Mutual Information Regularized Feature-Level Frankenstein for Discriminative Recognition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models require discriminative representations for improved recognition.
• Separating nuisance factors from essential features is crucial for robust performance.
• Existing methods struggle to explicitly disentangle various data attributes.

Purpose of the Study:

• To propose a novel approach for disentangling discriminative representations (d), latent variations (l), and semantic labels (s).
• To enforce independence and complementarity between d, l, and s for enhanced recognition.
• To develop a framework that achieves tolerance to nuisance factors in recognition tasks.

Main Methods:

• Utilizing an adversarial game in the latent space of an auto-encoder.
• Employing an end-to-end conditional adversarial network to decompose input samples.
• Incorporating uniform-target and entropy regularization for effective disentanglement.
• Implementing a collaborative mutual information regularization framework to ensure stable adversarial training.

Main Results:

• The proposed method successfully decomposes input samples into independent and complementary representations.
• Achieved top performance on diverse recognition tasks, including digit classification.
• Demonstrated state-of-the-art results on large-scale face recognition datasets (LFW, IJB-A).
• Showcased enhanced face recognition tolerance to variations like lighting, makeup, and disguise.

Conclusions:

• The novel framework effectively disentangles representations, leading to improved recognition accuracy.
• The method provides a discriminative representation with desired tolerance properties.
• This approach offers a significant advancement for robust and accurate deep learning recognition systems.