This paper proposes an adaptive network architecture called Hybrid SOLAR (Supervised One-shot Learning Algorithm for Real number inputs), which is a hybrid algorithm between a one-shot network construction algorithm and an iterative pruning algorithm. Hybrid SOLAR determines a network structure in two stages. In the first stage, Hybrid SOLAR requires only a single presentation of training examples to construct the network and learning is finished. In the second stage, the network prunes redundant weights to improve the generalization ability. Thus, Hybrid SOLAR retains the advantages of those two algorithms. It needs only a single presentation of the training set for learning and the generalization ability is satisfactory.
Dynamic link architecture is self-organizing map applied to image recognition. Mapping formed from data image to sample image. It tends to link two points which have similar local features, and so, can be distorted to some extent. Similarity of two images is evaluated after transforming data image by the mapping. Thus, dynamic link architecture is tolerant to distortion or rotation of data images. We propose a mathematically tractable model of dynamic link architecture, and reduced it to simple phase dynamics, which clearly shows how the model restores rotated data images. Computer simulations verify the theory.
We analyse the dynamics of sparsely encoded associative memory models by means of the statistical neuro-dynamical method. Now the three methods are taken to control the activity of the neuron, (1) optimally controlled threshold at each step, (2) fixed optimal threshold which gives a maximal storage capacity, and (3) adding the fixed threshold to lateral inhibitions. We study the basin of attraction in each of three cases by the statistical neurodynamics and the computer simulations. And, we use continuous dynamics instead of discrete one to improve dynamical properties.