(T) One of the key goals of many data scientist teams is to develop models faster for new use cases and new applications. To that end, “transfer learning” which aims to train on one task, and to transfer that learning to a new task has been widely used in particular in computer vision, and language models.

However, two new techniques have emerged recently: “multi-task learning” which aims to train a system on many tasks, and transfer that learning to the system for a new task, and “meta-learning” which empowers a system to learn to learn from many tasks, and transfer that learning to the system for similar tasks.

Multi-task and meta-learning seem to have found a sweet spot in deep reinforcement learning applications such as robotics and games.

Well-known researchers, who are pioneering multi-task and meta-learning, include Professor Sergey Levine and Professor Pieter Abbeel from UC Berkeley, and Professor Chelsea Finn from Stanford University.

Following is a gentle introduction to multi-task and meta-learning: Multi-Task and Meta-Learning, and some videos from Professor Finn and Levin’s classes:

Professor Chelsea Finn, Stanford University, CS 229, Supervised multi-task learning, black-box meta-learning:

Professor Chelsea Finn, Stanford University, CS 229, Reinforcement learning primer, multi-task RL, goal-conditioned RL:

Professor Chelsea Finn, Stanford University, CS 229, Guest Lecture Meta-RL, learning to explore:

Professor Chelsea Finn, Stanford University, CS 229, Model-based RL for multi-task learning, meta model-based RL:

Professor Sergey Levine, UC Berkeley, CS294 – Transfer learning and introduction to multi-task learning:

Professor Sergey Levine, UC Berkeley, CS 294 – Multi-task learning and transfer learning:

Professor Sergey Levine, UC Berkeley, CS 294 – Meta-learning and parallelism:

Note: The picture above is “Image à La Maison Verte” a painting from René Magritte.

Copyright © 2005-2020 by Serge-Paul Carrasco. All rights reserved.
Contact Us: asvinsider at gmail dot com