Title

MPS: Metric-aware Pseudo-label Selection Framework via Reinforcement Learning

DGIST Authors

Seong Gyeom Kim ; Jin Ho Chang ; Sung Hoon Im

Advisor

장진호

Co-Advisor(s)

Sung Hoon Im

Issued Date

2025

Awarded Date

2025-02-01

Citation

Seong Gyeom Kim. (2025). MPS: Metric-aware Pseudo-label Selection Framework via Reinforcement Learning. doi: 10.22677/THESIS.200000839746

Type

Thesis

Description

Machine Learning

Table Of Contents

Ⅰ. Introduction
1 Introduction 1
ⅠI. Related work
2.1 Pseudo-label 3
2.2 Reinforcement learning 4
ⅠII. Method
3.1 Pre-trained teacher model optimize and pseudo labeling 8
3.2 Training data sampling by contribution 9
3.2.1 Task predictor optimization 9
3.2.2 Controller optimization 10
3.3 The reinforcement learning algorithm 10
3.4 Data contribution evaluation using reinforcement learning 14
IV. Experiments
4.1 Implement details 15
4.2 Result 17
4.2.1 SSL with selected data 17
4.2.2 Analysis 19
4.2.3 Discussion 20
V. Conclusion
5 Conclusion 22

URI

http://hdl.handle.net/20.500.11750/58098
http://dgist.dcollection.net/common/orgView/200000839746

DOI

10.22677/THESIS.200000839746

Degree

Master

Department

Artificial Intelligence Major

Publisher

DGIST

Show Full Item Record