MetaInv: Overcoming Iterative and Direct Method Limitations for Inverse Learning

27 Sept 2024 (modified: 05 Feb 2025)Submitted to ICLR 2025EveryoneRevisionsBibTeXCC BY 4.0
Keywords: Invertible neural networks, switchable Architectures, analytical inverse
Abstract:

Invertible neural networks (INNs) have gained significant traction in tasks requiring reliable bidirectional inferences, such as data encryption, scientific computing, and real-time control. However, iterative methods like i-ResNet face notable limitations, including instability on non-contractive mappings and failure in scenarios requiring strict one-to-one mappings. In contrast, analytical approaches like DipDNN guarantee invertibility but at the expense of performance, particularly in tasks demanding rich feature extraction (e.g., convolutional operations in complex image processing). This work presents a detailed analysis of the limitations in current invertible architectures, examining the trade-offs between iterative and analytical approaches. We identify key failure modes, particularly when handling information redundancy or strict bijections, and propose a meta-inverse framework that dynamically combines the advantages of both i-ResNet and DipDNN. Our framework adapts in real-time based on task-specific signals, ensuring both flexibility and guaranteed invertibility. Extensive experiments across diverse domains demonstrate that our hybrid approach outperforms existing methods in forward accuracy, inverse consistency, and computational efficiency. Our results highlight the utility of this meta-inverse strategy for critical applications where precision, stability, and adaptability are crucial.

Supplementary Material: zip
Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 11884
Loading

OpenReview is a long-term project to advance science through improved peer review with legal nonprofit status. We gratefully acknowledge the support of the OpenReview Sponsors. © 2025 OpenReview