Discovering shared interpretable operations in image compression autoencoders
This work addresses the need for interpretability and reduced complexity in deep learning-based image compression, though it is an incremental step.
The paper investigates whether universal interpretable operations exist in bias-free autoencoders for image compression via Jacobian analysis, aiming to extract them for designing low-complexity models. Results show that shared operations can be detected, potentially enabling simpler architectures.
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.