التفاصيل البيبلوغرافية
| العنوان: |
Equitable Data Valuation for Semantic Segmentation |
| المؤلفون: |
Tejomay, Abhiroop, Ramachandra, Vikas |
| بيانات النشر: |
Zenodo |
| سنة النشر: |
2023 |
| المجموعة: |
Zenodo |
| مصطلحات موضوعية: |
Data quality, Shapley, Artificial intelligence, machine learning, semantic image segmentation |
| الوصف: |
In this study, the Shapley value is used to evaluate training data for challenging supervised learning tasks, particularly semantic segmentation. The Shapley value, originating from cooperative game theory is appealing as it satisfies the set of properties required by a data value notion. But, the Shapley value suffers from exponential complexity compute time. Among the approximation algorithms for the Shapley value, the approximation using the K-nearest neighbor (KNN) classifier, called KNN Shapley, is the most efficient as it requires training the model only once while the other approximation algorithms require training the model multiple times. We propose a modification to the KNN Shapley value, to apply to semantic segmentation tasks, providing a measure to determine the quality of data to train semantic segmentation algorithms. We evaluate the values produced by KNN Shapley algorithm in the label noise correction scenario and see that the values produced by the modified algorithm can detect noisy labels |
| نوع الوثيقة: |
report |
| اللغة: |
unknown |
| Relation: |
https://doi.org/10.5281/zenodo.8298884; https://doi.org/10.5281/zenodo.8298885; oai:zenodo.org:8298885 |
| DOI: |
10.5281/zenodo.8298885 |
| الاتاحة: |
https://doi.org/10.5281/zenodo.8298885 |
| Rights: |
info:eu-repo/semantics/openAccess ; Creative Commons Attribution 4.0 International ; https://creativecommons.org/licenses/by/4.0/legalcode |
| رقم الانضمام: |
edsbas.70978AD8 |
| قاعدة البيانات: |
BASE |