News

• Workshop proceedings are now available as part of Springer CCIS: https://link.springer.com/book/10.1007/978-3-031-74640-6

• The workshop will be held in Room 1T @ Politecnico di Torino, Corso Castelfidardo, 39, Turin, Italy. See map & Useful information

• We welcome Data Science Seed as Bronze Sponsor!

• Paper decisions are out. Please upload final paper versions on CMT by August 4th (11.59PM AoE).

• The invited speakers list has been finalized!

• We welcome ELISE and CINI as a supporting organizations.

About

The 1st International Workshop “Adapting to Change: Reliable Learning Across Domains” will take place as part of ECML-PKDD 2023 in September 2023 in Turin, Italy.

More information about the ECML-PKDD main conference is available on the official website.

Abstract

Most Machine Learning algorithms assume training and test sets to be sampled from the same data distribution. Despite its convenience for analyzing generalization properties, such assumption is easily violated in real-world problems. As a result, the predictive performance of classical methods can be unreliable when deployed in the wild. This is a crucial limitation preventing the application of learning-based solutions to safety-critical settings (e.g., autonomous driving, robotics, medical imaging). At the same time, leveraging data from similar, yet distinct, domains can greatly reduce labeling costs of target applications. This allows powerful, data-hungry deep models to benefit fields with scarce data via pre-training on general-purpose datasets and fine-tuning on smaller problem-specific ones.

The growing demand for reliable and data-efficient learning methods able to generalize across domains has fueled research in Transfer Learning. This includes  Domain Adaptation (DA), which exploits few potentially unlabeled examples from a target domain to adapt models trained on a different source domain, and Domain Generalization (DG), with the purpose of enhancing model robustness to unseen target domain variability. Lastly, many applications require models able to deal with  continuously shifting target distributions, potentially with  novel tasks presented sequentially. This is typically tackled by Continual Learning (CL) methods. Importantly, DA, DG, and CL share many similarities with the Learning-to-Learn framework, which aims at optimizing a learner over a meta-distribution of domains to generalize to unseen ones.

In this workshop, we aim at bringing together researchers across the above fields and the broader ML community to present and discuss recent results in reliable learning across domains, fostering new connections between theory and practical methods, and identifying solutions targeting different modalities (images, videos, language, and more) and application areas.

Institutional Partners

Bronze Sponsor

Special Thanks