JPH0394366A - Neural circuit network - Google Patents

Abstract

PURPOSE: To easily develop an associative storage device for color selection by providing plural input nodes corresponding to each element of a color expressing device, and a plurality of output nodes corresponding to each element of the color expressing device, and connecting the input nodes through hidden nodes with the output nodes. CONSTITUTION: Each input node 801-804 is connected with each output node 805-808, and each input node is not connected with the output node corresponding to the input node itself. Four nodes are expanded to ten nodes, and the color of the ten interface elements of Macintosh display is expressed by using the two input nodes as one set and the ten output nodes. Thus, an associative storage device which selects the visually aesthetical combination of colors can be easily developed.

Description

[Detailed Description of the Invention] [Field of Industrial Application] The present invention relates to the field of solutions based on artificial intelligence for solving problems, and more specifically, the present invention relates to an associative memory device, particularly This paper relates to selecting visually aesthetic color combinations using neural convex tract networks.

[Conventional technology]

It is known in the art to use the ten thousand methods of each skill to complete "artificial intelligence" or "A" in a computer device. Although a strict definition of AI is not available, AI can generally be described as a device in which a computer is enabled to perform mathematical reasoning. For example, after the device has been "trained" it can draw conclusions on the basis of an instruction after it has been instructed by one or a set of rules or a body of experience.

Such definitions include the so-called heuristic search" model and the "associative memory" apparatus as well as the "connectionist"
(nectionist) J mode included. Each of these terms is defined in more detail below.

Heuristic Search Model A heuristic search model can be thought of as a device that solves a problem by following the branches of a tree. A set of decisions is made. Each decision corresponds to a branch point in the tree. The search follows a set of rules, or heuristics.
tics) Controlled by J. Heuristics allow the device to zero in on the branch of the tree that provides the best solution. Heuristic search models are commercially available for a large number of applications, such as gaming, natural language processing, so-called "expert systems" and "knowledge-based" applications.

An example of a major shortcoming of such a heuristic is
Daedalus Journal●Op the American●Academy●Op Arts and Sciences (DA
EDALUS Journal of theAme
rica'n Academy of Arts an
d Sciences) Volume 117, No. 1, 191-
David Elle Waltz (Davi) published in 212
dL. Prospects for Bu11d by Waltz)
ing a Truly Intelligsnt M
achine) J.

In this paper, whether a given object rxJ is a bird or not, t−
An application for a heuristic model for determining is described. For this application, one heuristic rule could be "If X is a bird, then X can fly." However, the above paper is based on ``certification questions (quaguchi f
cation problem) J continues to be written. That is, given a general rule, one can always devise fact patterns that require more tests than the rule. For example, in a world where penguins and ostriches are classified as "birds" and both are flightless species, the rule could be "If X is a bird, then If it is not one of the birds that exist, X can fly." I have to correct it. In a world where birds can have their wings cut off, birds can die, or oil can be spread around the environment and the wings can become covered in oil, it is important to reach the right outcome. requires further exceptions to that rule. The possible exceptions to that ``general rule'' are countless.

Heuristic search models have further drawbacks in applications where it is difficult to define a "rule of thumb."

It can be difficult to define the cognitive processes that people use to solve a particular problem, and so-called "experts" in the field are often limited to 1-M. Depending on what the ``rule'' is, 1 or what the ``rule'' must be,
Rules may escape definition because it can be difficult to define the cognitive processes that people use to solve problem combinations or combinations of them. As will be seen, it is primarily the latter drawback that has led to the present invention.

Content-Associative Memory A content-addressable memory is provided that provides a memory database containing a set of historical solutions to a problem or set of problems. The associative memory solves the current problem by examining its symptoms or features and comparing those symptoms to previous solutions to the problem. A solution is selected from a group of historical solutions if the selected solution has a set of symptoms that are close to the current problem.

This type of device is distinguishable from heuristic search models in that it uses a "search" process rather than searching through a set of rules. The item ``searched for'', ie, the solution, is not a rule, but represents a particular solution.

An example of such an associative memory is its application in diagnosing medical problems. Symptoms and patient characteristics, and 1 of patients with IVf foot.
Associative memory was designed by Waltz, in which, given a database of a set of symptoms and characteristics, the device finds the most similar past patient and the process by which the same diagnosis should be made for the new patient. It has been stated.

An associative memory can be constructed that can "learn" by adding new syndromes to its database, determining whether proposed solutions yield the desired results, and storing new syndromes and new solutions in the database.

God! ``Sonoro Network 1 Neural Network'' can be described as a specific realization of an associative memory device. In a neural network, nodes 1 or processing units are interconnected by links, each link having a weight or strength.

Neural networks are stimulated by electric currents that are considered for processing and storage within the human brain. (The brain is made up of a huge number of highly interconnected elements that send each other very simple excitation and inhibition messages, and use those messages to control the excitation of the element. ) In a neural network, each individual neuron can function as an input, as an output, and can be hidden. (Hidden neurons are connected to other neurons, but It cannot be accessed directly from the input or output paths, as will be explained in more detail below.The connection between any two particular neurons is strengthened each time a flow flows between those two particular neurons. That is, given a problem expressed by an input neuron, and a solution expressed by an output neuron is found, the connection between A and B is strengthened.

A further explanation of the neural network is given by Daedalus.
Journal●Op●The●American●Academy●Off゜●Arts●and●Sciences (DAEDALU
S Journal of the America
anAcademy of Arts and
117, No. 1, pages 85-121
D. Cowan) and David H.
By David H. Sharp [
Neural Netsand A
rtificial Intelligence)
This was done in a paper entitled.

The educational neural network undergoes one 'education' process while being taught by one set of inputs. During this educational process, it is hoped that a stable configuration will eventually be reached.

A stable configuration can be thought of as a state of the neural network in which a set of inputs to the neural network produces a desired set of outputs.

An automatic associative network is a one-game associative memory system that is trained by using a set of words. These examples are often called exemplars. The neural network is trained by using exemplars as inputs to it and using those exemplars as target outputs. The weight of the link between the input node and the output node is 1 such that by adding a set of exemplars to the input node, a response sufficiently close to the target response appears at the output node, and v! 4 will be adjusted.

A typical known auto-associative network consists of binary values (O and 1).
). of the network by using simple binary inputs. Stabilization and education can be done more easily.

Cowan et al. Hopfield (
Hopfteld) describes a particular automatic associative memory device known as a network. A Hopfield network is a network of neuron-like elements with symmetrical connections. (A symmetrical connection was previously defined as: If X and y are two neurons, then the weight of the connection from X to y is equal to the weight of the connection from y to X. (Then, the connection of
He discusses and points out that the Hopfield tract network is not the ``neural tract network'' of FX. More specifically, each neuron excites one nearby neuron and inhibits another neuron.

One problem with Hopfield networks is that they can easily become trapped in metastable configurations. To avoid the metastable configuration, an improved Hopfield network is proposed that uses a "Monte Carlo" method to allow the Hopfield network to escape from the metastable configuration.

Using the Monte Carlo method, the Hopfield circuit M4
performs essentially random configuration changes. If a change results in a more stable state, that change will be maintained.

If the change lowers the stability, the change is canceled. Although in this way a stable 41l is eventually found, the 111l can be very slow.

Cowan et al. also describe a type of one-way network known as a multilayer one-way network, see Figures 1(A) and 1(B). ,
And by comparing it with a single layer network, it can be described in a simple way. The Hopfield network described above is computationally equivalent to a single layer network.

Figure 1(A) shows three input nodes 101, 102, 10
3 and three output nodes 104, 105, 106. Each input node 101, 102103
are directly connected to each output b-path network 104, 105, 106.

Simple path networks such as that shown in FIG. 1(A) may be useful for certain pattern recognition and classification problems. In general, this type of one-way network can recognize patterns that can be classified and separated by hyperplanes. This drawback is explained by IEEE Koichi SB Magazine (IEEE ASS
P Magazine), April 1987 issue, 4-22
Richard Bea Lipman (Ric) on Beige Rr.
hard p. An Intro to Computing with Neural Networks (Lippmann)
duction to computing
Neural Networks) J.

A type of network that can recognize such complex patterns and solves many of the problems associated with single-way networks of the type shown in FIG. 1A is called a "multilayer network." These networks typically have input and output layers interconnected by "hidden" layers.

Lippmann describes a single layer network as one in which the 9@U region can be identified by a hyperplane. Multilayer networks are arbitrary domains whose complexity is limited by the number of nodes in the network.

Another way to examine the shortcomings of a single-layer, one-way network is to be able to recognize patterns that are close to the training pattern of the network, but to be relatively poor at recognizing patterns that are not similar to the training pattern. be.

Figure 1(B) shows three input nodes 110, 111.11
A multilayer network with two and three output nodes 116, 117, 118 is shown. Additionally, this one-way network also has three "hidden" nodes 113, 114, 115. The term "hidden" refers to the fact that the node is not directly accessible from outside the network because of its input or output, and is connected between the input and output nodes in the network.

Back propagation is one technique used for training multilayer networks. During training, including back-propagation, the network undergoes a forward phase in which inputs to the network are simulated and responses are recorded. During the packed word phase, the weights of the connections between the hidden units and the output units are adjusted, and after adjusting the weights of those connections, the weights of the connections between the input units and the hidden units are The pieces are arranged.

FIG. 1(C) shows such a network that can be trained with back propagation techniques. As previously mentioned, this circuitry receives input to the social canon 120.

The responses to output nodes 122, 123, and 124 are recorded. Then the connection weights 126, 127, 12
8 is adjusted and after that adjustment the weight of connection 125 is adjusted. This is where the name pack propagation comes from.

The example in FIG. 1(C) shows a route network with one input node. Other networks have many input nodes. Furthermore,
Other networks may have more trenches or fewer output nodes.

Color Selection Device The above discussion is about the history and development of known AI devices. The present invention centers on using an AI device for color selection. In particular, as discussed below, preferred embodiments of the present invention are useful for color selection for computer interfaces, and are useful for color selection for components of Apple@Macintosh computer interfaces. Especially used.

It is known to use AI devices for color selection. Rochester Ins
position of Technology), School of Computer Science and Technology (School of C
computerScience and Techno
1 Color Noclette by Penn' Bauersteld dated August 16, 1988 in ``Knowledge Based System'' Two Select Colors Color Palette (Coloy Palette: A Knowl
edge-Based System to Se
Lect Colors for a Palette
) Master Thesis entitled J.
) describes a knowledge-based color selection device.

This article describes a knowledge-based device used to advise users in selecting colors for palettes used in computer screen design. The knowledge base proposed in this paper describes the relationships between individual colors contained in a set of colors and other colors contained in that set, for interface design and color analysis. Reasoning based on a set of rules obtained from research on a theory. Pauersfeld states that "the skill required to make a color selection is cognitive and can be easily expressed by an expert" (page 10 of the above-mentioned graduation thesis).

A second device for color selection that can be pressed on the computer device is 1
Shitsugraph on User Interface Software held in Banff, Alberta, Canada, October 17-19, 1988.
Symposium (8 I GGRAPH Symposium)
Um on Uaer Interface Soft
Proceedings of the AC
M) “Barbara J. Mier” on pages 117-128
AC: A Color Ex Heart System
7 O User Inc. Face●Design (AC
E: A Color Expert System
for UserInterface Design
) is published in a paper entitled J.

This paper describes an expert system that uses a set of rules for color selection for computer interfaces, etc.

Bauers Fold
and Meir have proposed devices for color selection, but both rely on a set of rules, or heuristics, to implement their proposed color selection methods. . Although rules are available to assist in color selection, there are no rules that cover all possible combinations of color selections that meet all the criteria of the so-called I exno.

Accordingly, the present invention recognizes that any implementation technique based on a set of rules or heuristics is subject to drawbacks similar to the general drawbacks of the heuristics described above.

[Problem to be solved by the invention]

It is therefore an object of the invention to develop an associative memory '#crit for color selection.

Another object of the invention is to develop an associative memory particularly useful in the design of color interfaces for computer devices.

Yet another object of the invention is to develop an associative memory for implementing a color selection device.

[! 〕Means for solving the problem〕

This specification describes an apparatus and method for assisting in the selection of color combinations. The present invention discloses an associative memory particularly used in the design of color interfaces for computing devices.

With content addressable memory, it is provided by the user of the computing device. You can design color interfaces based on your selections. The use of associative memory for color selection methods has also been found to be advantageous over "rule-based" color selection devices.

The apparatus of the present invention is implemented in a computer system having multiple elements on each inter-78 screen. In particular, a preferred comparator device utilizes ten separate elements. The disclosed color selection method allows the user to specify a color for any element in the interface from a selection of available colors. After selecting a preferred color for the violet elements, Aeser invokes the content addressable memory of the present invention to assist in selecting a color scheme for all of the IO elements. A color scheme with user specified colors for b% selected elements is obtained by the associative memory.

The present invention provides a method for preserving color combinations, a method for restoring a retained color set, a method for inspecting and "designating" a color set by a user, and a method for setting a color interface to monochromatic mode. and other methods useful in selecting colors for a computer interface are also disclosed.

The present invention also discloses a neural network implementation for a color selection device. Neural network devices can be described overall as a hybrid of automatic associative memory devices and multilayer neural nets that use back propagation for instruction.

The neural network of the present invention has a plurality of input units. Each input unit corresponds to an element of the computer interface. This neurooral network also has multiple output units. Each output unit corresponds to an element of the computer interface 7A. Each input unit is coupled to all output units via subnetworks, except for the output unit corresponding to its own element. Each output unit is coupled to its corresponding input unit via a circular link. Therefore, this main circuitry can be considered as the eye movement associative memory network portion of the hybrid circuitry of the present invention.

Each sub-network has a plurality of input nodes corresponding to each element of the color representation device and a plurality of output nodes corresponding to the elements of the color representation device. Input nodes are coupled to output nodes via hidden nodes. In the preferred embodiment, the RGB color space is utilized, so there are three input nodes and three output nodes, corresponding to red, green, and blue, respectively. Each of the three input nodes is fully connected to the three hidden nodes, and each of the three hidden nodes is fully connected to the three output nodes.

Therefore, each sub-network can be considered a multi-layer single-way network portion of the hybrid network of the present invention.

The present invention discloses a new technique for network mitigation. Mitigation of the network of the present invention utilizes back activity mitigation. The buck activity mitigation technique of the present invention is a method of conserving power in a network (as opposed to traditional methods of relaxing weights) by mitigating network activity.

This specification describes a color selection device.

In the following description, numerous specific details are set forth, such as specific computer interfaces, specific colors, etc., in order to provide a thorough understanding of the invention. However, it will be apparent to one skilled in the art that the present invention may be practiced without these specific details. In other cases, the present invention may be used unnecessarily! In order to avoid this, a well-known circuit,
Structure and technology were not detailed.

〔Example〕

Introduction The present invention describes a color selection device that has particular application in color selection for computer interfaces. Although the preferred embodiment of the invention is implemented on an Apple Macintosh n computer, it will be apparent to those skilled in the art that the invention may be used with other computing devices and outside the realm of computer activity. .

The invention describes a preferred implementation of a color selection device with a neural network arranged as an auto-associative network with recirculating links from output to input. Although this network is described for the specific application of color selection mode, the network is useful for other applications as well. Furthermore, it will be apparent to those skilled in the art that other embodiments may utilize other associative memories to perform color selection.

In modern computing devices, color plays an increasingly important role. This can be used to store user information, such as color screens or presentation slides,
This is the case in the use of computer equipment for presentation, in the use of high resolution graphics for technical and other applications, and in the design of computer equipment interfaces.

In particular, the use of color in the design of computer interfaces can make the interface more pleasing to the eye and improve the interface's ability to convey information.

However, it would be desirable for the end user of the computer to be able to change the colors of the interface to suit his personal preferences or to convey information that is more or less useful to a particular user. One problem with allowing a user to "select" a color in his interface is that it is difficult for many users to achieve aesthetically pleasing results.

Accordingly, a device is desired that allows users to customize their interface and yet assist in color selection to achieve desired results. The analogy can be used for the use of interior designers to assist in selecting a color scheme for a home or office. Typically, the interior designer obtains information about the client's particular tastes. Using this information, interior designers try to design color combinations that are comfortable for the client.

One observation of the present invention is that although a color expert, such as an interior designer, works within a set of overall principles of color design, the skill of the color expert is more artistic than technical. .

These artistic skills are difficult, if not impossible, to reduce to the necessary set of rules, such as that required to develop a color selection device based on a heuristic model.
Have difficulty.

Associative Memory for Color Selection Accordingly, the present invention provides a color selection device that uses an associative memory to assist in color selection. An associative memory network does not need to control a set of rules; the network can be trained with a set of "good" solutions and acquire the ability to operate as if it knew the set of rules. do.

This is similar to a color expert acquiring information and developing new color combinations. That is, a color expert can observe and construct pleasing color combinations, but is often unable to state specific rules or reasons for selecting or approving a color combination.

In another preferred embodiment of the present invention, it is desirable to implement a color selection device that uses the features of a rule-based expert system in combination with the features of an associative memory. For example, the expert system is 1
Neural networks can be used to first apply a set of rules and then select color combinations that match those rules. Alternatively, the expert system first obtains the color combinations and then applies a set of rules that match the proposed color combinations to established rules. As an example, rules requiring a certain level of contrast between text and background elements can be applied. Importantly, associative memory devices, either alone or in combination with rule-based devices, offer significant advantages in the color selection process over traditional devices that use only rule-based color selection devices. is obtained.

The device of the invention and its application to computer interfaces will be detailed below.

A preferred embodiment of the invention is a Maci
ntosh) computer interface (the Macintosh computer interface is commonly referred to as a "desktop"), which assists Aeser in selecting colors for elements in the Macintosh computer interface (the Macintosh computer interface is commonly referred to as a "desktop").

Of course, the present invention has application to interfaces with other computer devices, and to other applications on a computer device, such as color selection for graphic designs developed on a color computer device. That is, of course, obvious. The invention can also be used for other color selections. Color selection for interior decoration, make-up and wardrobe are examples of such applications.

Preferred Embodiment Color Selection Method The preferred embodiment interface is shown in FIG. As is familiar to many users of Matsukin toss computers, the Matsukin tosshu interface consists of the desktop 201 and the back of the menu bar t202.
and. Menu bar text 203, menu text 204, and . Menu background 205, window text 206, and window highlight 207
And window par 20B! 209 and a scrollper 210. In a color interface, each of these elements can be assigned a different color.

(Of course, a particular interface can also assign the same color to multiple elements.) It should be noted that the pine interface includes other elements. In accordance with a preferred embodiment of the present invention, colors may be assigned to the elements described above, and in other embodiments may be assigned to the entire element set or to another subset.

In the preferred embodiment device, a particular screen may have more than one menu option in menu bar text 203 and therefore more than one set of menu text 204 and menu background area 205. In the preferred embodiment, the options that apply to menu bar text 203 appear in the same color.

Each menue text 204 appears in the same color and each menu background area 205 appears in the same color. Similarly, window text area 206 and window highlight 207
A number of windows can be provided, each having a window 208, a scroll 210, and a recess 209. Those areas of each window are, in the preferred embodiment, . From window to window K,
They also appear in the same color.

3-8 are flow diagrams illustrating the use of the preferred embodiment color selection method.

FIG. 3 illustrates a preferred embodiment method of accessing the color selection device. First, the user desires to change the computer's color interface (block 301).
). The user “submits” a wish for color selection (block 30)
2). To submit your wish, just submit or start! Select the symbol on the desktop that corresponds to your wish and double-click on the selected symbol. (The term "click" is familiar to users of Macintosh computers and is typically associated with the use of a mouse as a pointing device. To do this, use the mouse to move the cursor to the desired display location, then press and release the button on the mouse. Double-clicking involves pressing and releasing the button on the mouse twice in quick succession. ) As familiar to users of Matsukintoshu computers, a menu bar is available along the top of the Matsukintossie screen. SUMMARY OF THE PREFERRED EMBODIMENT FOR PERFORMING A COLOR SELECTION WISH According to another embodiment, a user selects a function from a menu bar by pointing to the menu bar item with a pointing device, such as a mouse. be made possible. The user then selects a color selection function from the selected menu by depressing the pointing device (eg, a button on a mouse) at the point where the color selection function is highlighted and then releasing the pointing device.

In response to the selection of the color selection device, a screen such as the screen shown in FIG. 2 appears on the display (block 3o3). This screen shows a standard Matsukinto Switch Inter 7A, with label boxes pointing to each element which can change its color. The screen also shows a color palette 221 and a set of function buttons 220. These function buttons can be selected by the user. Each function button 221 will be explained in detail later. The color palette 221 has 16×16 square colors (256 colors in total).

After the color selection screen appears, the user may select any of the features available in the color selection process (block 30).
4).

FIG. 4 shows the preferred embodiment process for selecting colors. The user randomly decides that he wants to change the color of the elements on the Inter7Ace (block 401).

As mentioned above, the color selection method of this preferred embodiment involves adding buttons to the color selection process to consider the user's selections and suggest color combinations based on those selections. Configured to help. Accordingly, the user supplies information to the trench device about the user's selections.

To terminate this input supply, the user selects an element (block 402). The elements to be selected are 10 types of elements in Matsukin Toushinter 7Ace, namely, the desktop 201, the back of the menu bar t202
, menu bar text 203, menu text 204, menu background 205, window text 20
6. Window number 1 illumination 20T1 It can be any one element among the window par 208, the knob 209, and the scroll par 210. The user selects a desired element by indicating a label associated with the element with the pointing device and clicking on the label.

After an element is selected, the user selects a color for the element from color palette 221 (block 403).

In a preferred embodiment, the user selects a color by pointing pointing device e at color palette 221 and using the pointing device to indicate a desired color and clicking on it. Alternatively, embodiments of the invention can be developed that allow the user to automatically place the cursor in the color palette area after selecting an element.

After selecting a color, the elements on the display screen appear in that color. In this way, the user can immediately receive direct feedback on how the selected colors interact and can vary the selected colors until the desired combination is obtained.

After selecting a color for the first element, the user can select a color for the additional element (branch 4 0 4). In this way, the user has information about the preferred color scheme for the desired 凛. Can be supplied.

After the user has selected a color for the desired element, the user can select the "Suggest Color" function by pointing to and clicking on the "Suggest Color" button. The device then uses a preferred embodiment of a neural network implementation to suggest a color for each element that the user did not previously select.

The "Suggest Color" function provides the user's selected color scheme as input to the network. The network assumes that the selected element has not changed its color. The circuit network is 10
Ten different colors are generated as output, corresponding to each type of element, and the color scheme is presented to the user on a display screen. Those 10 colors are the colors selected by the user,
Includes a color that closely "matches" the selected color and a color published by Ichiro Net.

After the colors suggested by the network are presented, the user selects a choice and changes the color scheme (block 407).
. The user does this by selecting colors and elements as described in connection with blocks 402 and 403.

The user can lock or unpin the color selection by selecting the "Lock Color" function.

To select the "Fix Color" function, the user points to and clicks the "Fix Color" button. In response to selection of the color fixation feature, the user is presented with a set of test boxes corresponding to each element in the interface. All elements that were previously explicitly selected (in the course of blocks 402 and 403) appear as tested. Colors selected by neural networks? Do not test all elements that have The user can test all devices that are not tested, and can untest any device that has been tested by clicking in the appropriate box. (The inspection box acts as a toggle switch that toggles between being inspected and not inspected.) Of course, the user can selectively pin and unpin color selections. There are several different ways to do this. One of these alternative methods utilizes a visual indicator placed near each label on the desktop, in the casing, or within the label.

An example of a visual indicator is a suppression symbol. A visual indicator then indicates to the user whether the color for a certain element is fixed (eg, a suppression symbol can be highlighted when the color is fixed). The user can toggle the element between pinned and unpinned by clicking the visual indicator.

Then the user can select the "color suggestion" function again,
The device suggests colors for all elements that are unpinned (block 409).

After the desired color combination has been achieved by performing the color selection process, the color fixation process and the ``color suggestion'' function process, the ``Keep'' button can be indicated and clicked. The color combination is retained.

The current color combination is then retained and used as the color combination for the Macintosh and Linux interface after exiting the color selection process.

The preferred embodiment of the present invention includes several other features in the color selection method that users of the color selection method have found useful.

FIG. 5 illustrates the preferred embodiment method utilized when a user desires to set the color of an Inter7Ace to a standard black and white value κ (block 501). The user says “
``Black and white'' button and click it.
The ``black and white'' function is selected (block 502). ] In response to the selection of the black and white function, the Matsukintoshie interface is reset to standard black and white values (block 503).

FIG. 6 illustrates a preferred embodiment method for allowing each sample color set to be examined by a user who desires to examine the designer color sets used to train the network (block 601). The user points to and clicks the “Designer Color Set” button.
Designer color set function can be selected again. It then allows the user to cycle through each color set to train the network (block 603). The development and selection of a set of colors for teaching one-way networks is detailed below.

It is worth noting that a "keep" function can be utilized after examining any designer color set in order to use a particular designer color set as a new color set for the Matsukin Totsushie color set. The "keep" function can also be used to keep the black and white color set as the color set for the interface. Furthermore, it is possible to fix the element, unfix the element, and select a new color for the element (see Figure 4). Either can be modified. The "color suggestion" function can then be used to generate a new color set.

Referring now to FIG. 7, after modifying and testing the new color set, the user may desire to return to the last retained color set (block 701).

To do this, the user points to the "Back 1" button and clicks the button (block 702). 1
In response to clicking the "Revert" button, the device reverts the interface image to the last retained color set.

The feature sets described above are not meant to limit the scope of the invention. Similarly, other features can be added to embodiments of the invention. For example, the presently described preferred embodiment of the invention provides the user with the ability to maintain one color set. Ability to allow users to maintain a large number of color sets and later select from the retained color sets for later use and modification'5r
．． Can be added. An educational feature can be added that allows the user to design new color sets to be used to further educate the network. Custom designer sets available.

for example. Designer sets can be designed and marketed under the name Designer. The purchaser of such a custom designer set can design a designer color set that is influenced by the "name"designer's selection. The device of this preferred embodiment is configured to read values for hardening connections from an external file. This aspect of the preferred embodiment functions to support the use of a single name designer color set.

Preferred Embodiment Associative Memory The design of the associative memory of the present invention involves evaluation of several associative memories. From research on known associative memory devices, it has been determined that the known devices do not provide a satisfactory solution to the problem posed by the desired color selection device.

The device of the invention includes a large number of possible combinations.

There are many possible colors. Select a specific color from those possible colors. The computer system of the preferred embodiment computationally supports 248 colors.

Second, any one of those colors can be added to any one of several elements. In the preferred embodiment, there are ten elements that can be assigned colors. Therefore, there are a total of 20 colors that the user can select. Even in the preferred embodiment, where the user is limited to 256 (28) possible color selections, there are 280 possible combinations.

A successful implementation of a color selection device must be able to support color combinations for which Iicti has not been taught. That is, a color selection device that simply reads color combinations from storage is impractical given the large number of possible combinations. A feature that can be generalized from a set of examples to provide a solution to a particular situation is called a "generalization."

Generalization is known from auto-associative networks. In general, an auto-associative network, like many one-way networks, is a set of educational examples (
be taught by example). Auto-associative networks can be generalized to store a large number of memory patterns (based on exemplars) with a large number of point attractors, and to identify similar patterns as one. (Instead of creating different patterns for each pattern of a kind, this network creates an "average" of the patterns.
Or create one attractor for "F Grototype". ) develops a device that allows a user to select a new, unknown color, and a circuitry uses its knowledge of the numerical relationships between ``one match'' of colors to determine the proper use of the new color; lfi! It is desirable to generalize. Achieving a network with this performance requires not only training the network to create multiple point attractors, but also training a single view around the attractor basin. A generalization of this kind can be described as a Seman-bound network using a single basin.

However, this kind of generalization is known and dangerous in auto-associative networks.

Another difficulty arises from the fact that the colors used in the drawings are represented as a set of continuous values.

In the preferred embodiment, colors are represented as RGB (red-green-blue) values. Each RGB value is represented by three 16-bit integers. Of those three integers, one is for red, one is for green, and the remaining 01 is for blue. Typical auto-association models store only binary patterns. Implementation using a binary encoded representation of continuous RGB values is prohibitive since that many input units and output units are required.

Accordingly, we disclose a one-way network that can be described as a two-layer backpropagation network with added circular links to model an asymmetric continuous auto-associative memory device.

FIG. 8 shows the basic configuration of the color selection circuitry of the present invention.

A set of 10 input nodes and 10 output nodes are used to represent the colors of the 10 interface elements of a Macintosh display.

FIG. 8 shows a desktop 801 with four input nodes.
menu text 802 and scrollper 803
, a trench 804, a desktop 805 which is four output nodes, a menu text 806, a scroller 807, and a box 80B. It will be clear to those skilled in the art how the illustration of FIG. 8, which shows four nodes, would be extended with a preferred embodiment implementation having ten nodes.

each person canode (for example, nodes 801, 802, 80
3.804) at each output node (e.g., node 805
．． 806, 807, 808). However, each input node is not connected to its corresponding output node. For example, input node desktop 881
is connected to the output node menu text 806 , scroller 807 , and thumb 808 . Thus, in the preferred embodiment, each input node is connected to nine output nodes. Input node desktop 8
01 is not connected to the output node desktop 805. Thus, in the preferred embodiment, each input node is connected to nine output nodes. The resulting network is a complete and connected 10-node Hopfield network, except for asymmetric interconnections.
field) can be thought of as a circuit network. (Without a connection from an input unit to its corresponding output unit, the network will not learn to copy the input directly to the output, but instead will do the mapping indirectly through other interface elements.) (It is worth noting that the network has to be learned manually.) This network also includes a recirculating connection from each output node to its corresponding input node. For example, output node desktop 805 is connected to input node desktop 801. This connection causes the state of the output node to be copied back to the input node during the relaxation process (the relaxation process will be explained in detail later).

Each individual inode-output node pair shown in FIG. 8 actually corresponds to a sub-network. FIG. 9 shows the input node menu text of FIG. 8}802 (input node 91 of FIG.
9) and the output node scroller 807 (shown as output node 912 in FIG. 9). This connection connects three input units 910 (i.e. red 901, green 902, blue 90
3) and three output units 912 (i.e. red 90
4, green 905, blue 906), three hidden units 911, and 18 weighted links. 9 of the links are 3 input units sio
Completely connect the hidden unit 911 to the hidden unit 911, and the remaining nine connect the hidden unit 911 to the output unit 91.
Fully connect to 2.

Thus, the preferred embodiment one-way network has a total of 30 input units and 30 output units. As explained above, each of the 30 input units and each of the 30 output units are grouped into a set of three units. There are therefore 10 input unit groups corresponding to 10 screen elements. Each of the ten input unit groups is connected to nine output unit groups, so there are a total of 90 connections between the input and output unit groups. The input units are connected to the output units via three hidden units. Therefore, the one-way network of the preferred European embodiment has an additional 270 (90×3) hidden units. and <, each of the three input units in each input unit group is connected to each of the three hidden units associated with it. Each of the three hidden units is connected to each of the three output units associated with them. In total, the device has 1620 forward links or connections (90 times the connections from the input group to the output group; each connection has 9 connections from the input group to the hidden units and 9 connections from the unit group to the output group, so there are 90 connections (
9+9)1290×18=1620).

As mentioned above, the device also has ten recirculating connections from the output color nodes to the corresponding input color nodes.

Each recirculation connection connects the red node of the output unit to the red node of the corresponding input unit, the green node of the output unit to the green node of the corresponding input unit, and the blue node of the output unit to the blue node of the corresponding input unit. Connecting.

That is, there are 3×10=30 connections in total.

Although the preferred embodiment of the invention is implemented using an RGB representation of color, alternative embodiments can be implemented using other color representation systems. For example, common color representation systems include HSV (hue, saturation, and value) and CMY (cyan, macenta, yellow).

Since RGB encoding is standard for Macintosh II hardware and software, a KGB encoding system was chosen for the preferred embodiment described herein. Additionally, the circuitry of the present invention has applications other than the field of color selection. This kind of one-way network. This can be useful in many applications where it is difficult for experts in a particular field to formulate rigorous rules, but it is easy to express opinions about what is best.

Education of the circuit network The circuit network of the present invention can be taught using a color set designed by a "color expert". Specifically, the preferred embodiment apparatus described herein was trained on a 36 color set designed by five different color experts. Back propagation was used to train the network with a color set provided by an "expert." The educational pattern was used as the input pattern and as the target pattern in the output layer.

Of course, you can use another color set to train the network. An example of another color set was described above in connection with the description of a custom designer set designed by a designer.

B-way Network Mitigation This preferred implementation network utilizes a technique termed "backactivity mitigation" for back-way network mitigation. (Relaxation can be thought of as a process that allows the network to find a set of well-matched colors, i.e., the solution at the lowest energy level.) The term energy refers to the desired or target output and the actual (Often used in neural networks to describe the level of error between the output of Find a set of weights.

Here, the error is defined as the difference between the output value and the target value. The theory underlying this is that a least squares error analysis yields a condition in which the color of each element most closely matches the color of the other elements. Essentially, this network can be viewed as ten separate sub-networks. One sub-network corresponds to each element. Each sub-network is trained to learn the relationship between the color of each sub-network element and the colors of the other nine elements. By learning the numerical relationships between those colors, this network can respond to colors that did not appear in the training set.

Conventional techniques include backactivity propagation techniques characterized as learning algorithms to reduce the energy of a network by relaxing the weights of connections between nodes. Typically, pack propagation algorithms integrate the overall energy of the circuit by parts with respect to the local weights of the connections.

The present invention provides a pack activity mitigation technique that takes the partial integral of energy with respect to the activity of the connection rather than the weight.

The backactivity propagation algorithm is detailed with reference to Appendix I.

What has been described above is a one-way network that can store color patterns that can be encoded in continuous values and can learn relationships between colors of interface elements. The network has a recirculating network with pack propagation for learning and backactivity mitigation for solving. The circuitry of this preferred embodiment can be used to model an asymmetric auto-associative memory device capable of storing continuous values. The ability to model such automatic associative memory devices is a novel aspect of the present invention.

In the present invention, one important aspect is that any given unit is not connected directly or through hidden units to a corresponding output unit.

It is theorized that by allowing an input unit to be connected to a corresponding output unit, the network will learn to convert the input to the output.

Appendix I Back Activity Mitigation Algorithm In particular, given any input pattern, energy is defined as the mean squared error between the target and output activity.

If we set , we can obtain . For units that are hidden to the output unit by the sigmoid logistic function, let J be the value of the unit l in the lower layer. I would like to have a measurement of how much value of a certain unit needs to change.

aB ゜゜i＝“aat Here, η is the adjustment rate.

net' = Σ11W 1j 1 In particular, δk is just θE of enit k from the next layer.
There is a / a n e t k te.

The output anyt and the network pack propagation δ'B from each input unit l receives gradients from its output links.

The above integrals are for standard forward feed networks only. The color network also has recirculation lines connected back from the output layer to the input layer to simulate an auto-associative network. An output node and its corresponding input node actually represent the same color, and the input value is also the target activity for the output unit. Since each color is only connected to its neighbor, the gradient back-propagated into the unit did not take into account the IA caused by the unit itself. It is necessary to add the second term to the integral by parts.

In particular, Ot is the activity value of the output unit corresponding to input unit l.

In this way, each input unit}1 adjusts its value by Δa, and the relaxation process is repeatedly seen in the case where the active state reaches an equilibrium point.

[Brief explanation of drawings]

Figure 1 (A) shows a conventional single layer network, Figure 1 (B)
The figures show a conventional multilayer network, FIG. 1(C) shows a conventional one-way network illustrating a back propagation technique, FIG. FIG. 4 is a flowchart showing the access method of the color selection method that can be used in the present invention. FIG. 5 is a flowchart that shows the color selection method that can be used in the present invention. FIG. 6 is a flow diagram illustrating the method for examining designer color sets; FIG.
FIG. 8 is a flowchart illustrating the method of the present invention for returning to a previously retained color set; FIG. 8 illustrates the wholesale network that the present invention can utilize; FIG. . 201 , 801 , 805 ●●●●Desktop, 202 ●●●● Menier back! , 203 ●
●●●Menier Vertext}, 204 ●●●●71
= U-text, 205 ●●●・Menu background,
206 ●●●● Window text, 20.7 ◆・
●●Window highlight, 208 ・●●●Window bar, 209, 804. 808 ・Each one is enriched, 210.803.807 ●●●●Scroll par 2
00 ●・●・Function button. Engraving of the drawings (no changes to the content) (Jumeiyama-gyo 2 Suni 9-Tests 7 Rows) Reh. ．Indication of the case Heibō Z year patent 2. Junichimoku name Mythology Kuchiro Nego 3. Amendment application No. 12'lO.5 Patent related to case No.

Claims (2)

[Claims] (1) at least one input element including a first plurality of nodes; at least one output element including a second plurality of nodes; and at least one set of hidden nodes; A neural network, wherein the first plurality of nodes are coupled to the hidden element, and the hidden element is coupled to the second plurality of nodes. (2) (a) a first input element including a first plurality of input nodes; (b) a second input element including a second plurality of input nodes; and (c) a first plurality of outputs. (d) a second output element including a second plurality of output nodes, the first input element being coupled to the second output element, and the first input element being coupled to the second output element; A neural network for color selection, wherein a second input element is coupled to the first input element.