EnAET: A Self-Trained Framework for Semi-Supervised and Supervised Learning With Ensemble Transformations.

Area of Science:

• Machine Learning
• Deep Learning
• Computer Vision

Background:

• Deep neural networks (DNNs) require extensive labeled data, which is costly and time-consuming to acquire.
• Semi-supervised learning (SSL) methods leverage limited labeled data alongside abundant unlabeled data to mitigate this challenge.
• Existing SSL methods primarily focus on prediction consistency and confidence maximization.

Purpose of the Study:

• To introduce a novel framework, EnAET (Enhanced Autoencoder with Transformations), for improving semi-supervised learning.
• To explore the integration of self-supervised representations as a regularization technique within SSL.
• To demonstrate the framework's ability to enhance existing state-of-the-art SSL algorithms.

Main Methods:

• The proposed EnAET framework incorporates self-supervised learning signals as a regularization term.
• The framework is evaluated by integrating it with MixMatch, a leading semi-supervised learning algorithm.
• Consistent hyperparameters are used across diverse datasets to ensure generalization.

Main Results:

• The EnAET framework significantly improves the performance of existing semi-supervised learning algorithms.
• The framework demonstrates substantial performance gains in supervised learning, particularly in low-data regimes (e.g., 10 images per class).
• Experimental results across multiple datasets validate the effectiveness and generalization capability of EnAET.

Conclusions:

• The EnAET framework offers a novel and effective approach to enhance semi-supervised learning by leveraging self-supervised representations.
• This method provides a versatile regularization strategy applicable to various SSL techniques.
• EnAET shows promise for improving deep learning models, especially when labeled data is scarce.