Stable and Fast Deep Mutual Information Maximization Based on Wasserstein Distance.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Deep Learning

Background:

• Deep learning excels but requires extensive labeled data, hindering applications.
• Unsupervised learning is crucial for advancing AI without labeled data.
• Deep InfoMax (DIM) learns deep representations but suffers unstable training due to adversarial methods.

Purpose of the Study:

• To address the training instability of the Deep InfoMax (DIM) method.
• To propose a more stable unsupervised learning approach for deep representation learning.
• To enhance the practical applicability of DIM in AI.

Main Methods:

• Proposed a Wasserstein distance-based Deep InfoMax (WDIM) method.
• Replaced the adversarial network in DIM with Wasserstein distance for stable training.
• Evaluated WDIM on CIFAR10, CIFAR100, and STL10 datasets for unsupervised classification.

Main Results:

• WDIM demonstrates improved training stability compared to the original DIM method.
• WDIM achieves faster model convergence during unsupervised learning.
• The proposed WDIM method maintains competitive classification accuracy, comparable to DIM.

Conclusions:

• WDIM offers a stable and efficient alternative for unsupervised deep representation learning.
• The method enhances the reliability of training deep learning models.
• WDIM provides a promising direction for future research in unsupervised image classification.