US5467428A — Artificial neural network method and architecture…

Claim 1

Independent Claim 394 words

An artificial neural network training architecture comprising: source means for supplying at least one input data signal representing an input vector; a plurality of input terminals connected to said source means; an input multiplexer connected to each of said input terminals; N arithmetic and storage units, representing N sibling nodes, connected to said input multiplexer; N weight multiplier units connected to said N arithmetic and storage units; wherein each arithmetic and storage unit generates a Fourier series using an input data signal received from said input terminal through said input multiplexer as a variable of the Fourier series, said Fourier series having as arguments of the sine and cosine terms said input data signal and not a probability distribution function of said input data signal, and supplies an output data signal from said arithmetic and storage unit to said weight multiplier units; an adder unit connected to every one of said arithmetic and storage units through a plurality of said weight multiplier units; a weight initializer connected to each of said weight multiplier units; a weight changer connected to each of said weight multiplier units; wherein output signals, representing synaptic link weight values from said weight changer and said weight initializer are stored in said adder unit; an output multiplexer connected to said adder unit; at least one output node connected to each of said N weight multipliers through said adder unit and through said output multiplexer for generating at least one activation at said output node; means for connecting each said output node to a mid input terminal when said artificial neural network training architecture is performing as a lowpass, highpass, and bandpass filter; means for applying to said output node, as a target value, a noise corrupted signal associated with a noise signal at each said mid input terminal when said artificial neural network training architecture is performing as an in-band filter; a comparitor and error generator unit connected to said output node, and to said weight changer for generating an error measure signal when said activation is different from said signal at said output node; means for indicating the completion of training when said error measure becomes equal to a preset value; and means for storing the synaptic link weight values of said weight multiplier units in a records file when training of said artificial neural network training architecture is completed.

Claim 2

Dependent Claim 394 words

An artificial neural network training architecture comprising: source means for supplying at least one input data signal representing an input vector; a plurality of input terminals connected to said source means; an input multiplexer connected to each of said input terminals; N arithmetic and storage units, representing N sibling nodes, connected to said input multiplexer; N weight multiplier units connected to said N arithmetic and storage units; wherein each arithmetic and storage unit generates a Fourier series using an input data signal received from said input terminal through said input multiplexer as a variable of the Fourier series, said Fourier series having as arguments of the sine and cosine terms said input data signal and not a probability distribution function of said input data signal, and supplies an output data signal from said arithmetic and storage unit to said weight multiplier units; an adder unit connected to every one of said arithmetic and storage units through a plurality of said weight multiplier units; a weight initializer connected to each of said weight multiplier units; a weight changer connected to each of said weight multiplier units; wherein output signals, representing synaptic link weight values from said weight changer and said weight initializer are stored in said adder unit; an output multiplexer connected to said adder unit; at least one output node connected to each of said N weight multipliers through said adder unit and through said output multiplexer for generating at least one activation at said output node; means for connecting each said output node to a mid input terminal when said artificial neural network training architecture is performing as a lowpass, highpass, and bandpass filter; means for applying to said output node, as a target value, a noise corrupted signal associated with a noise signal at each said mid input terminal when said artificial neural network training architecture is performing as an in-band filter; a comparitor and error generator unit connected to said output node, and to said weight changer for generating an error measure signal when said activation is different from said signal at said output node; means for indicating the completion of training when said error measure becomes equal to a preset value; and means for storing the synaptic link weight values of said weight multiplier units in a records file when training of said artificial neural network training architecture is completed.

Claim 3

Dependent Claim 65 words

The artificial neural network training architecture of claim 1, wherein said noise corrupted signal is sampled n consecutive times and applied in the form of a time series to said n input terminals, and further wherein a noise corrupted signal corresponding to a mid terminal sample appears at said output node when said artificial neural network is used as a lowpass, highpass, and bandpass filter.

Claim 4

Dependent Claim 36 words

The artificial neural network training architecture of claim 2, wherein an output signal of a first of N said arithmetic and storage units is independent of a corresponding input data signal and equals (1/2 0 .5).

Claim 5

Dependent Claim 34 words

The artificial neural network architecture of claim 3 wherein the output signals of the remaining N-1 units of each set of N said arithmetic and storage units is dependent on the input data signal.

Claim 6

Dependent Claim 78 words

The artificial neural network architecture of claim 4, wherein training time of said artificial neural network training architecture, when used as an in-band filter, is reduced by said weight initializer by initially assigning said weight values to generate at said output node an activation equal to 90% to 97.5% of the signal at said output node for the first training cycle and setting the remaining (N-1) weights of each set of N said weight multiplier units to zero.

Claim 7

Dependent Claim 120 words

The artificial neural network architecture of claim 4, wherein training time of said artificial neural network training architecture, when used as a lowpass, highpass, and bandpass filter, is reduced by said weight initializer, by initially assigning said weight values to said weights of the first of each set of N said weight multiplier units between said arithmetic and storage units and said adder unit to generate at said output node an activation equal to 90% to 97.5% of the signal at said output node in every said cycle until said artificial neural network training architecture completes its training function, and setting for the first cycle the remaining (N-1) weights of each set of N said weight multiplier units to zero.

Claim 8

Dependent Claim 83 words

The artificial neural network training architecture of claim 1, wherein a noise signal is sampled n consecutive times, forming a set of n consecutive samples and applied in the form of a time series to said n input terminals, further wherein a set of noise corrupted signals corresponding to a mid terminal sample of each set of n input terminals is used as said target values of said output node when said artificial neural network training architecture is used as an in-band filter.

Claim 9

Dependent Claim 43 words

The artificial neural network architecture of claim 7, wherein the output of said comparitor and error generator unit is the difference between the noise corrupted signal at said output node and the activations generated by the noise signal applied to said input nodes.

Claim 10

Dependent Claim 49 words

The artificial neural network architecture of claim 8, wherein said error at the end of said training becomes approximately equal to said uncorrupted signal corresponding to the mid input terminal sample and said network runs in a continual training mode to remove said noise from said noise corrupted signal.

Claim 11

Dependent Claim 28 words

The artificial neural network architecture of claim 1, wherein said training architecture is carried out by simulation of hardware implementation of the artificial neural network on a computer.

Claim 12

Dependent Claim 370 words

An artificial neural network testing architecture comprising: source means for supplying at least one input data signal representing an input vector; a plurality of input terminals connected to said source means; an input multiplexer connected to each of said input nodes; N arithmetic and storage units, representing N sibling nodes, connected to said input multiplexer; N weight multiplier units connected to said arithmetic and storage units; wherein each arithmetic and storage unit generates a Fourier series using said input data signal receive from said input terminal through said input multiplexer as a variable of the Fourier series, said Fourier series having as arguments of the sine and cosine terms said input data signal and not a probability distribution function of said input data signal, and supplies an output data signal from said arithmetic and storage unit to said weight multiplier units; wherein at least two arithmetic and storage units using the low frequency terms of the Fourier series are activated for use and all others are turned off when said network performs as a lowpass filter; wherein at least two arithmetic and storage units using the high frequency terms of the Fourier series are activated for use and all others are turned off when said network performs as a highpass filter; wherein at least two arithmetic and storage units using the mid-frequency terms of the Fourier series are activated for use and all others are turned off when said network performs as a bandpass filter; an adder unit connected to every one of said arithmetic and storage units through a plurality of said weight multiplier units loaded from a records file of synaptic link weight values; an output multiplexer connected to said adder unit; a plurality of output nodes connected to said N weight multipliers through said adder unit and through said output multiplexer; a competitor and error generator unit connected to said output nodes; means for displaying the output of the adder unit as the system output when said artificial neural network is used as a lowpass, highpass, or bandpass filter; and means for displaying the output of the comparitor and error generator unit as the system output when said artificial neural network is used as an in-band filter.

Claim 13

Dependent Claim 370 words

An artificial neural network testing architecture comprising: source means for supplying at least one input data signal representing an input vector; a plurality of input terminals connected to said source means; an input multiplexer connected to each of said input nodes; N arithmetic and storage units, representing N sibling nodes, connected to said input multiplexer; N weight multiplier units connected to said arithmetic and storage units; wherein each arithmetic and storage unit generates a Fourier series using said input data signal receive from said input terminal through said input multiplexer as a variable of the Fourier series, said Fourier series having as arguments of the sine and cosine terms said input data signal and not a probability distribution function of said input data signal, and supplies an output data signal from said arithmetic and storage unit to said weight multiplier units; wherein at least two arithmetic and storage units using the low frequency terms of the Fourier series are activated for use and all others are turned off when said network performs as a lowpass filter; wherein at least two arithmetic and storage units using the high frequency terms of the Fourier series are activated for use and all others are turned off when said network performs as a highpass filter; wherein at least two arithmetic and storage units using the mid-frequency terms of the Fourier series are activated for use and all others are turned off when said network performs as a bandpass filter; an adder unit connected to every one of said arithmetic and storage units through a plurality of said weight multiplier units loaded from a records file of synaptic link weight values; an output multiplexer connected to said adder unit; a plurality of output nodes connected to said N weight multipliers through said adder unit and through said output multiplexer; a competitor and error generator unit connected to said output nodes; means for displaying the output of the adder unit as the system output when said artificial neural network is used as a lowpass, highpass, or bandpass filter; and means for displaying the output of the comparitor and error generator unit as the system output when said artificial neural network is used as an in-band filter.

Annuity Stage	Window Opens	Payment Deadline	Surcharge Period Until	Estimated Fee	Status
3.5 Years (Year 4)	1998-11-13	1999-05-15	1999-11-14	$2000.00	Window Closed / Lapsed
7.5 Years (Year 8)	2002-11-13	2003-05-15	2003-11-14	$3760.00	Window Closed / Lapsed
11.5 Years (Year 12)	2006-11-13	2007-05-15	2007-11-14	$7700.00	Window Closed / Lapsed

Artificial neural network method and architecture adaptive signal filtering

Abstract

Inventors

Bibliographic Details

Claims Analysis

References Cited (Backward Citations)

USPTO Annuity & Maintenance Windows

Assigned Categories & Tags

Assign Category Tag