Recurrence is needed to account for the sharper orientation-tuned surround suppression for oblique versus cardinal orientations

Poster Presentation: Tuesday, May 21, 2024, 2:45 – 6:45 pm, Pavilion
Session: Spatial Vision: Machine learning, neural networks

Huiyuan Miao1 (), Frank Tong1; 1Psychology department, Vanderbilt University

Surround suppression can reduce both the perceived contrast and the strength of V1 neuronal responses to stimuli within the classical receptive field when the neighboring surround is visually similar due to inhibitory neural interactions. Our previous study measured the orientation tuning curves of surround suppression and found that surround suppression was more broadly tuned around cardinal than oblique orientations, revealing an inverted oblique effect when compared to human orientation discrimination (Miao, Coggan, & Tong, VSS2023). Recent studies have demonstrated surround suppression in feedforward deep neural networks (DNN) optimized for object recognition (Pan et al., 2023). Here, we asked: Do DNNs show a similar inverted oblique effect in surround suppression orientation-tuning? We compared two types of models: a feedforward network (AlexNet) and a recurrent network (CORnet-RT). We presented these networks with both cardinal and oblique bandpass-filtered patterns to evaluate the magnitude of surround suppression across varying center-surround orientation differences (0-90˚). Stimuli with three different spatial frequency ranges and eight different target sizes were evaluated. While we found surround suppression in feedforward AlexNet’s intermediate layers, no stable inverted oblique effect was detected. Is recurrence necessary for the inverted oblique effect? We evaluated CORnet-RT, which contains four computational blocks with recurrent processing analogous to V1, V2, V4, and IT. In each block, the model implemented convolution, group normalization (GN), and ReLU during feedforward processing, followed by recurrent processing in which the output of an area's last time point was used for processing in the next cycle. We found that while the network revealed a general effect of orientation-tuned surround suppression in first processing step of the V1 block, the inverted oblique effect appeared only after recurrent processing. Taken together, our results reveal that surround suppression leads to a counterintuitive oblique effect that can be attributed to recurrent processing in the human visual system.

Acknowledgements: Supported by NIH grants R01EY029278 and R01EY035157 to FT.