JPH1091604A - Function learning device - Google Patents

Abstract

(57) [Summary] [PROBLEMS] To provide a function learning device capable of maintaining continuity of synapse values during learning without increasing the scale of hardware and enduring long-term use of synapse values after learning. An input / output device that has a variable parameter and whose output value is determined by an input value and a value of the variable parameter, an output value of training data, and an output when an input value of training data is input to the input / output device A parameter changing device that changes the value of the variable parameter so that the distance from the value decreases.The function learning device includes an analog memory that can hold the value of the variable parameter in a learning mode with a learning accuracy. A value stored in the first storage means in the use mode or a value obtained by analog / digital conversion of the value in the use mode with an accuracy capable of maintaining an input / output relationship of the input / output device within an allowable error range; A second storage unit that can hold the value of the variable parameter.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a function learning device using a neural network.

[0002]

2. Description of the Related Art As a technique for estimating generally nonlinear input / output relations required in pattern recognition, control systems, time series prediction, and the like, from a plurality of sets of input / output data presented as training data by learning. Attention has been paid to a method using a neural network.

Here, learning means optimizing a variable parameter of a network called a synapse value so that a neural network realizes a desired input / output relationship. Change the connection weighting factor between all neuron elements, i.e., the synapse value, so that the square of the error between the output value of the network and the output value of the training data with respect to the input value of the training data consisting of a set of output values is reduced. Go to B
P (Back Propagation) learning is performed.

[0004] By repeating this learning, the input / output relationship of the neural network becomes closer to the desired input / output relationship, and a reasonable estimated output is obtained even for an output for an input value not presented as training data. Is obtained.
Numerous engineering applications are being considered.

[0005] However, the execution of the neural network algorithm requires a huge amount of calculation, especially in the learning process, and real-time software processing is difficult.
Development of a dedicated LSI is also desired in order to expand an application field.

[0006]

One of the points for realizing a neural network with an analog LSI is a method of realizing a storage means for holding a synapse value. In order to execute a neural network algorithm faithfully, it is necessary to hold synapse values as continuous values.However, since analog memories with high accuracy and long-term reliability are not present, digital memories are generally simulated. They are used in analog form. However, in this case, there is a problem that the hardware scale increases as the number of discrete gradations increases.

[0007] In order to avoid this, attempts have been made to reduce the number of gradations of the digital memory by lowering the precision. However, in this case, particularly in the learning process, learning cannot be performed because the number of states of synapse vectors as many as necessary to maintain the feature of neurocomputing of analog search in a synapse vector space is not maintained.

The present invention has been made in view of the above circumstances, and without increasing the hardware scale of the entire synapse value storage unit, while maintaining the continuity of synapse values desired during learning, It is an object of the present invention to provide a function learning device using a neural network that can withstand long-term use of a synapse value after learning.

[0009]

According to the present invention, there is provided an input / output device having a plurality of variable parameters, wherein a predetermined number of input values and a predetermined number of output values are determined by the values of the variable parameters.
When training data including a set of input / output values is given, the variable is set so that the distance between the output value of the training data and the output value when the input value of the training data is input to the input / output device is reduced. A parameter changing device for changing the value of the parameter, a learning mode for changing the value of the variable parameter with the presentation of the training data, and an output value when an arbitrary input value is input to the input / output device. A function learning device having two operation modes of use modes to be used, a first storage unit including an analog memory capable of holding a value of the variable parameter in a learning accuracy in the learning mode; In a use mode, a value held in the first storage means or a value obtained by A / D conversion of the value is maintained in an input / output relationship of the input / output device within an allowable error range. In performance accuracy, and further comprising a second storage means which can hold the value of the variable parameter.

Preferably, prior to restarting the learning mode, for example, by adding training data, the value stored in the second storage means or an approximate value thereof is stored in the first storage means. It is characterized by writing.

Further, the present invention comprises an input / output device having a plurality of variable parameters, wherein a predetermined number of input values and a predetermined number of output values are determined by the values of the variable parameters, and a set of input / output values. A parameter for changing the value of the variable parameter so that a distance between an output value of the training data and an output value when the input value of the training data is input to the input / output device is reduced when the training data is provided; A learning mode for changing the value of the variable parameter in accordance with the presentation of the training data, and a use mode for using an output value when an arbitrary input value is input to the input / output device. In a function learning device having an operation mode, a first device configured by an analog memory in the learning mode is provided.
And the value of the variable parameter is held with a learnable accuracy in the form of the sum or difference of the value held in the storage means and the value held in the second storage means different from the first storage means. A first storage unit that can store the value of the variable parameter by the second storage unit in the use mode with an accuracy capable of maintaining an input / output relationship of the input / output device within an allowable error range. Means and a second storage means.

According to the present invention, a continuous value can be treated as a synapse value by using an analog memory at the time of learning, and analog exploratory learning that cannot be performed by a digital memory having only a small number of gradations can be performed. When the neural network is used as an input / output device after learning, a synapse value as a learning result can be stably held using a digital memory or the like.

Further, since the analog memory is used at the time of learning, when the digital memory is used after the learning, the number of necessary gradations can maintain the input / output relationship of the neural network within a desired allowable error range. I just need.

Therefore, the number of gradations is smaller than when the entire synapse value storage section is constituted only by digital memory. In general, since analog memory can be created with a much smaller hardware scale than digital memory, synapse value can be reduced. The hardware scale of the entire storage unit is
Equivalent functions can be reduced compared to a case where only digital memories are used.

[0015]

Embodiments of the present invention will be described below with reference to the drawings. (First Embodiment) A first embodiment of the present invention will be described.

FIG. 1 shows the configuration of a function learning apparatus according to the present embodiment. It is assumed that the hierarchical neural network 10 and the synapse updating device 40 are analog circuits, and the input value x and the output values y, y ^* are continuous values.

This function learning device has a learning mode and a use mode. First, a learning mode of the function learning device will be described. An input value x (x ₁ ,..., X _N ) and an output value y
Given training data consisting of a set of ^* (y ^* ₁ ,..., Y ^* _M ), a hierarchical neural network is generated based on an input value x and a synapse value held in the first storage unit 20. 10
Calculates and outputs an output value y (y ₁ ,..., Y _M ). Note that, for example, a random value is used as the initial value of the synapse value.

Here, the first storage unit 20 is constituted by an analog memory, and the hierarchical neural network 10 is, for example, as shown in FIG.
, A weighted sum of input values x (x ₁ ,..., X _N ) is calculated using a weighting coefficient called a synapse,
The value is input and output to an element called a neuron (9 in FIG. 2) generally having a monotonically increasing and bounded non-linear input / output characteristic, and this is output as a parallel hierarchical hardware configuration as shown in FIG. Is output as a final output value y (y ₁ ,..., Y _M ).

[0019] Synaptic updating device 40, the output value y of the hierarchical neural network _{10 (y 1, ..., y} M) output value _{^{y * (y * 1, ...}} , y * M) of the training data becomes equal to As described above, the synapse update amount that reduces the distance between the two, for example, the square error, is stored in the value (for example, the output value of each neuron) of the calculation process performed in the hierarchical neural network 10 and the first storage unit 20. The calculation is performed using the stored synapse value or the like, and the calculated value is added to the synapse value stored in the first storage unit 20. Alternatively, a value obtained by adding the synapse update amount to the pre-update synapse value is stored in the first storage unit 20.

Here, since the first storage section 20 is constituted by an analog memory, it is possible to handle continuous values as synapse values, and it is considered that learning proceeds much more smoothly than in the case where the values are stored as discrete values in a digital memory. Can be

The learning method described here is generally an error back-propagation learning method (error back-propagation).
This is called “n learning method”, and by executing this for each repeatedly presented training data, the hierarchical neural network 10 approximates all the input / output relations presented as training data. Go.

The above procedure is repeated until the input / output relationship of the hierarchical neural network 10 reaches a desired approximation accuracy, for example, until the sum of squares or the absolute value of the output errors for all the training data becomes less than a predetermined value. The learning mode is completed at the point of repetition.

When an analog memory such as a capacitor is employed as the first storage unit 20, for example, the long-term stability is not generally obtained due to electric charge leakage. Therefore, the learning mode is terminated before the storage capacity is significantly deteriorated.

Immediately after the end of the learning mode, each value of the first storage unit 20 holding each synapse value is converted into a discrete gradation value by a data conversion unit 50 composed of, for example, an A / D converter. Are stored in a second storage unit 30, which is constituted by, for example, a digital memory.

Thereafter, in a use mode in which the hierarchical neural network 10 is used as an input / output device, a value held in the second storage unit 30 is used as a synapse value, for example, a D / A converter. The data is used after being converted into a continuous value by the data reverse conversion unit 51.

Here, the number of gradations of the digital memory constituting the second storage unit 30 is determined by the required accuracy of the input / output relationship. In other words, it is determined that the final error obtained by adding the error generated by the discretization of the synapse value to the error at the end of the learning can satisfy the required level.

If the required level can be satisfied, the second storage unit 30 may be constituted by a floating gate element or the like, which is an analog memory having long-term stability but difficult to obtain high precision. good. In this case, the data conversion unit 50 and the data inverse conversion unit 51 become unnecessary.

FIG. 1 is a logical conceptual diagram of a hardware configuration of a function learning apparatus according to the present embodiment. In an actual hardware configuration, each neuron of a hierarchical neural network 10 realized as hardware is shown. A first storage unit 2 corresponding to each synapse multiplying circuit connecting the elements is provided in the vicinity thereof.
0, the second storage unit 30, the synapse updating device 40, the data conversion unit 50, and the like.

When a digital memory is used as a storage unit for synapse values, it is possible to express the input / output relationship with the required accuracy after learning based on the number of gradations required in the process of learning the desired input / output relationship. The number of necessary gradations is small. Therefore, according to the present invention using the analog memory at the time of learning and the digital memory or analog memory capable of holding the synapse value stably for a long time after the learning, the analogity of the synapse value required at the time of learning can be realized with small-scale hardware. Therefore, the hardware scale can be reduced as compared with the case where the synapse storage unit is constituted only by the digital memory, and the learning result can be stably held for a long time which cannot be realized by the analog memory using the capacitor or the like.

(Second Embodiment) Next, a second embodiment of the present invention will be described. In the second embodiment of the present invention, after the learning mode and the conversion and transfer of the synapse value from the first storage unit 20 to the second storage unit 30 in the previous first embodiment are once completed, When training data consisting of a set of input / output values that require learning is added, re-learning can be easily performed.

FIG. 3 shows the configuration of the function learning device according to the present embodiment. The configuration of the present embodiment is the same as that of the
Except that 0 is added, it is basically the same as the first embodiment.

Now, once the learning is completed, when new training data requiring additional learning is presented, each value held as a synapse value in the second storage unit 30, for example, a digital memory, is stored. Is converted to an analog value by a data inverse converter 60 which is, for example, a D / A converter, and each value of the result is stored in a first analog memory.
Is transferred to the storage unit 20.

Immediately thereafter, a set of previously learned training data to which training data that needs to be newly added is repeatedly presented as a new training data set, and learning is performed in the same manner as in the first embodiment. . After that the first
The same process as described in the embodiment is performed.

Error back propagation learning method (error back)
-Propagation learning met
hod) is an optimization in which the sum of the square of the error, which is a function of the synapse value, for all training data is taken as a cost function (strictly, a stochastic descent method for the cost function)
Therefore, adding training data results in a change in the cost function. Therefore, the optimal synapse value changes accordingly, but the number of added training data is small or the input / output relationship does not deviate much from the interpolation of the input / output relationship of the trained training data. , The change amount of the optimal synapse value is small. Therefore, in such a case, the value held in the second storage unit 30 as an approximate value of the optimal synapse value for the learned training data is adopted as the initial synapse value to be given to the first storage unit 20 when learning is resumed. It is considered that the number of times of learning until the end of learning may be smaller when, for example, the random synapse value is set as the initial value.

Therefore, according to the present invention, even when the training data to be learned is added, the analogity of the synapse value required at the time of learning can be realized with a small-scale hardware, and the synapse storage unit is formed of only digital memory. In addition to reducing the hardware scale as compared with the case of configuring with, the long-term stable holding of the learning result that cannot be realized with an analog memory using a capacitor etc. is also possible, and the number of times of learning until the end of learning at the time of additional learning is small. It is possible to provide an initial synapse value that is sufficient.

When the second storage unit 30 is configured by an analog memory such as a floating gate element,
The data conversion unit 50, the data conversion unit 51, and the data conversion unit 60 become unnecessary.

(Third Embodiment) Next, a third embodiment of the present invention will be described. In the first and second embodiments, the synapse value in the learning mode is held in the first storage unit configured by the analog memory. However, in the present embodiment, the synapse value in the learning mode is configured by the analog memory. The value held in the first storage unit and the value held in the second storage unit constituted by a storage device having a long-term stable holding capability are held in the form of a sum or a difference. is there.

FIG. 4 shows the sum of the value held in the first storage unit constituted by the analog memory and the value held in the second storage unit constituted by the storage device having a long-term stable holding ability. 2 shows a configuration of a function learning device according to the present embodiment when synapse values are held in a form.

It is assumed that the hierarchical neural network 10, the synapse updating device 40, and the adder 70 (or a subtractor described later) are analog circuits, and the input value x and the output values y, y ^* are continuous values.

The first storage unit 20 is formed of an analog memory, the second storage unit 30 is formed of, for example, a digital memory, and the hierarchical neural network 10 has been described in the first embodiment with reference to FIG. It is the same as the one.

First, the learning mode of the function learning device will be described. An input value x (x ₁ ,..., X _N ) and an output value y
Given training data consisting of a set of ^* (y ^* ₁ ,..., Y ^* _M ), the input value x, the value calculated by the adder 70 and held in the first storage unit 20, and The hierarchical neural network 10 calculates and outputs an output value y (y ₁ ,..., Y _M ) based on a synapse value which is a sum with a value held in the corresponding second storage unit 30. In addition,
The value stored in the second storage unit 30 is, for example, D / A
Before being input to the adder 70, the data is converted into a continuous value by the data inverse converter 52 composed of a converter.

In the synapse updating device 40, the output value y (y ₁ ,..., Y _M ) of the hierarchical neural network 10 becomes equal to the output value y ^* (y ^* ₁ ,..., Y ^* _M ) of the training data. As described above, the amount of update of the values held in the first storage unit 20 and the second storage unit 30 is reduced in the calculation process performed in the hierarchical neural network 10 so as to reduce the distance between them, for example, the square error. Values (eg, output values of each neuron) and the first storage unit 20
And the value stored in the second storage unit 30 and the like, and the calculated value is stored in the first storage unit 20 and the second storage unit 30.
Add to the value stored in. Alternatively, a value obtained by adding the synapse update amount to the synapse value before update is stored in the first storage unit 20.
And in the second storage unit 30.

The synapse updating device 40 stores the first storage unit 2
The update amount of the value held in 0 and the update amount of the value held in the second storage unit 30 may be independently obtained. Alternatively, an updated value of the sum of the values held in the first storage unit 20 and the second storage unit 30 is obtained, and then the updated value held in the first storage unit 20 and the second storage The values may be distributed to the updated values held in the unit 30.

Here, when the value held in the second storage unit 30 is input, the data conversion unit performs D / A conversion of the value, and when the value is output to the second storage unit 30. A data conversion unit for A / D converting the value is stored in the second storage unit 3
0 and the synapse updating device 40, or built in the synapse updating device 40.

Here, since the first storage unit 20 is formed of an analog memory, even if the second storage unit 30 is formed of a digital memory, continuous values can be handled as synapse values, and only the digital memory can be used. It is considered that the learning proceeds much more smoothly than the case where it is held as a discrete value.

When the synapse value is stored in the form of a difference between the values stored in the first storage unit 20 and the second storage unit 30, the adder 70 in FIG. 4 may be replaced with a subtractor. . If it is considered that the value of the value stored in either the first storage unit 20 or the second storage unit 30 with the opposite sign is stored in that storage unit, it is logically stored as a sum. Is exactly the same as Therefore, the following discussion is common to both.

Now, the above-described procedure is performed by using the hierarchical neural network 1.
Until the input / output relationship of 0 reaches a desired approximation accuracy, for example,
The learning mode ends when the repetition until the sum of the squares or the absolute value of the output errors for all the training data becomes equal to or less than a predetermined value.

In the present embodiment, at the final stage of learning, a rough value of a finally desired synapse value for realizing a desired input / output relationship is held in the second storage unit 30,
At that time, it can be expected that a value corresponding to the update amount calculated for each piece of training data whose absolute value is considered to be small is held in the first storage unit 20.

Accordingly, when the value held in the second storage unit 30 is used as a synapse value as it is, the hierarchical neural network 1
If the number of gradations of the digital memory constituting the second storage unit 30 is selected so that the approximation error with respect to the desired input / output relationship due to the input / output relationship of 0 can satisfy the required level, the first embodiment It is not necessary to terminate the learning mode before the storage holding capacity of the first storage unit 20 is degraded as in the above, and the smooth use of the first storage unit 20 as an auxiliary memory for executing analog computation is always possible. You can expect learning.

Thereafter, in the use mode in which the hierarchical neural network 10 is used as an input / output device, the value held in the second storage unit 30 is used as the synapse value, for example, D
It is used after being converted into a continuous value by the data reverse conversion unit 53 composed of the / A converter.

Further, when training data including a set of input / output values that need to be newly learned is added after learning is once completed, re-learning is performed using the synapse value at that time as an initial value. it can.

If the required level of the above-described approximation error can be satisfied, as in the first embodiment, a floating memory, which is an analog memory which has long-term stability but is difficult to obtain high accuracy. The second storage unit 30 may be configured by a gate element or the like.

When inputting the value held in the second storage unit 30, the synapse updating device 40 converts the value into a D / A
Although a data reverse conversion unit for conversion and a data conversion unit for A / D converting the value when outputting the value to the second storage unit 30 are built in, the second storage unit 30 is also analog. In the case of using a memory, the data inverse conversion unit 5
2. The data conversion unit 53 and the data conversion unit described as being built in the synapse updating device 40 are not required.

FIG. 4 is a logical conceptual diagram of the hardware configuration of the function learning apparatus according to the present embodiment. In an actual hardware configuration, each neuron of the hierarchical neural network 10 realized as hardware A first storage unit 2 corresponding to the vicinity of each synapse multiplication circuit connecting the elements,
0, the second storage unit 30, the synapse updating device 40, the adder 70, and the like.

As described above, according to the present invention, the analog memory is used as the auxiliary memory at the time of learning, and the digital memory or the analog memory that can hold the synapse value stably for a long time after the learning is used. Can be realized with small-scale hardware, so that the hardware scale can be reduced compared to the case where the synapse storage unit is composed of only digital memory, and at the same time, learning results that cannot be realized with analog memory using capacitors etc. Long-term stable retention is also possible.

Each of the embodiments described above can be easily adapted to L
It can be implemented as SI. The present invention is not limited to the above-described embodiment, and can be implemented with various modifications within the technical scope.

[0057]

According to the present invention, the analogity of the synapse value required at the time of learning can be realized by small-scale hardware, so that the hardware scale is smaller than when the synapse storage unit is constituted only by digital memory. At the same time, it is possible to stably hold the learning result which cannot be realized by an analog memory using a capacitor or the like for a long period of time. Also,
Even when training data to be learned is added, it is possible to provide an initial synapse value that requires a small number of times of learning until the end of learning at the time of additional learning.

[Brief description of the drawings]

FIG. 1 is a diagram showing a configuration of a function learning device according to a first embodiment of the present invention.

FIG. 2 is a diagram illustrating a configuration example of a hierarchical neural network;

FIG. 3 is a diagram showing a configuration of a function learning device according to a second embodiment of the present invention.

FIG. 4 is a diagram showing a configuration of a function learning device according to a third embodiment of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 10 ... Hierarchical neural network 20 ... 1st memory | storage part 30 ... 2nd memory | storage part 40 ... Synapse updating apparatus 50 ... Data conversion part 51, 52, 53, 60 ... Data inversion part 70 ... Adder

Claims (3)

[Claims] An input / output device having a plurality of variable parameters, a predetermined number of input values and a predetermined number of output values determined by the values of the variable parameters, and training data comprising a set of input / output values are provided. A parameter changing device that changes the value of the variable parameter so that the distance between the output value of the training data and the output value when the input value of the training data is input to the input / output device is reduced. And has two operation modes: a learning mode in which the value of the variable parameter is changed according to the presentation of the training data; and a use mode in which an output value when an arbitrary input value is input to the input / output device is used. In the function learning apparatus, a first storage unit including an analog memory capable of holding a value of the variable parameter with a learning accuracy in the learning mode; In the mode, the value held in the first storage means or the value obtained by A / D conversion of the value is held as the value of the variable parameter with an accuracy capable of maintaining the input / output relationship of the input / output device within an allowable error range. Function learning device, comprising: a second storage unit capable of performing the function. 2. Prior to resuming the learning mode,
2. The function learning apparatus according to claim 1, wherein the value stored in the second storage unit or an approximate value thereof is written in the first storage unit. 3. An input / output device having a plurality of variable parameters, wherein a predetermined number of input values and a predetermined number of output values are determined by the values of the variable parameters, and training data comprising a set of input / output values are provided. A parameter changing device that changes the value of the variable parameter so that the distance between the output value of the training data and the output value when the input value of the training data is input to the input / output device is reduced. And has two operation modes: a learning mode in which the value of the variable parameter is changed according to the presentation of the training data; and a use mode in which an output value when an arbitrary input value is input to the input / output device is used. In the function learning apparatus, in the learning mode, a value held in a first storage unit constituted by an analog memory and a value held in a second storage unit different from the first storage unit The value of the variable parameter can be held in the form of a sum or a difference with a learnable accuracy, and the value of the variable parameter is stored in the use mode by the second storage means, and A function learning device, comprising: a first storage unit and a second storage unit that can be maintained with an accuracy that can be maintained within an allowable error range.