Search Papers | Poster Sessions | All Posters
Poster B85 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
Humans and deep neural networks can use perceptual features to determine physical stability
Lauren Aulet1 (), Evren Konuk1, Jessica Cantlon1; 1Carnegie Mellon University
There is extensive debate about whether judgments of physical stability (e.g., whether a stack of blocks will fall) rely on low-level perceptual features or mental simulation. In the present work, we evaluated whether deep neural networks (DNNs) trained on ImageNet, which are thought to rely only on low-level image features (and cannot perform simulations) can discriminate images of stable and unstable block towers. Moreover, we evaluated whether human adults were affected by the stability of a distractor block tower image (i.e., same or different stability category) when performing an exact match-to-sample task. We found that DNNs discriminated stable and unstable block towers significantly above chance, and did so across a variety of stimulus perturbations. Furthermore, we found that human participants were significantly influenced by the stability of the block tower images, even in a task where mental simulation was highly improbable. Taken together, these results suggest there are visual features that are diagnostic of physical stability and are ‘perceived’ by both DNNs and humans.
Keywords: intuitive physics visual perception deep neural networks psychophysics