JPH07109609B2 - Neural network time series learning method - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a learning method of a neural circuit (neural networks) element for learning the contents of an output signal output corresponding to an input signal.

[0002]

2. Description of the Related Art A conventional learning method for a neural circuit will be described. In the Bop Propagation (BP) method, the output to be output by the neural circuit with respect to the input is predetermined as a teacher signal, and the squared error between the output output from the neural circuit and the teacher signal is obtained in response to the input signal. The coupling coefficient of the neural circuit is adjusted so that this squared error is reduced. This method requires complicated calculation and takes a long time for learning. Furthermore, since the theory is based on fixed time, it is not possible to learn time series. Therefore, in order to handle time-series data, a recursive coupling type neural circuit with a feedback part of the circuit is used for information processing. Adapting the BP method for the recursive coupling circuit to learn results in a more complicated circuit, so a practical theory has not been completed.

A Hebb learning method is known as another learning method. This method enhances the coupling coefficient of the inputs involved when there is a large output. Although this method is very simple, there is no guarantee that efficient and correct learning can be done, and time is fixed unless a special circuit for recursive coupling is added because the time is fixed. The point that it cannot be performed is the same as the above-mentioned BP method. Further, in the case of the recursive coupling type learning, since the read output is reproduced at the same timing, there is a drawback that the time interval of input is restricted.

Time-series information is important for living things, and part of it has been known as conditioning learning for a long time. That is, if the electric shock input is repeated several times after the buzzer stimulus input, the preceding action that predicts the next electric shock is taken only by the buzzer input. Many attempts have been made to explain the mechanism of this learning, but none are still satisfactory.

For example, the method conventionally proposed as a synapse model of physiological nerve cells related to time series is the following method. In other words, the learning method in which the coupling coefficient changes depending on the potential of the receiving side that the coupling coefficient of the subsequent input is enhanced when the next input exists where the internal potential of the receiving side rises due to the previous signal input. Is.
Looking at this in FIG. 5, it is assumed that the input 1 produces a large output of A and then the input 2 produces a large output of B.
At this time, the coupling coefficient is enhanced by the coupling from B to A input when the potential rises. When such learning is performed, the learning is associated with the input 2 to the input 1, and the reinforcement learning reverses in time. As time-series learning, it is necessary to be able to perform a coupling that associates the input 1 with the input 2 that comes next, that is, the learning that enhances the coupling coefficient from A to B, but this method has not been found until now.

[0006]

As described above, there has been no learning / memory method for time-series information processing that can sufficiently cope with the insufficient performance of the neural circuit element itself. Therefore, the present invention configures (a) a circuit for correctly learning and storing the context of time of an event.

(B) The reading and reproduction can be performed at high speed regardless of the event occurrence time interval.

(C) Even with a complicated circuit, simple calculation can be performed with a simple structure and saving memory.

(D) Self-organizing a balanced circuit that has coupling coefficient enhancement and attenuation for learning.

(E) The storage capacity is improved by using one element element for a plurality of meanings.

It is an object of the present invention to provide a learning method of a neural circuit capable of doing the above.

[0012]

In order to achieve such an object, the invention of claim 1 holds a time history of inputs for each input in a neural circuit element having a large number of inputs, and the time history of the inputs is maintained. And the internal state of the neural circuit element are compared, and based on the comparison result, learning is performed to strengthen or attenuate the coupling coefficient or the transmission efficiency for each input according to the time history, and before and after the input and output times. The relationship is stored in the coupling coefficient or the transmission efficiency.

According to the invention of claim 2, a plurality of neural circuit elements according to claim 1 are combined in order to enable reading and reproduction faster than an actual occurrence time interval, so that the sequence of events occurring according to the flow of time can be changed. It is characterized by executing a proactive learning that jumps over.

The invention of claim 3 has a plurality of meanings for the neural circuit elements used as the elements in the combination of neural circuit elements of claim 2 for enabling crossing or interference of time-series associative memory elements. And increase the memory capacity of the entire neural circuit.

[0015]

According to the present invention, the history value of each input of the neural circuit element is held and the coupling coefficient is enhanced when the input history value is large when the receiving side makes a large response, and attenuated when the history value is relatively small. This is a new learning method for neural circuit elements that depends on the history value of the input. How this learning method stores the correct temporal order is explained in FIG. When the A element outputs a large output by the input 1, a large history value is held in the coupling portion of the input of the B element from the A element. Next, when the B element produces a large output by the input 2, the influence is transmitted to each input coupling portion and the coupling coefficient having a large hysteresis value is strengthened. In this way, since the coupling coefficient from A to B is strengthened, correct time series learning can be performed.

According to the first aspect of the invention, the relationship between the input and the output is learned in time series depending on the history value of the input, and
This learning calculation can be easily learned only by the relationship between the input and output of one element.

The input history values of the elements are not required to hold all the input history values (N × N pieces) for each N pieces of elements, but the values held as the output history values of each element (N pieces) ) Can be commonly used for the calculation as the history value of the next input, and the memory can be saved.

According to the second aspect of the present invention, the learning is advanced by self-organization by increasing and decreasing the coupling coefficient, and the memory of the time series is stored in the coupling coefficient. Therefore, the reproduction and reading are related to the learning time. It can be played at high speed without the need for advanced prediction playback.

According to the invention of claim 3, when the time series of two event inputs is expanded to the relationship with the inputs 3 and 4 ahead of them, the learning becomes higher dimensional time series, and one element can realize a plurality of meanings. You can realize the work you have.

[0020]

Embodiments of the present invention will now be described in detail with reference to the drawings.

The element circuits or elements handled here can be realized by using analog ones for input and output, but in order to make the features easy to understand, digital pulse waveforms (binary) are integrated, An explanation will be given by using a neural circuit element which internally performs an analog operation and fires when the internal potential exceeds a threshold value to output a pulse output. This is because this element can handle time information and can perform a synchronized calculation by setting a certain time width, which is convenient for simulating a neural circuit with a computer. This element is proposed as a "neural circuit element" in Japanese Patent Application No. 4-254189 (filed on August 28, 1992). In this embodiment, the neural circuit element will be described first.

The relationship between the input and output of the neural circuit element is shown in FIG. Multiple input pulses are integrated into the neural circuit element through the coupling. The decay value aV of the internal potential is inherited as the residual value of the present invention from before and is added to the integrated input to become the final integrated value U. When the integrated value U is larger than the threshold value T, a bit value "H" is output and the integrated value U minus a constant p becomes the internal potential V.

When the integrated value U is smaller than the threshold value T, the output has the bit value "L", and the integrated value U is directly taken over as the internal potential (residual value of the present invention) V. The input / output bit values “H” and “L” are binary logic values, which are “1” and “0”, or “1” and “−1”, respectively. This will be described using 1 ”and“ 0 ”.

The above operation can be expressed as the following expression by an operation processing expression for each discrete time.

[0025]

[Equation 1]

Only when the input pulse is the bit "1", the respective coupling coefficients W are added to form the first term on the right side of the equation (3). The second term on the right side shows the residual internal potential from the previous time, and indicates that V decreases with time at a constant a (0 <a <1). The output Xi at this time is determined as follows by the integrated value Ui and the threshold value T.

[0027]

## EQU00002 ## If Ui (t) .gtoreq.T, Xi (t) = 1 and Vi (t) =
If Ui (t) -p Ui (t) <T, then Xi (t) = 0 and Vi (t) = Ui (t) where p is a constant and approximately equal to T.

The above arithmetic processing is repeatedly executed in parallel for each neural circuit element. Information processing is performed by combining such neural circuit elements to form a neural network (neural network). An example of realizing the neural circuit element of FIG. 6 with a digital circuit is shown in the above patent.

Next, adding the time series learning method of this patent to this neural circuit element will be described. As shown in FIG. 1, the history value Hji (t) of the input of the neural circuit element is added with a pulse value every time there is an input from j by using an integrator having leakage, and as shown in FIG. It shall decay over time. Although the neural circuit element i does not always output the ignition output by the input j, it outputs the ignition output when the input is large and the internal potential Vi (t) is larger than the threshold value.

If the history value Hji (t) at that time is larger than the learning level G1 as shown in FIG. 3, the coupling coefficient Wjj from j (also called the transmission efficiency) is increased to be between the learning levels G1 and G2. The neural network element learns so that Wji is attenuated when the value is. The relationship between the passage of time and the change in the coupling coefficient is summarized from FIG. 2 and FIG. 3 as shown in FIG. That is, it is shown that when the time difference from the input to the firing is small, the coupling is enhanced as a meaningful input path, and when there is a time difference to some extent, the learning is performed as the path having a weak relationship and the coupling is attenuated.

With respect to the history value Hji (t), if the number of elements in the entire neural network is N, the history value of the input must be calculated and held for each neural circuit element on the receiving side. It appears that it requires N × N multiple memories for (t). However, the number of memories can be N by calculating the history value Hj (t) of the output of each element and using it in common during learning of each element.

As a calculation formula of the history value H (t),

[0033]

## EQU3 ## Hj (t) = d.Hj (t-1) + q.Xj (t) is used. Where d is the decay time constant of H (d <1), q
Is the history value constant of pulse output. The output of the neural circuit is
There is output (Xj (t) = 1) or no output (Xj (t) =
Since it is 0), it means that q is added only at the ignition output. As a simpler method, if H decays differentially,

[0034]

## EQU00004 ## Hj (t) = Hj (t-1) -do + q.Xj (t) where do is the damping constant for each time. This allows the history value to be calculated without using multiplication, and the change with time of the value becomes as shown by the dotted line in FIG. In addition, Hj
A simpler method for calculating the history value can be obtained by resetting (t) = q and using only the subsequent attenuation of the history value.

Regarding the learning of the coupling coefficient W, in the usual case, the coupling coefficient W is learned between the maximum Wmax and the minimum Wmin of. As shown in FIG. 3, the variation ΔW of the coupling coefficient W
Is set to a threshold value G1 that increases and a threshold value G2 that decreases. H now
When j (t) is larger than G1, the coefficient is enhanced,

[0036]

[Expression 5] ΔWji = k1 · (Wmax-Wji) · (Hj (t) -G1) When Hj (t) is a value between G1 and G2, the coefficient is attenuated.

[0037]

[Equation 6] ΔWji = k2 · (Wmin −Wji) · (Hj (t) −G1) · (Hj (t) −G2) is calculated. However, k1 and k2 are learning rate constants and are shown in FIG. If the increase / decrease function of the coupling coefficient is simplified as shown by the dotted line in FIG. 3, the equations related to the time series learning in Equations 5 and 6 are the following simple calculation equations,

[0038]

[Formula 7] ΔWji = k · (Wm−Wji) However, if Hj (t)> G1, then k = k1,
If Wm = Wmax and G1> Hj (t)> G2, then k
= K2, Wm = Wmin.

Finally, the time interval from the input to the firing of the neural circuit element and the increase / decrease relationship due to the learning of the coupling coefficient are summarized as shown in FIG. That is, in the learning method of the present embodiment, the coupling coefficient is enhanced when the neural network element on the receiving side has an ignition output at the time of input or within a short time, and the neural circuit element is activated after a while after the input. It is characterized in that the coupling coefficient is attenuated when ignited. The learning levels G1 and G2 shown above are related to the decay time constant d of the history, and determine how long the effect of the input lasts. When the learning level changes, the increase / decrease time interval width changes. For example, by decreasing the learning level, the influence of the time series can be extended to the forward second and third order, or the learning level can be increased to limit the relationship to only the first order for a short time. You can adjust the learning method with. The relationship of this coupling coefficient is shown in FIG.

Actually, on a computer, a neural network having neural circuit elements (D, E, R, S ...) As shown in FIG. 7 is mutually connected, and an input is given to the neural network, and the mathematical expressions 1 and After repeating the calculation of the response by 2 and the calculation of the time series learning by the formulas 3 and 7, it was confirmed how the time series associative memory could be read.

The 20 neural circuit elements used in the calculation have the decay time constant a of the internal potential a = 0.7, the threshold T = 1000, and the firing decay constant P = 1000, and the initial value of the mutual coupling coefficient Wji is set. The value is close to zero. The maximum coupling coefficient after learning is Wmax = 500, and the minimum coupling coefficient is Wmi.
It was set to n = -50. The addition constant when outputting history values is q = 1
000, damping time constant d = 0.6, learning level G1 =
The calculation was performed by setting 200 and G2 = 20.

For learning, a time-series pattern, for example, FIG. 9 is used.
Indicated by the symbols inside are (L-I-O-N) and (T-I-G).
-E-R), (D-E-S-K) and other combinations of patterns are input one by one to perform time-series learning in which the coupling coefficient is changed. The learning method is as follows. First, the neural circuit element corresponding to the start is fired, then the neural circuit element of code L is fired, the neural circuit element of code I, the neural circuit element of code O, and the neural circuit of code N. The elements were fired in order and the coupling coefficient change was learned at each time point. Learning rate constant k
When 1 = 0.05 and k2 = 0.01, about 80 times of learning was sufficient for learning.

After the learning of each pattern is completed,
The time series was read out. For reading, the neural circuit element of the start and one neural circuit element located at the beginning of the pattern such as L, T, and D are fired, and then the neural circuit element following it is, for example, (L-I-O-N). It has been proved that time series can be correctly associated by firing one after another. Looking at the internal structure of the coupling coefficient, the connection of the neural circuit element of the start and the elements of the symbols L, T, D, the coupling of the code L to the code I, the code I to the code O, and the code O to the code N are shown. Thus, it was confirmed that the binding was enhanced in the forward order. Unlike the conventional recursive coupling associative circuit, this reading is not related to the elapsed time during learning, but is related to only the strength of the input and the coupling coefficient, so that high-speed reading is possible.

Considering human memory, it is easy to read the stored contents in the forward direction, but it is difficult to read the stored contents in the backward direction, and the time interval between events is uncertain. The context of is very similar to what you can remember clearly regardless of time scale. From this, this time series learning method is considered to be a more natural method.

The above description is an example for learning only the relationship between an event and the next (primary) event (claim 1).
However, in actual association, associative learning that is related to the next (secondary) event is common. In this time series learning method, by adjusting the learning level, the
It is possible to easily learn the high-order relationship of time series association with secondary and tertiary events by expanding the related time width of. actually,
The parameters used in the above description have a relation that associates up to a third-order event, and the coupling coefficient can form a coupling relationship as shown in FIG. Since the higher-order time series relationship can be learned in this way, even when the O element in the middle of (LION) is missing, it is possible to read O skipping from L → I to N, which is very It was able to have the property of flexible associative memory. This corresponds to the invention of claim 2 as well as the high-speed reading can be performed regardless of the learning timing.

Further, by the high-order learning method shown in FIG.
What features were realized by forming a third-order time series relationship in the neural network is shown below. First, the patterns (L-I-O-N) and (T-I-G) of FIG.
Even if there is a common element (neural circuit element) included in different series like the neural circuit element with the code I in (E-R), if it starts with the neural circuit element with the code T (T-I-G) -E
-R) pattern, starting with the code L (L-I-O-
It was possible to read out crossing like the pattern of N). The neural circuit element with the code E of the patterns (D-E-S-K) and (T-I-G-E-R) could be reproduced in the same manner. Further, by placing the start element, it was possible to correctly reproduce even a series starting from the middle of other series such as (GAME). As a result, the same neural circuit element can be commonly used in a plurality of series with different meanings, and it is possible to provide a new feature capable of increasing the memory capacity.
This example corresponds to the invention of claim 3.

In the above description, one element is associated with one event, but memory also inputs stimuli from various sensory organs at the same time and is associated with the event, and the firing of multiple elements. Associative learning is carried out as a flow of patterns. In this time series learning method, associative learning can be performed by increasing the coupling coefficient between events that occur at the same time and associating them with the firing pattern of these multiple elements. In the nervous tissue of the living body,
Because the structure of nerves is stacked in layers,
It is possible to realize a self-organization of the associative memory circuit in time and space by using the time series learning method for the connection in each layer and the connection between layers by adopting a multilayer structure even in the artificial neural network. Therefore, it is a great feature that spatial and temporal associative memories can be formed in a unified manner.

[0048]

As described above, the present invention is a very simple and naturally physiological learning method. According to the invention of claim 1, a method of accurately learning and reading out a time-series event flow in the coupling coefficient of a neural circuit element can be realized, and thus a high-level neural network capable of handling a temporal dynamic phenomenon. Bring the effect of realizing. That is, according to the first aspect of the present invention, it is possible to construct a logic circuit in which a temporal causal relationship is introduced in the neural network, which enables more advanced information processing.

According to the second aspect of the present invention, the time series is learned in the coupling coefficient and read out regardless of the time interval, so that high-speed reading can be performed. This makes it possible to realize an automated device that can take action in anticipation of the previous phenomenon based on the learned experience.

According to the invention of claim 3, one element can be commonly used in a plurality of series, so that the memory capacity can be increased, and the position of the coupling coefficient of the element can be made diverse. , It is possible to realize a function closer to human memory. Furthermore, it is possible to realize new information processing that handles complicated learning and memory in a unified manner, such as when several events occur spatially at the same time or in time sequence.

[Brief description of drawings]

FIG. 1 is a block diagram showing the contents of a time series learning method.

FIG. 2 is an explanatory diagram showing a relationship between an input history Hj (t) and a lapse of time Δt.

FIG. 3 is an explanatory diagram showing a relationship between an input history Hj (t) and an increase / decrease ΔW of a coupling coefficient depending on a learning level.

FIG. 4 is a graph showing a change Δt in time and a coupling coefficient increase / decrease ΔW from input.
It is explanatory drawing which shows the relationship of.

FIG. 5 is a configuration diagram showing a relationship between time series learning of events and a coupling coefficient.

FIG. 6 is a configuration diagram showing the contents of a neural circuit element.

FIG. 7 is a configuration diagram showing a configuration of a neural circuit for learning a time series.

FIG. 8 is an explanatory diagram showing a relationship between a coupling coefficient and a high-order time series.

[Explanation of symbols]

Xj (t-1) j-element output at time t-1 (becomes an input from another j-element at time t) Xi (t) i-element output at time t Vi (t) time Internal potential Hj (t) of i element at time t History value of input from j element at time t Wji j Coupling coefficient from j element to i element or transfer efficiency G1 Threshold value for positive change of G2 Threshold value for negative change in W (G1> G2) K1 Learning rate constant when change in W is positive (K1>
0) Learning rate constant (K2 <when K2 W change is negative)
0)

Claims (3)

[Claims] 1. A neural circuit element having a large number of inputs holds an input time history for each input, compares the value of the time history with an internal state of the neural circuit element, and based on the comparison result, A neural circuit characterized by executing learning for strengthening or damping the coupling coefficient or transmission efficiency for each input according to the time history, and storing the context of input and output times within the coupling coefficient or transmission efficiency. Time series learning method. 2. A combination of a plurality of neural circuit elements according to claim 1 for enabling read-out and reproduction at a higher speed than the actual occurrence time interval, whereby a sequence of events that occur according to the flow of time jumps ahead in time. A time series learning method for a neural circuit characterized by executing learning. 3. A neural circuit element used as said element in a combination of neural circuit elements according to claim 2 for enabling crossing or interference of time-series associative memory elements, has a plurality of meanings. A time series learning method for a neural circuit, characterized by increasing the memory capacity of the entire circuit.