Data processing: artificial intelligence – Knowledge processing system
Reexamination Certificate
2000-01-18
2003-08-26
Starks, Jr., Wilbert L. (Department: 2121)
Data processing: artificial intelligence
Knowledge processing system
C706S023000, C706S025000
Reexamination Certificate
active
06611826
ABSTRACT:
BRIEF DISCLOSURE OF THIS INVENTION
This invention is a learning binary network system characterized, by consisting of an input layer, a coupling layer, a first binary gate layer with first similar logical elements, a second binary gate layer with second similar logical elements, and an output layer, in that each coupling condition between the adjacent layers limited to one way directing from their inlet side to the outlet side, and each layer has independent routes without mutual coupling conditions, the coupling layer having means for selecting either one of a direct coupling condition and a coupling condition routed through an inverter, relative to routes from the respective signal units in the input layer to the respective signal units in the first binary gate layer in the learning network.
DETAILED DESCRIPTION OF INVENTION
1. The Technical Field of this Invention
This invention relates to a binary learning system applied to such as character recognition, robot motion control and association memory.
2. The Prior Arts
A learning system has been known as a neural network. The neural network means a circuit consisting of suspected elements of neural cells (neuron) in a network in the same as basic information treating units for function of neural cells of a living body, the neural network applied to such as character recognition, robot motion control and association memory.
A neuron NE as one unit indicated in
FIG. 17
, consists of means for receiving input signal from the other neuron, means for changing the input signal under a determined rule in coupling conditions, means for limiting its changing threshold, and means for putting out the changed output. And in the coupling condition to the other neuron, a weight [Wij] indicating a coupling power is additionally attached to the unit.
The coupling condition of this neuron includes an exciting coupling (Wij>0)indicating a condition in that the more increase input of its self is then the more increase input from the other neuron, and a suppressing coupling (Wij<0) indicating a condition in that the more decrease input of its self is then the more increase input from the other neuron in reverse. Then the changing of this weight [Wij] and threshold &thgr;i course the changed constitution of the network.
FIG. 18
indicates neutral network consisting of neuron NE mentioned above, which includes an input layer, a medium layer and an output layer, each layer having no coupling condition therein, and the medium layer being capable of use of a plurality of layers. Such network is actuated so as to propagate the input signal of the input layer to the medium layer, and then the signal therein being changed with coupling coefficient, or weight and threshold, as resultant of propagating to the output layer. In the output layer, the signal is further treated as an output signal Z by addition of any weight and threshold.
Input NEX in the medium and output layers is counted by FORMULA 19 mentioned hereinafter.
FORMULA 19
Then, the neuron input NEX puts out after non-linear treatment. Further, output Yj or Zj in the medium and output layers is obtained by Sigmoid coefficient so indicated as to FORMULA 20 in general.
FORMULA 20
In this case, the leaning means to change the weight and threshold to decrease or put preferably to zero, an error between the real output Zj and a prescribed output Tj (monitor signal). This changing value is given in a manner to use error-propagating in reverse, and the changing value in the formula mentioned above is different to the neuron in the input layer or medium layer.
In analog circuits applied to the mentioned network, signal intensity of the input or output exists as a voltage, the weight of the neuron is a resistance existing on each neuron line, and the neuron output coefficient (Sigmoid coefficient) is a propagating coefficient of an amplifier. And in order to indicate the exiting and suppressing couples between the neurons, the output of the amplifier is separated to two outputs, so as to generate a plus or minus signal by reversing one of the outputs through an inverter.
The mentioned system including the analog circuits involves problems mentioned hereinafter.
a. Irregular operations of neuron elements are generated due to the temperature property.
b. As to control the error between the real output Z and monitor output T mentioned above, the error revisable circuit is complicated as well as the error introducing to zero is difficult.
c. The use of amplifier induces the further complication and large size of circuit, and the operating time is extended and causes in such difficulty that the corporation network is not produced.
Already, digital circuits for neutral network are proposed in the Japanese publication document of the patent application, its publication number being 108594/93. In this case, all neuron elements consists of logic elements without the irregular operations due to their temperature property.
However, the system having the digital circuits mentioned above involves problems described hereinafter.
a. For signal propagation between the respective layers, pulse uses so as to indicate pulse density (pulse counts per unit time) as amount of analog signal. Accordingly, this system is incapable of error control into zero and wastes long operation time.
b. The volume of each neuron element is increased, thereby causing the largest and expanded construction of the neutral network.
c. As learning, the changing value of the weight and threshold must be controlled in the respective medium and output layers under use of the prior constitution of the neutral network.
In order to solve the problems mentioned above, this inventor had proposed a new binary learning system consisting of logical elements as shown in
FIG. 1
, disclosed to U.S patent application Ser. No. 744,299/96.
The learning network consists of an input layer
21
having a plurality of binary input terminals X
1
, X
2
, . . . Xn, a coupling layer
22
, a first binary gate layer (AND layer)
23
with a plurality of similar logical elements (for instance, AND logic), a second binary gate layer (OR layer)
24
with a plurality of similar logical elements (for instance, OR logic), and an output layer
25
, the respective layers having no coupled therein and the coupling condition between the mutual layers being limited to a way only directed from the input layer to the output layer (Feed forward type).
The couple in the coupling layer, between each units of the input layer and each units of AND layer, is selected to coupling conditions mentioned hereinafter.
(1) direct coupling
(2) coupling through an inverter
(3) all [1] coupling
(4) all [0] coupling
The coupling layer applied to the coupling conditions mentioned above can consist of suspected neurons, and then the learning network is produced as shown in FIG.
14
.
In this case, one unit of OR layer
24
is only shown in
FIG. 14
for easy explanation, and the respective output terminal Z is only one.
The suspected neurons NE, as shown in
FIG. 12
, exist one input and one output, the weight Wij from the input is either one of 1 or −1 and the threshold &thgr;ij is selected to −1.5, −0.5, 0.5 and 1.5.
Then, the output Yij given by the input Xi, weight Wij and threshold &thgr;ij is all indicated in four coupling conditions mentioned above. And the output Yij is calculated by FORMULA 3 or FORMULA 4 mentioned hereinafter
FORMULA 3
or
FORMULA 4
As learning, error E between the real output Z and monitor output T can be obtained by FORMULA 5 as nextly mentioned.
FORMULA 5
In this case, the learning is accomplished with control of the weight Wij and threshold &thgr;ij, as similar to the prior idea. When the weight Wij and threshold &thgr;ij are controlled according to error E lowering downwards in the highest speed, their control values &Dgr;W and &Dgr;&thgr;are obtained by FORMULA 1 or FORMULA 6.
FORMULA 1
or
FORMULA 6
&egr;w, &egr;&thgr; are to plus and are calculated as mentioned hereinafter, by the learning rule under use of
Crowell & Moring LLP
Sowa Institute of Technology Co., Ltd.
Starks, Jr. Wilbert L.
LandOfFree
Method of learning binary system does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Method of learning binary system, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Method of learning binary system will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3084547