A Two-layer Feedback Vision Sparse Coding Model Based on Multilayer Perceptrons

Neural representation and computation in primary vision cortex are important and also uncovered questions in vision research. Though sparse coding and independent component analysis have been successfully used to explain the receptive field of simple-cell in VI, they are forward and static model. In this paper, we put forward a novel two-layer feedback sparse coding model (briefly as TLF-SC). TLF-SC model transforms the visual stimuli combining the independence principle and feedback of higher-layer perceptron. TLF-SC model is experimented on the simulation data set and natural image set, and the results demonstrate that outputs of independent components layer ( ICL ) in TLF-SC not only have good statistical independence, but also have good discriminability for classification task. In biological viewpoint, TLF-SC model is more accordance with the dynamic and interactive characteristics of primary visual cortex. In the same time for application, it also has good potential in feature extraction and classification problems