Flexible Neural Networks and the Frontiers of Meta-Learning
summary
Shows a concrete method — MAML — that learns an initialization so models fine-tune on one to five gradient steps and solve true few-shot problems (5-way 1-shot classification, colored/rotated MNIST, CIFAR100 classes added over time). Demonstrations include a real robot learning spatula manipulation from pixels but failing to generalize until meta-training, a six‑leg robot that adapts in milliseconds after a leg removal by learning dynamics, and placing objects from human videos. Theory claims logarithmic regret for an online meta-learning procedure under convexity; a 50-task manipulation benchmark is released.
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:24How can we enable agents to learn skills in the real world?
- 8:25Example: Few-Shot Image Classification
- 10:10The Meta-Learning Problem: The Mechanistic View
- 12:35The Meta-Learning Problem: The Probabilistic View Supervised Learning
- 16:10Few-Shot Learning
- 19:46Optimization-Based Inference
- 30:09Leverage data with previous objects to quickly adapt to new ones?
- 41:38Practical instantiation of FTML
- 42:47Experiments
related talks
46mIncreasing Participant Agency by Understanding the Impact of Training Data
47mMeta-learning of Optimizers and Update Rules
38mPractical Few-Shot Learning
01h 04mPredicting and optimizing the behavior of large ML models
10mMIND MELD: Personalized Meta-Learning for Robot-Centric Imitation Learning
55mNetworks of Spiking Neurons Learn to Learn and Remember
49mML: Teaching machine to learn from its own mistakes (...)
22mA Hybrid Meta-Learning Framework for Adaptive Safe Controller Synthesis of Dynamical(…)