Semantic and Generalized Entropy Loss Functions for Semi-Supervised Deep Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Modern datasets are large, and obtaining labeled data is challenging.
• Semi-supervised learning combines supervised and unsupervised approaches with constraints.
• This necessitates efficient methods for learning from partially labeled data.

Purpose of the Study:

• To present a novel methodology bridging artificial neural network outputs and logical constraints.
• To introduce semantic loss and generalized entropy (Rényi entropy) loss functions.
• To enhance semi-supervised multiclass classification accuracy.

Main Methods:

• Developed a methodology compatible with any feedforward neural network.
• Introduced semantic loss and generalized entropy loss as regularization terms.
• Evaluated on simulated, MNIST, and Fashion-MNIST datasets.

Main Results:

• Both proposed loss functions effectively guide neural networks.
• Achieved near state-of-the-art results on semi-supervised multiclass classification.
• Demonstrated the influence of parameters on classification accuracy.

Conclusions:

• The semantic and generalized entropy loss functions are effective for semi-supervised learning.
• These methods offer a flexible regularization approach for neural networks.
• The methodology shows strong performance on benchmark datasets.