文章基本信息

标题：Deep learning models fail to capture the configural nature of human shape perception
本地全文：下载
作者：Nicholas Baker ; James H. Elder
期刊名称：iScience
印刷版ISSN：2589-0042
出版年度：2022
卷号：25
期号：9
页码：1-17
DOI：10.1016/j.isci.2022.104913
语种：English
出版社：Elsevier
摘要：SummaryA hallmark of human object perception is sensitivity to the holistic configuration of the local shape features of an object. Deep convolutional neural networks (DCNNs) are currently the dominant models for object recognition processing in the visual cortex, but do they capture this configural sensitivity? To answer this question, we employed a dataset of animal silhouettes and created a variant of this dataset that disrupts the configuration of each object while preserving local features. While human performance was impacted by this manipulation, DCNN performance was not, indicating insensitivity to object configuration. Modifications to training and architecture to make networks more brain-like did not lead to configural processing, and none of the networks were able to accurately predict trial-by-trial human object judgements. We speculate that to match human configural sensitivity, networks must be trained to solve a broader range of object tasks beyond category recognition.Graphical abstractDisplay OmittedHighlights•Humans rely on configural relations between local shape features to recognize objects•Networks trained to recognize objects are insensitive to these configural relations•Training and architecture innovations do not lead to configural processing•Networks remain unable to account for human object shape perceptionBiological sciences; Neuroscience; Sensory neuroscience.