Learning Kernel for Conditional Moment-Matching Discrepancy-Based Image Classification.

Area of Science:

• Machine Learning
• Computer Vision

Background:

• Conditional Maximum Mean Discrepancy (CMMD) is effective for pattern classification using nonlinear kernel functions.
• CMMD struggles with complex distributions where kernel functions fail to distinguish intraclass and interclass similarities.

Purpose of the Study:

• To propose a novel kernel learning method (KLN) to enhance CMMD's discrimination performance.
• To improve classification accuracy on complex datasets by addressing limitations of traditional CMMD.

Main Methods:

• KLN iteratively operates with deep network features, utilizing CMMD loss and an autoencoder (AE) to learn an injective function.
• A compound kernel, combining the injective function with a characteristic kernel, enhances CMMD's data category description capabilities.
• The method simultaneously learns an expressive kernel and label prediction distribution in an end-to-end manner.

Main Results:

• KLN demonstrates improved effectiveness of CMMD for data category description.
• The approach enhances classification performance in both supervised and semisupervised learning scenarios.
• Experiments on MNIST, SVHN, CIFAR-10, and CIFAR-100 datasets show KLN achieving state-of-the-art classification results.

Conclusions:

• The proposed KLN method effectively addresses CMMD limitations on complex distributions.
• KLN offers a robust and adaptable solution for improving classification tasks through advanced kernel learning.
• The iterative, end-to-end learning of kernel-based similarities on deep features provides superior performance.