JPH0721140A - Time space pattern classifying device - Google Patents

Abstract

PURPOSE:To provide the classifying device of simple constitution which follows up vibration in external input while suppressing time astableness as to a classifying device for multidimensional time space patterns which are successive with time like a speech signal and an image signal. CONSTITUTION:A pseudo neural cell element 1 fires when the weighted sum of external inputs from external signal input means 2 exceeds a threshold value. When plural pseudo neural cell elements 1 fire, a pseudo neural cell element 4 fires to suppress the pseudo neural cell elements 1 so that the pseudo neural cell elements 1 do not fire. Therefore, only the pseudo neural cell element 1 which obtains the maximum weighted sum of external inputs stably fires following the external inputs. Pseudo neural cell elements 3 corresponding to classification categories fire only when the pseudo neural cell element 1 coupled by a signal transmitting means 6 stably fires, so plural elements do not fire at the same time.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a spatiotemporal pattern classifying apparatus for classifying temporally continuous and spatially multidimensional information such as voice signals and image signals into several categories.

[0002]

2. Description of the Related Art For example, in speech recognition, which is one of spatiotemporal pattern classification, a speech signal subjected to preprocessing such as Fourier transform can be represented as a temporally continuous multidimensional function. No division corresponding to phoneme is given to this voice signal, and it is difficult to directly classify the pattern into phonemes in real time. Therefore, conventionally,
The main method was to memorize a voice signal, cut out a part of it as a spatial pattern that does not include time, and classify it with a classification method that cannot handle time directly.

In recent years, it has been proposed to use a neural pseudo circuit or neural network to classify such spatiotemporal patterns. For example, Weibel et al. Have proposed a model for classifying speech signals into phonemes by inputting them into a multilayer perceptron, which is one of the neural network models, by collecting the preprocessed speech signals for a certain period of time in the past. . It's "I-E-E-Transactions on Acoustics Speech and Signal Processing 3"
7 "(1989) pp. 328-339 (IEEE TRA
NSACTIONS ON ACOUSTICS, SPEECH, AND SIGNAL PROCESS
ING, Vol.37 (1989) P.328-339). Similar to the conventional speech recognition method, this model is also classified by the neural network model that cannot directly handle time after converting the spatiotemporal pattern into the spatial pattern via the memory.

On the other hand, Amari et al. Proposed a model using a neural cell pseudo element of continuous time continuous information type as one using a competitive neural network capable of directly handling time. Mathematical "(Amari, Sangyo Tosho, 1978), pages 205 to 214.
Page. This model will be described with reference to the drawings.

FIG. 6 is a block diagram of a conventional model using a continuous time continuous information type nerve cell pseudo element proposed by Amari et al. In FIG. 6, 81 is an inhibitory nerve cell pseudo element, 82 is an excitatory nerve cell pseudo element, 83 is a signal transmission means from the excitable nerve cell pseudo element 82 to the inhibitory nerve cell pseudo element 81, and 84 is an inhibitory nerve cell. Signal transmission means from the pseudo element 81 to the excitatory nerve cell pseudo element 82, 85
Is a feedback signal transmission means of the excitatory nerve cell pseudo element, 86 is an external signal input means to the excitatory nerve cell pseudo element, 87 is an external signal input means to the inhibitory nerve cell pseudo element, and 88 is an excitatory nerve cell pseudo. The output means outputs the output signal of the element to the outside. Here, each of the signal transmitting means 83, 84 and 85 is an axon simulating means for connecting a nerve cell simulating element and is accompanied by a weight. In this model, the internal state u _i of each excitatory nerve cell pseudo element 82 is as follows (Equation 1)
It changes according to.

[0006]

[Equation 1]

Here, t is time, and w1 is signal transmitting means 8
5 is a positive weight, f (x) is a function that is 1 when x is 0 or positive, and is 0 when x is negative, w ₂ is a weight associated with the signal transmitting means 84, and g (x) Is a function that becomes x when x is positive and becomes 0 when x is negative, v is the internal state of the inhibitory neuron pseudo element 81, h ₁ is a threshold value, and s _i is transmitted by the external signal input means 86. It is an external input signal. In addition,
f (u _i ) becomes each output value of each excitatory nerve cell pseudo-element 82 output to the outside via each output means 88. The internal state v of the inhibitory nerve cell pseudo element 81 changes according to the following (Equation 2).

[0008]

[Equation 2]

Here, τ is a time constant, n is the number of excitatory nerve cell pseudo elements, h ₂ is a threshold value, and s ₀ is an external input signal from the external signal input means 87. Note that g (v) is the output value of the inhibitory nerve cell pseudo device.

Next, the operation of Amari et al. Will be described. Hereinafter, the firing of the nerve cell pseudo element means a state in which the nerve cell pseudo element outputs a positive output value. First, consider a state in which s ₀ is 0 and all nerve cell pseudo elements are not fired. At this time, the excitatory nerve cell pseudo device 82 receiving the external input s _i exceeding h ₁ tries to fire. When at least one excitatory nerve cell pseudo element 82 fires, the inhibitory nerve cell pseudo element 81 that receives the output fires. If there is only one excitatory nerve cell pseudo-element 82 that receives an external input of h ₁ or more, the excitable nerve cell pseudo-element 82 continues to fire stably, but a plurality of excitable nerve cell pseudo-elements 82 When is fired, the inhibitory nerve cell pseudo-element 81 strongly fires, so that each excitatory nerve cell pseudo-element 82 transmits the signal transmission means 8
4 is strongly suppressed. The inhibition by the inhibitory nerve cell pseudo element 81 and the excitatory nerve cell pseudo element 82
The maximum external input s while repeating the firing cycle of
Only the excitatory nerve cell pseudo element 82 that has received _i finally fires, and the remaining excitatory nerve cell pseudo element 82 cannot fire, and this model becomes stable.

Then, once one excitatory neuron pseudo element 82 comes to fire stably, an external input s
_{Even if the} external input s _i to the excitatory nerve cell pseudo element that is changing and firing is not maximum, the excitatory nerve cell pseudo element 82 continues to fire. In order to stop the firing of the excitatory nerve cell pseudo element 82 and reset the state of all nerve cell pseudo elements, a positive or negative reset signal s ₀
Is required via the external signal input means 87, or a strong external input s _i is given to the excitatory nerve cell pseudo element 82 which has not fired.

[0012]

However, in the method of temporarily storing a part of the spatiotemporal pattern and cutting it out as a spatial pattern that does not include time, a large memory for storing the pattern is required, and the necessary and sufficient size of the spatial pattern is large. It was necessary to determine the level. Also, a device for dividing the spatiotemporal pattern, a device for storing the spatiotemporal pattern, and a device for classifying the stored spatial patterns are required separately, and the device tends to be complicated.

On the other hand, the competing neural network model proposed by Amari et al. Requires special treatment for alternation of firing excitatory nerve cell pseudo-elements, that is, resetting of the network, and changes every moment. It was difficult to follow the external input. A method of inputting a reset signal periodically may be considered as a method of realizing tracking of an external input, but in this case, it is easily affected by noise at the time of inputting the reset signal, and the excitatory nerve that fires. In the vicinity of the boundaries of the categories where the cell pseudo elements alternate,
They tended to be unstable in time, in which the stimulating excitatory nerve cell pseudo element was frequently replaced.

The present invention is intended to solve the above conventional problems, and an object of the present invention is to provide a simple spatiotemporal pattern classification device that follows changes in external inputs while suppressing temporal instability of classification results. There is.

[0015]

In order to achieve this object, the spatiotemporal pattern classification device of the present invention firstly fires when data to be classified is input and the value indicating the internal state exceeds a threshold value. And a second nerve cell pseudo-element for suppressing the firing of a plurality of the first nerve cell pseudo-elements when a plurality of the first nerve cell pseudo-elements are firing, and each of the first nerve cell pseudo-elements and the signal transmission A plurality of third neuron elements that are combined by the means and that are connected by the signal transmitting means among the first nerve cell pseudo-elements are continuously fired for a certain period of time or more to indicate the classification result. And a nerve cell pseudo element of the above.

The spatiotemporal pattern classification device of the present invention is the second
In addition to the above-mentioned first configuration, means for transmitting a signal for promoting firing to the first nerve cell pseudo-element connected to the firing third nerve cell pseudo-element by the signal transmitting means is provided. It has a configuration provided.

[0017]

According to the spatiotemporal pattern classifying apparatus of the present invention having the above-mentioned first configuration, the second nerve cell pseudo-element is the first nerve cell pseudo-element when the first nerve cell pseudo-elements are fired more than once. Each of the third nerve cell pseudo-elements receives an input signal and outputs a signal so that a time period in which a plurality of the first nerve cell pseudo-elements simultaneously fires a plurality of times is suppressed. Since the first nerve cell pseudo element operates so as not to fire unless it is stably fired for a certain time or more, a plurality of third nerve cell pseudo elements cannot fire at the same time. Therefore, the spatiotemporal pattern classification device of the present invention can easily realize the classification of spatiotemporal patterns that follows changes in external input while suppressing temporal instability of classification results.

Furthermore, the spatiotemporal pattern classification device of the present invention has the above-mentioned second structure and, in addition to the above-mentioned first operation, the third structure.
When the nerve cell pseudo-element of the above is firing, the first nerve cell pseudo-element receiving the output signal thereof is more likely to fire, so that the first nerve cell pseudo-element and the third nerve cell pseudo-element
The neuron pseudo element of has a hysteresis,
The alternation of firing of the third nerve cell pseudo element becomes difficult to occur. Therefore, the spatiotemporal pattern classification device of the present invention can easily realize the classification of spatiotemporal patterns that follows changes in external input while suppressing temporal instability of classification results.

[0019]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS A first embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram of a spatiotemporal pattern classification device according to a first embodiment of the present invention. In FIG. 1, reference numeral 1 is a first nerve cell pseudo element, 2 is an external signal input means for inputting data (two-dimensional data in this embodiment) for classifying the first nerve cell pseudo element 1, and 3 is a third Reference numeral 3 is a nerve cell pseudo element, 4 is a second nerve cell pseudo element, 5 is a signal transmitting means for inputting the output signal of the first nerve cell pseudo element 1 to the second nerve cell pseudo element 4, and 6 is Signal transmitting means for inputting the output signal of the first nerve cell pseudo-element 1 to the third nerve cell pseudo-element 3, 7 denotes the output signal of the second nerve cell pseudo-element 4 as the first nerve cell pseudo-element 1 is a signal transmitting means for inputting the signal to 1, and 8 is a signal transmitting means for outputting the output signal of the third nerve cell pseudo element 3 to the outside.

FIG. 2 is a block diagram of a nerve cell pseudo device according to the first embodiment of the present invention, in which the above-mentioned first nerve cell pseudo device 1, third nerve cell pseudo device 3, and second nerve cell are used. The process flow of the pseudo element 4 is shown. In FIG.
21 is an input adder that calculates a weighted sum of a plurality of inputs,
22 is an internal state calculation unit that calculates a new internal state from the calculation result of the input addition unit 21 and the current internal state and updates the internal state; 23 is a calculation result of the internal state calculation unit 22; An output calculation unit for calculating the output of the device, 24 is a signal transmission means for inputting a plurality of signals from the outside of the nerve cell pseudo device, and 25 is an input addition unit 21.
The signal transmission means for inputting the calculation result of the above into the internal state calculation part 22, 26 is the signal transmission means for inputting the new internal state which is the calculation result of the internal state calculation part 22 into the output calculation part 23, and 27 is the output It is a signal transmission means for outputting the output of the nerve cell pseudo element, which is the calculation result of the calculation unit 23, to the outside.

The operation of the first nerve cell pseudo device 1 according to the first embodiment of the present invention will be described below with reference to FIGS. 1 and 2. In the case of the first nerve cell pseudo element 1, the signal transmission means 24 correspond to the external signal input means 2 and the signal transmission means 7. In the present embodiment, the calculation performed by the input adder unit 21 of the first nerve cell pseudo device 1 follows the following (Equation 3).

[0022]

[Equation 3]

In (Equation 3), N is the dimension of the data input by the external signal input means 2, and N is 2 in this embodiment. W _ij is a positive or negative weight by which each external signal is multiplied, I
_j is an external signal, _winh is a positive weight by which the output of the second neural cell pseudo element 4 input through the signal transmission means 7 is multiplied, Y is the output of the second neural cell pseudo element 4, and M is the first In the present embodiment, M is 4 in terms of the number of nerve cell pseudo elements 1.
The value calculated by the input addition section 21 according to (Equation 3) is transmitted to the internal state calculation section 22 by the signal transmission means 25. Then, the internal state calculation unit 22 calculates the internal state of the first neural cell pseudo device 1 according to the following differential equation (Equation 4).

[0024]

[Equation 4]

In (Equation 4), τ ₁ is a time constant, x _i is an internal state of the i-th first nerve cell pseudo element 1, and h ₃ is a threshold value. The second and third terms on the right side are the results calculated by the input addition unit 21 according to (Equation 3).
The weight of the internal state calculated by the internal state calculation unit 22 according to (Equation 4) is transmitted to the output calculation unit 23 by the signal transmission unit 26. The output calculator 23 calculates the output of the first nerve cell pseudo device 1 according to the following (Equation 5).

[0026]

[Equation 5]

In (Equation 5), X _i is the output of the i-th first neural cell pseudo device 1. That is, the output calculation unit 2
In No. 3, 1 is output when the internal state is 0 or positive, and 0 is output when the internal state is negative.

By configuring the first nerve cell pseudo device 1 as described above, the first nerve cell pseudo device 1 operates as follows. First, when the output of the input addition unit 21 calculated by (Equation 3) is less than or equal to the threshold value h ₃ of (Equation 4), the first
The nerve cell pseudo element 1 of is firing at that time, that is, the output is 1
However, it will stop firing after the time determined by the time constant τ ₁ .
Next, consider a case where the output of the input adder 21 calculated by (Equation 3) exceeds the threshold value h ₃ of (Equation 4). At this time, since the internal state is positive, the first nerve cell pseudo device 1 starts firing after a time determined by the time constant τ ₁ even if it does not fire at that time.

Next, the operation of the second neural cell pseudo device 4 according to the first embodiment of the present invention will be described with reference to FIGS. 1 and 2. In the case of the second nerve cell pseudo element 4, the signal transmitting means 5 corresponds to the signal transmitting means 24. In the present embodiment, the calculation performed by the input adder unit 21 of the second neural cell pseudo element 4 follows (Equation 6) below.

[0030]

[Equation 6]

In ( _Equation 6), w _exc is the signal transmission means 5
Is a positive weight to be multiplied to each output of the first nerve cell pseudo device 1 input by. The value calculated by the input addition section 21 according to (Equation 6) is transmitted to the internal state calculation section 22 by the signal transmission means 25. And the internal state calculation unit 2
In 2, the internal state of the second neural cell pseudo element 4 is calculated according to the following differential equation (Equation 7).

[0032]

[Equation 7]

In (Equation 7), τ ₂ is a time constant, y is an internal state, and h ₄ is a threshold value. The second term on the right side is the result calculated by the input addition unit 21 according to (Equation 6). It should be noted that the threshold value h ₄ depends on which first nerve cell pseudo element 1
Is larger than the output of the input adder 21 in the case where only one first nerve cell pseudo element 1 is fired or two or more first nerve cell pseudo elements 1 are fired. The output is set to be smaller than the output of the input addition unit 21 in the case. The internal state calculated by the internal state calculation unit 22 according to (Equation 7) is output by the signal transmission unit 26 to the output calculation unit 2
3 is transmitted.

The output calculator 23 calculates the output of the second neural cell pseudo device 4 according to the following (Equation 8).

[0035]

[Equation 8]

In the equation (8), Y is the output of the second neural cell pseudo device 4. That is, the output calculation unit 23 outputs y when the internal state y is 0 or positive, and outputs 0 when the internal state y is negative. By configuring the second nerve cell pseudo device 4 as described above, the second nerve cell pseudo device 4 operates as follows.

First, when the output of the input adder 21 calculated in (Equation 6) is less than or equal to the threshold value h _{4 in} (Equation 7), the second
Time constant τ ₂
Stops firing after the time determined by. The output of the input addition unit 21 calculated in (Equation 6) is the threshold value h _{4 of} (Equation 7).
When it exceeds, the internal state y goes to the positive side, so that even if the second nerve cell pseudo device 4 does not fire at that time, it starts firing after the time determined by the time constant τ ₂ .

That is, from the setting of the threshold value h ₄ , whether any of the first nerve cell pseudo elements 1 has not fired
If only the nerve cell pseudo-element 1 of No. 2 is firing, the second nerve cell pseudo-element 4 is not firing if it is not firing, and if it is firing, the time determined by the time constant τ ₂ elapses. Stop firing. Next, the first nerve cell pseudo device 1
, The second nerve cell pseudo-element 4 continues to fire if it is firing, and if it is not firing, the internal state y changes so that it fires, and the time constant τ _{When the} time determined by ₂ passes, it starts firing.

Next, the operation of the third neural cell pseudo device 3 in the first embodiment of the present invention will be described with reference to FIGS. 1 and 2. In the case of the third nerve cell pseudo element 3, the signal transmission means 6 corresponds to the signal transmission means 24. In the present embodiment, the calculation performed by the input addition unit 21 of the nerve cell pseudo element 6 follows the following (Equation 9).

[0040]

[Equation 9]

In (Equation 9), w _out is the signal transmission means 6
This is a weight to be multiplied by the output of the first nerve cell pseudo element 1 input by. Input addition unit 21 according to (Equation 9)
The value calculated in (1) is transmitted to the internal state calculating section 22 by the signal transmitting means 25. Then, the internal state calculation unit 22 calculates the internal state of the third neural cell pseudo element 3 according to the following differential equation (Equation 10).

[0042]

[Equation 10]

In (Equation 10), τ ₃ is a time constant, z _i is an internal state of the i-th third nerve cell pseudo element 3, and h ₅ is a threshold value. The second term on the right side is the result calculated by the input addition unit 21 according to (Equation 9). Further, the threshold value h ₅ is set to a positive value and smaller than w _out . The internal state calculated by the internal state calculation unit 22 according to (Equation 10) is transmitted to the output calculation unit 23 by the signal transmission means 26. The output calculator 23 calculates the output of the third neural cell pseudo device 3 according to the following (Equation 11).

[0044]

[Equation 11]

In (Equation 11), Z _i becomes the output of the i-th third nerve cell pseudo element 3. That is, the output calculation unit 2
In 3, as in the case of the first nerve cell pseudo device 1, 1 is output when the internal state is 0 or positive, and 0 is output when the internal state is negative. By configuring the third nerve cell pseudo element 3 as described above, the third nerve cell pseudo element 3 operates as follows.

First, when the output of the input adder 21 calculated in (Equation 9) is less than or equal to the threshold value h _{5 in} (Equation 10),
Even when the third nerve cell pseudo element 3 is fired, the firing is stopped after a time determined by the time constant τ ₃ . Next, when the output of the input adder 21 calculated in (Equation 9) exceeds the threshold value h ₅ of (Equation 10), the internal state z _i is positive, and
Even if the third nerve cell pseudo element 3 does not fire, it starts firing after a time determined by the time constant τ ₃ .

That is, by setting the threshold value h ₅ , when the i-th first nerve cell pseudo-element 1 is firing, the i-th third nerve cell pseudo-element 3 changes its internal state in the firing direction. Then, when the i-th first nerve cell pseudo-element 1 has not fired, the i-th third nerve cell pseudo-element 3 changes its internal state in a direction in which it does not fire. And
The transition time between firing and non-firing is (Equation 9), (Equation 1
It can be adjusted by each parameter of 0) and (Equation 11). Then, each of the above parameters is set so that the third nerve cell pseudo device 3 coupled thereto is fired only when the first nerve cell pseudo device 1 is stably firing.

Based on the operations of the first nerve cell pseudo element 1, the third nerve cell pseudo element 3 and the second nerve cell pseudo element 4 described above, the first embodiment of the present invention will be described with reference to FIG. The operation of the spatiotemporal pattern classification device in the example will be described.
In the present invention, each pair of the first nerve cell pseudo-element 1 and the third nerve cell pseudo-element 3 which are connected one-to-one by the signal transmitting means 6 corresponds to each category which is the classification result. That is, the category corresponding to the firing third nerve cell pseudo element 3 becomes the classification result at that time.

The spatiotemporal patterns to be classified are input to each first nerve cell pseudo element 1 by the external signal input means 2. The magnitude of the effect of the signal received by each first nerve cell pseudo element 1 from the second nerve cell pseudo element 4 on the internal state of each first nerve cell pseudo element 1 is the third term of (Equation 4). It is common to all the first nerve cell pseudo elements 1. Therefore, the magnitude of the added value of the input calculated in (Equation 3) is determined by the second term of (Equation 4), that is, (Equation 12) which is the weighted sum of the external signals.

[0050]

[Equation 12]

In the classification of the spatiotemporal patterns in the present embodiment, the first nerve cell pseudo device 1 corresponding to the inputted category obtains the maximum (Equation 12) (Equation 1).
Set W _ij in 2). Each first nerve cell pseudo element 1
Tries to fire when the value of (Equation 12) is larger than h _{3 of} (Equation 4) (hereinafter, Condition 1). At this time, when there is only one first nerve cell pseudo element 1 that satisfies the condition 1, only the first nerve cell pseudo element 1 fires.

Next, the case where there are a plurality of first nerve cell pseudo-elements 1 satisfying the condition 1 will be described. In this case, the first nerve cell pseudo device 1 that satisfies the condition 1 fires in the descending order of the value of (Equation 12). However, a plurality of first nerve cell pseudo-elements 1 are fired, and after a while, the second nerve cell pseudo-elements 4 are fired, and all the first nerve cell pseudo-elements 1 are fired.
Is suppressed through the signal transmission means 7. Therefore, this time, in the first neural cell pseudo-element 1 that is firing,
The first nerve cell pseudo device 1 which is firing in the ascending order of the value of 2) stops firing. Then, when only one of the first nerve cell pseudo-elements 1 is in the firing state, the second nerve cell pseudo-element 4 stops firing so that the last first nerve cell pseudo-element 1, that is, The first nerve cell pseudo device 1 having the maximum value of (Equation 12) continues to fire. 2 of the first nerve cell pseudo-elements 1 that have stopped firing
The first nerve cell pseudo-element 1 having the second largest value (Equation 12) tries to fire again when the second nerve cell pseudo-element 4 stops firing, but within a short time after firing. Since the second nerve cell pseudo element 4 fires again, it can fire only for a very short time.

The other first nerve cell pseudo device 1
Cannot fire again because the second largest neuron pseudo element 1 fires and the second neuron pseudo element 4 fires before firing. Note that, during that time, the first nerve cell pseudo device 1 that has the maximum value of (Equation 12) continues to fire. In other words, in the state where the external input does not change, the maximum value of (Equation 12) is obtained.
Only the nerve cell pseudo-element 1 of the above continues to fire, and the first nerve cell pseudo-element 1 having the second largest value of (Equation 12) repeatedly fires for a short time, and the other first nerves. The cell pseudo device 1 cannot fire.

When the first neural cell pseudo-element 1 that has the maximum value of (Equation 12) due to the change of the spatiotemporal pattern to be input next is changed to another first neural cell pseudo-element 1 think of. In this case, the first nerve cell pseudo element 1 that has newly obtained the maximum value of (Equation 12) is fired. Then, since the internal state of (Equation 4) becomes sufficiently large until the second nerve cell pseudo element 4 fires, when the second nerve cell pseudo element 4 fires, before it has already fired. The first nerve cell pseudo element 1 that has obtained the maximum value of (Equation 12) stops firing first. Therefore, the first neural cell pseudo device 1 that fires alternates.

Now, the third nerve cell pseudo element 3 fires only when the first nerve cell pseudo element 1 connected by the signal transmitting means 6 is firing. However, since a plurality of first nerve cell pseudo-elements 1 are fired only when the coupled first nerve cell pseudo-elements 1 are stably firing, a plurality of first nerve cell pseudo-elements 1 are more than h _{3 of} (Equation 4). Even when a large value of (Equation 12) is obtained, only the third neural cell pseudo-element 3 that is coupled to the first neural cell pseudo-element 1 that obtains the maximum value of (Equation 12) fires. However, the other third nerve cell pseudo elements 3 cannot fire. Further, since the input spatiotemporal pattern changes and the first nerve cell pseudo element 1 is not stabilized until the third nerve cell pseudo element 3 coupled thereto is fired, for example, the spatiotemporal pattern is noisy. Or the first nerve cell pseudo-element 1 that has the maximum value of (Equation 12) due to the superimposition of or in a very short time-temporal pattern change and is fired for a short time. Even if the device 1 is replaced for a short time, the third nerve cell pseudo device 3 is not replaced and stable firing can be continued. The third nerve cell pseudo-element 3 stops firing when the first nerve cell pseudo-element 1 connected to it stops firing, but the time taken for that is the first nerve cell pseudo-element connected to it. It is shorter than the time from the firing of 1 to the firing of the third neural cell pseudo element 3. Therefore, when the firing third nerve cell pseudo-elements 3 alternate, the plurality of third nerve cell pseudo-elements 3 do not fire at the same time.

Each of the above-described first nerve cell pseudo elements 1, 3 and
The operation of No. 4 will be described more specifically using an example with reference to FIG. Note that, in the example used in FIG. 3, for simplicity, it is assumed that only the two first nerve cell pseudo-elements 1 always obtain the value of (Equation 12) larger than h ₃ .

FIG. 3 is a weighted sum of external signals of the first nerve cell pseudo element 1 calculated by (Equation 12) in the first embodiment of the present invention, and the first nerve cell pseudo elements 1, 3 FIG. 5 is a diagram schematically showing the firing states of the nerve cell pseudo element 3 and the second nerve cell pseudo element 4. FIG. 3A is a diagram showing changes in the weighted sum of external signals calculated by (Equation 12) of the two first nerve cell pseudo elements 1, and the solid line and the broken line represent the magnitudes of the respective weighted sums. . FIG. 3B is a diagram showing the state of firing of the two first nerve cell pseudo-elements 1 in FIG. 3A, in which the portion surrounded by the solid line in FIG. ) Shows a state of firing of the first nerve cell pseudo-element 1 corresponding to the solid line, and a portion surrounded by a broken line in (b) and having an oblique line in a direction different from the oblique line is (a)
The state of firing of the first nerve cell pseudo-element 1 corresponding to the broken line is shown, and the overlapping portions of the diagonal lines show that the two first nerve cell pseudo-elements 1 are firing. Figure 3
(C) is a diagram showing a state of firing of the second neural cell pseudo device 4. Further, (d) of FIG. 3 is a diagram showing the firing state of the two third nerve cell pseudo elements 3 coupled to the two first nerve cell pseudo elements 1, and the solid line in (d) is First nerve cell pseudo device 1 corresponding to the solid line in (a)
The third nerve cell coupled to the first nerve cell pseudo-element 1 in which the dashed line in (d) corresponds to the dashed line in (a). The state of ignition of the pseudo element 3 is shown.

In the example of FIG. 3, first, the first nerve cell pseudo element 1 corresponding to the solid line stably fires, and as shown in FIG. 3B, the first nerve cell pseudo element corresponding to the broken line. Element 1 ignites for a very short time. Then, as shown in FIG. 3C, the first nerve cell pseudo device 1 corresponding to the broken line
The second nerve cell pseudo element 4 is fired each time is fired. During that time, as shown in FIG. 3D, only the third nerve cell pseudo element 3 corresponding to the solid line fires.

Next, when the first nerve cell pseudo element 1 for obtaining the maximum weighted sum of external signals changes from the one corresponding to the solid line to the one corresponding to the broken line, it corresponds to the solid line and the broken line at the time of the change. The first nerve cell pseudo-element 1 and the second nerve cell pseudo-element 4 are fired, and finally the first nerve cell pseudo-element 1 corresponding to the broken line is stably fired. When the first nerve cell pseudo element 1 corresponding to the solid line stops firing and the first nerve cell pseudo element 1 corresponding to the broken line comes to fire stably, the third nerve cell fires correspondingly. The pseudo element 3 is also replaced, but since it takes time for the third nerve cell pseudo element 3 to fire, the plurality of third nerve cell pseudo elements 3 do not fire.

Then, the first nerve cell pseudo-element 1 that stably fires changes, and for a while the first nerve cell pseudo-element 1 that corresponds to the broken line continues to fire stably, and the first nerve cell pseudo-element 1 that corresponds to the solid line Although the nerve cell pseudo-element 1 of 1 is fired at intervals, the second nerve cell pseudo-element 4 is also fired, so the firing is stopped in a short time. Therefore, as shown in FIG. 3D, only the third neural cell pseudo element 3 corresponding to the broken line continues to fire. When the first nerve cell pseudo-element 1 that obtains the maximum weighted sum of external signals changes from the one corresponding to the broken line to the one corresponding to the solid line, the first nerve cell pseudo-element 1 corresponding to the solid line and 3 and the first nerve cell pseudo-elements 1 and 3 corresponding to the broken line alternate the state of stable firing. With the above operation, the spatiotemporal pattern classification device according to the first embodiment can easily realize the classification of the spatiotemporal pattern that follows the change of the external input while suppressing the temporal instability of the classification result.

In this embodiment, the first nerve cell pseudo-element 1 and the third nerve cell pseudo-element 3 correspond to each other in a one-to-one correspondence. However, as shown in FIG. Pseudo element 1
On the other hand, even if there is one that corresponds to one third nerve cell pseudo element 3, the pattern classifying device can be similarly realized.

The second embodiment of the present invention will be described below with reference to the drawings. FIG. 5 is a block diagram of the spatiotemporal pattern classification device according to the second embodiment of the present invention. Figure 5
40 is a signal transmitting means for inputting the output signal of the third nerve cell pseudo-element 3 to the first nerve cell pseudo-element 1. This embodiment is the spatiotemporal pattern classification device described in the first embodiment. The signal transmitting means 40 is added to the above. Further, the first nerve cell pseudo element 1, the third nerve cell pseudo element 3, and the second nerve cell pseudo element 4 also have the configurations shown in FIG.

The operation of the first nerve cell pseudo device 1 will be described below with reference to FIG. The first nerve cell pseudo device 1 operates based on the same principle as that of the first embodiment. However, in the case of the first nerve cell pseudo device 1 of the second embodiment, the signal transmission means 24 correspond to the external signal input means 2, the signal transmission means 7, and the signal transmission means 40. Therefore, in the second embodiment, the calculation performed by the input addition unit 21 of the first nerve cell pseudo device 1 follows the following (Equation 13).

[0064]

[Equation 13]

In ( _Equation 13), _whys is a weight by which each output of the third nerve cell pseudo device 3 input by the signal transmitting means 40 is multiplied. Other symbols are based on (Equation 3) and (Equation 11). The value calculated by the input addition section 21 according to (Equation 13) is transmitted to the internal state calculation section 22 by the signal transmission means 25. The internal state calculation unit 22
Then, the internal state of the third neural cell pseudo element 3 is calculated according to the following differential equation (Equation 14).

[0066]

[Equation 14]

In (Equation 14), each symbol conforms to (Equation 4). The second term, the third term, and the fourth term on the right side are
This is the result calculated by the input addition unit 21 according to (Equation 13). The threshold value h ₃ is set as in the first embodiment. The weight of the internal state calculated by the internal state calculation unit 22 according to (Equation 13) is transmitted to the output calculation unit 23 by the signal transmission unit 26. The output calculator 23 calculates the output of the first neural cell pseudo device 1 according to (Equation 5) as in the first embodiment.

The first nerve cell pseudo device 1 is constructed as described above, and operates in the same manner as the first nerve cell pseudo device 1 in the first embodiment. Further, the third nerve cell pseudo-element 3 and the second nerve cell pseudo-element 4 have the same configurations as the third nerve cell pseudo-element 3 and the second nerve cell pseudo-element 4 of the first embodiment, respectively. Have and do the same operation.

The operation of the spatiotemporal pattern classifying apparatus according to the second embodiment of the present invention will be described below with reference to FIG. 5 in consideration of the operation of the spatiotemporal pattern classifying apparatus according to the first embodiment.

First, consider the case where the third nerve cell pseudo element 3 is not fired. In this case, (Equation 13) and (Equation 14) are equal to (Equation 3) and (Equation 4), respectively. Therefore, the spatiotemporal pattern classifying apparatus according to the second embodiment performs the same operation as the spatiotemporal pattern classifying apparatus according to the first embodiment.

Next, consider a case where the third nerve cell pseudo element 3 is firing. In this case, the signal transmission means 6 and the signal transmission means 40 are provided to the third nerve cell pseudo device 3 which is firing.
The first nerve cell pseudo-element 1 coupled to the second nerve cell pseudo-element 1 receives the input larger than that of the other first nerve cell pseudo-elements 1 by at least w _hys because the third nerve cell pseudo-element 3 outputs 1. Therefore, the spatiotemporal pattern that is input to the spatiotemporal pattern classification device changes and the third
The first neural cell pseudo-element 1 which is firing and is connected to the neural cell pseudo-element 3 of FIG. 2 by the signal transmission means 6 and the signal transmission means 40 obtains the maximum value of (Equation 12). The first nerve cell pseudo-element 1 in which the other first nerve cell pseudo-element 1 fires even if the nerve cell pseudo-element 1 disappears
The first neural cell pseudo device 1 which is firing continues firing until a value of ( _Equation 12) larger by at least _whys is obtained. Therefore, the firing third nerve cell pseudo element 3 also continues firing.

In other words, once the third nerve cell pseudo element 3 fires, the firing of the first nerve cell pseudo element 1 connected to it becomes more stable, and the first value that obtains the maximum value (Equation 12) is obtained. Even if the nerve cell pseudo element 1 of is replaced with another first nerve cell pseudo element 1, if the difference from the firing first nerve cell pseudo element 1 is not large, the classification result does not change, This spatiotemporal pattern classifier has hysteresis. The degree of this hysteresis can be adjusted by further adjusting _whys . Therefore, when noise is superimposed on the spatiotemporal pattern input by the external signal input unit 2 or at the boundary of the category, W of (Equation 13)
If the classification scale determined by _ij is insufficient, classification is performed based on past classification results.

With the above operation, the spatiotemporal pattern classifying apparatus according to the second embodiment can perform classification that follows changes in external input while suppressing temporal instability of classification results. In addition, in the second embodiment, _whys
Is set to the same value for each first nerve cell pseudo element 1,
Not limited to this, by changing the individual first nerve cell pseudo element 1, it is possible to operate so that each category has different hysteresis. Further, although the first nerve cell pseudo-elements 1 and the third nerve cell pseudo-elements 3 have a one-to-one correspondence as in the first embodiment, a plurality of first nerve cell pseudo-elements 1 are provided. On the other hand, even if there is one that corresponds to one third nerve cell pseudo element 3, the pattern classifying device can be similarly realized.

Further, in the first and second embodiments, the nerve cell pseudo device is constructed as shown in FIG.
It's not something you get stuck with. For example, the input addition unit 2
Nerve cell pseudo element model in which 1 is multiplied by the input, nerve cell pseudo element model configured to calculate the distance between the input and the pattern stored in advance in place of the input adder 21, and internal state calculation The same effect can be obtained by using a model in which the section 22 is discretely expressed by a difference equation instead of a differential equation.

The spatiotemporal pattern classifying apparatus according to the present invention can be realized on a computer by a program, and can also be realized as an integrated electronic circuit or an electric circuit.

[0076]

As described above, according to the present invention, firstly, the second nerve cell pseudo-element is the first nerve cell when a plurality of the first nerve cell pseudo-elements are firing. Each of the third neural cell pseudo-elements receives an input signal so that the pseudo-elements are suppressed so that the first neural-cell pseudo-elements fire a plurality of times at the same time so as to shorten the time for which each of the third neural-cell pseudo-elements receives the input signal. Since the first nerve cell pseudo element operates so as not to fire unless it is stably fired for a certain period of time or more, a plurality of third nerve cell pseudo elements cannot fire at the same time. Therefore, it is possible to easily realize the classification of the spatiotemporal pattern that follows the change of the external input while suppressing the temporal instability of the classification result.

Further, according to the present invention, secondly, in addition to the above-mentioned first effect, the first nerve cell pseudo element which receives the output signal of the third nerve cell pseudo element when the third nerve cell pseudo element is firing is more effective. Since the fourth signal transmission means operates so as to easily fire, the first nerve cell pseudo element and the third nerve cell pseudo element
The neuron pseudo element of has a hysteresis,
The alternation of firing of the third nerve cell pseudo element becomes difficult to occur. Therefore, it is possible to easily realize the classification of the spatiotemporal pattern that follows changes in the external input while suppressing temporal instability.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a spatiotemporal pattern classification device according to a first embodiment of the present invention.

FIG. 2 is a configuration diagram of a nerve cell pseudo device according to the first and second embodiments of the present invention.

FIG. 3 (a) is a graph showing the first embodiment of the present invention (Equation 1)
The figure which showed the example of the time change of the weighted sum of the external input signal of the 1st nerve cell pseudo | simulation element 1 calculated by 2). (B) The firing time of the 1st nerve cell pseudo | simulation element 1 in the Example. (C) A diagram showing an example of a temporal change in firing of the second neural cell pseudo element 4 in the same embodiment (d) A third neural cell pseudo element 3 in the same embodiment Figure showing an example of temporal changes in ignition

FIG. 4 is a configuration diagram when a plurality of first nerve cell pseudo-elements 1 and one third nerve cell pseudo-element 3 correspond to each other in the embodiment.

FIG. 5 is a configuration diagram of a spatiotemporal pattern classification device according to a second embodiment of the present invention.

FIG. 6 is a block diagram of a conventional model using a neural cell pseudo device of continuous time continuous information type.

[Explanation of symbols]

 1 1st nerve cell pseudo element 2 External signal input means 3 3rd nerve cell pseudo element 4 2nd nerve cell pseudo element 5 Signal transmission means 6 Signal transmission means 7 Signal transmission means 8 Signal transmission means 21 Input addition part 22 Internal state calculation unit 23 Output calculation unit 24 Signal transmission means 25 Signal transmission means 26 Signal transmission means 27 Signal transmission means 40 Signal transmission means

Claims (2)

[Claims] 1. A first nerve cell pseudo-element that is fired when data to be classified is input and a value indicating an internal state exceeds a threshold value, and a plurality of the first nerve cell pseudo-elements are fired. A second nerve cell pseudo-element for suppressing the firing and a plurality of third nerve cell pseudo-elements corresponding to the classification category are provided, and each of the first nerve cell pseudo-elements includes the plurality of third nerve cell pseudo-elements. Any one of the nerve cell pseudo elements is coupled by the signal transmitting means, and the third nerve cell pseudo element is coupled to the one of the first nerve cell pseudo elements by the signal transmitting means for a certain time or more. The spatiotemporal pattern classification device is characterized in that the classification result is represented by firing when it is being fired. 2. A means for transmitting a signal for promoting firing to a first nerve cell pseudo-element, which is connected to a firing third nerve cell pseudo-element by a signal transmitting means, wherein The spatiotemporal pattern classification device according to claim 1, wherein the one nerve cell pseudo element and the third nerve cell pseudo element have hysteresis.