Interpolation consistency training for semi-supervised learning.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Semi-supervised learning leverages limited labeled data with abundant unlabeled data.
• Deep neural networks (DNNs) require large labeled datasets for optimal performance.
• Existing semi-supervised methods face challenges in efficiency and effectiveness.

Purpose of the Study:

• Introduce Interpolation Consistency Training (ICT), an efficient algorithm for DNNs.
• Enhance DNN performance in semi-supervised learning settings.
• Improve decision boundary placement in classification tasks.

Main Methods:

• Developed ICT, a novel algorithm for semi-supervised learning.
• ICT enforces consistency between interpolated predictions and interpolated unlabeled data points.
• Applied ICT to standard neural network architectures.

Main Results:

• Achieved state-of-the-art performance on CIFAR-10 and SVHN benchmark datasets.
• Demonstrated ICT's effectiveness in improving DNNs within the semi-supervised paradigm.
• Showcased ICT's ability to move decision boundaries to low-density regions.

Conclusions:

• ICT is a simple yet powerful technique for semi-supervised learning.
• The algorithm offers significant improvements in DNN training efficiency and accuracy.
• Theoretical analysis indicates ICT acts as data-adaptive regularization, reducing overfitting.