Accession Number:
AD1043682
Title:
Budgeted Interactive Learning
Descriptive Note:
Technical Report,21 May 2015,20 May 2017
Corporate Author:
NATIONAL TAIWAN UNIVERSITY TAIPEI Taiwan
Personal Author(s):
Report Date:
2017-06-15
Pagination or Media Count:
12.0
Abstract:
This project had two questions it was seeking to answer how to enrich the protocols for interactive learning, and how to properly make multi-criteria decisions during the interactive learning process Towards answering the first question, the PIs team broke it into three subareas1 protocols that combine the benefits of online and batch learning, 2 protocols that improve interactive learning with other sources of information, and 3 protocols that allow extracting useful representations during interactive learning. Aligned with the three subareas, they have designed algorithms that allow selecting active learning approaches on the fly for 2 and transferring the selection experience to other active learning tasks for 1, 2, and 3. The selection scheme is implemented and released as an open-source active learning package. They have studied theories for designing algorithms for interactive learning with batch-like feedback for 1 and algorithms for online digestion of representation for 1 and 3. The team has also addressed real-world needs for considering concept drift during online learning for 2 and utilizing costs during deep learning, multi-label learning and active learning for 2 and 3. For the second question, the PIs team has seen promising results on 4 the annotation-budget-sensitive active learning, 5 rethinking deep learning models that trade trainingprediction time with performance in large-scale learning, and 6 label embedding models that trade time embedding length with performance.
Descriptors:
Subject Categories:
- Cybernetics