The probabilistic neural network processor (PNNP) is a custom neural network parallel processor optimized for the high-speed execution (three billion connections per second) of the probabilistic neural network (PNN) paradigm. The performance goals for the hardware processor were established to provide a three order of magnitude increase in processing speed over existing neural net accelerator cards (HNC, FORD, SAIC). The PNN algorithm compares an input vector with a training vector previously stored in local memory. Each training vector belongs to one of 256 categories indicated by a descriptor table, which is previously filled by the user. The result of the comparison/conversion is accumulated in bins according to the original training vector's descriptor byte. The result is a vector of 256 floating-point works that is used in the final probability density function calculations.<<ETX>>
[1]
William A. Fisher,et al.
A programmable analog neural network processor
,
1991,
IEEE Trans. Neural Networks.
[2]
D. F. Specht,et al.
Probabilistic neural networks for classification, mapping, or associative memory
,
1988,
IEEE 1988 International Conference on Neural Networks.
[3]
Donald F. Specht,et al.
Probabilistic neural networks
,
1990,
Neural Networks.