A comparison of statistical learning of naturalistic textures between DCNNs and the human visual hierarchy

作     者：LU XinCheng YUAN ZiQi ZHANG YiChi AI HaiLin CHENG SiYuan GE YiRan FANG Fang CHEN NiHong 

作者机构：Department of Psychological and Cognitive SciencesTsinghua UniversityBeijing 100084China IDG/McGovern Institute for Brain ResearchTsinghua UniversityBeijing 100084China Department of Computer Science and TechnologyTsinghua UniversityBeijing 100084China School of Psychological and Cognitive Sciences and Beijing Key Laboratory of Behavior and Mental HealthPeking UniversityBeijing 100871China Key Laboratory of Machine Perception(Ministry of Education)Peking UniversityBeijing 100871China Peking-Tsinghua Center for Life SciencesPeking UniversityBeijing 100871China IDG/McGovern Institute for Brain ResearchPeking UniversityBeijing 100871China State Key Laboratory of Brain and Cognitive ScienceInstitute of PsychologyChinese Academy of SciencesBeijing 100101China 

出 版 物：《Science China(Technological Sciences)》 (中国科学（技术科学英文版）)

年 卷 期：2024年第67卷第8期

页      面：2310-2318页

核心收录：

学科分类：12[管理学] 1201[管理学-管理科学与工程(可授管理学、工学学位)] 081104[工学-模式识别与智能系统] 08[工学] 0835[工学-软件工程] 0811[工学-控制科学与工程] 0812[工学-计算机科学与技术（可授工学、理学学位）] 

基　　金：supported by the National Natural Science Foundation of China (Grant Nos. 31971031, 31930053, and 32171039) the STI2030Major Projects (Grant Nos. 2021ZD0203600, 2022ZD0204802, and 2022ZD0204804) 

主　　题：CNN perceptual learning naturalistic texture psychophysics 

摘      要：The visual system continuously adapts to the statistical properties of the environment. Existing evidence shows a close resemblance between deep convolutional neural networks(CNNs) and primate visual stream in neural selectivity to naturalistic textures above the primary visual processing stage. This study delves into the mechanisms of perceptual learning in CNNs,focusing on how they assimilate the high-order statistics of natural textures. Our results show that a CNN model achieves a similar performance improvement as humans, as manifested in the learning pattern across different types of high-order image statistics. While L2 was the first stage exhibiting texture selectivity, we found that stages beyond L2 were critically involved in learning. The significant contribution of L4 to learning was manifested both in the modulations of texture-selective responses and in the consequences of training with frozen connection weights. Our findings highlight learning-dependent plasticity in the mid-to-high-level areas of the visual hierarchy. This research introduces an AI-inspired approach for studying learning-induced cortical plasticity, utilizing DCNNs as an experimental framework to formulate testable predictions for empirical brain studies.