Related Experiment Videos
Mislabel Identification Using Transfer Learning-Based Ensemble Method
Md Shariful Islam1, Min Jun Kim2, Prashanta Dutta1
1School of Mechanical and Materials Engineering, Washington State University, Pullman, WA 99164, USA.
IEEE Access : Practical Innovations, Open Solutions
|May 11, 2026
Summary
This study introduces a new ensemble method using transfer learning to accurately identify and correct mislabeled data in machine learning. The approach significantly improves data labeling accuracy for sensitive applications like virus classification and medical diagnostics.
Area of Science:
- Machine Learning
- Data Science
- Bioinformatics
Background:
- Accurate data labeling is crucial for supervised machine learning, yet it is labor-intensive and prone to errors.
- Existing datasets, including MNIST and ImageNet, contain mislabeled samples, impacting model reliability.
- Sensitive applications like medical diagnostics and virus classification demand high data integrity.
Purpose of the Study:
- To develop and validate a robust method for identifying and correcting mislabeled training data.
- To enhance the reliability of machine learning models in critical applications.
- To establish a superior framework for mislabel detection in complex datasets.
Main Methods:
- A transfer learning-based ensemble method employing majority and consensus filtering.
- Utilized fine-tuned deep neural networks: ResNet-50, ResNet-101, VGG-16, EfficientNet, MobileNet, and Inception.
- Validated on MNIST, synthetically corrupted datasets, and adeno-associated virus (AAV) nanopore data.
Main Results:
- The ensemble method detected approximately 751 label inconsistencies in the MNIST dataset.
- Achieved up to 100% recovery of synthetically injected mislabels using voting strategies.
- Successfully identified most mislabeled samples in AAV nanopore datasets, with perfect detection on balanced subsets.
Conclusions:
- The proposed ensemble method demonstrates superior accuracy, stability, and true-label recovery compared to existing techniques.
- Establishes a strong framework for mislabel detection, particularly for complex, fine-grained data like nanopore signals.
- Offers a reliable solution for improving data quality in machine learning pipelines.
