JPH03196250A - Neuron simulation circuit - Google Patents

Abstract

PURPOSE:To facilitate the wiring of a neuron simulation circuit by providing an oscillator to the output part of each neuron unit to oscillate the oscillation frequencies different for each unit together with a modulator which modulates the frequency of the oscillator output in response to the output of each unit and then providing an electric filter and demodulator to the input part of each neuron unit. CONSTITUTION:Each neuron unit 6 includes an oscillator 8 and a modulator 7 in its output part and applies the modulation of frequency to an electric signal in response to the output of the unit 6. Thus it is possible to obtain a signal which has a fixed maplitude and an error proportional to the output of the unit 6 against the center frequency. Then these units 6 are modulated with different center frequencies via the oscillators having different oscillation frequencies. Therefore the outputs of the units 6 are mixed together and can be distributed to the units 6 of another layer via a single line. In such a constitution, just a single electric signal line suffices to serve as the wirings set between the layers of a network constitution. Thus no cross of wirings is produced and the wiring of a neuron simulation circuit is facilitated.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a neuron imitation circuit such as a neural computer that imitates a neuron network.

Conventional technology In recent years, research has begun on neural computers that imitate the nervous system of living organisms and are capable of parallel processing and learning in order to deal with problems that are relatively difficult for conventional Neumann computers, such as character recognition, associative memory, and motor control. have been carried out and various models have been proposed.

Conventional neural networks are shown in FIGS. 5 to 7. FIG. 6 is a diagram showing one neuron unit 1, and FIG. 5 is a diagram showing this as a network. A single neuron connects with many other neuronal units, receives signals, processes them, and produces output. In the case of FIG. 5, the network is hierarchical, and receives signals from a unit in the previous layer and outputs them to a unit in the next layer.

This will be explained in more detail. First, in neuron unit l in Figure 6, what is called the coupling coefficient represents the degree of coupling between other neuron units and the own unit.
The coupling coefficient between the first unit and the Jth unit is generally denoted by T1. Connections include excitatory connections, where the larger the signal from the other neuron, the greater the neuron's own output;
Conversely, there is an inhibitory connection in which the larger the signal from the other neuron, the smaller the own output, but T I J >
0 represents excitatory connection, T, <0 represents inhibitory connection. Now, assuming that your unit is the first unit,
If the output of the i-th unit is yl, then T, y, which is obtained by multiplying this by the coupling coefficient T, becomes the input to the own unit. As mentioned above, each unit is connected to many units, so T + for those units is
Σ"r,,y, which is the result of adding up r yI , becomes the input to its own unit. This is called the internal potential and is represented by U.

u7 = ΣT,,y, ・・・・・・・・・・・・
(1) Next, this input is subjected to nonlinear processing and output. The function at this time is called a neuron response function, and a sigmoid function as shown in equation (2) and FIG. 7 is used as the nonlinear function.

f (x)=1/(1+e-x)...
・・・・・・・・・・・・・・・(2) When a network is created as shown in Figure 5, by giving each coupling coefficient TIJ and calculating equations (1) and (2) one after another, , which gives the final output.

Such a neuron unit 1 can be realized, for example, by an electric circuit as shown in FIG. That is, the intensity of input and output signals is expressed by voltage, and the coupling coefficient is realized by the resistance value of the resistor 2. If the coupling coefficient is negative, it cannot be expressed by the resistance value, so the inverting amplifier 3
This is achieved by using , and inverting the output. Further, an amplifier 4 is used as an equivalent to the sigmoid function shown in FIG. For example, various methods have been proposed for realizing neuron units using electric circuits.

Problems to be Solved by the Invention However, no matter which method is used, when trying to configure a network as shown in Figure 5, the number of wires between neuron units is large. Since the lines also intersect, wiring cannot be done on the same plane, making wiring difficult.

Means for Solving the Problem In a neuron imitation circuit in which a circuit network is formed by multiple electrical circuit-like neuron units that imitate the functions of neurons, the output part of each neuron unit has a different oscillation frequency for each unit. An oscillator and a modulator for frequency modulating the oscillator output according to the output of each unit were provided, and an electric filter and a demodulator were provided at the input portion of each neuron unit.

Effect Since each neuron unit has an oscillator and a modulator in its output section, by frequency modulating the electrical signal according to the output from the unit, the deviation from the center frequency at a certain amplitude is reflected in the neuron. A signal proportional to the output from the cell unit is obtained.

At this time, each unit is modulated at a different center frequency by an oscillator with a different oscillation frequency. As a result, the outputs from each unit can be mixed and distributed to neuron units in other layers via a single line. As a result, when a network is configured, wiring between each layer requires only one electrical signal line, and wiring crossover does not occur. Since each input section is provided with an electric filter and a demodulator, the multiplexed electric signals can be frequency-divided and received and inputted without any problem. In other words, input/output processing can be performed as an electrical frequency modulation/frequency division method.

Embodiment An embodiment of the present invention will be explained with reference to FIGS. 1 to 4.

First, FIG. 1 shows one unit circuit 5, in which a neuron unit 6 having an electric circuit configuration is provided. This neuron unit 6 expresses the intensity of input and output signals as a voltage value, and the function of the unit is (1) (
2) Any material that satisfies the formula may be used. In this example, the second
As will be described later with reference to the drawings, the structure is similar to that shown in FIG. 8. A modulator 7 and an oscillator 8 are provided at the output portion of the neuron unit 6 .

As a result, the modulator 7 frequency-modulates signals of different frequencies with a constant amplitude from the oscillator 8 in proportion to the output of the neuron unit 6. At this time, the center frequency of the oscillator 8 is set to be different for each neuron unit 6. After that, a bandpass filter 9 is provided to prevent interference with other center frequencies. The output part of each neuron unit 6 is configured like this,
The signals created by each unit are the electric signal line 10 of the book.
The signals are collected, multiplexed, and distributed to each neuron unit 6 in the next layer. As a result, when a network is configured as described later with reference to FIG. 3, the wiring between each layer becomes one electric signal line 1O, and the wiring becomes easy without crossing.

Corresponding to the configuration of the output section, the input section of the neuron unit 6 is provided with a plurality of bandpass filters 11 as electric filters and demodulators 12 corresponding to the number of inputs. The bandpass filter 11 used has a different pass frequency range for each input. This results in
The multiplexed signal is decomposed into each frequency component by each bandpass filter 11, and demodulated into a voltage value to be input into the neuron unit 6 by a demodulator 12.

FIGS. 2 and 4 show more specific configuration examples. FIG. 2 shows an example of the electrical circuit configuration of the neuron unit 6, including a resistor 13 provided for each input and representing a coupling coefficient for the input voltage value, and an inverting amplifier for when the coupling coefficient becomes negative. 14 (provided only when necessary) and one amplifier 15 for realizing a nonlinear function, which is a sigmoid function, for these signals. The amplifier 15 utilizes the saturation characteristics of commercially available amplifiers.

FIG. 4 conceptually shows a case where the unit circuit 5 configured in FIG. 2 is formed into a network.

This is an example of a three-layer hierarchical configuration indicated by A, B, and C.
For example, the number of units in the first layer A is 35, the number of units in the second layer B is 15, and the number of units in the third layer 0 is 10.
It is considered as an individual.

First, in the first A layer, since an unmodulated input signal is used, the input signal corresponds to the bandpass filter 11 and demodulator 12 in the input part of the unit circuit 5 shown in FIG. The one with the part omitted is used. In addition, the frequency modulation by the modulator 7 of the output part has a frequency range of 80 MHz to 25'OMHz with a center frequency of 5.
35 types of 80MHz, 8 MHz are available.
5 MI (Z, ~, 250 MHz), which is set so that each neuron unit 6 can obtain an output with a different center frequency.

Next, in the second B layer, since there are 35 types of frequencies of signals input from each neuron unit 6 of the first A layer, the bandpass filter 11 attached to the input section has all of the above 35 types. is attached to each neuron unit 6. Since the bandpass filter 9 in the output section has 15 units, 15 frequencies with as wide a frequency interval as possible are selected from the 35 frequencies mentioned above, and filters with corresponding frequencies are installed individually. ing. For example, 35 types of frequencies 80M
Hz, 85MHz, ~, 250MH2, the B layer output part is 80M) Iz, 90MHz, ~, 250MH2
There are different frequencies like z.

In the third layer 0, the input part of each neuron unit 6 has 1 frequency corresponding to the 15 types of frequencies in the output part of the 8th layer.
Five bandpass filters 11 are each attached. Regarding the output part, since the output is an unmodulated electrical signal, in the unit circuit 5 shown in FIG.
An output section in which parts corresponding to the modulator 7, oscillator 8, and bandpass filter 9 are omitted is used.

An example in which a circuit with such a network configuration is applied to character recognition will be explained. First, read handwritten numbers from "0" to "9" with a scanner, divide them into 5x11 meshes, define the meshes with lines as "rl", and the meshes without 4g as "rQJ", and use this as the first AFJ. It was input to each neuron (neuron unit 6). 1 of the third 0 layer
0 neurons correspond to numbers from "0" to [9], and the recognition result is that the output of the neuron corresponding to that number is "1".
'', and other neurons are trained in advance on a computer simulation using the pack propagation method so that their output is 0, and the coupling coefficient determined by this is prepared on the circuit as the resistance value of each resistor 13. did. Also,
For those with negative coupling coefficients, the inverting amplifier 14
Also installed.

Finally, as a result of operating this network, a recognition rate of 100% was obtained for characters that had been sufficiently learned.

Effects of the Invention Since the present invention is configured as described above, communication between each layer of the neuron unit can be performed by multiplex communication by frequency modulation and frequency division using electricity, and the number of transmission wires can be expressed as an I-tree. , it is possible to eliminate cross-wiring and create a network configuration with easy wiring.

[Brief explanation of drawings]

1 to 4 show an embodiment of the present invention,
Figure 1 is a configuration diagram of one unit, Figure 2 is a specific circuit diagram of one unit, Figure 3 is a conceptual diagram showing its network configuration, and Figure 4 is a conceptual diagram showing a more specific network configuration. , Fig. 5 is a conceptual diagram of a neural network showing a conventional example, Fig. 6 is a conceptual diagram showing the configuration of one unit, Fig. 7 is a graph showing a sigmoid function, and Fig. 8 is a conceptual diagram showing the configuration of one unit.
The figure is a specific circuit diagram of one conventional unit. 6... Nerve cell unit, 7... Modulator, 8...
Oscillator, 11...Electrical filter, 12...Demodulation type Applicant Rico Co., Ltd. [F] Figure 05 %60'' ○Figure

Claims (1)

[Claims] In a neuron imitation circuit in which a circuit network is formed by multiple neuron units that mimic the functions of a neuron, each neuron unit has an oscillator with a different oscillation frequency and the output of this oscillator in the output section of each neuron unit. 1. A neuron imitation circuit comprising: a modulator that performs frequency modulation according to the output of each unit; and an electrical filter and a demodulator at the input portion of each neuron unit.