Entropy-informed weighting channel normalizing flow for deep generative models

doi:10.1016/j.patcog.2025.112442

Entropy-informed weighting channel normalizing flow for deep generative models

2025-09-15

0

PRE

AI

DOI：10.1016/j.patcog.2025.112442

Original

Original request for help

Save

Abstract

En 中文

• We propose a normalizing flow model named “Entropy-Informed Weighting Channel Normalizing Flow” (EIW-Flow) for density estimation and image generation based on the proposed Shuffle operation. • We propose a regularized Shuffle operation to adaptively shuffle the channel feature maps of its input based on the feature information contained in each map, while maintaining the reversibility of normalizing flows. • We demonstrate the efficacy of the Shuffle operation from the perspective of entropy using the principles of information theory and statistics, such as the Central Limit Theorem and the Maximum Entropy Principle. • Extensive qualitative and quantitative experiments demonstrate that the EIW-Flow achieves state-of-the-art density estimation results and comparable sample quality on CIFAR-10, CelebA, ImageNet and LSUN datasets, with minimal computational overhead.

Entropy-informed weighting channel normalizing flow for deep generative models

Abstract

Journal

Organization