JP3367295B2 - Multi-valued neural network learning method - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a large scale logic circuit,
Learning of a multi-valued neural network that can obtain a desired output signal at high speed and stably by detecting a local minimum capture state in a neural network applicable to fields such as pattern recognition, associative memory, data conversion and image processing. It is related to the method.
The present invention particularly relates to an improvement of Japanese Patent Application No. 7-77168 (filed on Mar. 9, 1995).

[0002]

2. Description of the Related Art As one of neural networks,
Although there is a multilayer neural network, a back propagation algorithm is widely used as a learning method using a teacher signal of this neural network. In the learning process using this algorithm, after the weighting coefficient is initialized, the teacher signal T (teacher signal element, T ₁ , T ₂ , ..., T _M ) prepared in advance is set.
Then, the output unit signal from the output layer for the learning input signal input to the input layer is subtracted through a subtractor to obtain the error signal, and the error signal power is minimized based on the output unit signal and error signal of each layer. As described above, the weighting coefficient between layers is updated and learning is performed. The learning consisting of the weighting coefficient adaptive control is executed for all learning input signals and repeated until convergence.

In the learning process, when the error power is minimized, the error power completely converges, and the multi-valued output unit signal obtained by multi-valued the output unit signal with respect to the learning input signal by the multi-valued discrimination threshold becomes the multi-valued teacher signal. Match. However, once the error power drops to the local minimum (local minimum), if it is very stable, the learning does not proceed thereafter and does not become the minimum. There are problems that the number of times increases remarkably and that the convergence characteristic depends on the initial value of the weighting coefficient.

On the other hand, in order to improve the learning speed for the binary teacher signals T of 0 and 1, "Parallel Distributed Pr" is used.
In ocessing "DE Rumelhart, MIT Press., a new teacher signal T'of 0.1 and 0.9 is set for each element in the multilayer neural network using the conventional learning shown in FIG. Learning is started after initialization is performed by the control signal from the operation mode controller 9. From the teacher signal T ′ from the terminal 3, the terminal 2 is started.
The output unit signal of the multilayer neural network 1 with respect to the learning input signal is obtained through the subtractor 4 to obtain a subtraction error signal, which is input to the weighting factor controller 5, and the weighting factor is updated by the back propagation algorithm, The process of setting again in the multilayer neural network 1 is repeatedly learned. A binary output unit signal is obtained from the output unit signal through the binary threshold circuit 6 that performs binarization using the binary discrimination threshold, and a binary threshold circuit 7 that performs binarization using the binary discrimination threshold is provided. A binary teacher signal T is obtained from the teacher signal T'through the detection signal, and the coincidence detector 8 detects a state in which they are completely coincident with each other. Learning ends.

By setting the teacher signals to 0.1 and 0.9 in this way, it is stated that the number of learnings for convergence is reduced as compared with the case of binary teacher signals of 0 and 1. ing. This is because, in the sigmoid function, when the input approaches 0 or 1, the gradient becomes very small and the update rate of the weighting coefficient becomes small. Therefore, the learning speed is improved by setting the teacher signal to 0.1 and 0.9 and increasing the gradient. However, learning will not progress if it falls into a stable local minimum as well.

[0006] These conventional learning methods have the drawback that the state of the local minimum can be easily dropped as described above, and that the state of complete convergence cannot be easily reached because it is difficult to get out. In particular, in a three-layer or multi-layer neural network with a large number of input units, a design method for surely converging the multivalued teacher signal has not been clarified, and the initial value of the weighting factor is changed due to the initial value dependence. And increasing the number of hidden units (intermediate units) by trial and error. Also,
Until now, there has been no easy way to detect the capture state in the stable local minimum.

[0007]

DISCLOSURE OF THE INVENTION Problems to be Solved by the Invention As described above,
In the conventional learning process of the multi-layer neural network using the multi-valued teaching signal, when the weighting coefficient is updated so that the multi-layer neural network sends out the desired multi-valued output unit signal corresponding to the learning input signal, When the number of learning, that is, the number of learning iterations until the converged state in which the value output unit signal is transmitted becomes very large,
However, it has a drawback that the weighting coefficient has an initial value dependency such that it does not converge even if it is learned, that is, it falls into a very stable local minimum state and a desired multi-value output unit signal is not transmitted. In particular, in the case of a teacher signal having a large number of units in the input layer and a small number of output units and a distributed expression format, convergence due to learning becomes very difficult, and a desired multi-valued output unit signal can be obtained. A method for freely designing a multi-layer neural network for transmission has not been established. In addition, there is a method to significantly increase the number of intermediate units to make it easier to converge, but as the generalization capacity deteriorates, the amount of calculation of each increases naturally, and a very large hardware capacity or calculation capacity is required. It

An object of the present invention is to solve the above-mentioned problems, to converge with an extremely small number of learning times as compared with the conventional learning method of a multivalued neural network, and to increase from 10 times to 100 times.
A multi-valued neural network with a local minimum capture state detection function that can stably complete learning at double speed and can easily obtain a desired multi-valued output unit signal for the learning input signal. To provide a new learning method.

[0009]

In order to solve the above problems, in a multi-valued neural network for learning using a multi-valued teacher signal, the output unit signal of the correct output unit is subtracted from the multi-valued teacher signal. If the absolute value of the obtained error signal is less than or equal to the first threshold, a weighting coefficient update error signal that has a polarity opposite to that of the error signal and whose amplitude decreases as the distance from the multilevel teacher signal increases If the value is larger than the first threshold, the weighting coefficient update error signal having the same amplitude as the error signal or less and the same weighting coefficient having the amplitude less than the error signal is output in the incorrect output unit. An update error signal is generated, the weight coefficient is updated by using the weight coefficient update error signal, and the weight coefficient is updated by using at least the minimum absolute value of the error signal in the incorrect output unit. Detecting a Ruminimamu captivity, constituting the neural network learning method which is characterized by updating the amplitude or further adjusted weighting coefficient threshold of the first heavy viewed coefficient updating error signal.

[0010]

In the learning method of the multi-valued neural network of the present invention, an output unit giving an incorrect multi-valued output unit signal which is largely deviated from the multi-valued teacher signal is subtracted from the multi-valued teacher signal. Using the weighting factor update error signal with amplitude less than the obtained error signal, while
In the output unit in which the error signal is very close to the multi-valued teacher signal and gives the correct multi-valued output unit signal below the first threshold, the amplitude changes as the distance from the teacher signal having the opposite polarity to the error signal increases. In the output unit which uses the smaller weight coefficient update error signal and further provides the correct multi-valued output unit signal whose error signal is larger than the first threshold and is relatively close to the multi-valued teacher signal, the same polarity as the error signal is used. The weighting coefficient update error signal having a slightly smaller amplitude is used to modify the weighting coefficient.
At this time, among the incorrect output units, the output unit minimum error which is the minimum value of the absolute value of the error signal is detected, and the stable local minimum capture state can be easily detected from the comparison result with the given threshold, By controlling and learning the amplitude of the weighting coefficient update error signal or further the first threshold according to the detection result, it is possible to easily get out of a very stable local minimum and to quickly approach the optimum state with a small number of learnings. By using the weight coefficient update error signal whose polarity and amplitude are controlled as described above, it is possible to significantly reduce the number of learning without the dependency of the weight coefficient on the initial value, and the number of learning or the number of intermediate units (hidden units) or The number of intermediate layers can be significantly reduced.

As described above, since the learning method of the present invention can easily detect a very stable local minimum capture state, it is easy to get out of the local minimum of a multi-valued neural network, and the initial value of the weighting coefficient is higher than that of the conventional method. Since a desired multi-valued output unit signal can be easily obtained without any dependence and very quickly and surely, a large-scale multi-valued neural network can be freely designed. In addition, we have realized a logic system with a learning function that requires re-learning in real time, and a multi-valued logic system using an input signal with an extremely large number of input signal elements and a multilayer neural network with a large number of units. it can. Furthermore, it is possible to easily design and realize pattern recognition, image processing, etc., which were difficult to obtain a desired output signal because the conventional method could not converge stably and at high speed, and also data conversion etc. Becomes

[0012]

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of a multi-valued neural network using the learning method of the present invention will be described below, and its configuration and operation will be described in detail when a binary teacher signal is used.

FIG. 2 shows an embodiment of the learning process of a multi-valued neural network using the learning method of the present invention. The input signal from the terminal 2 is input to the input unit of the input layer, and the output unit signal is sent from the output unit of the output layer. Multilayer neural network 1, binary teacher signal, output unit signal, and output A weighting factor update error signal generator 10 which receives a signal as an input and outputs a weighting factor update error signal, a binary threshold circuit 6 which binarizes the output unit signal using a binary discrimination threshold to obtain a binary output unit signal, The weight coefficient controller 5 for updating and setting the weight coefficient of the multilayer neural network 1 by using the input weight coefficient update error signal, receives the binary teacher signal and the binary output unit signal, and inputs the binary teacher signal and the binary teacher signal.
Correct value in the output unit by comparing with the value output unit signal,
An output correct / wrong judgment device 11 for detecting an incorrect answer and sending a correct / wrong judgment signal, a correct / wrong judgment signal from the output correct / wrong judgment device 11, a binary teacher signal and an output unit signal are inputted, and an output unit signal is outputted at an incorrect output unit. And the binary teacher signal, the minimum absolute value of the difference, that is, the incorrect output unit minimum error detector 13 for detecting the incorrect output unit minimum error, and the output unit signal and the binary threshold circuit 6 in the correct output unit. Using the minimum absolute value of the difference from the binary discrimination threshold of, that is, the correct output unit minimum margin detector 14 for detecting the correct output unit minimum margin, and the incorrect output unit minimum error and the correct output unit minimum margin. Local minimum capture state detector 15, multi-layer neural network 1 and weighting factor controller 5 for detecting the local minimum capture state Each of the output correctness determination unit 11, the incorrect output unit minimum error detector 13, the correct output unit minimum margin detector 14, and the local minimum capture state detector 15 is initialized, and the operation mode controller controls the start and end of learning. 12 and.

Next, the operation in the learning process will be described. The multilayer neural network 1 performs learning using a learning input signal from the terminal 2 and a binary teacher signal from the terminal 3. In the weighting coefficient update error signal generator 10, for the output unit in which the incorrect binary output unit signal is detected by the correctness determination signal from the output correctness determination unit 11,
The error signal obtained by subtracting the output unit signal from the binary teacher signal is output as the weighting coefficient update error signal with the amplitude reduced by about Dm3 as shown in equation (3), while the correct output unit signal is given. In the output unit, when the absolute value of the error signal is less than or equal to the given first threshold value dm, the weighting coefficient whose amplitude becomes smaller than the constant Dm1 (however ≧ 0) as the distance from the teacher signal is increased with the opposite polarity. An updated error signal is generated and output, and if the absolute value of the error signal is larger than the threshold, D with the same polarity
A weight coefficient update error signal having an amplitude as small as m2 is generated and output.

Here, a method for generating one of the weighting coefficient update error signals in the case of using a binary or multi-valued teacher signal will be shown below by using an equation.

[0016]   If the multi-valued output unit signal of the output unit is correct,   When | Tm−Ym | ≦ dm, Em = Tm-Ym-Dm1 · sgn (Tm-Ym) (1)                               However, Dm1> dm   When | Tm-Ym |> dm, Em = Tm-Ym-Dm2 · sgn (Tm-Ym) (2)                               However, dm ≧ Dm2   If the multilevel output unit signal of the output unit is incorrect, Em = Tm-Ym-Dm3 · sgn (Tm-Ym) (3)   here, sgn (x) = 1: x ≧ 0             = -1, x <0, (4)

M is an output unit number (1≤m≤M), E
m is a weight coefficient update error signal of the output unit m, Tm is a multivalued teacher signal (0 or 1 in the case of binary) of the output unit m, Ym is an output unit signal of the output unit m, dm
Is a positive constant for the output unit m, Dm1 is 0 or a positive constant for the output unit m, Dm2 is a positive constant for the output unit m, and Dm3 is a positive constant for the output unit m.

Where dm is the first threshold,
If the first threshold is defined as the convergence region, it can be similarly applied to a neural network that handles a continuous amount of teacher signals. The weighting factor controller 5 uses the weighting factor update error signal to correct the weighting factor by the back propagation algorithm, and repeats learning so that the power of the error signal is minimized.

Further, in the course of learning, in the local minimum capture state detector 15, the incorrect output unit minimum error output from the incorrect output unit minimum error detector 13 becomes larger than the given second threshold. Then, it is regarded as a very stable local minimum, and the detection output is sent to the output terminal 16 as the local minimum capture state. At this time, if Dm1 and Dm3 in the weighting factor update error signal generator 10 or Dm2 and dm are further increased according to the detection output, it is possible to easily escape from a very stable local minimum.

Further, the local minimum capture state detector 1
5, the incorrect output unit minimum error which is the output from the incorrect output unit minimum error detector 13 or the correct output unit minimum allowance which is the output from the correct output unit minimum margin detector 14 is given by the second If the threshold is exceeded and then falls below the threshold, and the threshold is rapidly exceeded again, it is considered that a different stable local minimum capture state has been entered.

Further, after the stable local minimum capture state is reached, the incorrect output unit minimum error output from the incorrect output unit minimum error detector 13 exceeds the third threshold together with the learning count. It gradually becomes smaller than this third threshold, and when it approaches the multilevel discrimination threshold for multileveling, it is considered that the local minimum capture state is released and it is considered to be in the convergence process, and dm, Dm1, Dm2 and Dm3. The value of may be reduced.

After the error is eliminated and the multivalued teacher signal and all the multivalued output unit signals have reached the converged state in the output correctness / incorrectness determiner 11, the error is generated in the weighting coefficient update error signal generator 10. The signal is used as it is as the weighting factor update error signal, and the weighting factor controller 5
The weighting factor may be updated at. In addition, the output signal of the correct output unit minimum margin detector 14 is given in advance by the fourth signal.
When the threshold of is exceeded, the learning end signal is sent from the operation mode controller 12, and the multilayer neural network 1
The learning of may be ended. This can avoid unnecessary learning and increase generalization.

The learning method according to the present invention is similar to the above method in the neural network which handles a continuous amount of teacher signals in addition to the multi-valued neural network, by providing a region which is regarded as a correct answer by converging to the teacher signals. Can be widely applied to the learning of neural networks using general teacher signals.

In the above embodiments, the explanation has been made on the premise of the multilayer neural network, but a neural network other than the above may be used as long as it is a neural network which performs learning by using a teacher signal. Further, although a binary neural network has been described as an example here, although the description is omitted, the present learning method can also be applied to a neural network having a continuous amount of teacher signals.

[0025]

As described above, in the learning method of the neural network of the present invention, the number of errors between the multi-valued teacher signal and the multi-valued output unit signal does not change even if the learning is advanced, and it is a very stable local. In the case of falling into the minimum, this is detected through the local minimum capture state detector 15 using the incorrect output unit minimum error detection or the correct output unit minimum margin detection, and the weight in the output unit giving the correct answer. By adjusting the amplitude or further polarity of the coefficient update error signal, and further adjusting the amplitude of the weighting coefficient update error signal at the incorrect output unit, one can easily get out of such a stable local minimum. Therefore, the number of times of learning is remarkably reduced by 10 to 100 times as compared with the conventional method.
Further, since it is possible to stably converge to the optimum state by using the minimum required number of intermediate units or the number of intermediate layers, the hardware scale and the amount of calculation can be reduced.

The multi-valued neural network using the learning method of the present invention uses a smaller number of intermediate layer units or intermediate layers as compared with the conventional method, has no initial dependence, and has a low calculation accuracy. Can converge at high speed and stably, and a desired multilevel output signal can be transmitted.
From this, it is necessary to freely design and realize a large-scale multi-valued logic circuit or the like, which is difficult to be realized by the conventional technology, in a short time by the multi-valued neural network using the present invention, and rapid learning so far. In addition, it has a very wide range of effects such as a wide range of applications such as artificial intelligence systems and search systems, pattern recognition, data conversion, data compression, multi-valued image processing, and communication systems that require complete convergence. There is.

[Brief description of drawings]

FIG. 1 is a configuration example of a learning process in a multilayer neural network according to a conventional learning method.

FIG. 2 is a configuration example of a learning process of a multilayer neural network using the learning method of the present invention in the embodiment.

[Explanation of symbols]

1 Multi-layer neural network 2 Input signal input terminal 2 ₁ , 2 ₂ , 2 _N Input signal unit input terminal 3 Teacher signal input terminal 3 ₁ , 3 ₂ , 3 _M Teacher signal unit input terminal 4, 4 ₁ , 4 ₂ , 4 _M Subtractor 5 Weighting factor controller 6, 6 ₁ , 6 ₂ , 6 _M , 7, 7 ₁ , 7 ₂ , 7 _M Binary threshold circuit 8 Matching detector 9 Operation mode controller 10, 10 ₁ , 10 ₂ , 10 _M weight coefficient update error signal generator 11 learning determiner 12 operation mode controller 13 incorrect output unit minimum error detector 14 correct output unit minimum margin detector 15 local minimum capture state detector 16 local minimum capture state detector output terminal

─────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-4-15860 (JP, A) JP-A-4-251392 (JP, A) JP-A-3-218559 (JP, A) Yotaro Yatsuka, “Error Fast Convergence of Binary 3-Layer Neural Network by Perturbation ", Proc. Of IEICE General Conference 1995, Japan, The Institute of Electronics, Information and Communication Engineers, March 10, 1995, 1st volume, pp. 24 (58) Fields investigated (Int.Cl. ⁷ , DB name) G06N 1/00-7/08 G06G 7/60 JST file (JOIS) CSDB (Japan Patent Office)

Claims (4)

(57) [Claims] 1. A multi-valued neural network for learning using a multi-valued teacher signal, wherein:
In the output unit in which the multi-valued output unit signal of the multi-valued neural network matches the multi-valued teacher signal and is correct, the absolute value of the error signal obtained by subtracting the output unit signal from the multi-valued teacher signal is less than or equal to the first threshold. Then,
Generating a weighting coefficient update error signal having a polarity opposite to that of the error signal and having an amplitude that decreases as the distance from the multivalued teacher signal increases,
When the absolute value of the error signal is larger than the first threshold, a weighting coefficient update error signal of the same polarity having an amplitude equal to or smaller than the error signal is generated, and the multi-valued output unit signal becomes the multi-valued teacher signal. For output units that are incorrect and do not match,
A weight coefficient update error signal of the same polarity having an amplitude equal to or smaller than the error signal obtained by subtracting the output unit signal from the multi-valued teacher signal is generated, and the weight coefficient is updated using the weight coefficient update error signal. The minimum value of the absolute value of the error signal in the correct output unit is set as the incorrect output unit minimum error, and when it exceeds the second threshold, it is regarded as the local minimum capture state, and when this is detected, the weight coefficient update error signal is detected. A multi-valued neural network learning method characterized in that after the amplitude or the first threshold is increased, a weighting coefficient update error signal is obtained and the weighting coefficient is updated. 2. A multi-valued neural network for learning using a multi-valued teacher signal,
In the output unit in which the multi-valued output unit signal of the multi-valued neural network matches the multi-valued teacher signal and is correct, the absolute value of the error signal obtained by subtracting the output unit signal from the multi-valued teacher signal is less than or equal to the first threshold. Then,
Generating a weighting coefficient update error signal having a polarity opposite to that of the error signal and having an amplitude that decreases as the distance from the multivalued teacher signal increases,
When the absolute value of the error signal is larger than the first threshold, a weighting coefficient update error signal of the same polarity having an amplitude equal to or smaller than the error signal is generated, and the multi-valued output unit signal becomes the multi-valued teacher signal. For output units that are incorrect and do not match,
A weight coefficient update error signal of the same polarity having an amplitude equal to or smaller than the error signal is generated, the weight coefficient is updated using the weight coefficient update error signal, and the minimum absolute value of the error signal in the incorrect output unit Is the incorrect error output unit minimum error,
At the same time when this exceeds the second threshold, the minimum absolute value of the difference obtained by subtracting the output unit signal from the multilevel discrimination threshold for the correct output unit is defined as the correct output unit minimum margin, and the second threshold is set. When it exceeds, it is regarded as a local minimum capture state, this is detected, and the amplitude of the weighting factor update error signal or further the first
Multi-valued neural network learning method characterized by updating the weighting coefficient by obtaining the weighting coefficient update error signal after increasing the threshold of. 3. The learning of the multivalued neural network according to claim 1, wherein when the minimum error output unit minimum error becomes smaller than a third threshold after a local minimum capture state is entered, the weight is reduced. A multi-valued neural network learning method characterized in that the amplitude of the coefficient update error signal or further the first threshold is reduced to obtain a weight coefficient update error signal and the weight coefficient is updated. 4. The learning in the multi-valued neural network according to claim 1, 2 or 3, wherein after all of the multi-valued output unit signals match the corresponding multi-valued teacher signals, the correct output unit is selected. When the correct output unit minimum margin, which is the minimum absolute value of the difference obtained by subtracting the output unit signal from the multilevel discrimination threshold, becomes larger than the fourth threshold, the updating of the weighting coefficient in learning is terminated. Multi-valued neural network learning method.