Neural processing module with input architectures that make maximal use of a weighted synapse array
AssigneeIrvine Sensors Corporation
InventorsJohn C. Carson, Christ H. Saunders
Life of this patent
Estimating the term…
Abstract
A neural processing module is disclosed which combines a weighted synapse array that performs “primitive arithmetic” (products and sums) in parallel with a weight change architecture and a data input architecture that collectively maximize the use of the weighted synapse array by providing it with signal permutations as frequently as possible. The neural processing module is used independently, or in combination with other modules in a planar or stacked arrangement.
Claims17
An electronic neural processing module for convolving a first group of signals with a second group of signals, comprising:
- means for receiving a first group of signals;
- switching means for receiving a second group of signals and for creating successive groups of permutated signals from the second group of signals before a next group of second signals arrives;
- analog multiplying means for simultaneously multiplying each signal in the first group of signals with each signal in each successive group of permutated signals to form a plurality of products;
- and
- means for accumulating the plurality of products to produce a convolution output.
An electronic neural network image recognition system comprising:
- means for receiving a plurality of weights;
- means for receiving successive groups of image signals at a predetermined frame rate;
- switching means for creating successive groups of image permutation signals from each group of image signals before receiving a subsequent group of image signals;
- a weighted synapse array of analog synapse cells that simultaneously perform a plurality of calculations at a calculation rate, wherein the calculation rate is greater than the frame rate, the plurality of calculations comprising the multiplying of each weight with each signal in each group of image permutation signals to form a plurality of products;
- and
- means for summing the plurality of products to produce a convolution output with a value that represents a correlation quality between the weights and each successive group of image permutation signals.
Dependencies are inferred from the claim text.