Speaker
Kieran Watts
(University of Ottawa)
Description
Active learning (AL) is under-researched for regression problems, with a majority of publications focusing on classification tasks. There is also a need for research concerning AL methods for deep learning. These methods are often approximations by necessity, as AL requires the models to not only predict current performance, but also changes in future performance. Deep learning methods do not directly provide this information.
We investigate recent methods of AL data selection for regression including variance-based query-by-committee, Fisher information matrices, and feature-based largest cluster maximum distance selection. Within the query-by-committee methods, we compare two dropout pseudo-ensemble committee frameworks as well as a true ensemble setup. We test these methods for synthetic regression datasets that increase in complexity to recommend best use cases. We also use our findings of the methods' relative strengths and weaknesses to determine an optimal method for a high-dimensional nanophotonics inverse design project.
| Keyword-1 | Machine Learning |
|---|---|
| Keyword-2 | AI for photonics |
| Keyword-3 | Active learning |
Author
Kieran Watts
(University of Ottawa)
Co-authors
Dr
Colin Bellinger
(University of Ottawa)
Lora Ramunno
(University of Ottawa)