Room 02.124, Building 5, St Priest campus Machine Learning in Montpellier, Theory & Practice The use of deep neural networks in critical domains, such as aviation, is limited by our ability to ensure their correct behavior. Execution monitors are components aimed at identifying dangerous predictions and discarding them before they lead to catastrophic consequences. Several recent works on real-time monitoring have focused on detecting out-of-distribution (OOD) inputs, i.e., identifying inputs that differ from the training data. In this presentation, we will show that OOD detection is not a well-suited framework for designing effective execution monitors and that it is more relevant to evaluate monitors based on their ability to discard incorrect predictions. We call this paradigm out-of-model-scope (OMS) detection and discuss the conceptual differences with OOD. We will also present in-depth experiments to show that studying monitors in the OOD framework can be misleading: 1. very good OOD results can give a false sense of security, 2. an OOD-based comparison may not identify the best monitor for detecting errors. Machine Learning in Montpellier, Theory & Practice

Room 109, IMAG, Triolet campus Machine Learning in Montpellier, Theory & Practice We study multitask online learning in a setting where agents can only exchange information with their neighbors on an arbitrary communication network. We introduce MT-CO2OL, a decentralized algorithm for this setting whose regret depends on the interplay between the task similarities and the network structure. Our analysis shows that the regret of MT-CO2OL is never worse (up to constants) than the bound obtained when agents do not share information. On the other hand, our bounds significantly improve when neighboring agents operate on similar tasks. In addition, we prove that our algorithm can be made differentially private with a negligible impact on the regret when the losses are linear. Finally, we provide experimental support for our theory. Arxiv: https://isdm.umontpellier.fr/arxiv.org/abs/2310.17385 Machine Learning in Montpellier, Theory & Practice

Room 01.124, Building 5, St Priest campus Machine Learning in Montpellier, Theory & Practice ,, Machine Learning in Montpellier, Theory & Practice

Room 109, IMAG, Triolet campus Machine Learning in Montpellier, Theory & Practice Machine Learning in Montpellier, Theory & Practice

Room 01.124, Building 5, St Priest campus Machine Learning in Montpellier, Theory & Practice Machine learning models are frequently used when inputs are missing during training or prediction, potentially leading to increased bias or impractical models without imputing unobserved variables. Imputing missing values is often inadequate and hard to interpret, especially with complex functions. This talk addresses the challenge of predicting with missing data at test time, highlighting the need for interpretable and practical models, crucial in critical sectors like healthcare. In this talk, I present two novel approaches: the Shared Pattern Sparsity Model (SPSM) for scenarios with recurrent missing data patterns, which promotes efficient data use and interpretability without reliance on imputation; and MINTY, a sparse linear rule model, regularized to minimize dependence on features with missing values. This model allows a trade-off between goodness of fit, interpretability, and robustness to missing values at test time. Additionally, I'll share early results from a project developing a predictive model for sequential risk scores sensitive to missing values across time steps. In collaboration with the Traumabase network, we'll conduct a user study with clinical professionals to assess the effectiveness of interpretable models in handling missing data. Machine Learning in Montpellier, Theory & Practice

Room 109, IMAG, Triolet campus Machine Learning in Montpellier, Theory & Practice We will present the recent Active Clustering Problem (ACP). In this problem, a set of items can be partitioned into groups where items within the same group are characterised by the same multi-dimensional vector. A learner obtains noisy observations of these vectors, and we consider an active setting where the learner chooses the order and the number of observations. The objective is to recover the hidden partition of the items, using as few requests as possible.In the presentation, I will explain the ACP and answer two questions. Can we improve upon the number of requests of the simple uniform sampling algorithm, using the benefits of active sampling ? Is there a fundamental computation-information gap for clustering in high-dimension with repeated measurements? Machine Learning in Montpellier, Theory & Practice