Emergence of higher-level neuron properties using a hierarchical statistical distribution model

Essential to visual tasks such as object recognition is the formation of effective representations that generalize from specific instances of visual input. Neurons in primary visual cortex are typically hypothesized to efficiently encode image structures such as edge and textures from natural scenes. Here this paper proposed a novel hierarchical statistical distribution model to generalize higher-level neuron properties and encode distributed regularities that characterize local image regions. Two layers of our hierarchical model are presented to extract spiking activities of excitatory neurons decorrelated by inhibitory neurons and to construct the statistical patterns of input data, respectively. Trained on whitened natural images, parameters including neural connecting weights and distribution coding weights are estimated by their corresponding learning rules. To prove the feasibility and effectiveness of our model, several experiments on natural images are conducted. Adapting our model to natural scenes yields a distributed representation for higher-order statistical regularities. Comparison results provide insight into higher-level neurons which encode more abstract and invariant properties.