Learning Reversible Stochastic Tensor Transformations

About the paper

This study uses reversible coupling blocks to learn a prescribed stochastic image transformation with an explicit inverse. A compact, 44,296-parameter network transforms all RGBA channels and retains floating-point payloads. Across three experimental runs, every tested image was recovered exactly after byte rounding.

The paper develops a tensor-based formulation for future multimodal systems and a conditional proof sketch involving a one-time pad. The experiments evaluate image transformations and numerical recovery; multimodal operation and cryptographic security of the experimental transform remain research goals.