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

Poster F in Poster Session F: Thursday, August 6, 1:45 – 3:30 pm, Kimmel Center, Shorin & Rosenthal Rooms

Divergent effects of hierarchical perceptual features on confidence in humans and deep neural networks

Yunpeng Liu¹, Xiao Hu¹; ¹Beijing Normal University

Presenter: Yunpeng Liu

In perceptual decision-making, confidence can be shaped by perceptual features beyond their effects on accuracy. The present study tested a hierarchical weighting account in which confidence preferentially relies on features processed earlier in the visual hierarchy. Across three orientation discrimination experiments, we manipulated contrast, reliability, and boundary distance and measured confidence with a forced-choice paradigm. Human confidence showed a hierarchical pattern: contrast had the strongest influence, reliability an intermediate influence, and boundary distance the weakest. Convolutional neural networks (CNNs) trained on the same perceptual tasks showed a human-like representational hierarchy, but their confidence was modulated differently and tracked accuracy more closely. These findings suggest that human confidence draws on hierarchical sensory representations but also relies on additional inferential processes beyond a direct perceptual readout.

Topic Area: Computational Models of Vision & Visual Cortex