Timezone: »

 
Variational Task Encoders for Model-Agnostic Meta-Learning
Joaquin Vanschoren

@
in 5th Workshop on Meta-Learning »
Event URL: https://openreview.net/forum?id=dfYhf5IuMPE »

Meta-learning allows an intelligent agent to leverage prior learning episodes as a basis for quickly improving performance on novel tasks. A critical challenge lies in the inherent uncertainty about whether new tasks can be considered similar to those observed before, and robust meta-learning methods would ideally reason about this to produce corresponding uncertainty estimates. We extend model-agnostic meta-learning with variational inference: we model the identity of new tasks as a latent random variable, which modulates the fine-tuning of meta-learned neural networks. Our approach requires little additional computation and doesn't make strong assumptions about the distribution of the neural network weights, and allows the algorithm to generalize to more divergent task distributions, resulting in better-calibrated uncertainty measures while maintaining accurate predictions.

Author Information

Joaquin Vanschoren (Eindhoven University of Technology)
Joaquin Vanschoren

Joaquin Vanschoren is Associate Professor in Machine Learning at the Eindhoven University of Technology. He holds a PhD from the Katholieke Universiteit Leuven, Belgium. His research focuses on understanding and automating machine learning, meta-learning, and continual learning. He founded and leads OpenML.org, a popular open science platform with over 250,000 users that facilitates the sharing and reuse of machine learning datasets and models. He is a founding member of the European AI networks ELLIS and CLAIRE, and an active member of MLCommons. He obtained several awards, including an Amazon Research Award, an ECMLPKDD Best Demo award, and the Dutch Data Prize. He was a tutorial speaker at NeurIPS 2018 and AAAI 2021, and gave over 30 invited talks. He co-initiated the NeurIPS Datasets and Benchmarks track and was NeurIPS Datasets and Benchmarks Chair from 2021 to 2023. He also co-organized the AutoML workshop series at ICML, and the Meta-Learning workshop series at NeurIPS. He is editor-in-chief of DMLR (part of JMLR), as well as an action editor for JMLR and machine learning moderator for ArXiv. He authored and co-authored over 150 scientific papers, as well as reference books on Automated Machine Learning and Meta-learning.

More from the Same Authors