Search Papers | Poster Sessions | All Posters
Poster A102 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Beyond feedforward: Leveraging discrepancies between humans and convolutional neural networks reveals recurrent dynamics during object recognition
Pablo Oyarzo1 (), Johannes J.D. Singer1, Kohitij Kar2, Radoslaw M. Cichy1; 1Freie Universität Berlin, 2York University
Convolutional neural networks (CNNs) have emerged as leading models for primate object recognition, yet humans often outperform them, revealing misalignments with human behavior and brain responses. This discrepancy indicates unique brain-specific computations engaged when object recognition is challenging. Here, we leverage this gap to identify the human neural mechanisms driving these computations. Specifically, we compared EEG and fMRI responses to images on which a feedforward CNN (AlexNet) and humans perform on par versus images on which the CNN performs worse. We find that for images where the CNN performs worse, humans show delayed information processing and the specific recruitment of frontal brain areas, suggesting the involvement of additional top-down recurrent computations. These results pinpoint the neural mechanisms beyond feedforward processing engaged for robust object perception when vision is challenging.
Keywords: object recognition deep neural networks eeg fmri