Constraint-validated self-consuming VAE-GANs for safe synthetic data in high-stakes domains

Self-consuming loops can collapse when models train on their own outputs; Gillman et al. (ICML 2024) stabilize this via “correctors.” For domains with known constraints—spatiotemporal rules, anatomy, or physics—Zhang et al. (SIGSPATIAL 2022) show how VAEs can enforce validity with constrained optimization for trajectories. This project combines them:

• Train a VAE-GAN hybrid generator (Cai 2024; Revathi & Babu 2024) with a domain validator f that scores constraint satisfaction (e.g., spatiotemporal legality, anatomical plausibility, energy-theft usage patterns as in Sun et al., 2023).
• Use implicit differentiation or REINFORCE-style gradients to backprop through non-differentiable validators; attach a self-corrector g that maps candidate samples toward valid regions before reusing them for self-training.
• In privacy-limited settings, simulate cross-site collaboration by sharing the generator (not data) as in Szafranowska et al. (2022), evaluating whether validator-aware sharing improves downstream classifiers more than plain GAN sharing.
  The novelty is a general recipe for “constraint-validated self-consumption,” with validations acting as safety rails against drift and collapse. This should yield stable, high-utility synthetic augmentation in medicine and critical infrastructure—domains where vanilla self-training and unconstrained generators are risky.

References:

1. Self-Correcting Self-Consuming Loops for Generative Model Training. Nate Gillman, Michael Freeman, Daksh Aggarwal, Chia-Hong Hsu, Calvin Luo, Yonglong Tian, Chen Sun (2024). International Conference on Machine Learning.
2. Factorized deep generative models for end-to-end trajectory generation with spatiotemporal validity constraints. Liming Zhang, Liang Zhao, D. Pfoser (2022). SIGSPATIAL/GIS.
3. Enhancing capabilities of generative models through VAE-GAN integration: A review. Dongting Cai (2024). Applied and Computational Engineering.
4. Synthesizing Realistic Knee MRI Images: A VAE-GAN Approach for Enhanced Medical Data Augmentation. Revathi S A, B. S. Babu (2024). International Journal of Advanced Computer Science and Applications.
5. Sharing generative models instead of private data: a simulation study on mammography patch classification. Zuzanna Szafranowska, Richard Osuala, Bennet Breier, Kaisar Kushibar, Karim Lekadir, Oliver Díaz (2022). International Workshop on Breast Imaging.
6. Energy Theft Detection Model Based on VAE-GAN for Imbalanced Dataset. Youngghyu Sun, Jiyoung Lee, Soohyun Kim, Joonho Seon, Seongwoo Lee, Chanuk Kyeong, Jinyoung Kim (2023). Energies.