Can You Learn How To Teach?
3 main points
✔️ Learning to Teach accelerates learning
✔️ Meta-learning generates synthetic data that allows the network to learn faster
✔️ Learning of neural networks 9 times faster
Hierarchically Decoupled Imitation for Morphological Transfer
written by Donald J. Hejna III, Pieter Abbeel, Lerrel Pinto
(Submitted on 3 Mar 2020 (v1), last revised 31 Aug 2020 (this version, v2))
Comments: Accepted at ICML2020
Subjects: Machine Learning (cs.LG); Artificial Intelligence (cs.AI); Robotics (cs.RO); Machine Learning (stat.ML)
Can you learn to teach?
The paper presented in this article is a study that addresses this challenging question. This question is similar but different from general meta-learning. Much of the existing meta-learning is focused on "learning how to learn," as it is described as "learning to learn.
In other words, the biggest problem is "how well can we learn from the data we are given".
If I had to compare it to studying.
When a textbook is given to students, how much more knowledge can they gain from it?
When students are given a task, how well can they solve it?
That was the goal.
In other words, the emphasis was on refining the "learners". However, is it really only the learners who matter? As long as a student excels, does it matter if the textbook or assignment given to that student isn't excellent? Intuitively, I wouldn't say that's the case. Of course, a good student can gain knowledge successfully from a textbook, no matter how difficult it is to understand. However, students will still learn more efficiently if they are given a textbook that is easy to understand.
Generative Teaching Networks (GTNs), introduced in this article, are a true "learning how to teach" method that allows us to create "easy-to-understand textbooks" in this analogy.
The goal of GTN is to generate synthetic data that can be trained efficiently. The paper discusses the application of this technique to speed up Network Architecture Search (NAS). NAS is a technique that also learns the structure of the neural network itself and builds an optimal network for solving a task. In order to evaluate the performance of the generated network architecture, it is necessary to either actually train the data or prepare models to predict the performance of the architecture. Naturally, such architectural performance estimates are very expensive to learn. GTN generates synthetic data that can be trained on a small amount of data and still perform as well as if it were trained on a large amount of data. This can be used to reduce the amount of training required for performance estimation in NAS. Experiments in the paper showed that it speeds up neural network learning by a factor of nine.
To read more,
Please register with AI-SCHOLAR.OR
Categories related to this article