Typically neural network algorithms use either large (e.g. sigmnoidal) or local (e.g. Gaussian) activation functions. Here we propose a new architecture and training method that integrates activation functions from multiple scales while balancing their signal absorption in local and global scales.
FUrther information: G. Mountrakis , W. Zhuang+ (2012). Integrating Local and Global Error Statistics for Multi-Scale RBF Network Training: An Assessment on Remote Sensing Data. Public Library of Science (PLOS One), 7(8): e40093. doi:10.1371/journal.pone.0040093.[1.1MB pdf],[link]