Explore the cutting-edge concepts of flexible neural networks and meta-learning in this 50-minute lecture by Chelsea Finn from Stanford University. Delve into the challenges of enabling agents to learn skills in the real world, focusing on few-shot image classification as a key example. Examine the meta-learning problem from both mechanistic and probabilistic perspectives, and understand how supervised learning relates to few-shot learning. Discover optimization-based inference techniques and learn how to leverage data from previous objects for quick adaptation to new ones. Gain insights into the practical implementation of FTML (Few-Shot Task-Agnostic Meta-Learning) and analyze experimental results in this thought-provoking talk from the Simons Institute's "Emerging Challenges in Deep Learning" series.
Flexible Neural Networks and the Frontiers of Meta-Learning
Overview
Syllabus
Intro
How can we enable agents to learn skills in the real world?
Example: Few-Shot Image Classification
The Meta-Learning Problem: The Mechanistic View
The Meta-Learning Problem: The Probabilistic View Supervised Learning
Few-Shot Learning
Optimization-Based Inference
Leverage data with previous objects to quickly adapt to new ones?
Practical instantiation of FTML
Experiments
Taught by
Simons Institute
Related Courses
-
Introduction to Scientific Machine Learning
-
Unsupervised Meta-Learning for Few-Shot Image Classification
-
Transfer Learning Larochelle
-
Learning Exploration Strategies with Meta-Reinforcement Learning
-
IMAML- Meta-Learning with Implicit Gradients
-
MetaHKG: Meta Hyperbolic Learning for Few-shot Temporal Reasoning - M1.2
