A Plug-in Method for Representation Factorization in Connectionist Models.

Area of Science:

• Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Deep learning models like Generative Adversarial Networks (GANs) and deep autoencoders learn complex latent representations.
• Controlling these latent representations semantically is crucial for tasks like image manipulation and data analysis.
• Existing methods often require retraining models or lack fine-grained control over feature decomposition.

Purpose of the Study:

• To develop a method for decomposing latent representations into semantically controllable factors.
• To enable manipulation of features in pre-trained deep learning models without modification.
• To improve the interpretability and controllability of learned feature representations.

Main Methods:

• Proposed a Factors' Decomposer-Entangler Network (FDEN) as a plug-in module.
• Learned to decompose latent representations into mutually independent factors.
• Minimized total correlation between factors using information-theoretic principles.

Main Results:

• Successfully applied FDEN to existing networks for image-to-image translation (e.g., style transfer while preserving identity).
• Demonstrated effectiveness in few-shot object classification tasks.
• Validated through qualitative, quantitative, and statistical ablation studies.

Conclusions:

• FDEN effectively decomposes latent representations into interpretable and controllable factors.
• The method enhances semantic control in computer vision tasks without altering original models.
• Offers a flexible, plug-in solution for improving feature disentanglement and manipulation in deep learning.