EnAET

EnAET integrates self-supervised ensemble autoencoding transformations into semi-supervised learning to regularize models and improve representation learning from limited labeled and abundant unlabeled data.

Key Features:

• Self-Supervised Information Integration: Incorporates self-supervised representations as an explicit regularization term rather than relying solely on prediction consistency and confidence.
• Ensemble AutoEncoding Transformations: Employs an ensemble of autoencoding transformations to generate diverse feature representations from input data.
• Regularization Term: Uses the generated self-supervised representations as an additional regularizer in the learning objective.
• Generalization Across Datasets: Demonstrates performance gains using MixMatch as a baseline across multiple datasets while maintaining consistent hyper-parameters.
• Performance Improvement in Supervised Learning: Improves supervised learning outcomes, including performance with as few as 10 images per class.

Scientific Applications:

• Scarce-labeled data domains: Applicable to settings where labeled data is scarce but unlabeled data is abundant, to improve model training.
• Image recognition: Enhances representation learning and performance for image recognition tasks under limited label availability.
• Natural language processing: Supports semi-supervised or low-label learning scenarios in natural language processing tasks.

Methodology:

Applies ensemble autoencoding transformations to input data to generate diverse representations and incorporates those representations as a self-supervised regularization term in the semi-supervised learning objective; evaluated using MixMatch as the baseline.