Week 1 #
Topic: An Introduction of Model-Agnostic Meta-Learning: Principles, Mechanisms, and Variants
Keynote Speaker: Wang Ma
Time: Jun 21, 11:00 - 12:30 am
Venue: Business Hall, 314 (SUSTech)
Online Link: TencentMeeting
Compendium #
I. Introduction: The Concept and Mechanism of Meta-Learning (Credits: Prof. Hung-yi Lee’s Slides)
- Discussing the concept and mechanisms of Meta-Learning.
- Exploring its distinctive role and divergence from traditional Machine Learning approaches.
II. The Emergence of MAML
- Providing an overview of the Model-Agnostic Meta-Learning (MAML) Algorithm.
- Offering an intricate explanation of MAML’s workings, augmented by sketch illustrations.
- Demonstrating the versatility of MAML through application examples such as classification, regression, and reinforcement learning problems.
III. First-Order MAML (FO-MAML)
- Illuminating the presence of the second derivative in the MAML Algorithm through mathematical derivation.
- Introducing the First-Order Model-Agnostic Meta-Learning (FO-MAML), with a focus on its divergence from the MAML algorithm (which part of the originial algorithm is ignored).
- Rewriting the FO-MAML algorithm to better illustrate its structure and functionality.
IV. The Advent of Reptile
- Re-examining the MAML optimization problem and diving deeper into FO-MAML.
- Introducing Reptile, with a clear delineation of its distinction from FO-MAML.
- Rewriting the Reptile algorithm for a better understanding.
V. Feature Reuse in MAML & Introduction to ANIL (Almost No Inner Loop)
- Discussing why MAML’s efficiency is amplified by Feature Reuse rather than Rapid Learning, supplemented with the explanation and analysis of three experimental case studies.
- Explaining how the concept of feature reuse spearheaded the idea of ANIL.
- Providing a comprehensive explanation of ANIL, its unique algorithm, and how it differentiates from MAML.
VI. Conclusion
- Recapping the key concepts, techniques, and distinctions between MAML and its variants, emphasizing their potential impact on the future of machine learning.
Material #
The first week slide from Wang Ma.
Slide-ref is reference slide from hungyi-lee.
References #
Chelsea Finn’s blog on Learning to Learn
Chelsea Finn et al. Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks (MAML)
Alex Nichol et al. On First-Order Meta-Learning Algorithms (Reptile)
Aniruddh Raghu et al. Rapid Learning or Feature Reuse? Towards Understanding the Effectiveness of MAML
Chelsea Finn. Learning to Learn with Gradients