolas Dewolf (Ghent University)

• Jelle Hustinx (Ghent University)

• Cyprien Gillet (Nice University)

• Mohammad Hossein Shaker (Paderborn University)

• Maryam Tavakol (Eindhoven University)

• Vu-Linh Nguyen (Eindhoven University)

• Thomas Krak (Eindhoven University)

• Erik Quaeghebeur (Eindhoven University)

• Gennaro Gala (Eindhoven University)

• David Montalvan (Eindhoven University)

• Andrea Campagner (MIlan University)

• Yonatan Carlos Alarcon (Compiègne University)

• Mathieu Randon (Compiègne University)

• ...

List of talks (click on title to get the presentation):

• Thomas Mortier
  Title: "Efficient set-valued prediction in multi-class and hierarchical classification"
  Asbtract: "In the past, a couple of frameworks have been proposed for set-valued prediction in multi-class classification, where — in case of uncertainty — a classifier preferably returns a set of candidate classes instead of predicting a single class label with little guarantee. The idea of the talk is to discuss an extension of those frameworks to the setting where a hierarchy is given over the classes and motivate this extension by means of representational and computational complexity."

• Nicolas Dewolf (20 minutes)
  Title: "Well-calibrated prediction intervals for regression problems"
  Abstract: "Uncertainty has always been a central concept in statistics. However, in machine learning, uncertainty used to be considered more as a secondary notion. Except for probabilistic classifiers, the study of uncertainty is a rather recent development. The last couple of years some new methods were introduced that can estimate how confident they are in their predictions. In this talk we compare some of these methods based on how reliable and efficient the estimates are."

• Julian Lienen (30 minutes)
  Title: "Credal Target Modeling: From Label Relaxation to Credal Self-Supervised Learning"
  Abstract: "When training probabilistic classifiers, e.g., deep neural networks for image classification, conventional methods typically rely on precise probability distributions as target information, serving as surrogates for underlying ground-truth conditional class probabilities. As this commitment to single probabilistic surrogates is likely to entail biases, we suggest to represent our knowledge by more cautious yet reliable sets of distributions, i.e., credal sets. This talk discusses two concepts of modeling such credal labels, namely deliberately imprecising hitherto precise data in supervised learning (label relaxation) and successively precising initially vague beliefs in semi-supervised learning (credal self-supervised learning). "

• Mohammad Hossein Shaker
  Title: "Credal Uncertainty Representation in Machine Learning: an Empirical Comparison"
  Abstract: "The idea to distinguish and quantify two important types of uncertainty, often referred to as aleatoric and epistemic, has received increasing attention in machine learning research in the last couple of years. In this talk, we consider ensemble-based approaches to uncertainty quantification. Distinguishing between different types of uncertainty-aware learning algorithms, we specifically focus on credal set creation and credal based uncertainty quantification, which naturally suggest themselves from an ensemble learning point of view, in comparison with Bayesian methods. For both approaches, we address the question of how to quantify aleatoric and epistemic uncertainty. The effectiveness of corresponding measures is evaluated and compared in an empirical study on classification with a reject option."

• Titouan Lorieul
  Title: "Classification Under Ambiguity: When Is Average-K Better Than Top-K?"
  Abstract: "In presence of ambiguity, many labels are possible and choosing a single one can lead to low precision. A common alternative, referred to as top-K classification, is to choose some number K (commonly under 10) and to return the K labels with the highest scores. Unfortunately, for unambiguous cases, K > 1 is too many and, for very ambiguous cases, K <=5 (for example) can be too small. Another sensible strategy is to use an adaptive approach in which the number of labels returned varies as a function of the computed ambiguity, but must average to some particular K over all the samples. We denote this average-K classification. In this talk, we will formally characterize the ambiguity profile when average-K classification can achieve a lower error rate than a fixed top-K classification. Moreover, we will provide natural estimation procedures for both the fixed-size and the adaptive classifier and prove their consistency. Finally, we will report experiments on real-world image data sets revealing the benefit of average-K classification in practice. Overall, when the ambiguity is known precisely, average-K is never worse than top-K, and, in our experiments, when it is estimated, this also holds."

• Cyprien Gilet
  Title: "Learning minimax classifiers under label shift"
  Abstract: "Our objective is to build a new supervised classifier for addressing the following difficulties that commonly occur in several real application fields like precision medicine: imbalanced datasets, prior probability shifts, presence of both numeric and categorical features, and dependencies between some features. To this aim, we develop a novel minimax classifier that addresses all the previously mentioned issues. This classifier aims to minimize the maximum of the class-conditional risks and becomes robust face to prior probability shifts. In order to facilitate the task of dealing with both categorical and numeric features, we beforehand discretize the numeric attributes so that we only deal with discrete features. This allows us to analytically calculate the empirical Bayes risk over the simplex as a function of the priors. We then compute the least favorable priors that maximize this empirical Bayes surface using a projected sub-gradient algorithm for which the convergence is established. If the experts of the application domain are able to provide independent bounds on the uncertainty of some class proportions, our approach can take into account these constraints to decrease the global risk of error. We moreover show that our approach can be tuned ton compute a minimax regret classifier, which is particularly appropriate for dealing with prior probability shifts."

• Gennaro Gala (40 minutes)
  Title: "Disentangled Representation Learning"
  Abstract: "While it is clear that the usefulness of a representation learned on data heavily depends on the end task which it is to be used for, one could imagine that there exist properties of representations which are useful for many real-world (downstream) tasks simultaneously. In Disentangled Representation Learning, the key underlying assumption is that high-dimensional observations (such as images) are actually a manifestation of a low-dimensional set of independent ground-truth factors of variation. Therefore, according to this assumption, it is possible to identify a generative process which can synthesize new observations starting from few independent and interpretable factors of variation. But how we can learn disentangled representations? "

• Soundouss Messoudi (30 minutes)
  Title: "Conformal regression for multi-variate regression problems"
  Abstract: "There are relatively few works dealing with conformal prediction for multi-target regression. This talk presents our work that tackles this problem, by exploring multiple extensions of existing single-target non-conformity measures and proposing new ones, in order to provide valid multi-variate predictions. It mainly focuses on the use of copula functions for inductive conformal prediction, and how taking into consideration the dependence structure ensures efficiency and validity for multi-target regression."

• Cassio de Campos (40 minutes)
  Title: "Probabilistic Graphical Models: Tractability and Robustness"
  Abstract: "This talk will present a view on the theoretical and practical tractability of some probabilistic graphical models such as Bayesian and Markov networks. We will discuss on relations among different graphical models, including models that represent computations explicitly, such as probabilistic circuits. We will also dive into ideas of cautiousness and robustness in AI based on credal graphical models. Finally, I will give a (considerably biased) opinion on how artificial intelligence (AI) is evolving and what we can expect regarding probabilistic graphical models in the next generation of AI. "

• Henrik Bostrom (40 minutes)
  Title: "Conformal Regressors and Conformal Predictive Systems"
  Abstract: "Conformal prediction is a general framework that allows for controlling the error rate of any predictive model, by turning point predictions into set predictions. Approaches to applying the framework to regression problems will be outlined and some of their distinguishing features will be highlighted. Results from empirical investigations, primarily using random forests as the underlying model, will be presented. The recently proposed framework of conformal predictive systems, which output probability distributions rather than set predictions, will also be discussed and it will be shown how the framework generalizes that of conformal regression and even can be used for improving the point predictions of the underlying regression model."

• Willem Waegeman (40 minutes)
  Title: "Classifier rejection and set-valued prediction under distribution shift"

• Andrea Campagner (35 minutes)
  Title: "Rough Set-based Feature Selection for Weakly Labeled Data"
  Abstract: Supervised learning is an important branch of machine learning, which requires a complete annotation (labeling) of the involved training data. This assumption is relaxed in weakly supervised learning, where labels are allowed to be imprecise or partial. In this talk, I'll address the setting of superset learning, in which instances are assumed to be labeled with a set of possible annotations containing the correct one. We tackle the problem of learning from such data in the context of Rough Set Theory (RST). More specifically, we consider the problem of RST-based feature reduction as a suitable means for both data disambiguation (i.e., for the purpose of figuring out the most plausible precise instantiation of the imprecise training data) and feature selection (i.e. to reduce the dimensionality and avoid complexity or overfitting problem). To this end, we define appropriate generalization of basic RST notions (e.g. decision tables and reducts), using tools from generalized information theory and belief function theory. Moreover, we analyze the computational complexity and theoretical properties of the associated computational problems. Finally, I'll present the results of a series of experiments, in which we analyzed the proposed concepts empirically and compared our methods with a state-of-the-art dimensionality reduction algotrithm, reporting a statistically significant improvement in predictive accuracy.

• Mathieu Randon (20 minutes)
  Title: "Recalibrated Gaussian Process Regression"
  Abstract: In the industry, supervised learning builds models from field-observed data. However, models are often limited by the amount and quality of available data. In that case, additional knowledge, such as physical bounds or information regarding measurement noise, can be considered so as to improve the robustness of the model. This paper proposes to use such information in the case of Gaussian Process Regression. The resulting model can be seen as a correction of the initial model given the additional information. Experiments on estimating the energy consumption of a simulated electrified vehicle show the interest of our approach in order to improve robustness against scarce and noisy data.