JPH04662A - Learning machine - Google Patents

Abstract

PURPOSE:To improve the efficiency and the accuracy of learning by constituting the learning machine so that a weight change in a multi-input-output signal processing part in which magnitude of an error is below a threshold is not executed, and also, the threshold becomes smaller as learning advances. CONSTITUTION:In an error signal calculating part 11, an error in the uppermost layer is calculated from a difference of an actual output signal outputted from an output signal calculating part 1 and a teacher signal, and based on the calculated error, a weight change amount calculating part 12 calculates a change amount of a weight coefficient stored in a memory of the output signal calculating part 1. Subsequently, when it is decided that the error is smaller than a set threshold T, a weight change amount control part 13 sets the weight coefficient change amount to '0'. In this case, a threshold control part 16 switches the threshold T to a smaller value as learning advances, based on a result of decision by a learning progress deciding part 15. In such a way, the learning efficiency and the accuracy are improved.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application The present invention relates to a learning machine for a data processing device.Background ArtThe learning machine includes, for example, E. Rumelhart, G.E.
Hinton and RlJ, Williams, ``Learning A 117° Reset 1 Instances To 'I Hightsku 7' To 1 Case Tink Erras.''
(Learning Representations
by Back-Propagating Er
rors)”Nature (Nature), vol.
, 323. pp, 533-536. Oct.

9, 1986), this conventional learning machine consists of an output signal calculation section and a weighting coefficient updating section, as shown in Fig. 6, and the output signal calculation section has a layered structure, and mutual connections within each layer. It is composed of a plurality of multi-input-output signal processing units 600 that are network-connected so that the signal propagates only to the upper layer. - The output of the output signal processing unit 600 is multiplied by a weighting coefficient representing the degree of connection, and the sum is converted by a threshold numerical value, and then the value is transmitted as an output to the upper layer. 4. The weighting coefficient updating section causes the teacher signal generating section 602 to generate a desirable output signal for the input signal as the teacher signal tk in accordance with the signal input from the input section 601 of the output signal calculating section.In the error signal calculating section 603 The actual output signal □b(om?;t,, represents the output of the th multi-input-output signal processing unit of the highest layer in the output signal calculation unit) and the An error E = 0.5 (th ok) is calculated from the difference from the teacher signal, and this value is used to evaluate the performance of the network in the current connection state (size of weighting coefficient)4. Based on the error E, the weight change amount calculation unit 604 calculates the change amount ΔW l 1 of the weighting coefficient of the output signal calculation unit based on the following formula Δw+'=-taE/&wx Here, ε is It is a positive constant called the learning rate and is 4
By repeating the update of the weights as described above, the error is made smaller, and when the error becomes small enough, the output signal is considered to be sufficiently close to the desired value, and learning is terminated. Problem to be solved by the invention. However, in the above configuration, even in a multi-input-output signal processing unit where the error value is sufficiently small, the weights are updated when the error E value decreases, so the learning efficiency is poor and the time required for learning is long. In view of this, the present invention aims to provide a learning machine that requires less time for learning. , an output signal calculation section consisting of a plurality of multi-input-output signal processing sections connected to a network so that there is no mutual coupling within each layer and signals propagate only to upper layers, and an output obtained from the output signal calculation section. The multi-input/output signal processing unit includes a weighting coefficient updating unit that updates the value of the weighting coefficient of the first output signal calculation unit based on the signal, and the multi-input/output signal processing unit includes a memory that holds a plurality of weighting coefficients, and a memory that stores a plurality of data as input. an input unit that weights the manual data from the input unit with a weighting coefficient stored in the memory; an addition unit that adds together a plurality of pieces of data weighted by the multiplication unit; The weighting coefficient updating unit includes a threshold processing unit that limits the output to a value within a certain range; an error signal calculation unit that calculates an error with a teacher signal; a weight change amount calculation unit that calculates a change amount of the weighting coefficient stored in the memory according to an output of the error signal calculation unit; an error signal determination unit that determines whether the output is below a threshold; a learning progress determination unit that determines the progress of learning; and a value of the threshold according to the determination result of the learning progress determination unit. a threshold value control unit that gradually decreases the weighting coefficient; and a threshold value control unit that gradually decreases the weighting coefficient change amount when the error value is determined to be less than the threshold value by the error signal determination unit.
A second invention is a learning machine comprising: a weight change amount control unit which counts the outputs of the first error signal determination unit, and outputs a jump signal when all output signals are equal to or less than the threshold value. a learning progress determining unit that determines the progress of learning; a threshold control unit that gradually reduces the value of the threshold according to a determination result of the learning progress determining unit; and the skipping signal. According to the first aspect of the present invention, the learning machine has a weight change amount control unit that skips the weight change operation in the weight change amount calculation unit in accordance with the weight change amount calculation unit. Since the input-output signal processing section does not change the weights, it not only improves learning efficiency and shortens the time required for learning, but also reduces the threshold value as learning progresses, thereby improving output signal calculation. Since the output of the section can be made to match the teacher signal more precisely, the accuracy of learning is improved.

According to the invention of 1982, when the error of all the multi-input-output signal processing units is less than the threshold value for human data, the weight change operation is skipped, which not only improves the learning efficiency, but also improves the learning efficiency. Since the amount of calculation can be significantly reduced, the time required for learning can be shortened. In addition, by decreasing the threshold value as learning progresses, the output of the output signal calculation unit can be made to match the teacher signal more precisely.
FIG. 1 shows a configuration diagram of a learning machine in an embodiment of the first invention. In FIG. 1, 1 is an output signal calculation section, and 2 is an output signal calculation section. The weighting coefficient updating unit updates the value of the weighting coefficient of the output signal calculation unit 1 based on the output signal obtained in step 1.
As shown in Figure 2, it has a multi-stage circuit network configuration.
3 is a multi-input/output signal processing section; 4 is an input section of the output signal calculation section 1; the configuration of the multi-input/output signal processing section 3 constituting the output signal calculation section 1 is specifically shown. is also shown in Figure 3. In Figure 3, 5 is an input section of the multi-input-output signal processing section 3, 6 is a memory that stores weighting coefficients for weighting multiple inputs from the input section 5, and 7 is a weighting coefficient of the memory 6. 8 is a multiplier that multiplies the inputs from input section 5 and 8 is an adder that adds together the outputs of each multiplier 7.
9 is a threshold processing section that limits the output of the adder 8 to values within a certain range.

FIG. 4 shows the human output characteristics of the threshold processing section 9.

For example, the input/output characteristics of the threshold processing unit 9 that limits the output to the range (0, 1) are f (I) = 1/(1+ exp(-
I 10 )) It can be expressed mathematically as (1). Here, ■ is the input to the threshold processing section 9. As for the input/output characteristics of the threshold processing section 9, a threshold numerical value other than the above may be used. Shown in Figure 1.

10 is a teacher signal generation section; 11 is an error signal calculation section; 12 is a weight change amount calculation section; 13 is a weight change amount control section; 14 is an error signal judgment section; 15 is a learning progress judgment section; 16 is a threshold control section; The operation of the learning machine of the embodiment configured as above will be described below. When an input signal is input to the input section 4 of the output signal calculation section 1, each multi-input-output signal processing section 3 performs the following operations. The output of the lower-layer multi-input-output signal processing section 3 connected to the input-output signal processing section 3 is multiplied by a weighting coefficient representing the degree of connection stored in the memory 6 by a multiplier 7, The sum of the outputs of each of the multipliers 7 is calculated by the adder 8.
The threshold value processing unit 9 converts the signal and outputs the value to the multi-input/output signal processing unit in the upper layer. In other words, the multiple inputs shown in Figure 3 -
Output signal processing unit 3 (upper) sets the input value to the input unit 5 to 0” (output of the j-th multi-input-output signal processing unit in the lower layer), and sets the weighting coefficient stored in the memory 6 to “Wll” (i-th The coupling weight between the multi-input-output signal processing unit and the j-th multi-input-output signal processing unit in the lower layer) and (U01-f (ΣWIj OJ) (
2). According to the signal inputted from the input section 4 of the output signal calculation section 1, the teacher signal generation section 10 calculates a desired output signal for the input signal. teacher signal t
L generated as k Difference or aberration between the actual output signal Qk outputted from the output signal calculation section 1 and the teacher signal in the error signal calculation section 11 Error of the second multi-input-output signal processing section in the top layer E = 0.5 (tk-oh)”
(3) is calculated, and this value is used to evaluate the performance of the network in the current connection state (size of weighting coefficient).

Based on the error E calculated in this way, the weight change amount calculation unit 12 calculates the change amount ΔWll of the weighting coefficient stored in the memory 6 of the output signal calculation unit 1 based on the following formula. △wa”= −εaE/aw1+ (4
) Here, ε is a positive constant called the learning rate.The weight change amount control unit 13 determines that the error 1tb−○kl is the threshold value T set by the threshold control unit 16 in the error signal determination unit 14. If it is determined that the weight coefficient is smaller than the range, the weighting coefficient change amount of the top layer in the output signal calculation unit 1 is set to O. At this time, the threshold value control unit 16, based on the determination result of the learning progress determination unit 15, As learning progresses, the threshold value T is switched to a smaller value. Determination of the learning progress in the learning progress determination unit 15 (above) Determination based on the sum of errors of the multi-input-output signal processing unit 3 in the highest layer of the output signal calculation unit 1 Determination based on the number of learning times In the upper layer, either of the following methods may be used: determination based on the number of multi-input-output signal processing units that output an error greater than a threshold value, or determination based on the value of the maximum output in one learning of the error signal calculation unit. (By repeating the update of the weights as above,
As the error is made smaller, when the error becomes small enough, the output signal is assumed to be sufficiently close to the desired value, and learning is terminated.In this way, according to this embodiment, if the error is smaller than the threshold value T, Since the weights are not changed in the input-output signal processing section, the learning efficiency can be improved and the time required for learning can be shortened. However, by decreasing the value of the threshold T as learning progresses, Output signal calculation section 1
Since the output Ok can be matched with the teacher signal tm more accurately, the learning accuracy is improved. Figure 5 shows a configuration diagram of the learning machine in the embodiment of the second invention. 10 is a teacher signal generation section 11
is the error signal calculation i 12 is the weight change amount calculation unit 14 is the error signal judgment unit 15 is the learning progress judgment unit 16 is the threshold control device 23 is the weight change amount side The operation of the learning machine according to the embodiment of the second invention configured as follows will be described below.

When an input signal is input to the input section 4 of the output signal calculation section 1, each multi-input-output signal processing section 31 is connected to the multi-input-output signal processing section 3 in the lower layer. The output of the unit 3 is multiplied by a weighting coefficient representing the degree of connection stored in the memory 6 by a multiplier 7, and the sum of the outputs of each of the multipliers 7 is calculated by an adder 8. The threshold processing unit 9 converts the value and outputs it to the upper layer multi-input-output signal processing unit.In other words, the multi-input-output signal processing unit 3 shown in FIG. (
(output of the j-th multi-input-output signal processing unit in the lower layer), and the weighting coefficients stored in the memory 6 are connection weight with
(2) is calculated by the weighting coefficient updating unit 2 (above).In response to the signal manually inputted from the input unit 4 of the output signal calculation unit 1, the teacher signal generation unit IO determines the desired output for the input signal. The signal is the teacher signal t
The difference between the actual output signal Ok output from the output signal calculation unit 1 and the teacher signal in the error signal calculation unit 11 is the error of the second multi-input-output signal processing unit in the top layer. E=0.5 (tb oh)' (3)
is calculated, and this value is used to evaluate the performance of the network in the current connection state (size of weighting coefficient).

Based on the error E calculated in this way, the weight change amount calculation unit 12 calculates the change amount ΔW11 of the weighting coefficient stored in the memory 6 of the output signal calculation unit 1 based on the following formula. △w++ = -t a E / a wz
(4) Here, ε is a positive constant called the learning rate. The error signal determination unit 14 calculates the error 1tk-O
kl is the threshold T set by the threshold controller 16
It is checked whether the error is smaller than the threshold value, and if the error is larger than the threshold value, 0 is output to the interlaced judgment section 24, and if it is smaller, 1 is output to the jump judgment section 24. The interlacing judgment unit 24 counts the judgment results and outputs an interlacing signal when the error 1th-□ill of all the multi-input-output signal processing units in the top layer is equal to or less than the threshold T. 23 is a threshold control unit 16 ζ table which controls the skip determination unit 24 to skip the weight change operation in the weight change amount calculation unit 12 after the skip signal has been generated.
As learning progresses based on the judgment result of the threshold T
Switch to a smaller value. Further, the learning progress determination unit 15 determines the learning progress based on the total error of the multi-input-output signal processing unit 3 in the highest layer of the output signal calculation unit 1, the determination based on the number of times of learning, and the highest level of the output signal calculation unit 1. In the upper layer, either of the following methods may be used: determination based on the number of multi-input-output signal processing units that output an error greater than a threshold value, or determination based on the value of the maximum output in one learning of the error signal calculation unit. (By repeating the update of the weight at 11 degrees or more, the error is made smaller. When the error becomes small enough, the output signal is assumed to be sufficiently close to the desired value, and learning is terminated. According to this embodiment, if the error of all the Takuker output signal processing units with respect to input data is less than the threshold value T, the weight change operation in the weight change amount calculation unit is skipped, so that the learning efficiency is only improved. However, since the amount of calculation can be significantly reduced, the time required for learning can be shortened.
By decreasing the value of the threshold value T as learning progresses, the output Ok of the output signal calculation unit 1 can be made to match the teacher signal ti+ more accurately, so that the accuracy of learning is improved. According to the first invention, by providing the error signal judgment a, the weight change amount calculation unit, the learning progress judgment unit, and the threshold control unit, it is possible to change the weight in the multi-input-output signal processing unit where the magnitude of the error is less than or equal to the threshold value. This not only improves learning efficiency and shortens the time required for learning, but also reduces the threshold value as learning progresses, allowing the output of the output signal calculation unit to be used as the teacher signal more accurately. Since the learning accuracy can be matched, the accuracy of learning can be improved.According to the second invention, 1. When the error size of all multi-input-output signal processing units is less than the threshold value, the weight change operation in the weight change amount calculation unit is skipped, which not only improves learning efficiency but also significantly reduces the amount of calculations. Therefore, the time required for learning can be shortened.Also, by decreasing the threshold value as learning progresses, the output of the output signal calculation section can be more precisely matched to the teacher signal, improving the accuracy of learning. do.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of a learning machine in an embodiment of the first invention. FIG. 2 is a configuration diagram of an output signal calculation unit in the embodiment. FIG. 3 is a multi-input-output signal processing unit in the embodiment. Fig. 4 shows the input/output characteristics of the threshold processing section in the same embodiment. Fig. 5 shows a block diagram showing the structure of the learning machine in the embodiment of the second invention. Fig. 6 shows the structure of the conventional learning machine. This is a block diagram showing the configuration: output signal calculation section 2... weight change amount calculation section 3.
...Multi-input-output signal processing N1.4...Input a of the output signal calculation section 5...Input a of the multi-input-output signal processing section 6...Memory, 7...Multiplier 8...・Additional statue
9...Threshold processing 10...Teacher signal generation
11... Error signal calculation section 12... Weight change amount calculation K 13... Weight change amount calculation section 14... Error signal calculation section 24... Name of jump judgment agent Patent attorney Shigetaka Awano et al. 1 person figure

Claims (10)

[Claims] (1) An output signal calculation unit consisting of a plurality of multi-input-output signal processing units that have a layered structure and are network-connected so that there is no mutual coupling within each layer and signals propagate only to upper layers, and the output signal The multi-input/output signal processing unit includes a weighting coefficient updating unit that updates the value of the weighting coefficient of the output signal calculation unit based on the output signal obtained by the calculation unit, and the multi-input/output signal processing unit holds a plurality of weighting coefficients. a memory, an input section for inputting a plurality of data, a multiplication means for weighting the input data from the input section with a weighting coefficient stored in the memory, and adding together a large number of the plurality of data weighted by the multiplication means. The weighting coefficient updating unit includes an adding unit and a threshold processing unit that limits the output of the adding unit to a value within a certain range, and the weighting coefficient updating unit is configured to generate a teacher signal that provides a teacher signal as a desirable value of the output signal of the output signal calculation unit. a generation unit; an error signal calculation unit that calculates an error between the output signal and the teacher signal; and a weight change amount calculation unit that calculates a change amount of the weighting coefficient stored in the memory according to the output of the error signal calculation unit. an error signal determination unit that determines whether the output of the error signal calculation unit is below a threshold; a learning progress determination unit that determines the progress of learning; and a determination result of the learning progress determination unit. a threshold control unit that gradually decreases the value of the threshold value according to the threshold value; and a weight that sets the weighting coefficient change amount to 0 when the error signal determination unit determines that the error value is less than or equal to the threshold value. A learning machine characterized by comprising a change amount control section. (2) The learning machine according to claim 1, wherein the learning progress determination unit determines the progress of learning based on the total error of the multi-input-output signal processing unit in the highest layer of the output signal calculation unit. . (3) The learning machine according to claim 1, wherein the learning progress determination unit determines the learning progress based on the number of times of learning. (4) The learning progress determination unit is characterized in that the learning progress is determined by the number of multi-input-output signal processing units that output an error equal to or greater than a threshold in the top layer of the output signal calculation unit. The learning machine according to claim 1. (5) The learning machine according to claim 1, wherein the learning progress determining section determines the learning progress based on a maximum output value in one learning by the error signal calculating section. (6) An output signal calculation unit consisting of a plurality of multi-input-output signal processing units that have a layered structure and are network-connected so that there is no mutual coupling within each layer and that signals propagate only to upper layers, and the output signal The multi-input/output signal processing unit includes a weighting coefficient updating unit that updates the value of the weighting coefficient of the output signal calculation unit based on the output signal obtained by the calculation unit, and the multi-input/output signal processing unit holds a plurality of weighting coefficients. a memory, an input section for inputting a plurality of data, a multiplication means for weighting the input data from the input section with a weighting coefficient stored in the memory, and adding a large number of the plurality of data weighted by the multiplication stage. The weighting coefficient updating section includes a threshold processing section that limits the output of the adding means to a value within a certain range, and the weighting coefficient updating section is configured to generate a teacher signal that provides a teacher signal as a desirable value of the output signal of the output signal calculation section. a generation unit; an error signal calculation unit that calculates an error between the output signal and the teacher signal; and a weight change amount calculation unit that calculates a change amount of the weighting coefficient stored in the memory according to the output of the error signal calculation unit. an error signal determination unit that determines whether the output of the error signal calculation unit is below the threshold; and an error signal determination unit that counts the outputs of the first error signal determination unit, and if all output signals are below the threshold; a skip determination section that outputs a skip signal to the learning progress determination section; a learning progress determination section that determines the progress of learning; and a threshold control section that gradually decreases the threshold value according to the determination result of the learning progress determination section. and a weight change amount control unit that skips a weight change operation in the weight change amount calculation unit in response to the skipping signal. (7) The learning machine according to claim 6, wherein the learning progress determination unit determines the progress of learning based on the total error of the multi-input-output signal processing unit in the highest layer of the output signal calculation unit. . (8) The learning machine according to claim 6, wherein the learning progress determination unit determines the learning progress based on the number of times of learning. (9) The learning progress determination unit is characterized in that the learning progress is determined based on the number of multi-input-output signal processing units that output an error equal to or greater than a threshold in the highest layer of the output signal calculation unit. The learning machine according to claim 6. (10) The learning machine according to claim 6, wherein the learning progress determining section determines the learning progress based on a maximum output value in one learning by the error signal calculating section.