![[2602.20480] VINA: Variational Invertible Neural Architectures](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2602.20480] VINA: Variational Invertible Neural Architectures
Summary
The paper presents VINA, a framework for Variational Invertible Neural Architectures, addressing theoretical gaps in normalizing flows and invertible neural networks for generative modeling and inverse problems.
Why It Matters
This research is significant as it provides a unified framework that enhances the understanding and application of invertible neural networks and normalizing flows. By offering theoretical guarantees under realistic assumptions, it advances the field of machine learning, particularly in generative modeling and inference tasks, which are crucial for various applications in AI.
Key Takeaways
- Introduces a unified framework for invertible neural networks and normalizing flows.
- Provides theoretical performance guarantees for posterior and distributional accuracy.
- Offers practical guidelines and design principles based on extensive case studies.
- Addresses key gaps in existing literature regarding approximation quality.
- Demonstrates effectiveness through a case study on ocean-acoustic inversion.
Computer Science > Machine Learning arXiv:2602.20480 (cs) [Submitted on 24 Feb 2026] Title:VINA: Variational Invertible Neural Architectures Authors:Shubhanshu Shekhar, Mohammad Javad Khojasteh, Ananya Acharya, Tony Tohme, Kamal Youcef-Toumi View a PDF of the paper titled VINA: Variational Invertible Neural Architectures, by Shubhanshu Shekhar and 4 other authors View PDF Abstract:The distinctive architectural features of normalizing flows (NFs), notably bijectivity and tractable Jacobians, make them well-suited for generative modeling. Invertible neural networks (INNs) build on these principles to address supervised inverse problems, enabling direct modeling of both forward and inverse mappings. In this paper, we revisit these architectures from both theoretical and practical perspectives and address a key gap in the literature: the lack of theoretical guarantees on approximation quality under realistic assumptions, whether for posterior inference in INNs or for generative modeling with NFs. We introduce a unified framework for INNs and NFs based on variational unsupervised loss functions, inspired by analogous formulations in related areas such as generative adversarial networks (GANs) and the Precision-Recall divergence for training normalizing flows. Within this framework, we derive theoretical performance guarantees, quantifying posterior accuracy for INNs and distributional accuracy for NFs, under assumptions that are weaker and more practically realistic than those ...