| 摘要 |
<p>Plural-bit digital input signals to be subjected to weighted summation in a neural net layer are bit-sliced; and a number N of respective first through N<th> weighted summations of the bits of the digital input signals in each bit slice are performed, resulting in a respective set of first through N<th> partial weighted summation results. Weighted summations of the partial weighted summation results of similar ordinal number are then performed to generate first through N<th> final weighted summation results. Each weighted summation of a bit slice of the digital input signals is performed using a capacitive network that generates partial weighted summation results in the analog regime. The weighted summation to generate a final weighted summation result can be done in the analog regime, which facilitates the analog final weighted summation result being non-linearly processed in an analog amplifier with sigmoidal response. This non-linear processing generates an analog axonal output response for a neural net layer, which analog axonal output response can then be digitized. Alternatively, the weighted summation to generate a final weighted summation result can be done in the digital regime and the digital final weighted summation result can then be non-linearly processed in digital circuitry with sigmoidal response. Where the neural net layer is to be trained using a back propagation training algorithm, the digital circuitry with sigmoidal response can be of a type where response slope is readily determined. <IMAGE></p> |
