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

Poster C in Poster Session C: Wednesday, August 5, 9:30 – 11:15 am, Kimmel Center, Shorin & Rosenthal Rooms

More shape bias does not mean more human-like visual representations

Alban Flachot¹, Katharina Dobs¹; ¹Justus Liebig Universität Gießen

Presenter: Alban Flachot

Deep neural networks (DNNs) are widely used as mod- els of human perception, often evaluated based on their performance on specific behavioral benchmarks such as shape–texture cue-conflict tasks. However, it re- mains unclear whether alignment with human behavior on such tasks implies alignment in underlying represen- tations. Here, we introduce a representational screen- ing approach to identify diagnostic stimuli from a large naturalistic image set that maximize representational dif- ferences between two competing models. We apply this method to a standard ImageNet-trained DNN and a Stylized-ImageNet variant previously reported to exhibit a more human-like shape bias in object classification, and then measure human perceptual similarity using a multi-arrangement task. Contrary to expectations, human representational geometry aligns more closely with the standard ImageNet model on these diagnostic stimuli. A variance partitioning analysis further suggests that this alignment is substantially associated with shared color information. Together, these results suggest that similar- ity judgments and object recognition can rely on different image features, and that increasing shape bias in DNNs does not necessarily translate into representational align- ment. Moreover, our findings highlight the importance of diagnostic stimulus selection for evaluating computa- tional models of perception.

Topic Area: Computational Models of Vision & Visual Cortex