JP2018116420A - Perturbation learning device and method thereof - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a perturbation learning device capable of processing a weight learning at fast speed.SOLUTION: A perturbation learning device includes: a branching unit 10 that outputs not-to-be-learned data X acquired by branching, element by element, learning data containing n number of elements, and to-be-learned data Xacquired by branching, element by element, from the header of the learning data by m times for each j pieces; a weight computing unit 20 that calculates pieces of weight learning data Dto Dby multiplying each element of the not-to-be-learned data X by a corresponding weight coefficient; a perturbation learning unit 30 that calculates j pieces of perturbation weight learning data by multiplying the to-be-learned data Xby a corresponding perturbation weight coefficient; a second branching unit 40 that branches each piece of the weight learning data into j+1 pieces; a coupling unit 50 that calculates a first output value O summing all pieces of the weight learning data branched into j+1 pieces, and j number of second output values Oδacquired by substituting any one of the j+1 pieces of weight learning data by the perturbation weight learning data corresponding to such weight learning data; and a learning unit 60 that repeats a calculation of updating each weight coefficient for each j piece by a unit of m times.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a perturbation learning apparatus and method for performing large-scale parallel computation of perturbation learning at high speed.

  Perturbation learning is a learning algorithm for an analog computing unit. For example, a parameter correction amount is calculated by changing a parameter to be learned such as a synaptic weight of a neural network by a minute amount and monitoring an influence on an error function.

  For example, Non-Patent Document 1 discloses a simultaneous perturbation optimization method. Further, for example, Non-Patent Document 2 discloses a multi-frequency vibration learning 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 Hiroshi Miyao et al., "Multi-frequency vibration learning method for neural network and learning LSI using it", IEICE Transactions A, Vol.179-A, No.4, pp.917-929 April 1996

  In order to perform perturbation learning at high speed, parallel processing of weight update is important. In the method disclosed in Non-Patent Document 2, parallel processing is performed by frequency-multiplexing signals. However, the number of multiplexed frequencies is limited to several hundred due to problems such as bandwidth and crosstalk.

  On the other hand, perturbation learning of a neural network is performed by updating the weight of a synapse that connects neurons. In deep learning, which is one of machine learning technologies based on neural networks, the number of learning data and the number of synapses is large, and the number of synapses may exceed several hundred million.

  Therefore, there is a problem that sufficient speedup cannot be expected in the parallel processing of several hundred scales using the conventional frequency multiplexing.

  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.

  The perturbation learning device according to an aspect of the present embodiment includes learning non-target data obtained by branching learning data composed of n elements for each element and j elements from the beginning of the learning data divided into m times for each element. A branch unit that outputs the branched learning target data, a weight calculation unit that calculates weight learning data by multiplying a corresponding weighting factor for each element of the learning non-target data, and the learning unit data corresponding to the learning target data A perturbation weight calculation that generates a perturbation weight coefficient by adding a perturbation coefficient that corrects the weight coefficient to the weight coefficient, and multiplies the corresponding learning object data by the perturbation weight coefficient. , A second branching unit for branching each of the weight learning data into j + 1 pieces, a first output value obtained by summing all of the j + 1 pieces of weight learning data, and j + 1 weight learning data. A combining unit that calculates j second output values that are totaled with all other weight learning data by replacing any one of the above with the perturbation weight learning data corresponding to the weight learning data; The gist is provided with a learning unit that is provided and repeats the calculation of updating the weighting coefficient for every j pieces from the first output value and the second output value in units of m times.

  The perturbation learning device according to another aspect of the present embodiment includes wavelength multiplexed light including n wavelengths, multi-wavelength light including j + 1 wavelengths, and single-wavelength light having a wavelength not included in the multi-wavelength light. A learning target signal obtained by dividing the multi-wavelength light by j pieces of learning data from the beginning of n pieces of learning data and giving a weight corresponding to the single wavelength light. A learning data generation unit that generates a target signal; and an optical branching unit that splits the learning target signal into two pieces of j single-wavelength light having different wavelengths and j multi-wavelength light including other wavelengths; , A first weight calculator that calculates first weight learning data by multiplying each of the j multi-wavelength lights by a corresponding weight coefficient, and a weight coefficient corresponding to each of the j single-wavelength lights, Generating a perturbation weighting coefficient by adding a perturbation coefficient for correcting the weighting coefficient; A perturbation weight calculator that calculates the perturbation weight learning data by multiplying the corresponding single wavelength light by a number, and a second weight learning data that multiplies each of the learning non-target signals by a corresponding weight coefficient. A two-weight operation unit, j first weight learning data and j perturbation weight learning data as inputs, and a first combined value obtained by combining all the j first weight learning data; J second combination values in which any one of the first weight learning data is replaced with the perturbation weight learning data corresponding to the first weight learning data and combined with all the other first weight learning data; , A first output value obtained by combining the first combined value and nj second weight learning data, and each of the j second combined values and nj Optical coupling for generating a second output value by coupling the second weight learning data And j learning units that repeat the calculation for updating the weighting factor for each j from the first output value and the second output value in units of m times. .

  A perturbation learning method according to an aspect of the present embodiment is a perturbation learning method performed by the perturbation learning device, wherein the branching unit includes learning non-target data obtained by branching learning data including n elements for each element. , Output learning target data that is divided into m pieces j times from the beginning of the learning data, and the weight calculation unit multiplies the corresponding weight coefficient for each element of the learning non-target data. The weight learning data is calculated, and a perturbation weight calculation unit generates a perturbation weight coefficient obtained by adding a perturbation coefficient for correcting the weight coefficient to the weight coefficient corresponding to the learning target data, and the perturbation weight coefficient is Multiplying the corresponding learning target data to calculate j perturbation weight learning data, the second branching unit branches each of the weighted learning data into j + 1 pieces, and the combining unit branches into j + 1 weights Learning The first output value obtained by summing all of the data and j + 1 weight learning data is replaced with the perturbation weight learning data corresponding to the weight learning data and summed with all other weight learning data. The j second output values are calculated, and the j learning units update the weighting coefficient for each j from the first output value and the second output value. The gist is to repeat it in units of times.

  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 the connection relationship between the branch part-coupling | bond part of the perturbation learning apparatus shown in FIG. It is a figure which shows the example of a 1st output value and a 2nd output value. It is a figure which shows the more specific structural example of the learning part shown in FIG. It is a figure which shows the modification of the learning part 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. It is a figure which shows the more specific structural example of the learning data generation part shown in FIG. It is a figure which shows the more specific structural example of the branch part shown in FIG. It is a figure which shows typically the effect | action of the wavelength circulator shown in FIG. It is a figure which shows typically the effect | action of the optical coupler and optical splitter which comprise the optical coupling part shown in FIG. It is a figure which shows typically the effect | action of the optical coupler which comprises the optical coupling part shown in FIG. It is a figure which shows the comparison result with a comparative example, (a) shows the power consumption of a filter, FIG.5 (b) is a figure which shows the number of components.

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 weight calculation unit 20, a perturbation weight calculation unit 30, a second branching unit 40, a combining unit 50, and a learning unit 60. The perturbation learning apparatus 1 processes, for example, a learning operation of a hierarchical neural network at high speed.

The branching unit 10 includes learning non-target data X obtained by branching learning data composed of n elements for each element, and learning target data X _j branched for each element j times from the beginning of the learning data. Is output (step S1). The learning non-target data X is a vector having n components corresponding to pixels if the learning target is an image (formula (1)).

The weight calculator 20 calculates the weight learning data D _{1 to} D _n by multiplying the corresponding weight coefficient for each element of the learning non-target data X (step S2). The weight calculator 20 calculates a neuron value to be connected to the learning non-target data X. The synaptic weight representing the strength of connection between neurons is a vector having n components as in the learning non-target data X (formula (2)).

The perturbation weight calculation unit 30 generates a perturbation weight coefficient obtained by adding a perturbation coefficient δ for correcting the weight coefficient to a weight coefficient corresponding to the learning target data X _j , and the perturbation weight coefficient is converted to the corresponding learning target data X _By multiplying _j , j perturbation weight learning data Dδ _i are calculated (step S3). The perturbation weight coefficient is a vector having j components (formula (3)).

The second branching unit 40 branches each of the weight learning data D _{1 to} D _n into j + 1 pieces (step S4). That is, the second branching unit 40 branches one weight learning data D ₁ into J + 1 D ₁ ′. Similarly, the weight learning data D _i and the weight learning data D _n are branched into J + 1 D _i ′ and D _n ′.

The combining unit 50 includes a first output value O (equation (4)) obtained by summing all of the weight learning data D ₁ ′ to D _n ′ branched to j + 1 by the second branching unit 40 and j + 1 weight learning. one either data _D 1 '~D _n', and all other weights training data instead placed perturbation weight learning data Dδ ₁ ~Dδ _n corresponding to heavy saw training data _D 1 'to D _n' The total j second output values Oδ _{1 to} Oδ _j (formula (5)) are calculated (step S5).

The learning unit 60 is provided with the same number of j as the weighting coefficients simultaneously calculated by the perturbation weight calculating unit 30 and updates the weighting factor w _i every j from the first output value O and the second output value Oδ _i. Is repeated in units of m times (step S6). The calculation for _obtaining the weighting factor w _{i_new} to be updated is repeated until the value converges (No in step S7).

When the weight coefficient _{wi_new} converges, the weight coefficient is updated to the converged value (step S8). The processes in steps S1 to S8 described above are repeated in units of m times.

  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. Further, since the number of learning data to be learned at a time is limited to j, the number of component parts of the perturbation learning device 1 can be reduced. Hereinafter, each functional component will be described in more detail with reference to the drawings.

(Branch part)
In FIG. 3, the more specific example of a connection of the branch part 10-the coupling | bond part 50 is shown. The branching unit 10 includes n optical demultiplexers 10 ₁ to 10 _n and a multiplexer 11.

The optical demultiplexers 10 ₁ to 10 _n each branch the learning data composed of n elements into two parts, and output _one A _{1 to An} to the second branch unit 40 and the other A _{1 to An.} Is output to the multiplexer 11. The multiplexer 11 selects learning data A _{mj + 1 to} A _{(m + 1) j} in the range of mj + 1 to (m + 1) j according to the value of m given from the learning unit, and outputs it to the perturbation weight calculation unit 30.

For example, when n = 14 and j = 5, the learning target data X _j is X _j = x ₁ , x ₂ , x ₃ , x ₄ , x ₅ , X _j = x ₆ , x ₇ , x _8. , X ₉ , x ₁₀ , X _j = x ₁₁ , x ₁₂ , x ₁₃ , x ₁₄ are output in three times. m is an integer of 0 or more, and X _j can be expressed by the following equation.

Optical demultiplexer ₁₀ 1 to 10 _n, for example the light intensity of one _A 1 to A _n and j + 1 minute j other _{A mj + 1 ~A (m +} 1) the light intensity of _j to 1 j + 1 minute. This is because the weight calculation unit 20 in the subsequent stage refers to the learning data X redundantly, the number of times the learning target data X _j is generated.

  In the present embodiment, the number of component parts is reduced by reducing the number of learning data to be learned to j. A specific example of the reduced number of component parts will be described later.

(Weight calculation part)
The weight calculation unit 20 can be composed of, for example, n light attenuation filters 20 _{0 to} 20 _n using liquid crystal. In that case, the weight w _i is given by the transmittance of the light attenuation filter. Learning is input to the weight calculating unit 20 with respect to non-object data _A 1 to A _n, each of the learning data X by the transmittance of the light attenuating _{filter, w 1, w 2, ...} , w i, ..., w n times Convert to

Weight learning data output from the weight calculation section 20 is a product operation value of the learning data X and the weight w _i, is denoted as D _i (Equation (7)). i is an integer of 1 to n.

When an optical attenuation filter is used as the weight calculation unit 20, an electro-optic (EO) modulator, an acousto-optic (AO) modulator, or a MEMS mirror can be used. The transmittance value of the light attenuation filter is not affected by the light intensity of the input signal. Therefore, the transmittance can be controlled relatively easily, and the difference between the desired transmittance and the actually applied transmittance can be reduced to reduce the calculation error. In addition, it is impossible in principle that the transmittance of the light attenuation filter is 1 or more. Therefore, when w _i > 1, the conversion coefficient q> 1 may be used to convert 1> wi / q = Wi ′> 1 / q. Specifically, the signal level may be converted to q times by the learning unit 60 in the subsequent stage.

Further, the weight calculation unit 20 may be configured with an optical amplifier. In that case, a semiconductor optical amplifier (SOA) can be used. The optical amplifier amplifies a minute optical signal and outputs an observable light intensity. Therefore, even if the scale of the neural network increases, n increases, and the light intensity given to the elements of the learning data becomes minute, the product operation value of the learning data X and the weight w _i can be calculated. In addition, it is impossible in principle that the amplification factor of the optical amplifier is 1 or less. Therefore, when w _i <1, the conversion coefficient q ′ <1 is used. The signal level conversion in this case is the same as when the above-described optical attenuation filter is used.

(Perturbation weight calculator)
The perturbation weight calculation unit 30 multiplies the input learning target data X _j by a perturbation weight coefficient to calculate perturbation weight learning data Dδ _i (formula (8)). The perturbation weight coefficient is obtained by adding a perturbation coefficient δ for correcting the weight coefficient to the weight coefficient. The perturbation weight calculation unit 30 can be configured by, for example, an optical attenuation filter or an optical amplifier, similarly to the weight calculation unit 20.

The perturbation weight calculator 30 calculates perturbation weight learning data Dδ _i according to the value of m. For example, when n = 14 and j = 5, when m = 0, the perturbation weight learning data Dδ ₁ obtained by multiplying the learning target data x _{1 to} x ₅ by the perturbation weight coefficients w ₁ + δ to w ₅ + δ, respectively. to calculate the ~Dδ _5. When m = 1, perturbation weight learning data Dδ _{6 to} Dδ ₁₀ are calculated by multiplying the learning target data x _{6 to} x ₁₀ by the perturbation weight coefficients w ₆ + δ to w ₁₀ + δ. When m = 2, perturbation weight learning data Dδ _{11 to} Dδ ₁₄ are calculated by multiplying the learning target data x _{11 to} x ₁₄ by the perturbation weight coefficients w ₁₁ + δ to w ₁₄ + δ.

(Second branch)
The second branching unit 40 branches each of the weight learning data D _* into j + 1 pieces (step S4). The second branching unit 40 branches the weight learning data D ₁ into J + 1 D ₁ ′. Similarly, the weight learning data D _i and the weight learning data D _n are branched into J + 1 D _i ′ and D _n ′.

  The second branching unit 40 can be configured by an optical demultiplexer as with the branching unit 10. As the optical demultiplexer, a quartz planar optical waveguide (PLC), a flexible polymer optical waveguide, a fusion-stretch fiber coupler, or the like can be used.

(Joining part)
The combining unit 50 includes a first output value O (equation (4)) obtained by summing all of the weight learning data D ₁ ′ to D _n ′ branched to j + 1 by the second branching unit 40 and j + 1 weight learning. one one of the data D _*, heavy unlearned data D _* corresponding to perturbation weight learning data d? _* every place of the j-number of the sum of all other weighting the learning data D _* second output value Oderuta _* (Equation (5)) is calculated.

FIG. 4 shows an example of the first output value O and the second output value Oδ _* . FIG. 4 shows a case where the number n of elements of learning data X is n = 14 and the number j of learning object data to be learned at a time is j = 5.

In the case of this example, the combining unit 50 outputs the first output value O and the second output values Oδ _{1 to} Oδ ₅ in three times. The first output values O that are output in three steps are all the same value.

On the other hand, the second output value Oδ _i is different in all the values output in three steps. The first Oδ ₁ is a value obtained by replacing D ₁ (x ₁ w ₁ ) with Dδ ₁ and adding all the other weight learning data. In the first Oδ ₂ , D ₂ (x ₂ w ₂ ) is replaced with Dδ ₂ . Similarly, the first Oδ ₅ is a value obtained by replacing D ₅ (x ₅ w ₅ ) with Dδ ₅ and adding it to other weight learning data.

The second Oδ ₁ is a value obtained by replacing D ₆ (x ₆ w ₆ ) with Dδ ₆ and adding all the other weight learning data. In the second Oδ ₂ , D ₇ (x ₇ w ₇ ) is replaced with Dδ ₇ . Similarly, the second Oδ ₅ is a value obtained by adding D ₁₀ (x ₁₀ w ₁₀ ) to Dδ ₁₀ and adding it to other weight learning data.

The third Oδ ₁ is a value obtained by replacing D ₁₁ (x ₁₁ w ₁₁ ) with Dδ ₁₁ and adding all the other weight learning data. In the third Oδ ₂ , D ₁₂ (x ₁₂ w ₁₂ ) is replaced with Dδ ₁₂ . Similarly, the third Oδ ₅ has no output because D ₁₅ (x ₁₅ w ₁₅ ) does not exist in this example.

(Learning Department)
FIG. 5 shows a more specific configuration example of the learning unit 60 _i . The learning unit 60 _i can be configured by an analog electric circuit, for example.

The learning unit 60 _i includes three differential amplifiers 61 _i , 62 _i , 63 _i , 65 _i , and a calculation unit 64 _i . The learning units 60 _i are provided in the same number as, for example, the optical attenuation filters constituting the perturbation weight calculation unit 30.

The differential amplifier 61 _i calculates a first cost value E (OT) that is a difference between the first output value O output from the combining unit 50 and the teacher value T that represents the learning target input from the outside. . The differential amplifier 62 _i calculates a second cost value E _δi that is a difference between the second output value O _δi output from the combining unit 50 and the teacher value T. Differential amplifier 43 _i calculates a first cost value E and the difference E-E _.delta.i the second cost value E _.delta.i.

The calculation unit 64 _i receives, as inputs, a perturbation coefficient δ that corrects an emphasis weight w _i that is input from the outside, a learning coefficient η that represents the gradient of the learning speed, and the difference EE _δi that is output from the differential amplifier 63 _i . A change amount Δw _{i_new} for updating the weight w _i is calculated by the following equation.

Differential amplifier 65 _i outputs the weights _{w i - new} updating by adding the change amount [Delta] w _{J_new} once before the weight _{w i_old.}

It is also possible to provide a second multiplier 67 _i between the first multiplier 66 _i, and a differential amplifier 62 _i and 63 _i between the differential amplifier 61 _i and 63 _i, as shown in FIG. The first multiplier 66 _i converts the first cost value E (OT) to a positive value. The first multiplier 66 _i may be configured by a square calculator that multiplies the first cost value E (OT) by the same value, or when the first cost value E (OT) is negative. Alternatively, the multiplier may be multiplied by −1. The same applies to the second multiplier 67 _i .

By providing the first multiplier 66 _i and the second multiplier 67 _i , the calculation accuracy of the change amount Δw _{i_new} can be increased.

[Second Embodiment]
FIG. 7 shows a functional configuration example of the perturbation learning device 2 according to the second embodiment. The perturbation learning device 2 is a learning device that performs weight learning of a neural network, for example, using wavelength multiplexed light.

  The perturbation learning device 2 includes a learning data generation unit 70, an optical branching unit 12, a first weight calculation unit 21, a second weight calculation unit 22, a perturbation weight calculation unit 31, a wavelength circulator 80, an optical coupling unit 90, and a learning unit. 60.

  The learning data generation unit 70 branches the wavelength multiplexed light including n wavelengths into multi-wavelength light including j + 1 wavelengths and single-wavelength light for each wavelength not included in the multi-wavelength light. A learning target signal in which j pieces of learning data from the beginning of the n pieces of learning data are provided to the light divided m times and a learning non-target signal to which a weight corresponding to the single wavelength light is given are generated.

  Wavelength multiplexed light is generated using a light source having n wavelength components. As the light source, it is only necessary to generate light having n wavelength components. Therefore, n single-wavelength coherent light sources may be prepared, and the light emitted from each light source may be generated and combined by an optical coupler. In addition, a multi-wavelength light source or an incoherent light source may be used. The learning data generation unit 70 will be described in detail later.

The optical branching unit 12 receives a learning target signal including j + 1 wavelengths from j single-wavelength lights B _{1 to} B _j having a certain wavelength and j multi-wavelength lights A _{1 to} A _j including other wavelengths. Two branches. For example, the wavelength of j single-wavelength lights is the shortest wavelength λ ₀ , and the j multi-wavelength lights include wavelengths λ _{1 to} λ _j . The optical branching unit 12 can be composed of wavelength branching elements such as AWG, diffraction grating, prism, and WDM coupler, for example.

The first weight calculation unit 21 calculates the first weight learning data D _{1 to} D _j by multiplying the weight coefficient corresponding to each of the j multi-wavelength lights. The first weight learning data D _{1 to} D _j can be expressed by the following equation. Here, i is an integer of 1 to j.

Here, A ₁ is multi-wavelength light including wavelengths λ _{1 to} λ _j . Thus, for example, D ₁ is a value obtained by multiplying the multi-wavelength light by the weight w _{mj + 1} . This effect can be realized by, for example, an optical attenuation filter, similarly to the weight calculation unit 20 described in the first embodiment.

The perturbation weight calculation unit 31 generates a perturbation weight coefficient obtained by adding a perturbation coefficient δ for correcting the weight coefficient to a weight coefficient corresponding to each of the j single-wavelength lights, and the perturbation weight coefficient The perturbation weight learning data is calculated by multiplying the wavelength light. The perturbation weight learning data Dδ _{1 to} Dδ _j can be expressed by the following equations. Here, B ₁ is single wavelength light having a wavelength of λ ₀ .

The second weight calculation unit 22 calculates second weight learning data D _{j + 1 to} D _n by multiplying each learning non-target signal by a corresponding weight coefficient. For example, the second weight learning data D _{j + 1} can be expressed by the following equation. In the second weight learning data D _n , the subscript of Expression (12) changes from j + 1 → n.

The wavelength circulator 80 receives j pieces of first weight learning data D _{1 to} D _j and j pieces of perturbation weight learning data Dδ _{1 to} Dδ _j, and inputs the j pieces of first weight learning data D _{1 to} D _j . a first coupling value D1 which all bonds, one or of the j first weight learning data _D 1 to D _j, perturbation weight learning data d? ₁ corresponding to the first weight learning data _D 1 to D _j ˜Dδ _j are output, and _j second combined values D2 _{1 to} D2 j combined with all the other first weight learning data are output.

  The wavelength circulator 80 can be composed of, for example, AWG. When the multi-wavelength light is input, the wavelength circulator 80 branches the multi-wavelength light for each wavelength and outputs the multi-wavelength light to an output port at a distance that is an integer multiple of the wavelength away from the input port to generate multi-wavelength light. do. A product operation value to be summed is selected using this action. The wavelength circulator 80 will be described in detail later.

The optical coupling unit 90 includes a first output value O obtained by combining the first combined value D1 and n−j pieces of second weight learning data D _{j + 1 to} D _n , second combined values D2 _{1 to} D2 _j , and the first output value O. Second output values Oδ _{1 to} Oδ _j obtained by combining the two weight learning data D _{j + 1 to} D _n are generated. The first output value O and the second output value Oδ _i can be expressed by the following equations. However, it is a case of j <= n <= 2j.

  The final sum operation is performed by reading the light intensity of the optical signal optically coupled by the wavelength circulator 80 with a light receiver. The notation of the light receiver is omitted.

The learning unit 60 repeats the calculation of updating the weighting factor for each _j from the first output value O and the second output values Oδ _{1 to} Oδ _j in units of m times. The learning unit 60 is the same as that of the first embodiment as is apparent from the reference numerals.

  As described above, according to the perturbation learning apparatus 2 of the present embodiment, the weight update calculation is processed in parallel, so that the weight learning can be processed at high speed. Further, since the number of learning data to be learned at a time is limited to j, the number of component parts of the perturbation learning device 2 can be reduced. Hereinafter, each functional component will be described in more detail with reference to the drawings.

(Learning data generator)
FIG. 8 shows a more specific functional configuration example of the learning data generation unit 70. The learning data generation unit 70 includes a wavelength branching unit 71, a wavelength synthesizing unit 72, an optical branching unit 73, a first learning data filter 74, a second learning data filter 75, and an optical coupling unit 76.

The wavelength branching unit 71 demultiplexes wavelength multiplexed light including n wavelengths into single wavelength light for each wavelength (λ ₀ ,..., Λ _j ,... Λ _n ). The wavelength branching unit 71 is configured by a wavelength branching element such as AWG as described above.

For example, most single-wavelength light shorter to wavelength λ ₀ ~λ _{j J} + 1 single wavelength is inputted to the wavelength combining unit 72, n-j-number of single-wavelength light and the second learning to wavelength λ _{j + 1} ~λ _n Input to the data filter 75.

  The wavelength synthesizer 72 multiplexes single wavelength light from the shortest wavelength demultiplexed by the wavelength branching unit 71 to the j + 1st longest wavelength. The wavelength synthesizer 72 can be configured with a wavelength branching element in the same manner as the wavelength branching unit 71.

  It should be noted that the wavelength branching unit 71 and the wavelength synthesizing unit 72 may be a coupler which is an optical combiner instead of a single set of wavelength branching elements. In the case of using a photo combiner, since it is impossible in principle to combine without loss, the conversion coefficient q> 1 is used, and the signal may be multiplied by q in the subsequent learning unit 60.

  The optical branching unit 73 equally divides the signal output from the wavelength combining unit 72 into j signals. The optical branching unit 73 can be configured by a quartz flat optical waveguide (PLC), a flexible polymer optical waveguide, or the like.

The first learning data filter 74 generates a learning target signal in which learning data is assigned to each of the j signals output from the optical branching unit 73 by dividing the j learning data into m times from the beginning of the n learning data. . The learning data x _{1 to} x _j change as indicated by the following equation depending on m input from the learning unit 60.

When j = 5 and n = 14, when m = 0, x ₁ = x ₁ ,..., x ₅ = x ₅ , and when m = 1, x ₁ = x ₆ ,..., x ₁₀ = x ₅ , m = 2 Then, x ₁ = x ₁₁ ,..., X ₃ = x ₁₄ . In this way, the learning data is given j times m times.

The second learning data filter 75 generates a learning non-target signal in which n−j learning data is added to each of the n−j signals output from the wavelength branching unit 71. When m = 0, x ₆ learning data is assigned to single wavelength light of wavelength λ _{j + 1} and x _n learning data is assigned to single wavelength light of wavelength λ _n .

  The optical coupling unit 76 generates a learning target signal obtained by combining the j output signals of the first learning data filter 74. The learning target signal is output to the optical branching unit 12.

(Branching part of 2nd Embodiment)
FIG. 9 shows a more specific functional configuration example of the optical branching unit 12. Optical branching unit 12 is provided with the wavelength branching unit 12 _1, the wavelength combining unit 12 _2, the first optical branching unit 12 _3, and a second optical branching unit 12 _4.

Relationship between the wavelength branching unit 12 ₁ and the wavelength combining unit 12 ₂ is the same as the relationship between the wavelength branching unit 71 and the wavelength combining unit 72 of the learning data generating unit 70. In this example, the wavelengths of λ ₀ to λ _j included in the learning target signal in the wavelength branching unit 12 ₁ and the wavelength synthesizing unit 12 ₂ are the single wavelength light having the shortest wavelength λ ₀ and multiple wavelengths including other wavelengths. Branch to light.

The first optical splitter 12 ₃ branches the multiple wavelength light in the j multi-wavelength light _A 1 to A _j. Second optical branching unit 12 ₄ branches the single-wavelength light in the j single-wavelength light _B 1 .about.B _j.

(Wavelength circulator)
FIG. 10 schematically shows the operation of the wavelength circulator 80. The wavelength circulator 80 receives the first weight learning data D _{1 to} D _j and the perturbation weight learning data Dδ _{1 to} Dδ _j, and outputs the first combined value D1 and the second combined values D2 _{1 to} D2 _j .

The first combined value D1 is a value obtained by summing (combining) all of the j pieces of first weight learning data D _{1 to} D _j . The second combined value D2 ₁ is obtained by replacing the first weight learning data D ₁ with the perturbation weight learning data Dδ ₁ and other first weight learning data excluding the perturbation weight learning data Dδ ₁ and the first weight learning data D _1. It is a value obtained by summing D _{2 to} D _j .

The second combined value D2 _j replaces the first weight learning data D _j with the perturbation weight learning data Dδ _j , and other first weight learning data excluding the perturbation weight learning data Dδ _j and the first weight learning data D _j. It is a value obtained by summing D _{1 to} D _j−1 .

Thus, the wavelength orbiting unit 80 includes a first coupling value D1 bound all the j first weight learning data _D 1 to D _j, one of the j first weight learning data _D 1 to D _j one, to output the first weight learning corresponding to the data perturbation weight learning data every place the j second coupling value D2 ₁ ~ D2 coupled with all the other first weight learning data _j.

(Optical coupling part)
The optical coupler 90 includes an optical coupler 91, an optical splitter 92, and an optical coupler 93. FIG. 11 schematically shows the operation of the optical coupler 91 and the optical splitter 92. FIG. 12 schematically shows the operation of the optical coupler 93.

The optical coupler 91 combines (combines) the non-learning target signals D _{j + 1 to} D _n to which weights that are not learned this time are given. This time is one of m times.

The optical branching device 92 equally branches the combined non-learning target signals D _{j + 1 to} D _n into j + 1. j + 1 is the number of weights to be learned simultaneously + 1.

The optical coupler 93 couples the non-learning target signals D _{j + 1 to} D _n to the first coupling value D1 and the second coupling values D2 _{1 to} D2 _j to which the current learning weight passed through the wavelength circulator 80 is given. Thus, the first output value O and the second output values Oδ _{1 to} Oδ _j are generated.

The relationship between the first output value O and the second output values Oδ _{1 to} Oδ _j shown in FIG. 12 is the same as the relationship between both in the first embodiment shown in FIG. That is, the present embodiment is the same as the first embodiment except that the present embodiment is different only in that a learning weight is given to the wavelength multiplexed light.

(Contrast with comparative example)
According to the perturbation learning devices 1 and 2 described above, the weight learning can be speeded up. Therefore, for the purpose of confirming 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.

  The simulation conditions were as follows: when the neural network scale was 10,000 neurons, an operation clock of 3 GHz and an 8-core CPU were used. The number of weight rewrites by the CPU of the comparative example was estimated to be 1.8 MCUPS.

  In the configuration of this embodiment, the number of weights simultaneously calculated by the perturbation weight calculation unit 31 is 10 and simulated. The conditions are: the time constant of the spatial light modulator used as the input filter is 0.1 ns, the time constant of the spatial light modulator used as the weight filter is 0.1 ns, the time constant of the avalanche photodiode used as the light receiver, and the weight change amount The delay of the analog circuit to calculate is assumed to be 0.1 ns, and the delay of the controller that commands the weight update amount is assumed to be 1 ns.

  The number of weight rewrites in this embodiment is estimated to be 100 MCPUS. The speed could be increased 50 times or more compared to the comparative example.

  Further, according to the present embodiment, low power consumption can be achieved. FIG. 13 shows an example of the comparison result. FIG. 13A shows an example in which the power consumption of the filter is compared, and FIG. 13B shows an example in which the number of components is compared.

  The scale of the neural network was compared with 10,000 neurons. It was assumed that the number of weights to be learned simultaneously was 10, the filter size was 800 × 600 pixels, the power consumption was 50 VA, and the filter size per data component was 8 pixels.

  Under these conditions, the power consumption of the filter of the comparative example was estimated to be 183.3 kVA. On the other hand, this embodiment was estimated to be 16.7 kVA. Compared with the comparative example, the power consumption could be reduced to about 1/10.

  Moreover, the perturbation learning apparatuses 1 and 2 of this embodiment can also reduce the number of components constituting the apparatus. As shown in FIG. 13B, the number of component parts of this embodiment can be reduced by an order of magnitude compared to the comparative example without depending on neurons. In the second embodiment, since the wavelength circulator 80 is used, the number of parts increases (two-dot chain line). However, the increment is not more than the number of neurons and is constant at a dozen. Therefore, the perturbation learning device 2 can sufficiently reduce the number of component parts.

  As described above, according to the perturbation learning devices 1 and 2 of the present embodiment, weight learning can be speeded up, reduced in power consumption, and reduced in size.

  The perturbation learning device 2 has been described with an example 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. For example, the branching unit 10 of the perturbation learning device 1 has been described with an example of unequal branching, but may be configured with a branching unit 10 that branches equally. In the case of equal branching, an attenuator that attenuates the signal to 1 / j may be provided between the perturbation weight calculation unit 30 and the coupling unit 50. Moreover, you may comprise the perturbation learning apparatus 1 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. The coupling unit 50 can also be configured by 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, 12: Branching unit 10 ₁ , 10 _i , 10 _n : Optical demultiplexer 11: Multiplexer 12, 73: Optical branching unit 12 ₁ : Wavelength branching unit 12 ₂ : Wavelength synthesizing unit 12 ₃ : First optical branching unit 12 ₄ : second optical branching unit 20: weight calculation units 20 ₀ , 20 _i , 20 _n : optical attenuation filter 30: perturbation weight calculation unit 40: second branching unit 50: coupling units 60 and 60 ₁ , 60 _j : learning unit 61 _j , 62 _j , 63 _j , 65 _i : differential amplifier 64 _j : calculation unit 70: learning data generation unit 74: first learning data filter 75: second learning data filter 80: wavelength Circulators 76, 90: optical coupler 91: optical coupler 92: optical splitter

Claims (8)

a learning non-target data obtained by branching learning data consisting of n elements for each element; a branching unit that outputs learning target data branched for each element j times from the beginning of the learning data; and
A weight calculation unit that calculates weight learning data by multiplying a corresponding weighting factor for each element of the learning non-target data,
A perturbation weight coefficient obtained by adding a perturbation coefficient for correcting the weight coefficient to the weight coefficient corresponding to the learning target data is generated, and the perturbation weight coefficient is multiplied by the corresponding learning target data to obtain j perturbation weights. A perturbation weight calculator for calculating learning data;
A second branching unit for branching each of the weight learning data into j + 1 pieces;
The first output value obtained by summing all the j + 1 weight learning data and any one of the j + 1 weight learning data are replaced with the perturbation weight learning data corresponding to the weight learning data. A combining unit that calculates all the weight learning data of j and the total of j second output values;
a perturbation learning comprising: a learning unit which is provided with j and repeats the calculation of updating the weighting coefficient for each j from the first output value and the second output value in units of m times. apparatus. The learning data is an optical signal,
The branching unit sets the light intensity of the n learning non-target data as a signal having a light intensity of j + 1 / j, and the light intensity of the learning target data as a signal having a light intensity of j + 1 /. The perturbation learning device according to claim 1.   The learning unit is provided in the same number as the perturbation weight calculation unit, and the teacher value is subtracted from each of the first cost value obtained by subtracting the teacher value representing the learning target from the first output value and the second output value. The calculation for obtaining the amount of change for updating the weighting coefficient from the second cost value, the perturbation coefficient, and the learning coefficient representing the gradient of the learning speed is repeated in units of m times. The perturbation learning device described in 1. When the learning unit sets the first cost value as E, the second cost value as E _δi , the perturbation coefficient as δ, and the learning coefficient as η,

4. The perturbation learning device according to claim 3, wherein the change amount [Delta] _{wi_new} is calculated by the above formula.   5. The perturbation learning according to claim 4, further comprising: a first multiplier that converts the first cost value into a positive value; and a second multiplier that converts the second cost value into a positive value. apparatus. Wavelength multiplexed light including n wavelengths is branched into multi-wavelength light including j + 1 wavelengths and single-wavelength light having a wavelength not included in the multi-wavelength light, and n learning data is included in the multi-wavelength light. A learning data generation unit that generates a learning target signal to which j pieces of learning data from the top of the learning data are given m times, and a learning non-target signal to which a weight corresponding to the single wavelength light is given,
An optical branching unit that bifurcates the learning target signal into j single-wavelength lights having different wavelengths and j multi-wavelength lights including other wavelengths;
a first weight calculator that calculates first weight learning data by multiplying each of the j multi-wavelength lights by a corresponding weight coefficient;
A perturbation weight coefficient is generated by adding a perturbation coefficient for correcting the weight coefficient to a weight coefficient corresponding to each of the j single-wavelength lights, and the perturbation weight coefficient is multiplied by the corresponding single wavelength light to perturb. A perturbation weight calculator for calculating weight learning data;
A second weight calculator that calculates second weight learning data by multiplying each of the learning non-target signals by a corresponding weight coefficient;
The j first weight learning data and the j perturbation weight learning data are input, and a first combined value obtained by combining all the j first weight learning data, and the j first weight learning data. Wavelength circulator for outputting j second combined values obtained by replacing any one of them with the perturbation weight learning data corresponding to the first weight learning data and combining with all the other first weight learning data When,
The first output value obtained by combining the first combined value and nj second weight learning data, and the j second combined value and nj second weight learning data are combined. An optical coupler for generating a second output value,
a perturbation learning comprising: a learning unit which is provided with j and repeats the calculation of updating the weighting coefficient for each j from the first output value and the second output value in units of m times. apparatus. The learning data generation unit
a wavelength branching unit that demultiplexes wavelength multiplexed light including n wavelengths for each wavelength;
A wavelength combining unit that combines wavelength components from the shortest wavelength demultiplexed by the wavelength branching unit to the j + 1st longest wavelength;
An optical branching unit for equally branching the signal output from the wavelength multiplexing unit into j pieces;
A first learning data filter that assigns the learning data to each of j signals output from the optical branching unit by dividing j times from the beginning of n learning data into m times;
A second learning data filter that generates a learning non-target signal obtained by adding nj learning data to each of the nj signals output from the optical branching unit;
The perturbation learning device according to claim 6, further comprising: an optical coupling unit configured to generate a learning target signal obtained by combining the j output signals of the first learning data filter. A perturbation learning method performed by a perturbation learning device,
A branching unit outputs learning non-target data obtained by branching learning data composed of n elements for each element, and learning target data branched for each element j times from the beginning of the learning data,
A weight calculation unit calculates weight learning data by multiplying a corresponding weighting factor for each element of the learning non-target data,
A perturbation weight calculation unit generates a perturbation weight coefficient obtained by adding a perturbation coefficient for correcting the weight coefficient to the weight coefficient corresponding to the learning target data, and multiplies the corresponding learning target data by the perturbation weight coefficient. To calculate j perturbation weight learning data,
A second branching unit branches each of the weight learning data into j + 1 pieces;
The combining unit adds any one of the first output value obtained by summing all the weight learning data branched into j + 1 pieces and the j + 1 weight learning data to the perturbation weight learning data corresponding to the weight learning data. Replacing all other weight learning data with the summed j second output values,
A perturbation learning method characterized in that j learning units repeat the calculation of updating the weighting factor for each j from the first output value and the second output value in units of m times. .

Images