Title :
A scheme for implementation of neural networks with replicated receptive fields
Author_Institution :
MIT, Cambridge, MA, USA
Abstract :
The authors shows how neural networks with local receptive fields and replicated weights can be mapped efficiently onto a CCD parallel processing architecture. Implementation of the neocognitron, a neural network for feature extraction and classification, on the CCD architecture was simulated. A modified training procedure for the neocognitron that improves its ability to extract features when using the CCD architecture is presented
Keywords :
charge-coupled device circuits; feature extraction; image recognition; learning (artificial intelligence); neural chips; parallel algorithms; parallel architectures; self-organising feature maps; CCD parallel processing architecture; RWNN algorithm; feature classification; feature extraction; implementation; local receptive fields; modified training procedure; neocognitron; neural networks; replicated receptive fields; replicated weights; Charge coupled devices; Computer architecture; Delay lines; Feature extraction; Finite impulse response filter; Image processing; Multi-layer neural network; Neural networks; Parallel processing; Pixel;
Conference_Titel :
VLSI Technology, Systems, and Applications, 1991. Proceedings of Technical Papers, 1991 International Symposium on
Conference_Location :
Taipei
Print_ISBN :
0-7803-0036-X
DOI :
10.1109/VTSA.1991.246709
