Search Papers | Poster Sessions | All Posters
Poster A136 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Configural processing as a fundamental mechanism for robust visual recognition in neural networks across varied viewing conditions
Hojin Jang1,2 (), Pawan Sinha2, Xavier Boix3; 1Department of Brain and Cognitive Engineering, Korea University, 2Department of Brain and Cognitive Sciences, MIT, 3Fujitsu Research
A hallmark of face recognition is configural processing; that said, the underlying neurocomputational mechanisms are still elusive despite decades of research. Moreover, despite a few previous studies hinting at the benefits of configural processing under challenging conditions, detailed insights were scarce and mostly empirical. The study posits that recognizing faces through configural cues is more effective than using local features across variations in viewing conditions. This hypothesis was tested using face-like digit stimuli and comparing neural network models trained to recognize them with either local or configural cues. The findings demonstrate that neural networks can indeed discern configural cues, which notably enhance performance against geometric alterations like rotation and scaling. Additionally, when both types of cues are present, models prefer configural over local cues. Our results offer new neurocomputational evidence of the advantages of configural processing in reliably recognizing faces across diverse conditions.
Keywords: configural processing face recognition robustness