Naturalness Predicts but Does Not Cause Transferability in Image Encodings of Real-World Streams
Faruk Alpay, Barış Başaran · 2026
Read Naturalness Predicts but Does Not Cause Transferability in Image Encodings of Real-World Streams by Faruk Alpay, Barış Başaran (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-25844.txt.
Description
Turning a data stream into a picture lets photo trained vision models read it, and pictures that look natural tend to score better. Faruk Alpay and Barış Başaran ask in 2026 whether that naturalness causes the success, or merely predicts it. They build WorldStream, two hundred ninety nine live series from weather, markets, quakes, and space weather, cut into thousands of windows for a nine way source recognition task. Across seven image encodings and six frozen backbones, distance from natural images predicted accuracy strongly. Then came controlled interventions.
An invertible encoder let them tune only the image's spectrum. Accuracy stayed flat and poor while naturalness peaked exactly where theory said it should. Scrambling phases instead, holding the spectrum fixed and wrecking local structure, dragged accuracy and naturalness down together. The correlation, they conclude, runs through local structure, not spectra. Full fine tuning never closed the gap either.
Catalog record
| CREATOR | Faruk Alpay, Barış Başaran |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 20 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.25844