| 摘要 |
A semiconductor neural network includes a coupling matrix having coupling elements arranged in a matrix which couple with specific coupling strengths internal data input lines to internal data output lines. The internal data output lines are divided into groups. The neural network further comprises weighting addition circuits provided corresponding to the groups of the internal data output lines. A weighting addition circuit includes weighing elements for adding weights to signals on the internal data output lines in the corresponding group and outputting the weighted signals, and an addition circuit for outputting a total sum of the outputs of those weighting elements. The internal data output lines are arranged to form pairs and the addition circuit has a first input terminal for receiving one weighting element output of each of the pairs in common, a second input terminal for receiving the other weighting element output of each of the pairs in common, and sense amplifier for differentially amplifying signals at the first and second input terminals. The neural network further includes a circuit for detecting a change time of an input signals, a circuit responsive to an input signal change for equalizing the first and second input terminals for a predetermined period, and a circuit for activating the sense amplifier after the equalization is completed. The information retention capability of each coupling element is set according to the weight of an associated weighting element. This neural network can provide multi-valued expression of coupling strength with fewer coupling elements.
|
