Campus St Priest (860 Rue Saint Priest 34095 Montpellier Cedex 5), bat. 5, Room: 02.124
Machine Learning in Montpellier, Theory & Practice
Tiffany Ding (UC Berckley)

In the first part of the talk, I will present some reflections on the purpose of prediction sets and the role that statistics can play in forming useful prediction sets. In particular, I will discuss how prediction sets fit into a decision making pipeline and the different kinds of decisions one may make using a prediction set. In the second part of the talk, I will describe a particular statistically motivated set-generating procedure for the classification setting called clustered conformal prediction, which gives all classes an equal chance of being correctly included in the prediction set (“class-conditional coverage”). This procedure can be useful in situations where it is important to identify instances of all classes, even the rare ones. We demonstrate the performance of this method on ImageNet and other image classification datasets.

LabéliséHallesIA

Campus St Priest (860 Rue Saint Priest 34095 Montpellier Cedex 5), bat. 5, Room: 02.124
Machine Learning in Montpellier, Theory & Practice
Jacob Feitelberg

We study the problem of distributional matrix completion: Given a sparsely observed matrix of empirical distributions, we seek to impute the true distributions associated with both observed and unobserved matrix entries. This is a generalization of traditional matrix completion where the observations per matrix entry are scalar-valued. To do so, we utilize tools from optimal transport to generalize the nearest neighbors method to the distributional setting. Under a suitable latent factor model on probability distributions, we establish that our method recovers the distributions in the Wasserstein metric. We demonstrate through simulations that our method (i) provides better distributional estimates for an entry compared to using observed samples for that entry alone, (ii) yields accurate estimates of distributional quantities such as standard deviation and value-at-risk, and (iii) inherently supports heteroscedastic distributions. In addition, we demonstrate our method on a real-world quarterly earnings predictions dataset. We also prove novel asymptotic results for Wasserstein barycenters over one-dimensional distributions.

LabéliséHallesIA

Inria Montpellier, St-Priest Campus, Building 5, Room 02/022
Machine Learning in Montpellier, Theory & Practice
Toby Hocking

Data.table is one of the most efficient open-source in-memory data manipulation packages available today. First released to CRAN by Matt Dowle in 2006, it continues to grow in popularity, and now over 1500 other CRAN packages depend on data.table. This talk will start with data reading from CSV, discuss basic and advanced data manipulation topics, and finally will end with a discussion about how you can contribute to data.table.

https://github.com/tdhock/2023-10-LatinR-data.table?tab=readme-ov-file#source-files-for-latinr-datatable-tutorial-slides

LabéliséHallesIA

Room Nadir, Maison de la Télédétection (500 rue Jean François Breton)
Machine Learning in Montpellier, Theory & Practice - Lars Van der Laan

We introduce efficient plug-in (EP) learning, a novel framework for the estimation of heterogeneous causal contrasts, such as the conditional average treatment effect and conditional relative risk. The EP-learning framework enjoys the same oracle-efficiency as Neyman-orthogonal learning strategies, such as DR-learning and R-learning, while addressing some of their primary drawbacks, including that (i) their practical applicability can be hindered by loss function non-convexity; and (ii) they may suffer from poor performance and instability due to inverse probability weighting and pseudo-outcomes that violate bounds. To avoid these drawbacks, EP-learner constructs an efficient plug-in estimator of the population risk function for the causal contrast, thereby inheriting the stability and robustness properties of plug-in estimation strategies like T-learning. Under reasonable conditions, EP-learners based on empirical risk minimization are oracle-efficient, exhibiting asymptotic equivalence to the minimizer of an oracle-efficient one-step debiased estimator of the population risk function. In simulation experiments, we illustrate that EP-learners of the conditional average treatment effect and conditional relative risk outperform state-of-the-art competitors, including T-learner, R-learner, and DR-learner. Open-source implementations of the proposed methods are available in our R package hte3.

LabéliséHallesIA

Room Nadir, Maison de la Télédétection (500 rue Jean François Breton)
Machine Learning in Montpellier, Theory & Practice - Martin Atzmüller, Scientific Director at DFKI and Full Professor at Osnabrück University (Germany)

In many applications, modeling complex data is of utmost importance, requiring the use of advanced machine learning models and approaches. However, in many domains users require insight into models and/or their decisions, which is not necessarily provided by the respective models per se. Explainable and interpretable learning approaches can facilitate such insights for making sense of models and decisions. The talk presents examples of such approaches in complex modeling domains, including interpretable as well as explainable deep-learning-based methods, and a neuro-symbolic architecture including domain knowledge for facilitating explainability.

LabéliséHallesIA

Inria Montpellier, St-Priest Campus, Building 5, Room 02/124
Machine Learning in Montpellier, Theory & Practice - Vincent Mallet

Learning on the 3D structure of biomolecules resulted in several
breakthroughs, for instance the de-novo design of potent binders or
functional luciferases, eventually leading to a Nobel Prize. Most of
this work relies on modeling the structure of protein as
graphs, with well-established tools.

IA&Experts, LabéliséHallesIA

Campus St Priest (860 Rue Saint Priest 34095 Montpellier Cedex 5), bat. 5, Room: 02.124
Machine Learning in Montpellier, Theory & Practice - Olivier Cappé (CNRS / ENS / Université PSL)

Performative learning addresses the increasingly pervasive situations in which algorithmic decisions may induce changes in the data distribution as a consequence of their public deployment. We propose a novel view in which these performative effects are modelled as push-forward measures. This general framework encompasses existing models and enables novel performative gradient estimation methods, leading to more efficient and scalable learning strategies. For distribution shifts, unlike previous models which require full specification of the data distribution, we only assume knowledge of the shift operator that represents the performative changes. Focusing on classification with a linear-in-parameters performative effect, we prove the convexity of the performative risk under a new set of assumptions. We also establish a connection with adversarially robust classification by reformulating the minimization of the performative risk as a min-max variational problem.

IA&Experts