Pattern recognition neural network with saccade-like operation

Claim 2

A multi-layered neural network for pattern recognition, comprising: an input layer for mapping into a two-dimensional scan window, said scan window defining an input space that can be scanned over objects, and which input contains at least two adjacent ones of said objects; an output layer comprised of at least a first output node with an associated activation energy that is activated to represent the presence of a desired object substantially centered in said scan window, the activation energy thereof having a variable level, which level represents the proximity of the object to the substantial center of said scan window, and a second output node with an associated activation energy that is activated to represent a desired distance from a frame of reference in said scan window to another and adjacent one of said objects when the activation energy thereof is raised above a predetermined threshold; and a hidden layer having local receptor fields and interconnected with said input layer and said output layer for mapping said input layer to said output layer, said hidden layer providing a representation of the position of the desired one of said objects relative to the substantial center of said scan window such that said first output node is activated in response thereto, said hidden layer providing a representation of the desired distance when the adjacent one of said objects is separated from said frame of reference in said scan window by substantially said desired distance such that said second output node is activated in response thereto.

Claim 3

The neural network of claim 1, wherein said output layer comprises a plurality of said first output nodes each with associated activation energies and each corresponding to a plurality of said objects which exist in a predetermined sequence, and which objects can be disposed adjacent each other and overlapping, and wherein said hidden layer contains a representation of the position of each of said objects relative to the substantial center of said scan window, such that when any of said objects are disposed within said scan window, activation of the associated one of said first output nodes results.

Claim 4

The neural network of claim 2, wherein an additional first output node is provided in said output layer representing the presence of the center between two adjacent ones of said objects, and said hidden layer contains the representation of the center between said two adjacent objects when substantially centered within said scan window.

Claim 5

The neural network of claim 2, wherein said plurality of objects comprise letters of the alphabet.

Claim 6

The neural network of claim 2, wherein said plurality of objects comprise numbers from "0" through "9".

Claim 7

The neural network of claim 1, wherein said output layer comprises a plurality of said second output nodes each corresponding to a plurality of desired distances between said frame of reference and one of said objects within said scan window, and with said desired distances being different along a single dimension, and wherein said hidden layer contains a representation of each of said desired distances, such that when any of said desired distances exist between said frame of reference and one of said objects within said scan window, activation of the associated one of said second output nodes results.

Claim 8

The neural network of claim 6, wherein the activation energy for each of said second nodes varies in intensity from a maximum at the desired distance associated with said second output node to a minimum at a predetermined distance on either side of said desired distance in accordance with a predetermined profile such that a distance less than or greater than said associated desired distance results in an activation energy that is lower than the maximum.

Claim 9

The neural network of claim 7, wherein the activation energy curve for each of said second output nodes is associated with distances that overlap distances associated with the activation energy of other of said second output nodes.

Claim 10

The neural network of claim 1, wherein said frame of reference is the substantial center of said scan window.

Claim 11

The neural network of claim 1, wherein said representations of the desired one of said objects and the desired distance are learned by backpropagation learning techniques.

Claim 12

The neural network of claim 1, wherein said hidden layer comprises a first hidden layer having local receptor fields having shared weights and a second hidden layer having local receptor fields and no shared weights.

Claim 13

The neural network of claim 1, wherein said objects in said scan window are normalized in size to said scan window.

Claim 14

A multi-layered neural network system for pattern recognition, comprising: an input field containing a plurality of adjacently disposed objects; a scanning mechanism for generating a two-dimensional scan window and scanning said scan window over said objects such that said scan window contains at least two of said adjacent objects; an input layer for mapping into said two-dimensional scan window; an output layer comprised of at least a first output node with an associated activation energy that is activated to represent the presence of a desired one of said objects substantially centered in said scan window, the activation energy thereof, and a second output node with an associated activation energy that is activated to represent a desired distance from a frame of reference in said scan window to another and adjacent one of said objects when the activation energy is raised above a predetermined threshold; a hidden layer having local receptor fields and interconnected with said input layer and said output layer for mapping said input layer to said output layer, said hidden layer providing a representation of this position of the desired object relative to the substantial center of said scan window such that said first output node is activated in response thereto, said hidden layer providing a representation of the desired distance when said adjacent object is separated from said frame of reference in said scan window by substantially said desired distance such that said second output node is activated in response thereto; and a processor for receiving the output of said first output node and recognizing when the first output node is activated indicating the presence of said desired object centered in said scan window and providing an output in response thereto, and said processor operable to receive said second output node and controlling said scanning mechanism to move by a distance substantially corresponding to said associated desired distance when said second node is activated.

Claim 15

A multi-layered neural network system for pattern recognition, comprising: an input field containing a plurality of adjacently disposed objects; a scanning mechanism for generating a two-dimensional scan window and scanning said scan window over said objects such that said scan window contains at least two of said adjacent objects; an input layer for mapping into said two-dimensional scan window; an output layer comprised of at least a first output node with an associated activation energy that is activated to represent the presence of a desired one of said objects substantially centered in said scan window, the activation energy thereof, and a second output node with an associated activation energy that is activated to represent a desired distance from a frame of reference in said scan window to another and adjacent one of said objects when the activation energy is raised above a predetermined threshold; a hidden layer having local receptor fields and interconnected with said input layer and said output layer for mapping said input layer to said output layer, said hidden layer providing a representation of this position of the desired object relative to the substantial center of said scan window such that said first output node is activated in response thereto, said hidden layer providing a representation of the desired distance when said adjacent object is separated from said frame of reference in said scan window by substantially said desired distance such that said second output node is activated in response thereto; and a processor for receiving the output of said first output node and recognizing when the first output node is activated indicating the presence of said desired object centered in said scan window and providing an output in response thereto, and said processor operable to receive said second output node and controlling said scanning mechanism to move by a distance substantially corresponding to said associated desired distance when said second node is activated.

Claim 16

The neural network system of claim 13, wherein said output layer comprises a plurality of said first output nodes each with an associated activation energy and each corresponding to one of said plurality of objects which exist in a predetermined sequence, and which objects can be disposed adjacent each other and overlapping, and wherein said hidden layer contains a representation of the position of each of said objects relative to the substantial center of said scan window, such that when any of said objects is disposed within said scan window, activation of the associated one of said first output nodes results, said processor operable to receive the output of all of said first output nodes and recognize when any of said first output nodes is activated and then selectively generate a separate output for each of said first output nodes in response thereto.

Claim 17

The neural network system of claim 14, wherein an additional first output node is provided in said output layer representing the presence of the center between two adjacent ones of said objects, and said hidden layer contains the representation of the center between said two adjacent objects when substantially centered within said scan window.

Claim 18

The neural network system of claim 13, wherein said plurality of objects comprise letters of the alphabet.

Claim 19

The neural network system of claim 13, wherein said plurality of objects comprise numbers from "0" through "9".

Claim 20

The neural network system of claim 13, wherein said output layer comprises a plurality of said second output nodes each corresponding to a plurality of desired distances between said frame of reference and one of said objects within said scan window, and with said desired distances being different along a single dimension, and wherein said hidden layer contains a representation of each of said desired distances, such that when any of said desired distances exist between said frame of reference and an object within said scan window, activation of the associated one of said second output nodes results, said processor operable to receive the output of all of said second output nodes and recognize when any one of said second output nodes is activated, said processor controlling said scanning mechanism to move said scan window by a distance corresponding to the desired distance associated with the activated one of said second output nodes.

Claim 21

The neural network system of claim 18, wherein the activation energy for each of said second nodes varies in intensity from a maximum at the desired distance associated with said second output node to a minimum at a predetermined distance on either side of said desired distance in accordance with a predetermined profile such that a distance less than or greater than said associated desired distance results in an activation energy that is lower than the maximum.

Claim 22

The neural network system of claim 19, wherein the activation energy curve for each of said second output nodes is associated with distances that overlap distances associated with the activation energy of other of said second output nodes.

Claim 23

The neural network system of claim 13, wherein said frame of reference is the substantial center of said scan window.

Claim 24

The neural network system of claim 13, wherein said representations of the desired one of said objects and the desired distance are learned by backpropagation learning techniques.

Claim 25

The neural network system of claim 13, wherein said hidden layer comprises a first hidden layer having local receptor fields having shared weights and a second hidden layer having local receptor fields and no shared weights.

Claim 26

The neural network system of claim 13, wherein said objects in said scan window are normalized in size to said scan window.

Claim 29

The method of claim 25, wherein the output layer comprises a plurality of first output nodes, each with associated activation energies, that correspond to a plurality of objects that exist in a predetermined sequence, and which objects can be disposed adjacent to each other and overlapping and comprising the steps of: storing representations of each of the plurality of objects in the hidden layer that each represent position of the associated object relative to the substantial center of the scan window; each of the representations associated with one of the plurality of first output nodes; and activating the associated one of the first output nodes when the associated one of the objects or a substantially similar representation thereof is disposed within the scan window.

Claim 30

The method of claim 26 and further comprising: providing an additional first output node with an associated activation energy in the output layer; storing in the hidden layer a representation of the presence of the center between two of the objects; and activating the additional output node when the center between the two of the objects, corresponding to the stored representation of the center between the two of the objects, is disposed in substantially the center of the scan window.

Claim 31

The method of claim 25 wherein the frame of reference is the substantial center of the scan window.

Claim 32

The method of claim 25, wherein the plurality of objects comprises letters of the alphabet.

Claim 33

The method of claim 25 and further comprising training the neural network through backpropagation learning techniques to store the representations of the desired object and the desired distance in the hidden layer.

Claim 34

The method of claim 25, wherein the step of providing the hidden layer comprises: providing a first hidden layer having local receptor fields, having shared weights; and providing a second hidden layer having local receptor fields and no shared weights.

Claim 35

The method of claim 25 and further comprising normalizing objects that are disposed within the scan window to a normalized size.

Claim 36

The method of claim 25, wherein the output layer comprises a plurality of second output nodes, each having an associated activation energy, each second output node corresponding to one of a plurality of desired distances that exist between the frame of reference and one of a plurality of objects, the desired distances being different, and comprising the steps of: storing representations of each of the desired distances in the hidden layer that represent the desired distances; each of the representations associated with one of the second output nodes; and activating the associated one with the second output nodes when one of the objects disposed in the scan window is separated from the frame of reference by substantially the associated desired distance.

Claim 39

The method of claim 34, wherein the output layer comprises a plurality of first output nodes, each with associated activation energies, that correspond to the plurality of objects that exist in a predetermined sequence, and which objects can be disposed adjacent to each other and overlapping and comprising the steps of: storing representations of each of the plurality of objects in the hidden layer that each represent the position of the associated object relative to the substantial center of the scan window; each of the representations associated with one of the plurality of first output nodes; and activating the associated one of the first output nodes when the associated one of the objects or a substantially similar representation thereof is substantially centered in the scan window.

Claim 40

The method of claim 35 and further comprising: providing an additional first output node with an associated activation energy in the output layer; storing in the hidden layer a representation of the presence of the center between two of the plurality of objects; and activating the additional output node when the center between the two of the plurality of objects, corresponding to the stored representation of the center between the two of the plurality of objects, is disposed in substantially the center of the scan window.

Claim 41

The method of claim 34 wherein the frame of reference is the substantial center of the scan window.

Claim 42

The method of claim 34 and further comprising training the neural network through backpropagation learning techniques to store the representations of the desired object and the desired distance in the hidden layer.

Claim 43

The method of claim 34 and further comprising normalizing objects that are disposed within the scan window to a normalized size.