Mon, Dec 05, 2022 @ 10:00 AM - 11:30 AM
Thomas Lord Department of Computer Science
Committee: Maja Mataric, Fei Sha, Salman Avestimehr, Jesse Thomason, Stefanos Nikolaidis.
Title: Quickly solving new tasks, with meta-learning and without
- Date: Monday 12/5 at 10am PT on
- Abstract (shortened):
The success of modern machine learning (ML) stems from the unreasonable effectiveness of large data. But what about niche tasks with limited data? Some methods are able to quickly solve those tasks by first pretraining ML models on many generic tasks in a way that lets them quickly adapt to unseen new tasks. Those methods are known to ``learn how to learn\'\' and thus fall under the umbrella of meta-learning. While meta-learning can be successful, the inductive biases that enable fast adaptation remain poorly understood.
This thesis takes a first step towards an understanding of meta-learning, and reveals a set of guidelines which help design novel and improved methods for fast adaptation. Our core contribution is a study of the solutions found by meta-learning. We uncover the working principles that let them adapt so quickly: their parameters partition into three groups, one to compute task-agnostic features, another for task-specific features, and a third that accelerates adaptation to new tasks.
Building on those insights we introduce several methods to drastically speed up adaptation.
We propose Kronecker-factored meta-optimizers which significantly improve post-adaptation performance of models that are otherwise too small to meta-learn. We also show how to apply our insights to a visual reinforcement learning setting where meta-learning is impractical. Freezing task-agnostic parameters and adapting task-specific ones with policy-induced self-supervision enables adaptation to unseen tasks with large feature extractors pretrained on generic vision datasets.
Audiences: Everyone Is Invited
Contact: Lizsl De Leon