Keynotes | K&Ts | GACs | Talks | Posters | Search

Poster A in Poster Session A: Tuesday, August 4, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

Rethinking the inversion effect as a graded phenomenon across object categories

Jakob Winkler¹, Katharina Dobs¹; ¹Justus Liebig Universität Gießen

Presenter: Jakob Winkler

The inversion effect - poorer recognition of upside-down stimuli - has often been reported for faces and interpreted as evidence for face-specific processing. But is inversion sensitivity truly face-specific, or does it reflect more general visual properties? Here, we use deep neural networks (DNNs) to measure inversion sensitivity across a broad range of object categories and identify the features that predict it. We find that in-version sensitivity in DNNs varies continuously and re-liably across categories, with faces falling toward the upper range but not constituting qualitative outliers. This graded pattern was more pronounced in trained than in untrained DNNs, indicating a key role of visual experience. Feature-encoding based on image proper-ties, particularly horizontal and vertical symmetry, strongly predicted inversion sensitivity. These findings suggest that inversion effects in DNNs, and potentially in brains, emerge as graded phenomenon across ob-jects based on the natural visual input statistics.

Topic Area: Computational Models of Vision & Visual Cortex