JPH03260785A - Signal processing method and its network - Google Patents

Abstract

PURPOSE:To improve the function of a neural cell simulation unit by processing a signal by digital logic and enabling a coupling coefficient to be read and written from outside. CONSTITUTION:An input and an output signal, an intermediate signal, a tutor signal, etc., are all represented as binary pulse trains of 0 and 1, and the quantities of the signals are represented as pulse density. Calculations in a neural cell unit are carried out by logical arithmetic among the pulse trains, the pulse train of the coupling coefficient is placed on a memory, and learning is attained by rewriting the pulse train. Further, a function for accessing the contents of a memory stored with the coupling coefficient and its sign from outside is added. Consequently, data which is already learnt can be written, data which is learnt by hardware can be taken out and stored, and the learning process can be monitored from outside to improve the convenience.

Description

[Detailed description of the invention]

The present invention relates to a signal processing circuit method and its circuit network, and more particularly, to a neural computer imitating a neural network, and is applicable to, for example, character and figure recognition, motor control of robots, and associative memory. The present invention relates to a preferred signal processing method and a circuitry for implementing the method. By imitating the functions of nerve cells (neurons), which are the basic unit of information processing in living organisms, and by creating a network of these "neuron mimicking elements", we aimed to parallelize information processing. It is a neural network. Many things, such as character recognition, associative memory, and motor control, are easily accomplished in living organisms, but are difficult to achieve with conventional Neumann computers. Many attempts are being made to solve these problems by imitating the nervous system of living organisms, especially functions unique to living organisms, such as parallel processing and self-learning. However, many of these attempts have been carried out using computer simulations, and parallel processing is required to achieve the original function, and for this purpose it is necessary to implement neural networks in hardware. Although some attempts have already been made to turn it into hardware, the self-learning function, which is one of the characteristics of neural networks, has not been realized, which has been a major bottleneck. Additionally, most of them are implemented using analog circuits, which pose problems in terms of operation. FIG. 9 is a diagram for explaining a conventional neural network model. In the figure, A represents one neuron unit, and FIG. 10 shows it configured into a network. A2 and A3 each represent a neuron unit. A single neuron connects with many other neuronal units, receives signals, processes them, and produces an output. In the case of FIG. 10, the network is hierarchical, and unit A2 receives a signal from unit hA1 in the previous layer,
Output to unit A3 of the next layer. First, to explain neuron unit A in FIG. 9, the coupling coefficient T represents the degree of coupling between other neuron units and its own neuron unit, and is 1. The coupling coefficient between the th unit and the jth unit is generally denoted by TI4. There are two types of connections: excitatory connections, in which the larger the signal from the other neuron, the greater the own output; and conversely, inhibitory connections, in which the larger the other neuron's signal, the smaller the own output. O represents an excitatory connection, and T I J < O represents an inhibitory connection. If you are the jth unit and the input from the jth unit is y, then TiJ is multiplied by 'rIJy. is the input to your unit. As mentioned above, each unit is connected to many other units, so the sum of Tta3'1 for those units, that is, ΣTl, yl, becomes the input to its own unit. This is called the internal potential and is expressed as uj. uj=ΣTtj3/+ (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 below is used as a nonlinear function. f(x)=
(2) 1+e Figure 11 shows the above sigmoid function. When connecting the above units to create a network,
The final output is obtained by giving each coupling coefficient and calculating equations (1-2) one after another. FIG. 12 is a diagram showing an example of the above-mentioned network realized by an electric circuit (Japanese Patent Application Laid-Open No. 62-295188). Coupling coefficient TIJ between units
The value of is realized by the resistance value. That is, the 12th
In the figure, the plurality of amplifiers 53 have an inverting output 53a and a non-inverting output 52b, and the input of each amplifier 53 has means 52 for supplying an input current, and a preselected first an interconnection matrix 51 connecting the output of each of the amplifiers to the input with a conductance (T I J ) that is the value of or a second preselected value. said Tij represents the transconductance between the output of the j-th amplifier and the input to the j-th amplifier, said T, J conductances being chosen to produce a plurality of local minima on which the network is balanced; I am trying to minimize an energy function that has multiple local minima. Coupling coefficient T
In the case of LJ, a negative resistance value cannot be achieved, so the amplifier 5
This is achieved by inverting the output using 3. Further, an amplifier 53 is used as a function corresponding to the sigmoid function shown in FIG. However, these circuits have the following problems. ■When signal strength is expressed as an analog value such as electric potential or current, and internal calculations are also performed in an analog manner, the value changes due to temperature characteristics, drift immediately after power-on, etc. ■Since it is a network, a large number of elements are required, but it is difficult to match the characteristics of each element. ■When the accuracy or stability of a single element becomes a problem, new problems may occur when it is made into a network, and the behavior of the entire network cannot be predicted. - Since T and j are fixed, there is no way to use values learned in advance by other methods such as simulation, and self-learning is not possible. On the other hand, as a learning law used in numerical calculations, there is the following one called pack propagation. First, each coupling coefficient is given randomly at first, but when input is given in this state, the output result is not necessarily desirable. For example, in the case of character recognition, if a handwritten character "1" is given, the output result will be "This character is "1"...! Although it is a desirable result to obtain l, this is not necessarily a desirable result if the coupling coefficients are random. Therefore, a correct answer (teacher signal) is given to this network, and each coupling coefficient is changed so that the correct answer is obtained when the same input is received again. At this time, the algorithm for determining the amount by which the coupling coefficient is changed is called pack propagation. First, let y be the output of the j-th neuron in the final layer,
d the teacher signal for that neuron. Then, E=Σ(ap-y,+) 2 (3
) is used to change Ti3 so that E expressed by is minimized. More specifically, when calculating the coupling coefficient between the output layer and the layer immediately preceding it. δJ= (dJ)#)X f'(ua)
Find δ (error signal) using (5) (however,
f' is the first derivative of f). When calculating the coupling coefficient between layers further before that, δ3=ΣδITljxf'(ut) (6
), find δ (error signal), and use
7), and change TL. Here, η is a learning constant, and α is a stabilization constant. Each of these is something that cannot be determined theoretically, but must be determined empirically. By repeating this learning many times, T is determined so that a desired result can be obtained for a given input. Now, if this learning method were to be implemented in hardware in some way, the learning would require a large amount of four arithmetic operations, which would be difficult to implement. The learning method itself is also unsuitable for hardware implementation. 13 to 15 are diagrams showing an example of realizing such a neural network with a digital circuit. FIG. 13 is a diagram showing an example of a circuit configuration of a single neuron, 60 is a synaptic circuit, 61 indicates a dendritic circuit, and 62 indicates a cell body circuit. FIG. 14 shows a configuration example of the synaptic circuit 60 shown in FIG. 13, and FIG. 15 shows a configuration example of the cell body circuit 62 shown in FIG.
In FIG. 14, f is an input signal, W is a weighting value, and a is a magnification (1 or 2) to be multiplied by Feet Rose No. 94n. This represents the input and output of a neuron unit as a pulse train, and the pulse density represents the amount of signal. The coupling coefficient is expressed as a binary number and stored in memory. By inputting the input signal to the rate multiplier clock and inputting the coupling coefficient to the rate value, the pulse density of the input signal is reduced in response to the rate value. This corresponds to the T and y parts of the equation of the BP (pack propagation) model. Next, the Σ portion of ΣT, , yL is realized by an OR circuit represented by a dendrite circuit 61. Since connections have excitatory and inhibitory properties, they are divided into groups in advance and marked with ○R for each group. 13th
In the figure, Fo indicates excitatory output and F2 indicates inhibitory output. An output is obtained by inputting these two outputs to the up side and down side of the counter 62 shown in FIG. 15 and counting them. Since this output is a binary number, it is converted into a pulse density using a rate multiplier again. A neural network can be realized by forming a network of these neuron units. Regarding learning,
This is achieved by inputting the final output into an external computer, performing numerical calculations inside the computer, and writing the results into the coupling coefficient memory. Therefore, there is no self-learning function at all. Furthermore, the circuit configuration is complicated, as the pulse density signal is first converted into a numerical value using a counter, and then converted back into the pulse density. As mentioned above, in the conventional technology, (1) Analog circuits do not operate reliably. ■Learning methods based on numerical calculations also require complicated calculations and are not suitable for hardware implementation. ■Digital circuits have complex circuit configurations. ■Self-learning is not possible on hardware. There were drawbacks such as. The present invention was made in view of the above-mentioned circumstances, and
In particular, we will enable the adoption of digital circuits with reliable operation, provide simple signal processing and learning methods that can be easily implemented in hardware, and aim to actually implement self-learning on hardware. It was done for a purpose. SUMMARY OF THE INVENTION In order to achieve the above object, the present invention has a coupling coefficient variable means and a coupling coefficient generating means that generates a variable coupling coefficient value based on an error signal with respect to a teacher signal and a variable coupling coefficient value of the coupling coefficient variable means. A learning means is formed, this self-learning means is attached to a neuron imitation element to form a neuron imitation unit, and these self-learning means and the neuron imitation element are formed in a network to form a plurality of neuron imitation means. In the signal processing method for processing signals using digital logic, the coupling coefficients can be read and written externally, or a signal processing circuit that implements the above method in hardware, that is, a variable coupling coefficient circuit; A self-learning circuit is configured with a coupling coefficient generation circuit that generates a variable coupling coefficient value of the coupling coefficient variable circuit based on an error signal with respect to a teacher signal, and this self-learning circuit is attached to a neuron imitation element to form a neuron imitation circuit. In a signal processing circuit configured to constitute a plurality of neuron imitation circuit networks by connecting these self-learning circuits and neuron imitation elements in a network using digital logic circuits, a circuit capable of reading and writing the coupling coefficients from the outside. It is characterized by having the following. Hereinafter, the present invention will be explained based on examples. First, to explain the basic idea of the present invention,
The basic concept of the present invention is as follows: (1) Input/output, intermediate signals, coupling coefficients, teacher signals, etc. are represented by a pulse train expressed as a binary value of 0.1. ■The amount of signal is expressed by pulse density ([
1”). ■Calculations within a neuron unit are expressed by logical operations between pulse trains. ■Place the coupling coefficient pulse train in memory. ■Learning is achieved by rewriting this pulse train. (2) Regarding learning, calculate the error based on the given teacher signal pulse train, and change the coupling coefficient pulse train based on this. At this time, the calculation of the error and the calculation of the change in the coupling coefficient are all performed by logic operations using a pulse train of 0.1. There is a particular thing. Examples that embody the above idea of the present invention will be described below. [Signal calculation] Figure 1 is a diagram showing a portion corresponding to one nerve cell,
To create a network using this, the same hierarchical structure as before is used, as shown in FIG. Input/output is done 1.
The data that has been binarized to 0 and further synchronized is used. The intensity of the input signal yI is expressed as a pulse density, for example, the number of states of 1 within a certain period of time, and is expressed as the following (8)
It is expressed as the formula. Input signal L--1"-"-=4/G (8
) Sync signal 1"-Up"-Sat"- This shows the signal representing 4/6, and the sync pulse 6
Among the signals, there are 4 1 signals and 2 0 signals. At this time, it is desirable that the 1's and O's are arranged randomly, as will be described later. On the other hand, the coupling coefficient TIJ is similarly expressed in terms of pulse density, expressed as a bit string of O's and 1's as shown in the following equation (9), and prepared in advance on the memory. Coupling coefficient = 3/6 (9) Synchronization signal L''-upper''-earth''- This represents rlololoJ = 3/6, and in this case as well, it is preferable that the arrangement of 0's and 1's be random. The specific method for determining this will be described later. Then, this bit string is read out sequentially from the memory according to the synchronous clock,
The AND circuit shown in FIG. 1 performs AND with the input pulse train (ytnTtJ). This is the input to neuron j. (8), (9) above
To explain using an example of a formula, the input signal is rlollol
When J is input, a bit string is called up from memory in synchronization with this, and AND is performed sequentially. Input signal L-1m = 4/6 Coupling coefficient = 3/6 (10) 3/ t nT t b = 2 /
6 is obtained, which indicates that the input y is converted by TIJ and the pulse density becomes 276. The pulse density of the output of the AND circuit is approximately the product of the pulse density of the input signal and the pulse density of the coupling coefficient, and has the same function as the coupling coefficient of the analog method. This function becomes closer to a product as the string of signals becomes longer and the arrangement of 1's and 0's becomes more random. Incidentally, if the pulse train of the coupling coefficient is shorter than the input pulse train and there is no more data to be read, it is sufficient to return to the beginning of the data and repeat the reading. Since one neuron unit has multiple inputs, there are many ANDJs between input signals and coupling coefficients, as explained earlier, and then OR these.Since the inputs are synchronized,
If the first data is "1o1000J, 2nd note' - evening is roloooo" (7), the OR of both is rllll
It becomes oooJ. This is calculated simultaneously for multiple inputs as shown in equation (11) and output. t, n1'0, L-nee 1111:j + n T t, -nee 1111 (11) U (y L n TIj) LLL-111 This part is the calculation of the sum in analog calculation and a nonlinear function (sigmoid function). When the pulse density is low, the ORed pulse density approximately matches the sum of the respective pulse densities. As the pulse density increases. Since the output of the OR gradually becomes saturated, the result does not match the sum of the pulse densities, and nonlinearity appears. In the case of OR, the pulse density is never greater than 1, never less than O, and is a monotonically increasing function, approximately similar to a sigmoid function. Now, coupling has excitatory and inhibitory properties, and in the case of numerical calculations, it is expressed by the sign of the coupling coefficient, and in the case of analog circuits, as mentioned earlier, when TIJ is negative (inhibitory coupling)
In this case, an amplifier is used to invert the output and connect it to other neurons with a resistance value corresponding to TIJ. On the other hand, in the present invention, each connection is first divided into two groups, excitatory connections and inhibitory connections, depending on the sign of T and j, and then, ``OR of the ANDJ of the pulse train of the input signal and the connection coefficient is Calculate separately. Then, when the output of the excitatory group is 1 and the output of the inhibitory group is O, the output is set to 1, and when the output of the excitatory group is O and the output of the inhibitory group is 1, the output is set to 0. .When both are 1 or when both are O, the output can be O or 1, and the probability is 1.
/2 and the output may be set to 1. For example, the output is output only when the excitatory group's output is 1 and the inhibitory group's output is O, or the excitatory group's output is O and the inhibitory group's output is output. In order to realize this function of outputting an output when (15)), in the latter case (N0T of the output of the inhibitory group) and the output of the excitatory group may be ORed (Equations (16) to (19)). Excitatory group output LL--LL- Inhibitory group output -LL--L- (12) a” U (Y t n T t J) (T = excitatory) (13) iE excitatory b”U (yinTt, +) (T=inhibitory) iC inhibitory (14) y4=anb (15) Excitatory group output 11--1''-inhibitory group output -LL-one-(16) a = U (3/1nTtj) (T==excitability
(17) iE excitability b=LJ (': I t n T t s ) (T
= inhibitory) (18) iE inhibitory y j = a U b
(19) The network of neuronal units is BP
Hierarchical type similar to (pack parogation) (Figure 10)
shall be. If the entire network is synchronized, each layer can be calculated using the functions described above.

[Learning] ■Error signal in the final layer First, calculate the error signal in each neuron in the final layer. In the present invention, the error signal is defined as follows. When error is expressed numerically, it can generally take both + and -, but since pulse density cannot express both positive and negative at the same time, two signals are used: one representing the ten commandments and the other representing the - component. to express the error signal. δ', E (yJ XORd,) AND d. Error signal -' (23) δ゛, = (ya XORd,) AND y. In other words, the error signal 0 is a pulse that exists on the teacher signal side among the different parts (1, 0, or 051) between the teacher signal pulse and the output pulse, and the error signal - is a pulse that similarly exists on the output side. That is, by adding ten error signal pulses to the pressure pulse and removing the error signal minus pulse, the teacher signal pulse is obtained. The coupling coefficient is changed based on this error signal pulse, which will be described later. (2) Error signal in intermediate layer Furthermore, the error signal is back-propagated, and not only the coupling coefficient between the final layer and the layer immediately before it but also the coupling coefficient of R9 before it changes. Therefore, it is necessary to calculate the error signal at each neuron in the hidden layer. In the same way as propagating a signal from a neuron in the middle layer to each neuron in the next layer, we collect the error signals in each neuron in the next layer and calculate our own error. Signal. This can be done in the same manner as the calculations in equations (8) to (19) within the neuron unit. That is, first, depending on whether the connection is excitatory or inhibitory,
Divided into two groups, the multiplication part is expressed by AND, and the Σ part is expressed by OR. However, the difference from equations (8) to (19) within a neuron unit is that y is one signal, whereas δ has two signals representing positive and negative, and both signals need to be considered. Therefore, it is necessary to differentiate into four cases: positive/negative of T and positive/negative of δ. First, to explain the case of excitatory connection, the error signal +(δ°,) at the th neuron in the next layer and the coupling coefficient between that neuron and itself (T, +k) (7
) A N D is obtained (δ'mnTti) for each neuron, and the OR of these is calculated (U (δ'hnTti)). This is then taken as the error signal of this layer. By taking the AND of the next error signal and the coupling coefficient, and then ORing these two, , Similarly, let the error signal of this layer be -. δ-□nTJ□ δ-h n T J, knee-1- (25) δ-1-bit''-Next, to explain the case of inhibitory coupling, the AND of the error signal of one ahead and the coupling coefficient and then O between these
Take R. δ-0nT,, -LL-L-"- δ-* n T J k-LL-"- Also, AND is taken with the next error signal 0 and the coupling coefficient, and then OR is taken between them. This is then used as the error signal for this layer. δ+ , nT, 1 δ" hn T Jk Since some neurons connect to another neuron through excitatory connections and others through inhibitory connections, we calculated it using equation (24). The OR of δ" and δ" obtained using equation (26) is taken, and this neuron is set to 61.Similarly, the OR of δ- obtained using equation (25) and δ- obtained using equation (27) is taken. , is the δ of this neuron.To summarize the above, δ't=(Ll(δ”knT,+k))U(IJ(δ
−hnTjk)) kE excitatory kE inhibitory δ−a=(U(δ−hnT,+m))u(u(δ” h
nTg)) kE excitatory kE inhibitory (28). Furthermore, a function corresponding to the learning rate may be provided. When the rate is less than 1 in numerical calculation, the learning ability further increases. This can be achieved by thinning out the pulse train in the pulse train calculation. I thought of this as a counter and came up with something like the following. For example, when η=0.5, every other pulse train of the original signal is thinned out, y=0.3
In 3, every second is left, and in y=0.67, every second is thinned out. Note that even if the pulses of the original signal are not equally spaced,
The original pulse train can be thinned out as shown in example (2). Example 1) When the original signal η = 0.5, when η = 0.33, η = o, when 67, Example 2) When η = 0.67 (29) In this way, the error signal is thinned out. By this, it is possible to provide the functions of learning relays 1 to 1. (2) Changing each coupling coefficient from the error signal The error signal is obtained by the method described above and each coupling coefficient is varied, which will be described below. The signal flowing through the line to which the coupling coefficient to be changed belongs and the error signal are ANDed (δny). However, in the present invention, since there are two error signals, + and -, calculations are performed for each. δ' jn 'j 1 ΔT Lj δ' 3 n '/t
ΔT°,. Let the two signals obtained in this way be ΔTi. Since Tlj in the present invention is an absolute value component, the original Tlj
Differentiate the cases depending on whether LJ is excitatory or inhibitory, and create a new T.
Find i. Increase the component of 8T'tj to the TLj of Shozamo Hakuryuen Ikumoto, and make 8T.
``Reduce the component of 9. TIJ ΔT 1j Δ'l'' I J ``-'' - 1 u - [- 11111111 - (32) 'I'rs [L north - 11-11] For the TIJ of υ屹金原, reduce the component of Δ-1-+,
Increase the component of Δ゛r-1. Tlj ΔT・Ij L-1"-m- ΔT"lj-111"-" (33) Based on the above learning side after learning, calculate the network. [Circuit configuration] Based on the above algorithm, actual circuits are shown in Figures 2 to 4.The entire network is shown in Figure 10.
Similar to the figure, the circuit of the part corresponding to line B in Fig. 10 is shown in Fig. 2, and the part corresponding to A1-A3 of Fig. 10 is shown in Fig. 3. Figure 4 shows the part for calculating the error signal at the final no from the teacher signal and the teacher signal. By making these Figures 2 through 4 into a network as shown in Figure 10, a digital method that allows self-learning can be created. It is possible to realize a neural network of
Save it in shift register 8. 8A is the outlet and 8B is the inlet. In this case, other devices may be used as long as they have the same function as the shift register, such as a RAM+address controller. Further, although not shown, a synchronization signal expressed by equation (9) is given to the shift register, address controller, etc. 9 is a circuit corresponding to equation (10), which performs an AND operation between the input signal and the coupling coefficient. This output is interesting to combine! Although it is necessary to divide the output into groups depending on whether it is IIi-like or suppressive, it is more versatile to prepare outputs 4.5 for each group in advance and switch which way to output them. For this reason, a signal indicating whether the connection is excitatory or inhibitory is stored in the memory 14 and switched by the circuit 13. The OR circuit 11 in FIG. 2 corresponds to equation (]-1) for processing each input. Furthermore, formula (12)! )l[′#i
The circuit 17 in FIG. 3 outputs an output only when the sexual croup is ↓ and the inhibitory croup is O. Also, formula (16)
This case can also be easily realized using logic circuits. Next, the error signal will be explained. FIG. 4 shows how to create the error signal in the final layer. This is equations (20) to (23)
corresponds to Error signals 22 (Gpj), (G
Make Mj). 1 in Fig. 2 has been made into a circuit using formulas (24) to (27) for calculating d1 of the error number 4B at intermediate j-.
It is 0. Cases are differentiated depending on whether the binding is excitatory or inhibitory. It is the circuit 12 that performs this switching by a bit set in the 9-warning circuit 14. Further, the part of calculation formula (28) that collects the error signal is 18 in FIG. Also, the part of equation (29) corresponding to the learning rate is the third
This is the frequency dividing circuit of 19 in the figure. This can be easily realized by using a flip-flop or the like. Finally, the part that calculates new coupling coefficients from the error signal will be explained. In the equation, it is represented by (30) to (33), and this part is 11 in FIG. Similar to the above, cases must be differentiated depending on the excitatory and inhibitory properties of the connections, and this is realized by the circuit 12 in FIG. Now, in this circuit, the coupling coefficient and its sign are stored in memory. However, the contents of the memory are lost when the power is turned off, so it is necessary to save the contents on another recording medium. Furthermore, if already learned coupling coefficients and their codes are available, writing them into memory eliminates the need for relearning from the beginning. Therefore, it would be very convenient if the contents of this memory could be freely accessed from outside. Therefore, in the present invention, a function has been added that allows external access to the contents of the memory that stores the coupling coefficients and their codes. This is shown in Figure 6.
As shown in FIG. 7, it is possible to draw out the exit (8A in FIG. 2) and the entrance (8B in FIG. 2) of the shift register 8 to the outside. Should I pull out the exit 8A as it is, or should I use the external signal 3 for the entrance 8B when reading and writing from the outside?
3, when not reading or writing, 4i1 No. 33' from inside
Since it uses (J
. Switching is performed using the selector 31 using the number 34. At this time, a clock 36 is required to shift the contents of the shift register 8, but this uses the synchronizing signal of the internal clock 35 formula (9) that operates this signal processing circuit as it is, and 35" to the outside, and the contents of the shift register 8 may be read and written externally in synchronization with this clock.
Alternatively, as shown in FIG. 7, a tarot card 37 may be applied from the outside and reading and writing may be performed in synchronization with this. However, when giving Tarokku 37 from the outside,
It is necessary to switch the clock 36 given to the shift register 8 with the internal clock 35. This can be done by using the selector 31a in the same way as switching the input signals to shift 1 to register 8. Furthermore, when a RAM+atrest coater is used instead of the shift register 8, random access from the outside becomes possible by drawing out the address bus, data bus, and control signal line for reading and writing to the outside. In the case of a memory with two or more ports, one may be used for external reading and writing, and the other for internal processing. In the case of a memory with one port, when accessing from the outside, it is sufficient to switch the address bus etc. to the line from the internal atrestecoder. This can be done by replacing the tarlocks 35, 36, and 37 in FIGS. 6 and 7 with address buses and control signal lines for reading and writing. Next, with reference to FIG. 8, a case will be described in which a plurality of memories are read and written using a common bus. Figure 8 shows only the parts necessary for explanation in Figures 6 and 7. First, a unique address is set for each memory in advance, and the address given from the outside and the own address are set in advance. Determination circuit 4 for determining whether or not a unique address matches
2, and a chip is selected based on the determination result (44). When reading and writing from the outside, by giving this address 40 from the outside, only one memory is selected, and any memory can be read and written. Also, when using a memory with only one port, do not read or write externally.
If you use selenium 1 to all memories when not in use, and also prevent data from flowing to the common bus for external read/write (43), you can avoid the inherent parallel processing of this signal processing circuit. Becomes a force exporter. In addition, although an example of the case where the present invention is implemented as a heart has been described above, the whole of these may be incorporated in one computer, or only a part thereof may be incorporated, and furthermore, each of these may be incorporated into a single computer. It is easy to understand that these may be combined with separately constructed parts having a single function to form a whole. A self-learning character recognition device using the network described above will be explained. First, handwritten characters are read with a scanner, divided into 16 x 16 meshes, and the character parts are separated as shown in Figure 5. The data (256 pieces) were input into the network, with meshes with meshes set as 1 and meshes without meshes set as O, and the recognition result was set to the position of the unit with the largest output among the five units. Therefore, when a number from "1" to "5" is input, the number corresponding to that number is learned to have the largest output. Next, the network configuration is such that the first layer is 25
The second layer was composed of 6 neuronal units, the second layer was composed of 20 neuronal units, and the third layer was composed of 5 neuronal units. If each coupling coefficient is initially set at random, the output result will not necessarily be the desired one. Using the self-learning function of this circuit, each coupling coefficient is newly determined and this is repeated several times to obtain the desired output. In this embodiment, since the input is 0 or 1, the input pulse train is always a simple LOW level or HIGH level. Further, the output was connected to an LED through a transistor, and was turned on when it was at a LOW level (4) and turned on when it was at a HIGH level. Since the synchronization clock was set to 1000 kHz, the brightness of the LED changes to the human eye depending on the pulse density, so the brightest part of the LED becomes the answer. A recognition rate of 0% was obtained for the characters that had been trained. As is clear from the above explanation, conventional neural networks rely mostly on computer simulations, and some attempts have been made to convert these networks into circuits. Serial processing is performed using computer simulation. In contrast, according to the present invention, the functions of the neural network 1-work, including the learning function, can be performed in parallel on hardware. Significant speed increase. Furthermore, since the coupling coefficient and its sign can be accessed freely from the outside, it is possible to write already learned data, extract and save data learned by hardware, or monitor the learning process from the outside. This makes it possible to improve convenience. There are advantages to having a hat.

[Brief explanation of drawings]

FIG. 1 is a diagram showing one of the neural circuit units, FIGS. 2 to 4 are diagrams showing circuit configuration examples, FIG. 5 is a diagram showing an example of self-learning character recognition, and FIG. and FIG. 7 is a diagram for explaining an example of reading and writing, and FIG. 8 is a diagram for explaining an example of reading and writing using a common bus.
FIG. 9 is a diagram showing an example of neuron units 1 to 10.
The figures are for explaining an example of a hierarchical neural network, Figure 11 is a diagram showing a sigmoid curve, and Figure 12 is a diagram showing a sigmoid curve.
1 to 15 are diagrams for explaining examples of conventional neural network circuits. 1...Input signal, 2.3.6.7...Error signal, 4
. . . Activation signal, 5. Inhibition signal, 8. Shift register, 20. Input signal, 21. Teacher signal. 22.2;3 error signal.

Claims (1)

[Claims] 1. A self-learning means is formed by a coupling coefficient variable means and a coupling coefficient generating means that generates a variable coupling coefficient value of the coupling coefficient variable means and an error signal with respect to a teacher signal. A learning means is attached to the neuron imitation element to form a neuron imitation unit, and these self-learning means and the diameter cell imitation element means are formed in a net shape to form a plurality of neuron imitation units, and a plurality of neuron imitation units are formed by digital logic. A signal processing method for processing a signal, characterized in that the coupling coefficient can be read and written from outside. 2. A self-learning circuit is configured with a variable coupling coefficient circuit and a coupling coefficient generation circuit that generates a variable coupling coefficient value of the variable coupling coefficient circuit based on an error signal with respect to the teacher signal, and this self-learning circuit is configured to imitate neurons. In the signal processing circuit network in which a neuron imitation circuit is attached to an element, and a plurality of neuron imitation circuit networks are configured by connecting these self-learning circuits and the neuron imitation element in a network form by a digital logic circuit, A signal processing circuit network characterized by having a circuit that allows coupling coefficients to be read and written from the outside.