Meta-Learning for Versatile Artificial Intelligence
Event details
| Date | 17.07.2019 |
| Hour | 14:00 › 16:00 |
| Speaker | Arnout Devos |
| Location | |
| Category | Conferences - Seminars |
EDIC candidacy exam
Exam president: Prof. Patrick Thiran
Thesis advisor: Prof. Matthias Grossglauser
Co-examiner: Prof. Martin Jaggi
Abstract
Humans can learn new concepts from only a few examples and quickly adapt to unforeseen circumstances. To this end, they build upon their prior experience and prepare for the ability to adapt. Most machine learning systems, however, lack the versatility to learn new tasks efficiently and effectively when little data is available. Learning to learn, or meta-learning, is a paradigm which successfully trains machine learning models for effective adaptation with little new task data. In this work we discuss fundamental and recent advances in meta-learning and few-shot learning. The model-agnostic meta-learning (MAML) method and the differentiable plasticity method are discussed in detail. We review our published and ongoing work on meta-learning, reaching state-of-the-art performance, and propose possible future research directions.
Background papers
A Closer Look at Few-shot Classification, by Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, Jia-Bin Huang, ICLR 2019.
Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks, by Chelsea Finn, Pieter Abbeel, Sergey Levine, ICML 2017.
Differentiable plasticity: training plastic neural networks with backpropagation, by Thomas Miconi, Kenneth Stanley, Jeff Clune, ICML 2018.
Exam president: Prof. Patrick Thiran
Thesis advisor: Prof. Matthias Grossglauser
Co-examiner: Prof. Martin Jaggi
Abstract
Humans can learn new concepts from only a few examples and quickly adapt to unforeseen circumstances. To this end, they build upon their prior experience and prepare for the ability to adapt. Most machine learning systems, however, lack the versatility to learn new tasks efficiently and effectively when little data is available. Learning to learn, or meta-learning, is a paradigm which successfully trains machine learning models for effective adaptation with little new task data. In this work we discuss fundamental and recent advances in meta-learning and few-shot learning. The model-agnostic meta-learning (MAML) method and the differentiable plasticity method are discussed in detail. We review our published and ongoing work on meta-learning, reaching state-of-the-art performance, and propose possible future research directions.
Background papers
A Closer Look at Few-shot Classification, by Wei-Yu Chen, Yen-Cheng Liu, Zsolt Kira, Yu-Chiang Frank Wang, Jia-Bin Huang, ICLR 2019.
Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks, by Chelsea Finn, Pieter Abbeel, Sergey Levine, ICML 2017.
Differentiable plasticity: training plastic neural networks with backpropagation, by Thomas Miconi, Kenneth Stanley, Jeff Clune, ICML 2018.
Practical information
- General public
- Free