JP2018097407A - Perturbation learning apparatus and method thereof - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a perturbation learning apparatus capable of processing weight learning at high speed.SOLUTION: A perturbation learning apparatus comprises: a branching part 10 which branches learning data in each element; a propagation part 20 which calculates filter data by multiplying each element by a filter factor with different weight for each position of the element due to the category of the learning data, and calculates a first output value O by adding a value obtained by multiplying an emphasis weight for each element of the filter data; a perturbation propagation part 30 which creates a perturbation weight by adding a perturbation factor correcting the emphasis weight, and is provided correspondingly to each filter data, and calculates a second output value Oby multiplying the filter data by the perturbation weight; and a learning part 40 which repeats calculation for obtaining a variation to update the emphasis weight, from a first cost value O obtained by subtracting a teacher value representative of a learning objective from the first output value, a second cost value obtained by subtracting the teacher value from each of the second output value O, the perturbation factor δ, and a learning factor η representative of the slope of learning speed.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a perturbation learning apparatus and method for processing learning operations at high speed.

  In recent years, as an environment for acquiring a large amount of data from the Internet and various sensors is constructed, research and business for analyzing the data and performing highly accurate knowledge processing and future prediction are actively performed. One of the analysis techniques that has attracted particular attention in this trend is called deep learning.

  Deep learning is a machine learning technique based on a neural network, and has a function of discriminating and judging by learning data in an artificial neural network in which neurons are arranged in multiple layers.

  In general, neural network learning is performed by updating the weights of the synapses that connect the neurons. However, with the recent deep learning, the number of learning data and the number of synapses is large, so it takes a lot of time for learning. It takes.

  Under such a background, an analog arithmetic unit replacing a digital arithmetic unit such as a CPU or a GPU has been attracting attention. In particular, optical computing using light as a signal transmission medium is considered to be promising as an analog computing unit that accelerates deep learning.

  A perturbation learning method has been proposed as a learning algorithm for an analog computing unit. The perturbation learning method is a method of calculating a parameter correction amount by changing a parameter to be learned, such as a synaptic weight, by a small amount and monitoring an influence on an error function.

  For example, Non-Patent Document 1 discloses a simultaneous perturbation optimization method.

Hiroshi Maeda et al., “Additional control of 2-axis robot arm by neural network using simultaneous perturbation learning rule”, Denki Theory C, Vol. 123, No. 9, 2003

  However, the configuration of a device that calculates by the conventional perturbation learning method is a method in which the output value is compared with the teacher value, and the weights are sequentially learned one by one so as to minimize the error function. Therefore, when the scale of the neural network is large, there is a problem that enormous time is required for convergence.

  The present invention has been made in view of this problem, and an object of the present invention is to provide a perturbation learning apparatus and method that can process weight learning at high speed.

  A perturbation learning device according to an aspect of the present embodiment includes a branching unit that branches learning data composed of a plurality of elements for each element, and a filter coefficient having a different weight for each position of the element depending on the category of the learning data. Multiplying each of the elements to obtain filtration data, a weighting unit for emphasizing the filtration data, a propagation unit for obtaining a first output value obtained by adding a value multiplied for each element of the filtration data, and the enhancement weights, A perturbation propagation unit that generates a perturbation weight by adding a perturbation coefficient for correcting the enhancement weight, is provided corresponding to each of the filtered data, and multiplies the filtered data by the perturbation weight to obtain a second output value; The same number as the perturbation propagation unit, a first cost value obtained by subtracting a teacher value representing a learning target from the first output value, a second cost value obtained by subtracting the teacher value from each of the second output values, A learning unit that repeats calculation for obtaining a change amount for updating the enhancement weight from the perturbation coefficient and a learning coefficient that represents a gradient of the learning speed, and obtaining the enhancement weight by adding the change amount to each of the enhancement weights. The gist is to include a weight updating unit.

  A perturbation learning method according to an aspect of the present embodiment is a perturbation learning method performed by the perturbation learning apparatus, wherein learning data including a plurality of elements is branched for each element, and the elements are classified according to the category of the learning data. A filter coefficient having a different weight for each position is multiplied by each of the elements to obtain filtration data, and a first output value obtained by adding a value obtained by multiplying the emphasis weight for emphasizing the filtration data by the element of the filtration data is added. And generating a perturbation weight obtained by adding a perturbation coefficient for correcting the enhancement weight to the enhancement weight, provided corresponding to each of the filtration data, and multiplying the filtration data by the perturbation weight to obtain a second output value. The first cost value obtained by subtracting the teacher value representing the learning target from the first output value, the second cost value obtained by subtracting the teacher value from each of the second output values, the perturbation coefficient, and the learning speed From the learning coefficient representing the gradient, repeating the calculation for obtaining the amount of change to update the emphasis weights to each of the enhancement weight, and summarized in that for obtaining the enhancement weight by adding the amount of change.

  According to the present invention, it is possible to provide a perturbation learning apparatus and method that can process weight learning at high speed.

It is a figure which shows the function structural example of the perturbation learning apparatus which concerns on 1st Embodiment of this invention. It is a figure which shows the operation | movement flow of the perturbation learning apparatus which concerns on this embodiment. It is a figure which shows typically the structure of the propagation part of the perturbation learning apparatus shown in FIG. It is a figure which shows the structural example of one learning part which comprises the learning part of the perturbation learning apparatus shown in FIG. It is a figure which shows the function structural example of the perturbation learning apparatus which concerns on 2nd Embodiment of this invention.

Hereinafter, embodiments of the present invention will be described with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated.

[First Embodiment]
FIG. 1 shows a functional configuration example of the perturbation learning device 1 according to the first embodiment. FIG. 2 shows an operation flow of the perturbation learning device 1.

  The perturbation learning device 1 includes a branching unit 10, a propagation unit 20, a perturbation propagation unit 30, a learning unit 40, and a weight update unit 50. The perturbation learning apparatus 1 processes, for example, a learning operation of a hierarchical neural network at high speed.

  The branching unit 10 branches the learning data X composed of a plurality of elements for each element (step S1). The learning data X is given by, for example, a coherent light source. The branching unit 10 is an optical branching device in this example.

  It is assumed that the learning data X is, for example, image data of a specific person's face of 100 pixels. For example, the branching unit 10 branches 1 × 100 pieces (n pieces) of 100-pixel learning data.

  The branch part 10 is constituted by, for example, a quartz flat optical waveguide (PLC), a flexible polymer optical waveguide, or the like. The optical signal emitted from the branch unit 10 is input to the propagation unit 20 as parallel light by a collimator lens provided at each branch output. In FIG. 1, the description of the collimator lens is omitted. The wavelength of the optical signal is, for example, 1550 nm.

  The propagation unit 20 obtains filtered data by multiplying each of the elements by a filter coefficient having different weights for each element position depending on the category of the learning data X, and sets an emphasis weight for emphasizing the filtered data for each element of the filtered data. A first output value O obtained by adding the multiplied values is obtained (step S2). The filtered data is output to the perturbation propagation unit 30. The category of the learning data X is the type of the image (for example, a person, a cat, a dog, etc.) when the learning data X is image data, for example.

For each element position, the branched learning data x ₁ ,..., X _j ,..., X _n may be represented by a subscript j on behalf of n pieces. ) Each of them. Each branch output of the branch unit 10 has a one-to-one correspondence with an element (pixel) of the learning data X.

If the learning data X is, for example, a human face, the filter coefficients have different weights for each position of the human face such as eyes, nose, mouth, ears, etc., and coefficients f ₁ ,. f _j ,..., f _n . That is, x _j × f _j is filtered data that makes it easy to filter the characteristic portion to be learned.

For each element of the filtered data, the propagation unit 20 obtains a first output value O obtained by adding a value multiplied by an emphasis weight w _j that emphasizes the filtered data. Details of the propagation unit 20 will be described later.

Perturbations propagating section 30 is stressed to the weight w _j, generates a perturbation weights w _{j +} [delta] plus perturbation coefficient for correcting [delta] a reinforced adjustment weights w _j, corresponding to each of the filtration data (x _j × f _j) Provided. That is, the perturbation propagation unit 30 includes n perturbation propagation units 30 ₁ to 30 _n . Hereinafter, the reference numeral of the perturbation propagation unit 30 is expressed as 30 _j .

The perturbation propagation unit 30 _j multiplies each filtered data (x _j × f _j ) by a perturbation weight w _j + δ to obtain a second output value O _δj (step S3). The n second output values O _δj are output to the learning unit 40.

The learning unit 40 includes the same number of learning units 40 ₁ to 40 _n as the perturbation propagation unit 30 _j . Learning unit 40 _j is first cost value E obtained by subtracting the teaching value T representing the learning objectives from the first output value O, the second cost value by subtracting the teaching value T from each of the second output value O _.delta.j E _.delta.j, A change amount Δw _{j_new} for updating the emphasis weight w _j is obtained from the perturbation coefficient δ and the learning coefficient η representing the gradient of the learning speed (step S4). The enhancement weighting _{w j} to obtain the change amount [Delta] w _{J_new} updating process is repeated until emphasized weights _{w j} is converged by a control unit (not shown) (NO in step S5).

The weight updating unit 50 obtains the enhancement weight w _jnew by adding the change amount Δw _{j_new} to each of the enhancement weights w _j (step S6). The calculation in step S6 is repeated until the emphasis weight w _j converges.

  According to the perturbation learning device 1 of the present embodiment described above, the weight update calculation is processed in parallel, so that the weight learning can be processed at high speed. Hereinafter, each functional component will be described in more detail with reference to the drawings.

(Propagation part)
FIG. 3 shows a more specific configuration example of the propagation unit 20. The propagation unit 20 includes an input filter 21, an enhancement filter 22, an optical coupler 23, and a light receiving unit 24.

The input filter 21 converts the light intensity of each of the learning data w _j to f _j times according to the transmittance of the filter. This effect can be realized by using, for example, an optical attenuation filter using liquid crystal. Since it is impossible in principle that the transmittance of the light attenuating filter is 1 or more, when f _j > 1, the conversion coefficient a> 1 may be used and the conversion may be made as 1> f _j / a. In this case, it may be converted to a times, for example, by the learning unit 40 _j or the like at a later stage.

  The input filter 21 can be composed of an electro-optic modulator, an acousto-optic modulator, a MEMS mirror, or the like that is an optical attenuator. An optical amplifier may be used. In that case, a semiconductor optical amplifier (SOA) which is an optical amplifier can be used.

The input filter 21 outputs filtered data obtained by multiplying the learning data x _j by a filter coefficient f _j having a different weight for each position of the learning data element depending on the category of the learning data X. The input filter 21 is provided in the propagation unit 20 in advance corresponding to the category of the data X to be learned. Alternatively, the attenuation factor of the electro-optic modulator may be set every time learning is performed.

The enhancement filter 22 can be composed of an electro-optic modulator, which is an optical attenuator, like the input filter 21. The enhancement weight w _j of the enhancement filter 22 is a value _obtained by adding the change amount Δw _{j_new} obtained by the weight update unit 50 to the previous enhancement weight w _jord .

The initial weight w _j is set in advance. The initial weight w _j is updated repeatedly and changes to the convergence value.

The optical coupler 23 performs a sum operation on data (x _j × f _j × w _j ) obtained by emphasizing (product operation) the filtered data output from the enhancement filter 22. In this case, for example, data obtained by product operation by wavefront coupling using LCOS (Liquid Crystal On Silicon) may be optically coupled. Or you may couple | bond using optical waveguides, such as PLC.

  The optical coupler 23 outputs a first output value O expressed by the following equation.

  The light receiver 24 converts the first output value O into a current value.

The perturbation propagation unit 30 _j can be configured by deleting the input filter 21 and the optical coupler 23 of the propagation unit 20. That is, the perturbation propagation unit 30 _j obtains the second output value O _δj by multiplying each filtered data by the perturbation weight w _j + δ. Therefore, the perturbation propagation unit 30 _j can be configured by a combination of the enhancement filter 22 and the light receiving unit 24. The notation of the perturbation propagation unit 30 _j corresponding to FIG. 3 is omitted.

(Learning Department)
FIG. 4 shows a more specific configuration example of the learning unit 40 _j . The learning unit 40 _j can be configured by an analog electric circuit, for example.

The learning unit 40 _j includes three differential amplifiers 41 _j , 42 _j , 43 _j , and a calculation unit 44 _j . As is clear from the reference number _j , the same number of learning units 40 _j as the perturbation propagation units 30 _j are provided.

The differential amplifier 41 _j calculates a first cost value E (OT) that is a difference between the first output value O output from the propagation unit 20 and the teacher value T representing the learning target input from the outside. . The differential amplifier 42 _j calculates a second cost value E _δj that is a difference between the second output value O _δj output from the perturbation propagation unit 30 _j and the teacher value T. Differential amplifier 43 _j calculates the first cost value E and the difference E-E _.delta.j the second cost value E _.delta.j.

The calculation unit 44 _j receives, as inputs, a perturbation coefficient δ for correcting the emphasis weight w _j input from the outside, a learning coefficient η representing the gradient of the learning speed, and a difference EE _δj output by the differential amplifier 43 _j . A change amount w _{j_new} for updating the emphasis weight w _j is calculated by the following equation.

[Second Embodiment]
FIG. 5 shows a functional configuration example of the perturbation learning device 2 according to the second embodiment. The perturbation learning device 2 is different from the perturbation learning device 1 (FIG. 1) in that a determination unit 60 is provided.

The determination unit 60 outputs a match signal indicating that the category of the learning data X matches the category of the filter coefficient f _j when the first cost value O is smaller than the threshold, and the first cost value O is less than the threshold. When it is larger, a mismatch signal indicating that the category of the learning data X does not match the category of the filter coefficient f _j is output.

  According to the perturbation learning device 2, it is possible to determine whether or not the input learning data X matches the category to be learned. Therefore, the perturbation learning device 2 can be used for image identification, for example.

(Learning speed simulation)
In order to confirm the effect of the present embodiment, the learning speeds of the comparative example and the perturbation learning device 1 (FIG. 1) were compared. The learning speed was compared by comparing the number of parameter updates CUPS (Connections Updated Per Second) that can be rewritten per second in the hardware of the neural network.

  As simulation conditions, the neural network was assumed to have a scale of 1000 neurons and 10000 neurons with a total of 10 layers. The number of clocks required to execute the four arithmetic operations is assumed to be 2 clocks for sum operation, 2 clocks for subtraction operation, 4 clocks for product operation, 8 clocks for division operation, and 16 clocks for exponent operation. The number of weight rewrites by the CPU of the comparative example was estimated to be 1.8 MCUPS.

On the other hand, there are 10 learning units 40 _j , the delay time of the input filter 21 is 0.1 ns, the delay time of the enhancement filter 22 is 0.1 ns, the delay time of the light receiving unit 24 is 1 ns, and the delay time of the learning unit 40 _j is 0.1 ns. The delay time of the control unit that controls the repetitive calculation is assumed to be 1 ns. In this case, the learning speed of the perturbation learning device 1 was calculated as 100 MCUPS.

  Thus, the perturbation learning device 1 of the present embodiment was able to increase the learning speed by about 55 times.

  As described above, according to the perturbation learning device 1 of the present embodiment, weight learning can be processed at high speed. Further, according to the perturbation learning device 2 of the present embodiment, it can be determined whether or not the learning data X matches a predetermined category.

  The perturbation learning devices 1 and 2 have been described using an example in which they are configured using optical components. When light is used as a carrier, an optical wiring can be formed in a three-dimensional space, and there are few demerits such as a short circuit and wiring.

  Note that the present invention is not limited to this example. You may comprise the perturbation learning apparatuses 1 and 2 of this embodiment, without using an optical component. Light is a type of electromagnetic wave. Therefore, for example, the branching unit 10 can be configured by an electromagnetic wave branching device. Moreover, the optical coupler 23 can also be comprised with an electromagnetic wave coupler. Thus, the present invention is not limited to the above-described embodiment, and can be modified within the scope of the gist thereof.

  The present embodiment can be applied to, for example, a learning device for a hierarchical neural network, and can be used in fields such as optical computing.

1, 2: Perturbation learning device 10: Branching unit 20: Propagating unit 21: Input filter 22: Weight filter 23: Optical coupler 24: Light receiving unit 30, 30 ₁ , 30 _j , 30 _n : Perturbation propagation unit 40, 40 ₁ , 40 _j , 40 _n : Learning unit 50: Weight update unit

Claims (3)

A branching unit for branching learning data composed of a plurality of elements for each element;
Filtered data is obtained by multiplying each of the elements by a filter coefficient whose weight for each element position varies depending on the category of the learning data, and an emphasis weight for emphasizing the filtered data is multiplied for each element of the filtered data. A propagation unit for obtaining a first output value obtained by adding the values;
A perturbation weight obtained by adding a perturbation coefficient for correcting the emphasis weight to the emphasis weight is generated, provided corresponding to each of the filtered data, and a second output value is obtained by multiplying the filtered data by the perturbation weight. A perturbation propagation section;
The same number as the perturbation propagation unit, a first cost value obtained by subtracting a teacher value representing a learning target from the first output value, a second cost value obtained by subtracting the teacher value from each of the second output values, A learning unit that repeats calculation for obtaining a change amount for updating the weighting weight from a perturbation coefficient and a learning coefficient that represents a gradient of the learning speed;
A perturbation learning device comprising: a weight update unit that obtains an enhancement weight by adding the amount of change to each of the enhancement weights. A determination unit for determining the magnitude of the first cost value;
The determination unit outputs a match signal indicating that the category of the learning data matches the category of the filter coefficient when the first cost value is smaller than a threshold value, and the first cost value is less than the threshold value. The perturbation learning apparatus according to claim 1, wherein when it is large, a perturbation learning device is output that indicates that the category of the learning data does not match the category of the filter coefficient. A perturbation learning method performed by a perturbation learning device,
Branching learning data consisting of a plurality of elements for each element,
A value obtained by multiplying each of the elements by a filter coefficient whose weight for each position of the element differs depending on the category of the learning data, obtaining filtration data, and multiplying an emphasis weight for emphasizing the filtration data for each element of the filtration data To obtain a first output value obtained by adding
A perturbation weight obtained by adding a perturbation coefficient for correcting the enhancement weight to the enhancement weight is generated, provided corresponding to each of the filtered data, and a second output value is obtained by multiplying the filtered data by the perturbed weight. ,
A first cost value obtained by subtracting a teacher value representing a learning target from the first output value, a second cost value obtained by subtracting the teacher value from each of the second output values, the perturbation coefficient, and a learning speed gradient. From the learning coefficient, the calculation for obtaining the amount of change for updating the weighting is repeated,
A perturbation learning method, wherein the enhancement weight is obtained by adding the change amount to each of the enhancement weights.

Images