Tutorial on understanding meta-learning for fast adaptation
Kavya Ranjan Saxena, Vipul Arora · 2023
Deep learning methods provide excellent performance when sufficiently large annotated data is available. But their performance degrades when the models are applied to different target domains with very different data distributions as compared to that of the source domain. The performance of such models can be improved by domain adaptation. In this tutorial, we discuss two domain adaptation techniques, i.e. fine-tuning and meta-learning, with a methodical analysis of meta-learning based adaptation. Meta-learning works in a manner similar to how humans take in a handful of examples and learn new skills very quickly and efficiently. Generally, data-level and parameter-level approaches are considered when solving a few-shot learning problem. Meta-learning is one of the parameter-level approaches. The tutorial is targeted toward those who wish to gain comprehensive insight and practical experience in applying meta-learning for fast adaptation. We show applications of these techniques through Jupyter notebooks, demonstrating how core concepts can be translated into empirical work. For those who are new to this area, we will provide a hands-on session and online pedagogical guide that can serve as the basis for fostering future research. We hope to convey a comprehensive understanding of recent advances and current state-of-the-art approaches to those who have prior knowledge of the areas. We would expect some basic experience in the execution of Jupyter notebooks and some basic knowledge of concepts related to machine learning as a prerequisite for participation.