Application of Directional Vectors for Independent Subspaces in the Bio-Inspired Networks

Machine learning, deep learning and neural networks are extensively developed in many fields, in which neural network architectures have shown a variety of applications. However, there is a need for explainable fundamentals in complex neural networks. It is important to know how the sensory information in neural networks develops to the higher-level processing for classification and learning. In this paper, it is shown that bio-inspired networks are useful for the explanation of network functions. First, the asymmetric network is created based on the biological retinal network with nonlinear functions. Second, the classification performance of the asymmetric network is compared to the conventional symmetric network. Prominent characteristic in the biological networks is sensitive to the motion intensity changes in their visual environments. Here, it is shown that the adjacent neurons create sensory directional information in the movements. Further, directional vectors are generated on the activities of the adjacent neurons caused by the intensity changes of the input. These vectors are useful for the generation of independent subspaces, which connect from the sensory information to the higher-level functions in networks.