Catching the Imposter: Self-Supervised Learning of Physical Coherence with Cross-Entity Feature Permutations
The paper tests “imposter” feature swaps as a self-supervised signal for learning physically coherent scientific representations.
The method replaces some features of an entity with real values from another entity, then trains the encoder to detect which features were swapped.
Because the inserted values are plausible on their own, the model has to learn cross-feature physical dependencies to solve the task.
The authors evaluate it on ERA5-Land data with 21 environmental variables and seven downstream land-surface tasks.
They report that no single pretext task dominates, but imposter adds complementary information when combined with other SSL objectives.
ArXiv · AI/CL/LG's note
The method replaces some features of an entity with real values from another entity, then trains the encoder to detect which features were swapped.
Because the inserted values are plausible on their own, the model has to learn cross-feature physical dependencies to solve the task.
The authors evaluate it on ERA5-Land data with 21 environmental variables and seven downstream land-surface tasks.
They report that no single pretext task dominates, but imposter adds complementary information when combined with other SSL objectives.
ArXiv · AI/CL/LG's note
score 5