A neural network weight pattern study with ECG pattern recognition
Q. Xue, Yu Hen Hu, Willis J. Tompkins · 2003
Singular-value decomposition (SVD) is used to analyze the weight pattern of a back-propagation (BP) model for efficient classification of ECG waveforms. It is found that the rank of a matrix formed by all the weights, which can be determined by SVD of that matrix, is a good indicator of the number of hidden nets needed. The SVD is also used to analyze the relationship between the weight patterns and the learned features of the input patterns.>