Discovering shared interpretable operations in image compression autoencoders
2026-07-06 • Multimedia
Multimedia
AI summaryⓘ
The authors study deep learning models called autoencoders, which are used for image compression but often become very complex and hard to understand. They focus on simpler versions without bias and analyze their internal math (using something called the Jacobian) to see if common patterns appear. If these universal patterns exist, they could help design simpler and more efficient image compression models inspired by complex deep learning systems. The work aims to better understand how these models work inside rather than just making them bigger and more complicated.
deep learningautoencoderimage compressionrate-distortion trade-offblack-box modelJacobian analysisbias-freemodel complexity
Authors
Caroline Mazini Rodrigues, Nicolas Keriven, Thomas Maugey
Abstract
With the increasing adoption of deep learning for applications such as image compression, improvements in the rate-distortion trade-off have been achieved at the cost of increasingly larger and more opaque ''black-box'' models. Autoencoders are among the most widely used architectures for this task; however, without a clear understanding of their internal behavior, these models tend to grow in complexity to achieve more performance gains. In this paper, we investigate whether universal behaviors can be detected from the internal operations of bias-free autoencoders through Jacobian analysis. If such behaviors exist, they may be extracted to design low-complexity image compression models inspired by high-complexity deep learning architectures.