Water flow neural network theory reproduction device

Abstract

To provide a device for reproducing a theory such as a structure and movement of a neural network by utilizing water that is one of the most familiar things.SOLUTION: A water flow neural network theory reproduction device includes: a plurality of input-side water tanks 110 for storing water to be used as input data; a first water channel unit α installed below each of the input-side water tanks and receiving water discharged from the input-side water tank 110; a second water channel unit β that is installed below the first water channel unit α and receives water discharged from the first water channel unit α; a third water channel unit 180 that is installed below the second water passage unit β and receives water discharged from the second water passage unit β; and an output-side water tank 190 that is installed below the third water passage unit 180 and receives water discharged from the third water passage unit 180.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] The present invention relates to a device that reproduces the theory of neural networks at a visual level by utilizing water flow. [Background technology]

[0002] A neural network is a mathematical model that mimics the structure of the neural circuits in the human brain, and has been widely used in fields such as voice recognition, translation, and image recognition, especially in recent years, with the rapid development of AI technology.

[0003] This neural network is conceptually shown in Figure 4. Reference numeral 10 denotes the input value, reference numeral 20 denotes the input layer (input unit), reference numeral 50 denotes the intermediate layer (intermediate unit), and reference numeral 80 denotes the output layer (output unit). Reference numerals 40 and 70 denote edges connecting the layers (units), and reference numeral 90 denotes the output value. Here, a neural network with a three-layer structure is shown, but it may of course have more layers, and each layer may have a greater number of units (shown as circles composing each layer).

[0004] As shown in Fig. 4, in a fully connected neural network, the units on the input side connect edges to all of the units on the output side. Specifically, the three units that make up the input layer 20 connect two edges 40 to each of the two units that make up the next-stage hidden layer 50. Similarly, the two units that make up the hidden layer 50 connect three edges 70 to each of the three units that make up the next-stage output layer 80. If there were ten units that make up the hidden layer, ten edges would connect to each of the units that make up the input layer. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] IBM website: "What is a neural network?" https: / / www.ibm.com / jp-ja / topics / neural-networks Summary of the Invention [Problem to be solved by the invention]

[0006] However, the neural networks we usually see are merely presented as desk theories in the form of text and diagrams as above, and we have not been able to visually grasp their movements as the movements of actual "objects." For example, if we could visually grasp the flow of input information being processed and propagated, and the results being output, we could imagine using them as educational toys or exhibited as art pieces, which could be an opportunity to increase the number of people interested in artificial intelligence.

[0007] The present invention was made to solve these problems, and its objective is to provide a device that reproduces theories such as the structure and behavior of neural networks using water, one of the most familiar materials. [Means for solving the problem]

[0008] In order to solve the above problems, the present invention provides a water supply system including a plurality of input water tanks for storing water to be used as input data, a first water channel section disposed below each of the input water tanks and receiving water discharged from the input water tanks, a second water channel section disposed below the first water channel section and receiving water discharged from the first water channel section, a third water channel section disposed below the second water channel section and receiving water discharged from the second water channel section, and an output water tank disposed below the third water channel section and receiving water discharged from the third water channel section. The first water channel section receives water discharged from the input tanks, The second water channel section comprises a first receiving section that receives water released from the first water channel section, a first branch section that is continuous from the first receiving section and branches off into a predetermined number, and a predetermined number of first branch water channel sections that branch off from the first branch section, and further comprises a first separation structure at the tip of each first branch water channel section for reproducing weighting, and the second water channel section comprises a second receiving section that receives water released from the first water channel section, a second branch section that is continuous from the second receiving section and branches off into a predetermined number, and a predetermined number of second branch water channel sections that branch off from the second branch section, and further comprises a second separation structure at the tip of each second branch water channel section for reproducing weighting.

[0009] This configuration makes it possible to visualize the neural network and understand it as the flow of water. In other words, you can visually grasp the process by which water flows through a gutter, branching, gathering, and being discharged, and goes through other physical operations before outputting a result.

[0010] Moreover, the second water channel portion acting as an intermediate layer may be provided in a plurality of stages as required.

[0011] With this configuration, more dynamic movements can be reproduced.

[0012] Moreover, each of the water channels is formed in a gutter shape that is "V-shaped" in cross section.

[0013] With this configuration, water always flows through the center of the waterway, making it easy to branch the water as intended regardless of the amount of water flowing through the waterway. Also, the flowing water can be easily visually grasped from the outside.

[0014] Moreover, the weighting in the first isolation structure and the second isolation structure is realized by the installation angle of the partition wall portion that protrudes upward from the apex of the bottom surface of the "V-shape" trough.

[0015] By configuring in this way, it becomes possible to reproduce the weighting more accurately. Effect of the Invention

[0016] By applying this invention, it is possible to provide a device that reproduces theories such as the structure and behavior of neural networks using water, one of the most familiar materials. [Brief description of the drawings]

[0017] [Figure 1] 1 is a schematic overall configuration diagram of a water flow type neural network theory reproduction device according to an embodiment of the present invention; [Diagram 2] 1A and 1B are diagrams illustrating a separation structure provided at the tip of each water channel section, in which (a) is a perspective view and (b) is a front view. [Diagram 3] FIG. 1 is a diagram showing a first embodiment of a water flow type neural network theory reproduction device. [Figure 4] FIG. 1 is a diagram illustrating the concept of a neural network. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0018] Below, we will explain the water flow type neural network theory reproduction device 100, which is one example of an embodiment of the present invention, with reference to the attached drawings. Please note that in order to make the drawings easier to understand, some parts are exaggerated in size and dimensions, and some parts do not necessarily match the actual product. Furthermore, each drawing should be viewed in the direction of the symbols, and the directions are expressed as up, down, left, right, front, and back based on the relevant directions.

[0019] <Configuration of a water flow type neural network theory reproduction device> 1, the water flow type neural network theory reproduction device 100 shown as an example of the embodiment of the present invention comprises three input side water tanks 110 for storing water used as input data, a first water channel section α installed below each input side water tank 110 and receiving water discharged from the input side water tank 110, a second water channel section β installed below the first water channel section α and receiving water discharged from the first water channel section α, a third water channel section 180 installed below the second water channel section β and receiving water discharged from the second water channel section β, and three output side water tanks 190 installed below the third water channel section 180 and receiving water discharged from the third water channel section 180. Reference numeral 112 denotes a hose-like structure for guiding water from the input side water tank 110 to the corresponding first water channel section α.

[0020] The water flow type neural network theory reproduction device 100 described here is configured as a device based on the conceptual diagram shown in FIG.

[0021] The first water channel section α includes a first receiving section 120 (corresponding to the input layer unit 20 in FIG. 4) that receives water discharged from the input side tank 110, a first branching section 130 that is continuous with the first receiving section 120 and branches into a predetermined number (two in this embodiment), and a predetermined number (two in this embodiment) of first branching water channel sections 140 (corresponding to the edge 40 in FIG. 4) that branch off from the first branching section 130. That is, in this embodiment, one first water channel section α includes two first branching water channel sections 140. Furthermore, a first separation structure 146 for reproducing weighting is provided at the tip of each first branching water channel section 140. In this embodiment, a total of three first water channel sections α are provided corresponding to the input side tank 110.

[0022] The second water channel section β includes a second receiving section 150 (corresponding to the intermediate layer unit 50 in FIG. 4) that receives water discharged from the first water channel section α, a second branching section 160 that is continuous with the second receiving section 150 and branches into a predetermined number (three in this embodiment), and a predetermined number (three in this embodiment) of second branching water channel sections 170 (corresponding to the edge 70 in FIG. 4) that branch off from the second branching section 160. That is, in this embodiment, one second water channel section β includes three second branching water channel sections 170. Furthermore, the tip of each second branching water channel section 170 includes a second separation structure 176 for reproducing weighting. In this embodiment, a total of two second water channel sections β are included.

[0023] The second receiving portion 150 of the second water channel portion β is configured to be able to receive water from each of the first branch water channel portions 140 of the three first water channel portions α.

[0024] The third water channel section 180 is configured to be able to receive water from each of the second branch water channel sections 170 of the two second water channel sections β. All of the water that enters the third water channel section 180 flows directly into the corresponding output side water tank 190 installed below.

[0025] In this embodiment, both the first water channel section α and the second water channel section β are configured as gutters having a “V-shape” in cross section.

[0026] Next, the first separation structure 146 provided at the tip of each first branch water channel section 140 will be described with reference to FIG. 2. The first branch water channel section 140 is composed of a gutter having a "V-shape" in cross section, and a partition wall section 148 protruding upward from a bottom apex 147 of this "V-shape" makes it possible to separate the flowing water. One part of the separated water flows to the next stage (second water channel section β), while the other part is discharged to the outside and does not flow to the next stage. The angle of the partition wall 148 is set to an appropriate angle by performing a computer simulation in advance. Of course, a variably adjustable structure may be adopted.

[0027] The second separation structure 176 provided at the tip of each second branch water channel section 170 has a structure similar to that of the first separation structure 146 .

[0028] <Actions and Functions of the Water Flow Type Neural Network Theory Reproduction Device> The configuration shown in FIG. 1 is a simplified explanation of the relationship between the structures of the various parts so as not to complicate the diagram, so the operation of the water flow type neural network theory reproduction device 100 will be explained with reference to FIG. 3, which shows the operation in a more understandable manner.

[0029] 3 is composed of a total of 100 input side water tanks 110 arranged in a 10x10 grid. Furthermore, the input side water tanks 110 are filled with water only in the area corresponding to the number "5." This means that the number "5" is recognized as an image of a grid.

[0030] When the water flow type neural network theory reproducing device 100 is operated, the water flows sequentially through the first water channel section α and the second water channel section β, and finally flows into the ten output side water tanks 190 arranged at the bottom. In the process of flowing through the first water channel section α and the second water channel section β, the water is branched, some of it is discharged, and so on, before flowing into the ten output side water tanks 190. At this time, the ten water tanks are each assigned a number from "0" to "9," and the number assigned to the output side water tank 190 with the largest amount of water stored is the answer calculated by the water flow type neural network theory reproducing device 100.

[0031] In order to obtain an accurate answer, it is necessary to carry out a simulation using a computer in advance and set the angle of the partition portion of each separation structure.

[0032] As described above, the present invention includes a plurality of input water tanks 110 for storing water to be used as input data, a first water channel section α disposed below each of the input water tanks and receiving water discharged from the input water tanks 110, a second water channel section β disposed below the first water channel section α and receiving water discharged from the first water channel section α, a third water channel section 180 disposed below the second water channel section β and receiving water discharged from the second water channel section β, and an output water tank 190 disposed below the third water channel section 180 and receiving water discharged from the third water channel section 180. The first water channel section α is a first receiving section that receives water discharged from the input tanks 110. The second water channel section β comprises a first receiving section 120, a first branch section 130 that is continuous from the first receiving section 120 and branches off into a predetermined number of sections, and a predetermined number of first branch water channel sections 140 that branch off from the first branch section 130, and further comprises a first separation structure 146 at the tip of each first branch water channel section 140 for reproducing weighting, and the second water channel section β comprises a second receiving section 150 that receives water released from the first water channel section α, a second branch section 160 that is continuous from the second receiving section 150 and branches off into a predetermined number of sections, and a predetermined number of second branch water channel sections 170 that branch off from the second branch section 160, and further comprises a second separation structure 176 at the tip of each second branch water channel section 170 for reproducing weighting.

[0033] This configuration allows the neural network to be visualized and understood as the flow of water. In short, the process of water flowing through a gutter, branching, gathering, and being discharged along the way, and the result being output, can be visually grasped. In particular, sensors, electronic devices that are electrically driven, and electric devices are not required. Another advantage is that the main part of the neural network from input to output has no moving parts, making it less likely to break down. Furthermore, it does not use electricity, making it ecological and economical. In addition, since it is a physical device, it has a higher sense of realism in terms of vision and other senses than a neural network on a computer program, making it easier to understand the theory intuitively. In addition, since water flows over time, it is easy to visually see and understand the behavior of a neural network.

[0034] Furthermore, the second water channel section β, which plays the role of an intermediate layer, may be provided in a plurality of stages as necessary, and such a configuration makes it possible to reproduce a more dynamic movement.

[0035] In addition, each of the water channel sections αβ is formed in a V-shape in cross section. As a result, water always flows through the center of the water channel section, making it easy to branch the water as intended regardless of the amount of water flowing through the water channel section. In addition, the flowing water can be easily visually grasped from the outside.

[0036] Furthermore, since the weighting in the first isolation structure 146 and the second isolation structure 176 is realized by the installation angle of the partition portion 148 protruding upward from the bottom apex 147 of the "V-shape," it is possible to reproduce the weighting more accurately.

[0037] Other configuration examples In the above, the water passage is shown as a "gutter", but it may be in the shape of a pipe. [Explanation of symbols]

[0038] 100···Water flow neural network theory reproduction device 110 Input water tank 112 Pipe 120...1st receiving part 130...First branch 140 First branch waterway section 146...1st separation structure 148...bulkhead 150...Second receiving part 160...Second branch 170...Second branch waterway section 176...Second separation structure 180...3rd waterway section 190···Output water tank

Claims

1. A plurality of input water tanks for storing water to be used as input data; a first water channel portion disposed below each of the input water tanks and configured to receive water discharged from the input water tanks; a second water channel portion disposed below the first water channel portion and configured to receive water discharged from the first water channel portion; a third water channel section disposed below the second water channel section and configured to receive water discharged from the second water channel section; an output side water tank disposed below the third water channel portion and configured to receive water discharged from the third water channel portion; the first water channel section includes a first receiving section for receiving water discharged from the input tank, a first branching section continuing from the first receiving section and branching into a predetermined number of sections, and a predetermined number of first branching water channel sections branching off from the first branching section, and further includes a first separation structure at a tip of each first branching water channel section for reproducing weighting; The second water channel section includes a second receiving section that receives water discharged from the first water channel section, a second branch section that is continuous with the second receiving section and branches into a predetermined number of second branch water channel sections, the predetermined number of which branch off from the second branch section, and further includes a second separation structure at the tip of each second branch water channel section for reproducing weighting. A water flow type neural network theory reproduction device characterized by the above.

2. In claim 1, The second water channel portion acting as an intermediate layer may be provided in a plurality of stages as required. A water flow type neural network theory reproduction device characterized by the above.

3. In claim 1 or 2, Each of the water channels is formed in a V-shaped gutter in cross section. A water flow type neural network theory reproduction device characterized by the above.

4. In claim 3, The weighting in the first isolation structure and the second isolation structure is reproduced by the installation angle of the partition wall portion protruding upward from the top of the bottom surface of the "V-shape" trough. A water flow type neural network theory reproduction device characterized by the above.

Images