Dynamic hierarchical self-organizing neural networks

In this paper, we integrate the concept of self-organizing feature maps and the architecture of dynamic trees to develop dynamic hierarchical self-organizing neural networks. The proposed network is capable of allocating new neurons dynamically during the learning process and then determining its own topology. A fuzzy membership function which measures the similarity of the input data is employed in the weight updating procedure and used as an criterion to allocate new neurons. Unlike the conventional self-organizing neural networks, the training samples are not input in sequence to update the weights. Instead, the weight updating procedure takes into account all the training data at one time. All these provide the proposed networks several advantages including revelation of hierarchical structure in data, dynamic allocation of neurons, short learning time, and short search time. To demonstrate their capabilities, the proposed networks are applied to solve the image segmentation problems.