Improving Dataset Volumes and Model Accuracy with Semi-Supervised Iterative Self-Learning.

Area of Science:

• Machine Learning
• Computer Vision
• Artificial Intelligence

Background:

• Deep learning models require large labeled datasets for optimal performance.
• Unlabeled data is abundant but challenging to integrate effectively into training.
• Semi-supervised learning (SSL) offers a promising approach to leverage unlabeled data.

Purpose of the Study:

• Introduce a novel semi-supervised learning (SSL) technique.
• Enhance the performance of deep classification models using unlabeled data.
• Demonstrate the broad applicability of the proposed SSL method.

Main Methods:

• A simple iterative learning cycle is employed.
• Learned thresholding techniques are integrated.
• An ensemble decision support system is utilized.
• The approach is model-agnostic, independent of specific architectures or loss functions.

Main Results:

• State-of-the-art model performance is achieved.
• Increased training data volume is effectively utilized through unlabeled data.
• The method demonstrates robustness on standard SSL benchmarks.
• Successful evaluation on challenging datasets like CIFAR-100 and ImageNet subsets.

Conclusions:

• The proposed SSL technique effectively improves deep classification models.
• The method's versatility allows application across diverse machine learning tasks.
• Leveraging unlabeled data via this SSL approach enhances performance and scalability.