JPH0736180B2 - Neurocomputer operation method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial field of application] The present invention calculates the product of a correlation matrix and a state vector of a neuron, and selects the most input vector from the vector information stored in advance for the input vector. The present invention relates to a calculation method of a neuro computer that outputs similar vector information.

[Conventional technology]

Figure 5 shows, for example, Applied Optics, 1985, Volume 24, Issue 10.
FIG. 1 is a structural conceptual view showing an optical neurocomputer device which is a neurocomputer using the conventional arithmetic method shown in pages 1469 to 1475, in which 1 is a light emitting element array for generating an optical signal, and 2 is an optical modulator. 3 is a photodetector array for receiving an optical signal, and 7 is a threshold element including a comparator for comparing the output of the photodetector array 3 with a threshold value.

Next, the operation will be described. This device is based on the neural net model proposed by Hopfield (eg, Proceedings of National Academy of Science, USA).
1982, Vol. 79, pp. 2554-2558). According to this model, M pieces of previously stored N-dimensional vector information Si
^(m) (i = 1 ... N, m = 1, ... M, Si ^(m) ∈ {1, -1} ^N fov ^∀ m An association that selects the most similar one to the input vector Si ^(mo) The memory can be configured.

The accumulated information is given by the correlation matrix mask 2. If the correlation matrix is Tij, Tij is given by the following equation.

Here, for convenience of explanation, it is assumed that there is almost no correlation between the stored information vectors. That is, the accumulated information vectors are assumed to be orthogonal to each other. That is And now, assuming an incomplete input vector Si ^(mo) close to Si ^(m) , the product of the correlation matrix and Si ^(mo) is It is shown that O (x) is a minute amount of order x.

Therefore, a strong non-linear action is performed, and by further inputting this and feeding it back, the stored information vector closest to the input vector is output. Here, the threshold element 7 is used as a non-linear action, and Vi = (Si + 1) / 2 (4) is used to represent the neuron state by turning the light emitting element array 1 ON / OFF using a unipolarized vector. ing. Also, the product Ui of this neuron state vector and the correlation matrix Is expressed as the output of the light receiving element array 3. This Ui
Threshold, input and feed back. I.e. At this time, the relationship between the number of accumulated information M and the number of neurons N is expressed by M to N / (4logN) (7).

[Problems to be Solved by the Invention]

Since the operation method in the conventional neural computer is configured as described above, in order to increase the number M of stored information, the number N of neurons must be increased according to the equation (7). The correlation matrix becomes large and the entire system becomes large,
In addition, there is a problem that alignment of the optical system becomes difficult.

The present invention has been made to solve the above problems, and an optical neurocomputer capable of substantially increasing the number of neurons without changing the size of the entire system and thereby increasing the number of accumulated information is provided. The purpose is to get.

[Means for Solving the Problems]

The control unit 1A in the operation method of the neurocomputer according to the present invention divides the correlation matrix into a number of sub-matrices, generates the sub-matrices in time series, and also generates partial vectors of the state vector in time series, It is characterized in that the matrix and the vector of each part are calculated to perform the product operation of the correlation matrix and the state vector of the neuron as a whole.

Further, the control unit 1B in the operation method of the neurocomputer according to the present invention, the submatrix and the partial vector of the state vector when the correlation matrix is divided into several submatrices are frequency-multiplexed by modulating at different frequencies, It is characterized in that the product operation of the state vector and the correlation matrix is performed in a multiplexed state, the product operation is performed, and then frequency separation is performed to obtain the product of the entire state vector and the correlation matrix.

[Action]

The control unit 1A divides the correlation matrix into several sub-matrices, generates the sub-matrices in time series, and also generates the partial vector of the state vector in time series. Then, the control unit 1A performs the product operation of each submatrix and the partial vector to perform the product operation of the correlation matrix and the state vector as a whole.

In addition, the control unit 1B divides the correlation matrix into some small matrices,
Frequency modulation is performed by modulating each sub-matrix with a different frequency, frequency-multiplexing is also performed for the state vector, and each state vector is also modulated with a different frequency. Then, the control unit 1B calculates the small matrix and the partial vector in the frequency-multiplexed state, performs the product calculation, and then performs the frequency separation to obtain the product of the state vector and the correlation matrix as a whole.

Embodiments of the Invention Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a block diagram showing a first embodiment, in which 8 is an accumulator constituted by an integrating circuit using a capacitor, 9 is a light emitting element array 1, a correlation matrix mask 2 and a light receiving element array 3. It is a control circuit for controlling. Here, the accumulator 8 and the control circuit 9 constitute a control unit 1A.

Next, the operation will be described.

The number of virtual neurons is N, and the number of neurons actually displayed (the number of light emitting / light receiving element arrays) is P. Neuron state vector And the correlation matrix T (membrane potential) And Further, the correlation matrix T is divided into small matrices as follows.

Becomes

In FIG. 1, P = 2 is shown.

Now, at time t = 1, the control circuit 9 causes the correlation matrix mask 2
The partial correlation matrix T (1,1) is set to
Partial vector of state vector above Is displayed. This allows matrix vector operations Is the output of the light-receiving element array 3, and the accumulator 8 also performs matrix vector operation. Is stored. At time t = 2 (= P), the partial correlation matrix T (2,1) is set on the correlation matrix mask 2 and the partial vector is set on the light emitting element array 1. Is displayed, and in the light-receiving element array 3, matrix vector operation And thus in the accumulator 8 Is stored. At time t = 3, the control circuit 9 outputs the accumulator 8 and the threshold value element 7 performs threshold processing. And the result of this processing is used as an input to the light emitting element array 1.

At time t = 4, the partial correlation matrix T (1,2) and the partial vector And are displayed, and at time t = 5, the partial correlation matrix T (2,2)
And partial vector By displaying, the accumulator 8 To store. At time t = 6 (= 2P + 2), output the accumulator 8 and perform threshold processing. And perform feedback to the light emitting element array 1. The above process is repeated until the output of the threshold element 7 does not change. Thereby, the same operation as the number N of virtual neurons can be realized. FIG. 2 is a table showing the flow of operations.

In this case, the correlation matrix mask 2 may be a spatial light modulator such as a liquid crystal panel whose transmittance can be controlled from the outside, and the elements of each matrix may be expressed by the transmittance to express the partial correlation matrix.

Next, a second embodiment will be described.

FIG. 3 is a block diagram of a neurocomputer using the arithmetic system according to the second embodiment of the present invention. In FIG. 3, components corresponding to those shown in FIG. 1 are designated by the same reference numerals, and their description will be omitted. In FIG. 3, 10, 11, 12 are FMs for demodulating the output of the light receiving element array 3 by FM.
Demodulator, 13 is an adder for adding the outputs of the FM demodulator 12, 14 is a control circuit for controlling the correlation matrix mask 2, 15 is a comparator for comparing the output of the adder 13 with a threshold value, 16 is a light emitting element It is a modulator for performing the modulation operation of the array 1. FM demodulator 10,11,12,
The adder 13, the control circuit 14, the comparator 15, and the modulator 16 constitute the control unit 1B as a whole.

Next, the operation will be described. The number of neurons is N, and the number of neurons actually displayed (the number of light emitting element / light receiving element arrays) is
Let P be the number of divisions and S be the number of divisions. Note that N = S × P. Neuron state vector in light emitting element array 1 Each of the components is frequency-modulated by the modulator 16 at the angular frequency ω _h , and is multiplexed and appears simultaneously. Ie Is expressed as

The correlation matrix T in the correlation matrix mask 2 is divided into S × S sub-matrices under the control of the control circuit 14, and the respective components are modulated at the angular frequencies Ω _k and Ω _l ′ and appear at the same time. Ie Is expressed as

Then the neuron state vector And the correlation matrix T Is represented as follows.

However, ω _h >> Ω _k >>>> Ω _l .

Is the output of the light receiving element array 3, the first light receiving element output U (1) will be considered first. The signal ▲ U (1) ^ω ₁ ▼, ..., ▲ U (1) ^ω _s ▼ of the central angular frequency ω ₁ , ..., ω _s of U (1) after passing through the FM demodulator 10 is expressed by the following equation. become.

Then _{^{▲ U (1) ω 1 ▼}} , ..., ▲ U (1) the center angular frequency Ω _¹ ω _s ▼, respectively, ..., through the FM demodulator 11 of the Omega _s when the ▲ U ₍₁₎ ^Ω 1 ▼ , ..., U (1) ^Ω _s ▼ is as follows.

Then, let ▲ U (1) ^Ω ₁ ▼ be the central angular frequency Ω ' ₁ , ...,
_{'Of s, ..., ▲ U (1} ) Ω s ▼ center angular frequency Omega _a' Omega _1, ..., through the FM demodulator 12 of the Omega _'s. The outputs ▲ U (1) ^{Ω '} ₁ ▼, ..., ▲ U (1) ^Ω' _s ▼ are as follows.

When ▲ U (1) ^{Ω '} ₁ ▼, ..., ▲ U (1) ^Ω' _s ▼ are respectively led to the adder 13 and added. Signal is obtained. This signal U (1) is thresholded by the comparator 15 and then fed back to the input V (1) of the first light receiving element of the light emitting element array 1 The same operation as U (2), ..., U By performing for (P), the same operation as in the case of the number of neurons N can be realized.

Although the following sub-matrix is displayed at a timing faster than the response speed of the light receiving element array 3 in the above-described embodiment, an integrator is provided at the rear stage of the light receiving element array, and You may perform the calculation of. Also, a memory and an adder may be provided. Further, although the vector / matrix calculation is performed by the optical crossbar method, it may be performed by another method, for example, a systolic calculation method.

Further, in the above-described embodiment, the case of the associative characteristic has been described, but it may be the case of the optimum value problem or the case of introducing the learning function, and the same effect as the above-described embodiment is obtained.

Further, in the above embodiment, the case of the Hopfield neural network model has been described, but other neural network models such as a back propagation model may be used, and the same effect as that of the above embodiment is obtained.

In the above-mentioned embodiment, the case of the Hopfield model is shown, but it is also applicable to a multilayer neural network.

Figure 4 is a block diagram of this multilayer neural network.
In the figure, 21 is an input layer, 22 is a hidden layer, and 23.
Represents the output layer. Further, 24 is an input unit, and each input unit is divided into P sub-units 25, and each of them is frequency-modulated with frequencies ω ₁ , ω ₂ , ... ω _P. The input signal is binary, and is 1 when the frequency modulation signal exists and 0 when it does not exist. This signal is multiplexed at the input unit 24 and its output propagates to each hidden unit 26. The hidden layer 22 is also divided into partial units. At the input unit 26 of the hidden unit, the signals from the respective input units are added, and the output is ω ₁ , ω ₂ ,
The frequency is separated by a bandpass filter 27 having a center frequency of ω _P and thresholded by a comparator 28. When the output of each comparator 28 is 1, it is modulated by the modulator 29 at a frequency of ω ₁ ... ω _P and the signal is transmitted to the output layer 23. The output unit is also divided into partial units, and the output unit 30
The outputs are frequency-separated again by the bandpass filter 31, peak-held by the peak-hold circuit 32, and their outputs are collected in the comparator 33, thresholded and output. In this way, if there are P modulation frequency signals,
This is equivalent to the fact that there are P times as many units as the number of units N, and the capacity of the network is significantly improved. In this example, a normal multilayer neural network model is shown, but it can be easily extended to, for example, a back propagation model (neural network model having hierarchies).

Although the light emitting element (neuron state vector) is modulated by the frequency multiplexed signal in the above embodiment, the same effect can be obtained even if the sensitivity of the light receiving element is modulated by the frequency multiplexed signal. This can be achieved by using, for example, a phototransistor as the light receiving element.

In the above embodiment, the case of ω _k >> Ω _q is shown, but
Conversely, Ω _q >> ω _k may be used.

Further, although the case of the associative characteristic of the Hopfield model has been described in the above embodiment, it may be an optimum value problem such as a traveling salesman problem, and the same effect as that of the above embodiment is obtained.

〔The invention's effect〕

As described above, according to the present invention, the correlation matrix is divided into several sub-matrices, the sub-matrices are generated in time series, and the partial vectors of the state vector are generated in time series. Since the calculation of each sub-vector is performed and the product of the correlation matrix and the state vector of the neuron is performed as a whole, the number of virtual neurons can be increased without increasing the system scale, and thus the number of accumulated information can be increased. The effect is obtained.

Further, according to the present invention, when the correlation matrix is divided into a number of sub-matrices, the sub-matrix and the partial vector of the state vector are respectively modulated at different frequencies and frequency-multiplexed. The product of the correlation matrix is calculated, and after performing the product calculation, frequency separation is performed to obtain the product of the entire state vector and the correlation matrix. Therefore, the number of virtual neurons can be increased without increasing the system scale. The effect that the number of information can be increased can be obtained.

[Brief description of drawings]

FIG. 1 is a block diagram showing a neurocomputer using the arithmetic system according to the first embodiment of the present invention, FIG. 2 is a table showing the operating procedure of this embodiment, and FIG. 3 is a second view of the present invention. FIG. 4 is a block diagram showing a neuro computer using an arithmetic method according to the embodiment, FIG. 4 is a block diagram showing a multilayer neural network using the above embodiment, and FIG. 5 is a neuro computer using the conventional arithmetic method. It is a block diagram shown. 1 ... Light emitting element array, 2 ... Correlation matrix mask, 3 ...
Photodetector array, 7 ... Threshold element, 8 ... Accumulator, 9,14 ... Control circuit, 10, 11, 12 ... FM demodulator,
13 ... Adder, 15 ... Comparator, 16 ... Modulator, 1A, 1B ...
… Control unit.

Claims (1)

[Claims] 1. When selecting the vector information most similar to the input vector from among the vector information stored in advance for the input vector, the product of the correlation matrix giving the vector information and the state vector of the neuron. In a calculation method of a neurocomputer that performs a calculation, a product of a correlation matrix and a state vector as a whole based on a submatrix obtained by dividing the correlation matrix into several submatrices and a partial vector of the corresponding state vector There is provided a control unit for performing control for obtaining the sub-matrix of the correlation matrix and the partial vector of the state vector in time series, and each sub-matrix and each partial vector at each time timing. Is performed, and the result is sequentially stored in an accumulator composed of an integrating circuit etc. Or, the above-mentioned small matrix and the above-mentioned partial vector are respectively modulated at different frequencies to be frequency-multiplexed, and a product operation of the state vector and the correlation matrix is performed in a state where they are simultaneously appeared and multiplexed, and then frequency-separated, An operation method of a neuro computer characterized by performing control for adding all components.