JPH04260160A - Neurocomputer - Google Patents

Abstract

PURPOSE:To shorten the learning time in the case of executing a new learning processing by adding a pattern to be learned and adding the number of input signals, in the neurocomputer. CONSTITUTION:At the time of leaning of a neuronetwork based on a back propagation algorithm, a convergence coefficient is determined as a function of a weight coefficient for taking such a value as the nearer to '1' an absolute value of a value of the weight coefficient between each neuron is, the nearer to '0' it is, and the nearer to '0' the above value is, the nearer to '1' it is. In such a state, for instance, even if the number of digits of an input bit pattern consisting of a binary number increases from first 3 digits, while the number of neurons of each layer of the network is being increased automatically in accordance with magnitude of the sum total of the absolute value of the weight coefficient between the neurons of each layer, additional learning is executed by only a new leaning pattern, and also, new learning is executed so as not to exert the influence on the network part learned already to the utmost.

Description

[Detailed description of the invention]

0001

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to the architecture of a neurocomputer that executes operations such as pattern recognition using a network of neurons.

0002

[Prior Art] Neurocomputers are based on conventional AI (
This is a computer system that can easily realize a system that recognizes multiple sensor outputs in real time and performs the necessary control, which has been difficult to realize with expert systems based on techniques such as artificial intelligence. A neurocomputer is a system that realizes a network of neurons using electronic circuits. Through learning, a neurocomputer can autonomously change its network structure so as to obtain the optimal output for input. Due to its own generalization effect, it has excellent advantages such as being able to provide a valid output even for input patterns intermediate to the basic input pattern used during learning.

FIG. 7 is a diagram showing a mathematical model of a neuron that has been studied in physiology. According to this model, the relationship between the input signal yi(t) and the output signal x(t) is expressed by the following equation. can. However, "t" in parentheses indicates that each is a function of time.

0004

[Math 1]

Here, wi is a weighting coefficient, and the nonlinear function f usually has characteristics as shown in FIG.
A moid function is used. The following equation is often used as the sigmoid function.

[0006]

[Math 2]

[0007] x is the argument of the function f, and e is the base of the natural logarithm. Next, FIG. 9 shows a typical example of a neural network realized using the above-mentioned neurons as constituent elements.

The network in the figure consists of an input layer, each having a number of neurons (each neuron corresponds to FIG. 7), hidden layers I and II, and an output layer (in each vertical column in the figure). All neurons in a certain layer are connected to all neurons in the layer to the left thereof and to all neurons in the layer to the right thereof by weighted directed links. The direction of the directed link is such that the input is the neuron in the left layer, and the output is the neuron in the right layer. The weight of the link from neuron j to neuron i is wi
It is expressed as j. Also, neurons in the same layer do not connect to each other.

Each neuron of the input layer corresponds to each pixel of image data representing, for example, visual information and tactile information regarding an object as a two-dimensional input pattern, and each neuron has each pixel of image data expressed as a two-dimensional input pattern. The strength of the stimulus is given according to the pixel value.

[0010]The output value of each neuron in the rightmost output layer becomes the output of the neuronetwork. That is, for example, a neuron that senses softness outputs a value corresponding to the softness of the object that generated the input pattern,
Neurons that sense size output a value corresponding to the size of the object. Further, for example, one neuron that senses a shape is provided corresponding to each shape such as a sphere, a cube, etc., and the output value of the neuron corresponding to the shape of the object becomes a large value. One set of each of these outputs is obtained for one set of input patterns.

[0011] By inputting an input pattern obtained as an output from a sensor etc. to a neural network having the above configuration, the neural network performs recognition processing on the input pattern, and outputs the recognition result appropriately. Obtained as a pattern.

Expanding the above-mentioned equation 1 under such a neural network, we get the following. That is, the output of the i-th neuron in a certain layer (this is called the s-th layer) for the p-th input pattern is xip, and the j-th neuron in the layer one layer to the left (s−1 layer) Let yjp be the output of neurons in , and let n be the number of neurons in each layer.

[0013]

[Math 3]

[0014]

[Math 4]

[0015] In order to produce an ideal output when an input pattern is given to the neural network in Fig. 9 in which such a relational expression holds, the weighting coefficient wij, that is, the strength of the connection between neurons, must be adjusted through learning. Needs to be updated.

[0016] To actually do this, Rumelhar
t{NatureVol. 323-9 (1986)} proposes the following learning algorithm called backpropagation. First, it is assumed that there are t patterns used for learning, and in the n neurons of the sth layer, the output value (the most desirable output value) from the teacher for the pth pattern in the i-th neuron is d.
ip, and the actual output value is xip. Using the Euclidean norm, the error Ep for the pth pattern in the sth layer and the error E for all t patterns are evaluated by the following equations 5 and 6.

[0017]

[Math 5]

[0018]

[Math 6]

To minimize E with respect to wij, ∂Ep
/∂wij (partial differential of Ep with respect to wij) is calculated, and wij is updated using the following equation.

[0020]

[Math 7]

[0021] Here, wij(t-1) is the value of the weight wij before updating, and wij(t) is the value of the weight wij after updating.
is the value of here,

[0022]

[Math. 8]

[0023] From equations 7 and 8, wij(
It can be seen that t) is obtained by changing the weight wij(t-1) before updating by the step width η in the direction of the steepest descent of the error Ep.

In the above equation 8, Δp wij=−η∂E
p/∂wij is obtained by performing partial differentiation and chain rule on wij,

[0025]

[Math. 9]

However, the method for determining δip is slightly different depending on whether the sth layer is the output layer or not. First, when the sth layer is the output layer, δip in Equation 9 is

[0027]

[Math. 10]

On the other hand, if the sth layer is not the output layer, the subscript of the neuron in the s+1th layer is k, and

0
029]

[Math. 11]

[0030] However, fi' (zip) is
The first derivative of fi (zip) is shown. From this, we can sequentially apply equations 8 to 11 in the reverse direction of the network.
By calculating the change in j and iteratively adjusting the strength of the connection between neurons using Equation 7 for each layer, when a pattern p is finally given, the signal dip from the teacher
It is possible to construct a neuronetwork that outputs a value close to .

[0031] The above algorithm is programmed into a general-purpose or dedicated computer system, and learning patterns (sets of visual information, tactile information, etc.) are input to a neural network realized as software or dedicated hardware. However, by executing the above algorithm, the network can be trained. in this case,
Once programmed, there is no need to program again even if the pattern to be learned changes in the future. In other words, it is possible to construct a system that can automatically change its internal structure in response to external stimuli.

[0032]

[Problem to be solved by the invention] The above-mentioned Rumelha
The backpropagation algorithm of rt is a relatively simple algorithm, and many examples of its application to various industrial fields have been proposed.

However, if the pattern to be learned is clear in advance, the learning operation may be performed based on the above-mentioned formula 7, but if the pattern to be learned is added later, or if the input signal is If the number (number of neurons in the input layer) is added later, the previously memorized pattern (weighting coefficient wij, that is, the value of the strength of the connection between neurons) is often forgotten when relearning is performed. Therefore, it was necessary to combine both the previously learned pattern and the newly added pattern and start learning again from the beginning.

[0034] In this case, the learning algorithm based on Equation 7 described above generally has a slow convergence speed, and retraining a network once converged is extremely uneconomical in terms of computer usage efficiency. It had a point.

[0035] In the present invention, when performing new learning processing by adding patterns to be learned or adding the number of input signals,
The purpose is to make it possible to shorten the learning time.

[0036]

Means and Effects for Solving the Problems The present invention has an input layer consisting of at least one neuron that inputs an input pattern, and an output layer consisting of at least one neuron that outputs an output pattern corresponding to the input pattern. , is assumed to be a neurocomputer that executes calculations of a neuronetwork including at least one intermediate layer connected between the input layer and the output layer, each of which is composed of a plurality of neurons. In this case, the weighting coefficient between each neuron before updating is changed by a predetermined convergence coefficient in the direction of the steepest descent of the error between the output pattern and the teacher pattern for the learning input pattern, thereby updating the weighting coefficient between each neuron. The neural network is trained based on a backpropagation algorithm that includes a step of updating weighting coefficients.

The present invention is characterized in that the following processing is performed during learning of a neural network based on the above-mentioned backpropagation algorithm. First, the above-mentioned predetermined convergence coefficient is closer to 0 as the absolute value of the weighting coefficient between each neuron is closer to 1, and as it is closer to 0, it is 1.
is determined as a function of the weighting coefficient that takes a value close to .

That is, for example, the convergence coefficient η in Equation 8 or Equation 9 described above is wi indicating the connection strength of neurons.
It is determined as a function of wij such that the closer the absolute value of the component of the j matrix is to 1, the closer it is to 0, and the closer it is to 0, the closer it is to 1.

Specifically, such a convergence coefficient η is defined, for example, by the following equation.

[0040]

[Math. 12]

[0041] Here, η0 and b are constants. wi
A large absolute value of j indicates that the degree of neural connectivity is strong and that the nerves have become thicker due to learning. Note that the initial value of wij in Equation 7 above is generally 0, and in that case, in the initial state, nerve cells (neurons)
Even if they exist, it means that they are not connected to each other.

According to Equation 12 above, additional learning can be performed by forming a network between cell groups where wij is 0, so that previously learned network parts are not affected as much as possible. In other words, the pattern once learned will not be forgotten.

In order to achieve this, the present invention further increases the number of neurons in each layer of the neural network in accordance with the total sum of absolute values of weighting coefficients between neurons in each layer. This means that as the total mass of nerve cells (neurons) increases, the number of nerve cells (neurons) used also increases, and human brain cells increase the number of connections between neurons through learning. This corresponds well to the well-known mechanism.

FIG. 1 shows a neural network consisting of three layers numbered 1 to 3 at the beginning, from the three neurons in the input layer indicated by the bottom solid circle.
When an input bit pattern consisting of a binary number of digits is input, if it indicates an even value, one neuron in the output layer indicated by the solid circle at the top outputs 0,
The operation of the present invention will be described with respect to an example in which learning is performed to output 1 if an odd value is indicated.

As shown in the figure, as a result of learning, the nerves connecting neurons (corresponding to the value of the weighting coefficient wij) become thicker. In this case, based on the algorithm described above, the number of neurons in each layer is automatically increased, but in addition to such behavior, now 3
Assume that after learning for an input bit pattern consisting of a binary number of digits is completed, it becomes necessary to change the input bit pattern to a pattern consisting of a binary number of four digits.

In this case, first, a neuron indicated by a broken line circle numbered 4 is added to the neurons indicated by the solid line circles numbered 1 to 3 in FIG. According to the above algorithm, the cells of the newly added neuron are not connected to other neurons, so the broken line arrows connecting the neuron number 4 in Figure 1 and each of the other neurons numbered 1 to 3 Although the nerve weighting coefficients wij shown are 0, they gradually take on values from 0 to 1 by being updated based on Equation 12 described above. In other words, the nerves gradually become thicker.

By repeating the above-described operations and sequentially increasing the number of input digits, it is possible to construct a neuronetwork in which the input has an arbitrary number of digits M, as shown in FIG.

As described above, in the present invention, the number of neurons in each layer of the neural network is automatically increased according to the sum of the absolute values of the weighting coefficients between the neurons in each layer, and only a new learning pattern is created. Additional learning is performed, and in this case new learning is performed so as to have as little influence as possible on previously trained network parts. Therefore, when new learning processing is performed by adding patterns to be learned or adding the number of input signals,
This makes it possible to shorten the learning time.

[0049]

Embodiments Hereinafter, embodiments of the present invention will be described with reference to the drawings. Explanation of First Embodiment FIG. 2 is a perspective view showing a first embodiment in which the present invention described above is applied to a heterogeneous sensor integration system using a neuronetwork.

In this system, first, two visual sensors (CCD cameras) 211 placed on the floor,
It is assumed that the visual sensor signal processing device 201 that converts its output 214 into image data is not installed, and therefore, the visual information 204 is not input to the host computer 207.

[0051] Then, the object 213 (ball in the figure)
is recognized by the tactile sensor 212 attached to the gripper 210 of the hand section 209 at the tip of the robot hand 208, and based on the recognition result, the robot hand 20
8, and gripping of the object 213 by the gripper 210 attached to the hand section 209 is performed.

[0052] The tactile sensor 212 applies force on the substrate, for example.
It has a structure in which electrical signal converters are arranged in a matrix and the surface is covered with skin rubber. The output 215 of the tactile sensor 212 is converted into image data expressed in several gradations by the tactile sensor signal processing device 202, and sent to the host computer 207 as tactile information 205.

Further, the hand portion 209 to which the gripper 210 is attached is provided with an encoder (not shown) for detecting the angle at which the gripper 210 is opened (hereinafter referred to as the opening degree). The encoder output 216 is sent to the host computer 207 as angle information 206 via the encoder interface 203.
sent to.

Next, FIG. 3 is a block diagram of the host computer 207 of FIG. 2. Interface (I/F) 30
6 and 307 (assuming that 305 is not initially installed) are the tactile information 205 from the tactile sensor signal processing device 202 and the encoder interface 20 in FIG. 2, respectively.
input the angle information 206 from 3 and send them to bus 308
The data is transferred to the RAM 302 via the .

On the other hand, on the RAM 302, each of the above-mentioned tactile information 205 and angle information 206 is input.
A neuronetwork 303 is configured to detect the softness, size, shape, etc. when the robot hand 208 grips the object 213. Further, the auxiliary storage device 304 stores coefficients used when configuring the neuronetwork 303 on the RAM 302. These will be described later.

[0056] The CPU (central control unit) 301 is connected to the bus 3
Overall operation of each part connected to 08, especially RAM3
Controls the operation of the neuronetwork 303 configured on 02.

Note that the mechanism for controlling the movement of the robot hand 208 shown in FIG. Since there is no difference from the control technology, it will be omitted.

Next, the neuronetwork 303 configured on the RAM 302 will be explained. This neuronetwork initially consists of tactile information 205 and angle information 2.
It has a function of detecting the softness, size, shape, etc. when the robot hand 208 grasps the object 213 by inputting each of 06, and its basic configuration is the same as that described above with reference to FIG. .

Next, this neuro network 303
Regarding the learning procedure for the object 21 and a specific example of the recognition operation of the actual object 213 (Fig. 2) using this network, we will explain the learning procedure for the object 21
An example will be explained in which 3 is a tennis ball, which is an example of a spherical and soft object.

First, the robot hand 208 is made to hold a tennis ball in various postures, and the tactile information 2 in each posture is
05 and the angle information 206 are stored in the auxiliary storage device 304 in FIG. 3 or the like.

Next, the CPU 301 in FIG. 3 inputs each of the above patterns to the neural network 303 in FIG. In order to make each output of the neuron, the size-sensing neuron, and the shape-sensing neuron equal to the teacher signal, according to the back propagation algorithm shown in Equations 7 to 12 above, Equations 5 and 6 are applied. The weighting coefficient wij is adjusted so that the error determined by the equation becomes sufficiently small. When the object 213 is a tennis ball, the output value of the neuron that senses softness is, for example, 0.7, the output value of the neuron that senses size is, for example, 0.5, and the output value of the neuron that senses shape is spherical. For example, the output value of a neuron that senses is 1 and the others are 0. The final weighting coefficient wij value for each neuron in each layer in FIG. 9 obtained in this way is stored in the auxiliary storage device 304 in FIG.

In addition to this, learning is performed by giving a baseball or a golf ball as the object 213, and similar learning is also performed with objects 213 other than spherical shapes such as building blocks.

Next, when performing a recognition operation for the target object 213, the CPU 301 in FIG.
302, and the output signal is calculated using the algorithm shown in Equations 2 to 4 described above. According to this,
In addition to recognizing tennis balls, baseball balls, etc. that were used for learning, even if a softball that was not used for learning is used as the object 213, appropriate values such as neurons that feel softness, neurons that feel size, etc. The output value of the neuron that senses the spherical shape is 0.
5, 1, 1, etc. will be output.

Once the learning has been completed and the system has been established, as shown in FIG. Consider a case in which a sensor signal processing device 201 is additionally installed and an interface 305 for inputting visual information 204 from the visual sensor signal processing device 201 is installed in the host computer of FIG.

As described above, according to the present invention, even if the number of sensor inputs increases, the number of neurons can be automatically adjusted accordingly. It becomes possible to distinguish between new objects, such as spherical objects and cylindrical objects, which are difficult to distinguish using only a tactile sensor, by simply teaching the cylindrical object, without making the user forget the values (values) as much as possible. In the case of the embodiments shown in FIGS. 2 and 3, the learning process based on the present invention described above is performed by the CPU 3 shown in FIG.
01 executes. Explanation of Second Embodiment FIG. 4 shows a second embodiment in which the present invention described above is applied to a heterogeneous sensor integration system using a neuronetwork. This embodiment uses two visual sensors 401 and 4
02 and an ultrasonic sonar 403 with multiple ultrasonic sensors arranged
This is an example in which the present invention is implemented in a mobile robot equipped with.

[0066] This robot has caterpillars 404 on the left and right.
Because of this, it can be moved freely and is used, for example, to transport parts and products within a factory. Since this robot is equipped with a right visual sensor 401, a left visual sensor 402, and an ultrasonic sonar 403, which are different types of sensors, the present invention is applied to integrate these different types of sensors.

Right visual sensor 401 and left visual sensor 40
The output 405 of the right visual sensor and the output 406 of the left visual sensor detected in each of the above-mentioned cases are input to a detection circuit 408 of the right visual sensor and a detection circuit 409 of the left visual sensor, respectively. Furthermore, the ultrasonic sonar output 407 output from each ultrasonic sensor in the ultrasonic sonar 403 is input to an ultrasonic sonar detection circuit 410 .

The detection circuit 4 of these right visual sensors
08, each output of the left visual sensor detection circuit 409 and the ultrasonic sonar detection circuit 410 is input to the neuronetwork unit 411.

Here, the neuronetwork unit 411 is actually configured in the RAM inside the host computer as shown in FIG. 3, as in the case of the first embodiment, but here, for simplicity, Neuro network section 41
It is conceptually shown as 1. Each of the above-mentioned detection circuits 408
, 409 and 410 are input to the host computer and input to the neuronetwork in the RAM, the operations are almost the same as in the first embodiment. Therefore, in the following explanation, only how each of the above outputs is processed on the neuronetwork unit 411 will be explained. The same applies to the third and fourth embodiments described later.

As described above, the inputs of the neuronetwork unit 411 include the detection circuit 408 of the right visual sensor, the detection circuit 409 of the left visual sensor, and the detection circuit 4 of the ultrasonic sonar.
Ten outputs are provided. Here, the right visual sensor 40
The 1 and left visual sensors 402 output values of a plurality of pixels expressed in gradations as image information. For this reason, the input layer of the neuronetwork unit 411 (see FIG. 9) originally requires a number of neurons corresponding to the number of pixels, but in FIG.
Only one neuron is shown for each of 01 and left visual sensor 402. Similarly, ultrasonic sonar 4
03, the number of neurons corresponding to the number of ultrasonic sensors arranged in an array should be prepared in the input layer, but in FIG. 4, only one neuron is shown for the ultrasonic sonar 403. . In addition, along with the simplified representation of neurons in these input layers, hidden layers (see Figure 9, Figure 4
The number of neurons (consisting of only one layer) is also shown smaller than the actual number. These simplified expressions are
The same applies to third and fourth embodiments to be described later.

Learning of the robot shown in FIG. 4 is performed as follows. Now, it is assumed that only the outputs of the detection circuits 408 and 409 of the visual sensor are initially input to the neural network unit 411 to perform learning. In this case, first, the computer learns to output the size of the obstacle and the distance to the obstacle in the range of 0 and 1. That is, the value of the output 412 representing the size of the obstacle output from one neuron of the output layer of the neuronetwork unit 411 is set to 1 if the obstacle is large, and about 0.5 if the obstacle is medium. to perform learning. In addition, learning is performed so that the value of the output 413 representing the distance to the obstacle output from another neuron in the output layer of the neuronetwork unit 411 becomes a value closer to 1 as the obstacle is closer. Let it happen.

Next, as described above, the visual sensor 40
In order for a robot that has learned only by 1 and 402 to learn even when it is placed in an environment where smoke or fog is generated, sensor output 407 by ultrasonic sonar 403 is also required. However, in this case, as in the case of the first embodiment, even if the number of sensor inputs increases, the number of neurons can be automatically adjusted accordingly, so that the previous learning content ( Additional learning can be performed without forgetting the values of the weighting coefficients wij as much as possible. Explanation of Third Embodiment FIG. 5 shows a third embodiment in which the present invention described above is applied to a heterogeneous sensor integration system using a neuronetwork. This embodiment is an example in which the present invention is implemented in a freshness evaluation system for fruits and vegetables 507 moving on a belt conveyor 509.

The sensor used is the image sensor 50.
1. Phototransistor 502 and color sensor 503
It is. The respective outputs are input to the neuronetwork section 513 via an image sensor detection circuit 510, a phototransistor detection circuit 511, and a color sensor detection circuit 512. Neuro network section 51
3 is represented in a simplified manner as in the case of the second embodiment of FIG.

The image sensor 501 has one CCD element.
It is arranged in a dimension and is originally used to read barcode information. The phototransistor 502 detects scattered light 508 that is generated as a result of the laser light 505 emitted from the laser light source 504 being once reflected by the polygon mirror 506 and irradiated onto fruits and vegetables 507 . Since the polygon mirror 506 is rotated by the motor 516, the laser beam 505 is scanned over the surface of the fruits and vegetables 507. Therefore, the one-dimensional scattering phenomenon on the surface of the fruit or vegetable 507 can be captured by the phototransistor 502.

[0075] The color sensor 503 has red, blue, yellow, green,
It is a sensor that distinguishes eight colors: pink, light blue, white, and black. This sensor can also capture the one-dimensional color distribution on the surface of the fruit or vegetable 507.

In the system of FIG. 5, the value of an output 514 representing the freshness of fruits and vegetables outputted from one neuron of the output layer of the neuronetwork unit 513, and the value of output 515 representing the taste of fruits and vegetables outputted from other neurons similarly.
The freshness and taste of the fruits and vegetables 507 can be automatically determined based on the value of .

Here, the learning of the system shown in FIG. 5 is performed as follows. First, neuro network section 5
Learning is performed so that the value of an output 514 representing the freshness of fruits and vegetables outputted from one neuron of the 13 output layers becomes a value closer to 1 as the fruits and vegetables are fresher. In addition, an output 515 representing the taste of fruits and vegetables output from another neuron in the output layer of the neuronetwork unit 513
Learning is performed so that the value of is closer to 1 the more delicious the fruit or vegetable is. The basic learning method is the same as in the first embodiment.

[0078] If a new sensor of a different type is added to further improve the discrimination accuracy of such a system, additional learning will be required. In this case, as in the case of the first or second embodiment, even if the number of sensor inputs increases, the number of neurons can be automatically adjusted accordingly. Weighting factor wij
Additional learning can be done without making students forget the values of each value). Fourth Embodiment FIG. 6 shows a fourth embodiment in which the present invention described above is applied to a heterogeneous sensor integration system using a neuronetwork. This embodiment is an example in which the present invention is implemented in a bill validator.

The sensor used is the color sensor 601.
, magnetic heads 602 and 603, and an image sensor 604. Each output is sent to the color sensor detection circuit 60.
7. The signal is input to the neuronetwork unit 611 via the magnetic head detection circuits 608 and 609 and the image sensor detection circuit 610. The neuronetwork unit 611 is represented in a simplified manner as in the case of the second embodiment shown in FIG.

In the banknote validator shown in FIG. 6, the recognition output 612 for 1,000 yen, the recognition output 613 for 5,000 yen, or the recognition output 614 for 10,000 yen output from the three neurons of the output layer of the neuronetwork section 611, Can identify banknotes.

Here, the learning of the banknote validator shown in FIG. 6 is performed as follows. First, when a thousand yen bill 605 is moved in the bill movement direction indicated by 606 in FIG. Have them learn to become 1. Similarly, the banknote 60 of the 5,000 yen bill
In the case of 5, learning is performed so that the recognition output 613 of 5,000 yen output from another neuron in the output layer of the neuronetwork unit 611 becomes 1. Furthermore, in the case of a 10,000 yen bill 605, the neuronetwork unit 61
Learning is performed so that the recognition output 614 of 10,000 yen output from yet another neuron in the output layer 1 becomes 1. Then, the above learning is performed using various banknotes such as new banknotes, old banknotes, damaged banknotes, and faded banknotes. The basic learning method is the same as in the first embodiment.

[0082] Many conventional banknote validators use only magnetic heads, but the use of color sensors and image sensors makes it possible to eliminate counterfeit banknotes. There was also a machine. However, depending on the output of these different types of sensors,
The algorithms created to detect 10,000 yen and counterfeit bills have become complex, and the number of steps in the computer program that implements the algorithms has also become enormous. Furthermore, every time a new note was issued, it became necessary to review all the algorithms that had been created.

On the other hand, according to the fourth embodiment, the neuronetwork unit 611 automatically acquires the ability to discriminate between various types of banknotes through learning, so that a new algorithm is created when new banknotes are issued. It becomes possible to significantly reduce the time and effort required for creation. In addition, when relearning the neural network unit 611 when new banknotes are issued, the learning time for the old banknotes can be shortened because the learning effect of the old banknotes is not impaired as much as possible.

[0084] The first to fourth embodiments described above are examples in which a neuronetwork algorithm is incorporated into a general-purpose computer. However, considering the acceleration of the development of neurochips in recent years, By using a neurochip, embodiments in which a neuronetwork is configured on hardware can be easily implemented, and this is particularly effective for robot control and system control that require real-time performance.

[0085]

Effects of the Invention According to the present invention, it is possible to provide a reasonable output even for inputs intermediate to the basic patterns used for learning, which conventional neurocomputers have, and to provide sensor inputs or cell inputs. While maintaining the complementary effect due to neuron damage and being suitable for parallel processing, it is possible to shorten the learning time when a new pattern to memorize occurs.

Furthermore, when the present invention is applied to a sensor fusion system consisting of a plurality of different types of sensors, even when the number of sensors used increases, it is possible to respond more flexibly than with conventional neurocomputers. Become.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of the principle of the present invention.

FIG. 2 is a perspective view showing a first embodiment of a heterogeneous sensor integration system using a neuronetwork.

FIG. 3 is a configuration diagram of a host computer according to the first embodiment.

FIG. 4 is a perspective view showing a second embodiment of a heterogeneous sensor integration system using a neuronetwork.

FIG. 5 is a perspective view showing a third embodiment of a heterogeneous sensor integration system using a neuronetwork.

FIG. 6 is a perspective view showing a fourth example of a heterogeneous sensor integration system using a neuronetwork.

FIG. 7 is a diagram showing a mathematical model of a neuron.

FIG. 8 is a diagram showing a sigmoid function.

FIG. 9 is a diagram showing an example of a neuronetwork.

[Explanation of sign]

Claims (1)

[Claims] [Claim 1] At least one for inputting an input pattern
an input layer consisting of one neuron, an output layer consisting of at least one neuron that outputs an output pattern corresponding to the input pattern, and an output layer connected between the input layer and the output layer, each consisting of a plurality of neurons. A neurocomputer that executes calculations of a neural network consisting of at least one intermediate layer, wherein the weighting coefficients between each neuron before updating are in the direction of the steepest descent of the error between the output pattern and the teacher pattern with respect to the learning input pattern. In the neurocomputer, the neural network is trained based on a backpropagation algorithm including a weighting coefficient updating step in which the weighting coefficients between the neurons are updated by changing the weighting coefficient by a predetermined convergence coefficient to the neuron. When the neural network is trained based on a backpropagation algorithm, the predetermined convergence coefficient is closer to 0 as the absolute value of the weighting coefficient between each neuron is closer to 1, and the closer it is to 0, the closer the predetermined convergence coefficient is to 1. The number of neurons in each layer of the neural network is increased according to the sum of the absolute values of the weighting factors among the neurons in each layer. neurocomputer.