Search Papers | Poster Sessions | All Posters
Poster C7 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
Contrasting computational models of task-dependent readout from the ventral visual stream
Johannes J.D. Singer1 (), Radoslaw M. Cichy1, Tim C. Kietzmann2, Sushrut Thorat2; 1Department of Education and Psychology, Freie Universtität Berlin, 2Institute for Cognitive Science, Osnabrück University
Humans categorize visual information based on features of different complexity. While some tasks require high-level information, others rely on low-level cues. This raises the question of how these features, originating from different parts of the visual system, are integrated for perceptual decisions. Here, we test three potential mechanisms: single readout, direct access, and attentional routing. The mechanisms were implemented and contrasted based on ANN models, equipped with different readout strategies. These networks were trained to perform two tasks that require access to low- or high-level visual features, respectively, and subsequently evaluated in terms of performance and their ability to predict human reaction times of participants performing the same tasks. We found that the direct access and attentional routing models performed as well as humans in both tasks, while the single readout model did not perform well. Importantly, the attentional routing model best predicted human reaction times overall. These results indicate that neither a readout only from high-level visual cortex nor direct access to upstream regions might be sufficient to explain human categorization behavior across tasks, and suggest attentional modulations along the ventral visual stream as a critical mechanism that enables flexible readout through high-level visual cortex.
Keywords: visual categorization perceptual readout deep neural networks attentional modulation