Sign in

Circuit employing logical gates for calculating activation function ...

 David G. Stork et al
A circuit employing logical gates for calculating activation function derivatives on stochastically encoded signals. In one embodiment a two input nonlinear circuit calculates neuron activation functions suitable for gradient-descent learning.
Inventors: David G. Stork, Ronald C. Keesing
Assignees: Ricoh Corporation, Ricoh Company, Ltd.

U.S. Classification
307/201; 307/490; 395/27

International Classification
G06G 700

View patent at USPTO

Citations

Patent NumberTitleIssue date
4893255Spike transmission for neural networksJan 9, 1990
4912651Speeding learning in neural networksMar 27, 1990
4916635Shaping command inputs to minimize unwanted dynamicsApr 10, 1990
4972363Neural network using stochastic processingNov 20, 1990
5003490Neural network signal processorMar 26, 1991
5046020Distributed parallel processing network wherein the connection weights are generated using stiff differential equationsSep 3, 1991
5052043Neural network with back propagation controlled through an output confidence measureSep 24, 1991

Referenced by

Patent NumberTitleIssue date
5479574Method and apparatus for adaptive classificationDec 26, 1995
5581662Signal processing apparatus including plural aggregatesDec 3, 1996
5619617Neuron unit, neural network and signal processing methodApr 8, 1997

Claims

What is claimed is:

1. A nonlinear circuit for calculating activation function derivatives on stochastically encoded signals, said circuit comprising

an OR gate for receiving at least first and second inputs representative of first (A) and second (B) stochastically encoded signals each having a spike frequency coding over a given interval, said first gate forming as an output the logical OR function f(A+B),
a delay gate for receiving said f(A+B) function to form an inverted and delayed output 1-f(A+B),
an AND gate for receiving said f(A+B) and 1-f(A+B) signals to form the approximate derivative quantity output signal f(1-f).

2. The circuit as in claim 1 wherein said delay gate delays one of said f(A+B) signals by at least one clock cycle so as to decorrelate said signals.

3. In a circuit for calculating activation function. derivatives on stochastically encoded signals, the method comprising the steps of

receiving at least first and second inputs representative of first (A) and second (B) stochastically encoded signals each having a spike frequency coding over a given interval to form as an output the logical OR function f(A+B),
receiving said f(A+B) function to form an inverted and delayed output 1-f(A+B), and
receiving said f(A+B) and 1-f(A+B) signals to form the approximate derivative quantity output signal f(1-f).