Systematic image perturbations reveal persistent gaps between human and machine vision
Mugihiko Kato, Biyu J. He · 2026
Read Systematic image perturbations reveal persistent gaps between human and machine vision by Mugihiko Kato, Biyu J. He (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/pmc-13639169.txt.
Description
Mugihiko Kato and Biyu He compare how humans and machines see objects. Their 2026 paper uses a crafted image set that pulls apart three cues: global shape, inner parts, and texture. Deep networks are often cast as models of human vision. Whether they lean on the same cues was unknown.
More than 200 networks were tested, across many designs, training diets, and goals. None matched the human cue profile, not even recurrent or specially trained ones. Fine-tuned image-text models came closest, but lost that match when global shape was broken. All networks fell far behind humans when shape alone carried the answer. Brain recording fit did not predict behavior fit. The authors call the gap systematic and persistent.
Catalog record
| CREATOR | Mugihiko Kato, Biyu J. He |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 114 KB |
| VERSION | 1.0 |
| LICENSE | CC BY 4.0 |
| RIGHTS | CC BY |
| SOURCE | Europe PMC |