Investigation Of The Neural Representation To Visual Stimuli’s Hierarchy Geometry In Ventral Pathway Based On FMRI

Topological perception theory believes that the basic function of the visual system is the perception of topological properties.The perception of topology is earlier than the perception of other geometric properties.The time dependence of perceiving form properties is systematically related to their structural stability under change.A large number of behavioral experiments show that the visual system’s perception of geometric properties is consistent with the hierarchy structure of the Erlangen Program.However,there is a lack of research on the neural mechanism of geometric structures,so there is a lack of more direct evidence that the visual system’s neural representation of geometric structures also has the characteristics of Erlangen’s program.Researches on the visual mechanism show that the early visual cortex is related to the representation of the local properties in the target object,while the higher-level visual cortex,such as the Lateral occipital complex(LOC),plays a key role to represent global properties of the target object.Therefore,this article mainly focuses on: 1.Recognition of the corresponding geometric properties of visual stimuli by the response patterns of visual cortex;2.Representation structure of hierarchical geometric properties of visual cortex;3.Similarity between the representation structure of visual cortex for hierarchical geometric properties and the Erlangen Program;4.Explain the representative of the hierarchical geometric properties of visual cortex through computer vision models.In this paper,a four-quadrant identification experiment(random interval,event-related)is designed,combined with functional magnetic resonance imaging technology and multi-voxel pattern analysis,to explore the representation of hierarchical geometric properties of visual cortex.First,through Searchlight analysis,it is found that the ventral visual pathways located in the occipital lobe,including the V1 and LOC,participate in the representation of hierarchical geometric properties.In this area,the accuracy of the four-category multi-voxel classification is significantly higher than the random level.We further applied multi-voxel pattern classification to V1 and LOC.The four classification results show that the classification accuracy of Euclidean properties in the V1 is higher than that of other geometric properties,and the projective properties cannot be predicted correctly.On the contrary,the classification accuracy of the topological properties of the LOC is higher than that of the other three local geometric properties.Therefore,the V1 is more sensitive to Euclidean property,and the LOC is more sensitive to topological property.We further explore the representative structure of V1 and LOC for hierarchical geometric properties through representative similarity analysis.Category-clustering of Euclidean response patterns can be observed in the representative distance matrix(RDM)of V1,Category-clustering of local geometric properties such as Euclidean,affine,and projective can be observed in RDM of LOC.In the RDM of V1 and LOC,the representative distance between topological properties and other local geometric properties is relatively large.The results of representative similarity analysis show that V1 and LOC have different representative structures for hierarchical geometric properties.The representative structure of the V1 region reflects the representative difference between Euclidean properties and other geometric properties.The representative structure of the LOC reflects the representative difference between topological properties and local geometric properties.Through transfer classification,we find that geometric properties with higher stability are easily predicted by the classifier as similar and less stable geometric properties.This result shows that the occipital region’s representation of high stability geometric properties also includes pattern information of geometric properties with lower stability.Although the computational object vision model is constantly improving,it has not yet reached the performance of human beings.Here,we studied some computational vision models(such as gist,ssim,and deep convolutional network),and tested their ability to explain the representative structure of V1 and LOC.The deep convolutional network trained by millions of natural pictures has a similar representation structure as V1 and LOC.The results show that although the deep neural network can not fully explain the representative structure of V1 and LOC,it’s performance is better than other singlelayer feature extraction models.