Search Papers | Poster Sessions | All Posters
Poster C4 in Poster Session C - Friday, August 9, 2024, 11:15 am – 1:15 pm, Johnson Ice Rink
Generative perceptual inference in deep neural network models of object recognition induces illusory contours and shapes
Tahereh Toosi1 (), Kenneth D. Miller1; 1Columbia University
Illusory contours and shapes highlight the striking gap between how natural and artificial vision perceive the world. In this study, we show that a pattern recognition model embodies a generative model that integrates perceptual priors and the sensory processing. We introduce a novel perceptual algorithm, Generative Perceptual Inference (GPI), which iteratively updates the activations by accumulating propagated error in the early layers. Given a Kanizsa square as input to a deep neural network (DNN) optimized for robust object classification, our results show that running GPI led to the emergence of edge-like patterns in the area of the perceived 'white square'. Moreover, when GPI is applied to the same DNN with Rubin's vase image as input, it creates a vase-like pattern, while GPI in a DNN with the same architecture but optimized for face recognition creates face-like patterns. Thus, we found the direct link between natural image prior and perception of illusory contours and shapes, through an image-computable algorithm that captures experimental findings regarding processing of illusions in animals and humans. More broadly, this work reconciles the views of the visual cortex as both a pattern recognition and a generative model in a unified framework.
Keywords: illusions object recognition deep neural networks perceptual inference