Design of a partially activated neural network
Dong Hyuk Choi, Won Ho Choi · 2002
The authors designed a partially activated neural network to reduce the amount of computation in pattern classification with many classes. The structure of the proposed net is the hierarchical association of the unsupervised competitive growing (UCG) and the supervised competitive growing (SCG). The role of UCG is to restrict the number of active nodes in SCG by prediction. The hierarchical association of UCG and SCG is represented by a matrix. The minimum distance node in UCG selects a row of the matrix, and the selected row activates the nodes in SCG partially. To evaluate the partially activated SCG, a performance criteria function whose variables are loss in classification rate and gain in computational load is introduced. The network was applied to Korean character recognition for the experiments.