Title :
Towards continuously learning neural networks
Author :
Ayestaran, H.E. ; Prager, R.W.
Author_Institution :
Dept. of Eng., Cambridge Univ., UK
Abstract :
A modular three layer variable structure feedforward network capable of learning by stages is proposed. It consists of a first layer of threshold units, and two subsequent layers of logical gates. The threshold units have a vectorial threshold (instead of a simple scalar), which gives them a spatial reference within the input space. They are trained using the method of centroids, developed by the authors. The modularity of this arrangement allows progressive learning, and the extra units are added as needed, to match the complexity of the problem. The system was tested both with real data and with artificially generated data, to assess its potential. Finally, ways of expanding on the present model are discussed.
Keywords :
feedforward neural nets; formal logic; learning (artificial intelligence); logic gates; centroids; continuously learning neural networks; feedforward neural network; input space; logical gates; modular neural network; spatial reference; supervised learning; three layer variable structure feedforward network; threshold units; vectorial threshold; Artificial neural networks; Backpropagation algorithms; Multidimensional systems; Neural networks; Probability; Resumes; Supervised learning; System testing;
Conference_Titel :
Neural Networks, 1993. IJCNN '93-Nagoya. Proceedings of 1993 International Joint Conference on
Print_ISBN :
0-7803-1421-2
DOI :
10.1109/IJCNN.1993.714180
