Filtering out mislabeled training instances using black-box optimization and quantum annealing.

Area of Science:

• Machine Learning
• Data Science
• Quantum Computing

Background:

• Mislabeled instances in training datasets degrade model generalization.
• Efficient noise-removal strategies are crucial for real-world applications.
• Existing methods may lack scalability or efficiency in handling noisy data.

Purpose of the Study:

• To propose a robust and efficient method for removing mislabeled instances from contaminated training datasets.
• To enhance the generalization capability of machine learning models by improving training data quality.
• To leverage quantum annealing for efficient sampling of high-quality training subsets.

Main Methods:

• Combining surrogate model-based black-box optimization (BBO) with postprocessing.
• Utilizing quantum annealing for efficient sampling of diverse training subsets with low validation error.
• Evaluating filtered training subsets based on validation loss and iteratively refining loss estimates.

Main Results:

• The proposed method effectively prioritizes the removal of high-risk mislabeled instances.
• Integration with D-Wave's physical quantum annealer demonstrated faster optimization and higher-quality subsets compared to simulated annealing.
• The approach offers a scalable framework for enhancing dataset quality in supervised learning.

Conclusions:

• The developed method is effective for supervised learning tasks, improving dataset quality and model generalization.
• Quantum annealing, particularly on physical hardware, provides advantages in optimization speed and subset quality.
• Future work includes applying the method to unsupervised learning, real-world datasets, and large-scale implementations.