Title :
A self-organizing neural network to categorize a random input sequence of multivalued vectors
Author :
Baraldi, A. ; Parmiggiani, F.
Author_Institution :
IMGA, CNR, Modena, Italy
Abstract :
The Simplified ART-based artificial Neural Network, SARTNN, is a flat, feedforward, unsupervised NN which detects statistical regularities in a random input sequence of multivalued vectors. SARTNN employs a competitive training procedure based on a winner-takes-all-strategy. In this paper, an improved SARTNN version, named SARTNN2, is presented. SARTNN2 introduces a bubble competitive strategy in place of the SARTNN winner-takes-all competition. The bubble strategy, allowing lateral cooperation among neurons belonging to the resonance domain centered on the winning neuron, renders the SARTNN2 mapping topologically correct. SARTNN2, featuring the typical SARTNN robustness and user-friendliness (only two user-defined parameters, both having an intuitive physical meaning, are necessary for the network to operate), improves the SARTNN performance in terms of discriminating ability and computation time
Keywords :
ART neural nets; feedforward neural nets; geophysical signal processing; geophysical techniques; optical information processing; remote sensing; self-organising feature maps; SARTNN; SARTNN2; Simplified ART-based artificial Neural Network; bubble competitive strategy; categorization; feedforward neural net; geophysical measurement technique; image processing; infrared imaging; land surface image segmentation; multidimensional signal processing; multispectral remote sensing; multivalued vector; optical imaging; random input sequence; self-organizing neural network; statistical regularities; terrain mapping; training procedure; unsupervised neural network; visible; winner-takes-all-strategy; Artificial neural networks; Computer networks; Electronic mail; Feedforward neural networks; Feedforward systems; Mathematical model; Neural networks; Neurons; Performance evaluation; Physics computing; Resonance; Robustness; Testing;
Conference_Titel :
Geoscience and Remote Sensing Symposium, 1995. IGARSS '95. 'Quantitative Remote Sensing for Science and Applications', International
Conference_Location :
Firenze
Print_ISBN :
0-7803-2567-2
DOI :
10.1109/IGARSS.1995.521719
