WO2020044407A1 - Learning device, learning method, and learning program - Google Patents

Abstract

This learning device 10 comprises: a first storage unit 11 which stores parameters of each unit used in inference processing for calculating in a predetermined order an output for the discrimination data of each unit of a discrimination model in which a plurality of layers each composed of one or more units are combined in a layer shape; and a second storage unit 12 which stores parameters to be updated in the learning processing for updating at least a part of the parameters of each unit on the basis of the output of each unit for learning data.

Description

Learning device, learning method and learning program

The present invention relates to a learning device, a learning method, and a learning program.

With the spread of machine learning, further innovations are required to cope with ever-changing situations. In order to cope with a situation that changes from moment to moment, it is necessary to incorporate various raw data acquired in an environment in which it is used into learning as learning data. The learning data is data used for learning the discriminant model.

In learning using machine learning data (machine learning), for example, parameters of arithmetic expressions and discriminants used in a predetermined learning device are adjusted based on the relationship between input and output indicated by the learning data. . The learning device is, for example, a discrimination model that performs discrimination regarding one or a plurality of labels when data is input.

As a relationship between computational resources and computational accuracy in machine learning, for example, Non-Patent Document 1 discloses an example of a learning computation circuit and a learning method for efficiently executing deep learning of a neural network, particularly with low power consumption. Has been described.

In Non-Patent Document 2, in deep learning in CNN (Convolutional Neural Network), a learning range is divided into a plurality of convolutional layers into a layer in which the weight is fixed and a layer in which the weight is updated (extended function layer). An example of a learning method for shortening the learning time by restricting the learning time is described.

{Also, Non-Patent Document 3 describes an optimization example of an accelerator design based on an FPGA (Field-Programmable {Gate} Array}) as an example of a circuit configuration for learning operation in machine learning.

Hereafter, the outline of the learning method will be described. FIG. 10 is an explanatory diagram showing an example of a general learning method and a circuit configuration for learning in a neural network including one or more intermediate layers between an input layer and an output layer.

In the example shown in FIG. 10, the large-scale learning circuit 70 learns the entire neural network as a predetermined discriminant model in order to support a learning algorithm for general use.

吹 き The speech balloon attached to the large-scale learning circuit 70 shown in FIG. 10 schematically describes the processing direction and the processing range in the learning process of the neural network. In the balloon, a unit 71 corresponding to a neuron in the neural network is represented by an ellipse.

{Circle around (7)} The line segment 72 (the line connecting the units 71 shown in FIG. 10) represents the connection between the units 71. Further, an arrow 73 (a rightward thick line arrow shown in FIG. 10) indicates an inference process and a range of the inference process. An arrow 74 (a thick left arrow shown in FIG. 10) indicates a parameter update process and a range of the parameter update process. The parameter update process is an example of a learning process.

FIG. 10 shows an example of a feedforward neural network in which the input to each unit 71 is the output of the unit 71 in the preceding layer. For example, when time-series information is held, the input to each unit 71 may include the output of the unit 71 in the previous layer at the previous time, as in a recurrent neural network.

When the input to each unit 71 includes the output of the unit 71 of the previous layer at the previous time, the direction of the inference processing is also considered to be the direction from the input layer to the output layer (forward direction). It is. The input to each unit 71 is not limited to the above example.

推 The inference processing performed in a predetermined order from the input layer is also called “forward propagation”. On the other hand, the direction of the parameter update processing is not particularly limited. For example, as in the parameter update process shown in FIG. 10, the direction of the parameter update process may be a direction (reverse direction) from the output layer to the input layer.

The parameter updating process shown in FIG. 10 is an example of a process executed by the back propagation method. However, the parameter update processing is not limited to the processing executed by the back propagation method. For example, the parameter update processing may be executed by STDP (Spike {Timing} Dependent {Plasticity}).

限 ら Not limited to neural networks, examples of model learning methods in deep learning include the following learning methods. First, after inputting learning data to the input layer, an inference process of calculating the output of each unit 71 in the forward direction in each layer up to the output layer is performed (forward propagation: see the arrow 73 shown in FIG. 10).

Next, based on an error calculated from an output (final output) from the output layer and a relation between the input and the output indicated by the learning data, a parameter used to calculate the output of each unit 71 in the layer is determined. A parameter update process for updating is performed (back propagation: see arrow 74 shown in FIG. 10). As shown in FIG. 10, the parameter update processing is performed by following each layer from the output layer to the first layer in the reverse direction. The parameter updating process is performed so that the calculated error is minimized.

As shown in FIG. 10, when the entire model is a learning target, the output of each unit 71 in all the layers (first to n-th layers) subsequent to the input layer is calculated by the parameter updating process. The parameters used for updating are updated. The parameter to be updated is, for example, the weight of the connection between the units 71 that connects each unit 71 in the layer and the unit 71 in another layer.

学習 The parameter updating process as described above is repeatedly performed a plurality of times while, for example, the learning data is changed, thereby generating a learned model having a high recognition rate. FIG. 10 shows a large-scale learning circuit 70 that performs the above-described inference processing and parameter updating processing with high calculation accuracy as an example of realizing an arithmetic circuit that performs learning.

FIG. 11 is an explanatory diagram showing an example of input / output of the unit 71 and connection with another unit 71 when focusing on one unit 71. FIG. 11A shows an example of input / output of one unit 71. FIG. 11B shows an example of a connection between the units 71 arranged in two layers.

As shown in FIG. 11A, when four inputs (x ₁ to x ₄ ) and one output (z) are given to one unit 71, the operation of the unit 71 is expressed by, for example, the formula (1A) ).

z = f (u) Expression (1A)
_{However, u = a + w 1 x} 1 + w 2 x 2 + w 3 x 3 + w 4 x 4 ··· formula (1B)

Note that f () in Expression (1A) represents an activation function. A in the formula (1B) represents an intercept. Further, w ₁ to w ₄ in the equation (1B) represent parameters such as weights corresponding to the respective inputs (x ₁ to x ₄ ).

On the other hand, as shown in FIG. 11B, when the units 71 are connected between the layers arranged in two layers, the input to each unit 71 in the layer (x ₁ the output of each unit 71 for _{~ x 4) (z 1 ~} z 4) , for example, be expressed as follows.

z _i = f (u _i ) Equation (2A)
_{However, u i = a + w i} , 1 x 1 + w i, 2 x 2 + w i, 3 x 3 + w i, 4 x 4 ··· formula (2B)

In the expression (2A), i is the identifier of the unit 71 in the same layer (i = 1 to 3 in this example). Further, the intercept a in the equation (2B) can be regarded as a coefficient of a constant term having a value of 1 (that is, one of the parameters).

In the following, equation (2B) is simplified,
u _i = Σ (wi _{, k} * x _k ) (2C)
May be written. In addition, in formula (2C), the intercept a is omitted. Further, k in the expression (2C) represents an input to each unit 71 in the layer, more specifically, an identifier of another unit 71 that performs the input.

When the input to each unit 71 in the layer is only the output of each unit 71 in the preceding layer,
u _i ^(L) = Σ (wi _{, k} ^(L) * x _k ^(L−1) ) Expression (2D)
It is also possible to write

Note that L in Equation (2D) represents a layer identifier. In addition, w _{i, k} in Expression (2D) represents a parameter of each unit i in the L-th layer. More specifically, w _{i, k} corresponds to the weight of the connection between each unit i and another unit k (connection between units 71).

In the following, a function (activation function) for determining an output value of the unit 71 may be simplified as z = Σ (w * x) without distinguishing the unit 71.

集合 The above set of weights is described in vector form as follows:

_{w i = [w i, 1} , w i, 2, ···, w i, k] T ··· formula (3)

Equation (3) is called a weight vector. Further, if an input vector x = [x ₁ , x ₂ ,..., X _k ] ^T , which is a set of inputs of a certain layer, and a weight matrix in which weight vectors are horizontally connected is W, the output vector z is f ( represented by W ^T x). Note that the following relationship holds between the output vector z and the activation function.

_{z = f (u) = [} f (u 1), f (u 2), ···, f (u n)] ··· Equation (4)

In the above example, the calculation in which a certain unit 71 calculates the output z from the input x corresponds to the inference processing in the unit 71. In the inference processing, parameters (for example, weight w) are fixed. The inference process is, for example, a process performed by an operating monitoring system or the like to determine whether an object in an image is a specific object. On the other hand, the calculation for obtaining the parameters of the unit 71 corresponds to the parameter updating process in the unit 71.

FIG. 12 shows an example of an inference apparatus that performs inference processing. FIG. 12 is a block diagram illustrating a configuration example of a general inference apparatus. The inference device 80 illustrated in FIG. 12 includes a weight memory 81, a weight load unit 82, and a calculation unit 83.

The weight memory 81 has a function of storing the weight matrix W. The weight loading unit 82 has a function of loading the weight matrix W stored in the weight memory 81 from the weight memory 81.

The weight loading unit 82 inputs the loaded weight matrix W to the calculation unit 83. The calculation unit 83 has a function of performing the above inference processing using the input weight matrix W.

Next, FIG. 13 shows an example of a learning device that performs a parameter updating process. FIG. 13 is a block diagram illustrating a configuration example of a general learning device. The learning device 90 illustrated in FIG. 13 includes a weight memory 91, a weight load unit 92, a calculation unit 93, and a weight storage unit 94.

The weight memory 91 has a function of storing the weight matrix W. The weight loading unit 92 has a function of loading the weight matrix W stored in the weight memory 91 from the weight memory 91.

The weight load unit 92 inputs the loaded weight matrix W to the calculation unit 93. The calculation unit 93 has a function of performing the above-described parameter update processing using the input weight matrix W.

The operation unit 93 inputs the weight matrix W updated in the parameter update process to the weight storage unit 94. The weight storage unit 94 has a function of writing the weight matrix W updated by the calculation unit 93 into the weight memory 91.

Specifically, the weight storage unit 94 updates the weight matrix W stored in the weight memory 91 to the input weight matrix W. In writing the weight matrix W, the weight storage unit 94 may have a function of temporarily storing the weight matrix W.

YHChen, et.al., "Eyeriss: an Energy-Efficient Reconfigurable Accelerator Deep Convolutional Neural Networks '', in IEEE Jornal of Slid-State Circuits, vol.52, no.1, anJan. 2017, pp.127-138 . Wei. Liu, et.al., “SSD: Single shot MultiBox Detector”, arXiv: 1512.02325v5, Dec. 2016. Chen Zhang, et.al., "Optimizing FPGA-based Accelerator Design for Deep convolutional Neural Networks", In ACM FPGA 2015, pp. 160-170.

を Consider designing a learning device that can execute the above inference processing and learning processing in parallel. When the learning device has only one weight memory that stores the weight matrix W, the learning device cannot execute the inference process and the learning process in parallel.

理由 The reasons are explained below. If there is only one weight memory, the learning device has only one place where the weight matrix W is stored. Therefore, there is a possibility that the learning device uses the updated weight matrix W for the inference processing.

た め Since the inference processing normally uses fixed parameters in the inference processing, if the learning apparatus uses the updating weight matrix W in the inference processing, normal inference processing is not executed. Non-Patent Documents 1 to 3 do not describe a method of executing inference processing and learning processing in parallel.

[Object of the invention]
Therefore, an object of the present invention is to provide a learning device, a learning method, and a learning program that can solve the above-described problem and can execute inference processing and learning processing in parallel.

The learning apparatus according to the present invention is used in an inference process of calculating an output for discrimination data of each unit of a discrimination model in which a plurality of layers each composed of one or more units are combined in a layered manner in a predetermined order. A first storage unit for storing the parameters of each unit, and a second storage unit for storing parameters to be updated in a learning process of updating at least a part of the parameters of each unit based on an output of each unit for learning data And characterized in that:

The learning method according to the present invention is used in the inference process of calculating the output for the discriminating data of each unit of the discriminating model in which a plurality of layers each composed of one or more units are combined in a layered manner in a predetermined order. The parameters of the units are respectively stored in the first storage unit, and the parameters to be updated in the learning process of updating at least a part of the parameters of each unit based on the output of the learning data of each unit are stored in the second storage unit. It is characterized by making it.

The learning program according to the present invention is an inference process in which a computer calculates, in a predetermined order, outputs for discrimination data of each unit of a discrimination model in which a plurality of layers each composed of one or more units are combined in a layered manner. Parameters to be updated in a first storage process for storing parameters of each unit to be used in the first storage unit and a learning process for updating at least a part of parameters of each unit based on an output of each unit for learning data. Is stored in the second storage unit.

According to the present invention, the inference processing and the learning processing can be executed in parallel.

1 is a block diagram illustrating a configuration example of a first embodiment of a learning device according to the present invention. FIG. 3 is a block diagram illustrating a configuration example of a calculation unit 140. FIG. 4 is a block diagram showing another configuration example of the first embodiment of the learning device according to the present invention. It is a flowchart which shows the operation | movement of the inference process and the learning process by the learning apparatus 100 of 1st Embodiment. It is a block diagram showing an example of composition of a 2nd embodiment of a learning device by the present invention. It is a block diagram showing another example of composition of a 2nd embodiment of a learning device by the present invention. It is a flow chart which shows operation of inference processing and learning processing by learning device 200 of a 2nd embodiment. FIG. 3 is an explanatory diagram illustrating an example of a hardware configuration of a learning device according to the present invention. It is a block diagram showing the outline of the learning device by the present invention. FIG. 3 is an explanatory diagram showing an example of a general learning method and a circuit configuration for learning in a neural network including one or more intermediate layers between an input layer and an output layer. FIG. 7 is an explanatory diagram showing an example of input / output of the unit 71 and coupling with another unit 71 when focusing on one unit 71. It is a block diagram showing an example of composition of a general inference device. It is a block diagram showing an example of composition of a general learning device.

Embodiment 1 FIG.
[Description of configuration]
Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram illustrating a configuration example of a first embodiment of a learning device according to the present invention.

As shown in FIG. 1, the learning device 100 includes an inference weight memory 110, an updating weight memory 120, a weight load selection unit 130, a calculation unit 140, a weight storage unit 150, and a weight update control unit 160. , A weight copy unit 170 and a control unit 180.

単 The unidirectional arrows in each block diagram indicate the direction in which data flows. However, the possibility of data flowing bidirectionally at the location where each arrow is described is not excluded.

解決 Since the learning device 100 executes the inference process and the learning process in parallel, the learning device 100 may have a plurality of weight memories as a solution. Hereinafter, the function of each component of the learning device 100 that performs the inference process and the learning process in parallel will be described.

The inference weight memory 110 has a function of storing a weight matrix W (parameter group) used for inference processing. The updating weight memory 120 has a function of storing the weight matrix W used for the learning process.

The weight load selection unit 130 has a function of loading the weight matrix W from one of the inference weight memory 110 and the updating weight memory 120. When the inference processing is performed, the weight load selection unit 130 loads the weight matrix W from the inference weight memory 110.

When the learning process is performed, the weight load selecting unit 130 loads the weight matrix W from the updating weight memory 120. The weight load selection unit 130 inputs the loaded weight matrix W to the calculation unit 140.

The operation unit 140 has a function of performing the above inference processing using the weight matrix W loaded from the inference weight memory 110. The arithmetic unit 140 has a function of performing the above-described learning process using the weight matrix W loaded from the updating weight memory 120. That is, since different weight matrices W are used, the arithmetic unit 140 can execute inference processing and learning processing in parallel.

Specifically, the arithmetic unit 140 calculates in a predetermined order the output for the discriminating data of each unit 71 of the discriminating model in which a plurality of layers each composed of one or more units 71 are combined in a layered manner. Execute the process.

The inference weight memory 110 stores the weight (weight matrix W) of each unit 71 used in the inference processing. The weight of each unit 71 is a parameter of each unit 71 in the present embodiment. The discrimination model is, for example, a neural network.

The arithmetic unit 140 executes a learning process of updating at least a part of the weight of each unit 71 based on the output of each unit 71 for the learning data. The updating weight memory 120 stores the update target weight (weight matrix W) in the learning process.

{The weight load selector 130 loads the weight from the inference weight memory 110 or the updating weight memory 120. The operation unit 140 performs inference processing or learning processing using the loaded weights.

The operation unit 140 inputs the weight matrix W updated in the learning process to the weight storage unit 150. The weight storage unit 150 has a function of writing the weight matrix W updated by the arithmetic unit 140 into the updating weight memory 120.

Specifically, the weight storage unit 150 updates the weight matrix W stored in the updating weight memory 120 to the input weight matrix W. When writing the weight matrix W, the weight storage unit 150 may have a function of temporarily storing the weight matrix W.

That is, the weight storage unit 150 stores the update target weight (weight matrix W) of each unit 71 in the learning process in the updating weight memory 120. The weight storage unit 150 stores the weight matrix W in the updating weight memory 120, so that the updated learning matrix W is used in the next learning process.

When the learning is completed, the arithmetic unit 140 notifies the weight update control unit 160 that the learning is completed. Upon receiving the notification, the weight update control unit 160 copies the weight matrix W stored in the updating weight memory 120 to the inference weight memory 110, and copies the weight matrix W stored in the inference weight memory 110. The weight copy unit 170 is instructed to replace the weight matrix W.

Upon receiving the instruction, the weight copy unit 170 copies the weight matrix W stored in the updating weight memory 120 to the inference weight memory 110, and is copied with the weight matrix W stored in the inference weight memory 110. Replace the weight matrix W.

The case where the learning is completed is a case where a learning process including a weight update process using a predetermined number of learning data is repeatedly executed a predetermined number of times. When the learning is completed, the weight copy unit 170 updates the weight stored in the inference weight memory 110 to the update target weight stored in the updating weight memory 120 corresponding to the weight.

That is, the weight copy unit 170 stores the weight (weight matrix W) of each unit 71 used in the inference processing in the inference weight memory 110. The weighted copy unit 170 stores the weight matrix W in the weight memory 110 for inference, so that inference processing after learning is completed uses the learned model.

The control unit 180 has a function of controlling the weight load selection unit 130, the calculation unit 140, and the weight update control unit 160. The control unit 180 notifies the weight load selection unit 130 and the calculation unit 140 which of the inference process and the learning process is to be executed based on the type of the input image.

(4) When the learning is completed, the control unit 180 notifies the weight update control unit 160 via the arithmetic unit 140 that the learning is completed.

FIG. 2 is a block diagram showing a configuration example of the arithmetic unit 140. As shown in FIG. 2, the arithmetic unit 140 includes an inference weight register 141, an updating weight register 142, an arithmetic unit weight load selection unit 143, an arithmetic unit 144, and an arithmetic unit weight storage unit 145.

When the configuration of the arithmetic unit 140 is the configuration illustrated in FIG. 2, the learning device 100 includes a plurality of arithmetic units corresponding to each unit 71.

The inference weight register 141 has a function of storing the weight w of the unit 71 corresponding to the operation unit 140 in the weight matrix W input from the operation unit 140 and loaded from the inference weight memory 110.

中 The updating weight register 142 has a function of storing the weight w of the unit 71 corresponding to the arithmetic unit 140, of the weight matrix W input from the arithmetic unit 140 and loaded from the updating weight memory 120.

The arithmetic unit weight load selection unit 143 has a function of loading the weight w from one of the inference weight register 141 and the updating weight register 142. When the inference processing is performed, the arithmetic unit weight load selector 143 loads the weight w from the inference weight register 141.

When the learning process is performed, the arithmetic unit weight load selector 143 loads the weight w from the updating weight register 142. The computing unit weight load selection unit 143 inputs the loaded weight w to the computing unit 144.

The arithmetic unit 144 has a function of performing the above-described inference processing operation using the weight w loaded from the inference weight register 141. The arithmetic unit 144 has a function of performing the above-described arithmetic operation for the learning process using the weight w loaded from the updating weight register 142.

The arithmetic unit 144 inputs the weight w updated through the operation for the learning process to the arithmetic unit weight storage unit 145. The arithmetic unit weight storage unit 145 has a function of updating the weight w stored in the updating weight register 142 to the input weight w.

As described above, the arithmetic unit 140 shown in FIG. 2 is an inference weight register 141 that stores a parameter (weight w) of the unit 71 corresponding to the arithmetic unit 140 used in the inference process, and is an update target in the learning process. An updating weight register 142 that stores a parameter (weight w) of the unit 71 corresponding to the arithmetic unit 140.

Note that the inference weight register 141 and the updating weight register 142 may store one column or the like of the weight matrix W instead of one weight w.

Since the arithmetic unit 140 includes the inference weight register 141 and the updating weight register 142, the number of times the weight matrix W is read from the inference weight memory 110 and the updating weight memory 120 is reduced. That is, since the power consumed for reading is reduced, the cost for switching between the inference processing and the learning processing is reduced.

The configuration of the calculation unit 140 of the present embodiment may be a configuration other than the configuration illustrated in FIG. For example, the arithmetic unit 140 may have only one register that stores the weight w of the unit 71 corresponding to the arithmetic unit 140 in the input weight matrix W. When the arithmetic unit 140 has only one register, the weight load selecting unit 130 loads the weight matrix W from the inference weight memory 110 or the updating weight memory 120 according to the process to be executed.

FIG. 3 is a block diagram showing another configuration example of the first embodiment of the learning device according to the present invention. The learning device 101 illustrated in FIG. 3 includes the components included in the learning device 100 illustrated in FIG. 1 and a fixed-layer weight memory 190.

In the parameter updating process (learning process), all parameters used for calculating the output of each unit 71 do not need to be updated. For example, a layer close to the input layer, that is, a shallow layer has a feature that the parameter does not largely change even if the parameter update processing is performed a plurality of times.

When the parameter (weight in this example) used to calculate the output of each unit 71 in some of the first to n-th layers is fixed, the fixed-layer weight memory 190 is fixed. The weight matrix related to the weight is stored. In addition, the inference weight memory 110 and the updating weight memory 120 store weight matrices for weights that are not fixed.

When the inference process is performed, the weight load selection unit 130 shown in FIG. 3 loads the weight matrices from the inference weight memory 110 and the fixed-layer weight memory 190, respectively.

[Description of operation]
Hereinafter, the operation of the inference processing and the learning processing of the learning device 100 of the present embodiment will be described with reference to FIG. FIG. 4 is a flowchart illustrating an operation of an inference process and a learning process performed by the learning device 100 according to the first embodiment.

First, an image is input to the control unit 180 of the learning device 100 (Step S101). The control unit 180 determines the type of the input image (Step S102).

When it is determined that the input image is a learning image (“learning image” in step S102), the control unit 180 determines that the process to be executed is a learning process by the weight load selecting unit 130 and the arithmetic unit 140. Is notified (step S103).

In step S103, the control unit 180 inputs the input learning image to the calculation unit 140. Alternatively, the control unit 180 inputs predetermined input data obtained from the learning image, such as the feature amount of the learning image extracted by the unit 71 of the previous layer, to the arithmetic unit 140.

Next, the weight load selector 130 loads the weight matrix W from the updating weight memory 120 (step S104). The weight load selection unit 130 inputs the loaded weight matrix W to the calculation unit 140.

Next, the arithmetic unit 140 performs a learning process using the loaded weight matrix W (step S105). The calculation unit 140 inputs the weight matrix W updated in the learning process to the weight storage unit 150.

Next, the weight storage unit 150 updates the weight matrix W stored in the updating weight memory 120 to the weight matrix W updated in the learning process (step S106).

Next, the control unit 180 determines whether the learning has been completed (step S107). If it is determined that the learning has not been completed (No in step S107), the learning device 100 ends the learning process on the input learning image.

If it is determined that the learning has been completed (Yes in Step S107), the control unit 180 notifies the weight update control unit 160 via the arithmetic unit 140 that the learning has been completed. Next, the weight update control unit 160 activates the weight copy unit 170 (Step S108).

Next, the weight copy unit 170 copies the weight matrix W stored in the updating weight memory 120 to the inference weight memory 110, and the weight matrix W stored in the inference weight memory 110 and the copied weight matrix. W is replaced (step S109). After the replacement, the learning device 100 ends the learning process on the input learning image.

When it is determined that the input image is the image for determination (“image for determination” in step S102), the control unit 180 determines that the process to be executed is the inference process by the weight load selection unit 130 and the arithmetic unit 140. (Step S110).

In step S110, the control unit 180 inputs the input determination image to the calculation unit 140. Alternatively, the control unit 180 inputs predetermined input data obtained from the discrimination image, such as the feature amount of the discrimination image extracted by the unit 71 of the previous layer, to the calculation unit 140.

Next, the weight load selecting unit 130 loads the weight matrix W from the inference weight memory 110 (Step S111). The weight load selection unit 130 inputs the loaded weight matrix W to the calculation unit 140.

The weight load selection unit 130 of the learning device 101 shown in FIG. 3 also loads a weight matrix from the fixed-layer weight memory 190 in step S111.

Next, the arithmetic unit 140 performs an inference process using the loaded weight matrix W (step S112). After executing the inference processing, the learning device 100 ends the inference processing on the input determination image. Note that the learning device 100 may collectively execute inference processing on a plurality of determination images.

The learning device 100 can execute the inference process and the learning process illustrated in FIG. 4 in parallel. Therefore, the weight load selection unit 130 may load the weight matrix W obtained from the weight memory when the process to be executed is switched. The timing at which the weight load selector 130 loads the weight matrix W is not particularly limited.

[Explanation of effects]
The learning device 100 of the present embodiment includes an inference weight memory 110 and an updating weight memory 120. Therefore, even while the weight matrix W stored in the updating weight memory 120 is being updated, the calculation unit 140 performs the inference process using the weight matrix W stored in the inference weight memory 110. I can do it.

Since the learning apparatus 100 includes the weight memory 110 for inference and the weight memory 120 during updating, the weight matrix W for inference processing is maintained in a highly reliable state. That is, the learning device 100 can execute the inference process and the learning process in parallel.

(2) In the case where the configuration of the arithmetic unit 140 is the configuration shown in FIG. 2, the power consumed for reading is reduced, so that the cost for switching between the inference process and the learning process is reduced.

Embodiment 2. FIG.
[Description of configuration]
Next, a second embodiment of the learning device according to the present invention will be described with reference to the drawings. FIG. 5 is a block diagram illustrating a configuration example of a second embodiment of the learning device according to the present invention.

As shown in FIG. 5, the learning device 200 includes a first weight memory 210, a second weight memory 220, a weight load selector 230, a calculator 240, a weight store selector 250, and a weight update controller 260. And a control unit 270.

同 様 Similar to the first embodiment, in order for the learning device 200 to execute the inference process and the learning process in parallel, the learning device 200 may have a plurality of weight memories as a solution. In the learning device 200 of the present embodiment, the roles of the plurality of weight memories are not fixed.

{That is, the learning device 200 has a configuration in which the roles of a plurality of weight memories are interchangeable. Hereinafter, the function of each component of the learning device 200 that performs the inference process and the learning process in parallel will be described.

The first weight memory 210 and the second weight memory 220 have a function of storing a weight matrix W used for inference processing or learning processing.

The weight load selection unit 230 has a function of loading the weight matrix W from one of the first weight memory 210 and the second weight memory 220.

The weight load selection unit 230 holds a memory address of a weight memory in which a weight matrix W for inference processing is stored and a memory address of a weight memory in which a weight matrix W for learning processing is stored. The held memory address is a memory address of either the first weight memory 210 or the second weight memory 220.

The weight load selection unit 230 loads the weight matrix W for inference processing from either the first weight memory 210 or the second weight memory 220 according to the stored memory address.

Similarly, according to the stored memory address, the weight load selecting unit 230 loads the weight matrix W for learning processing from either the first weight memory 210 or the second weight memory 220. The weight load selection unit 230 inputs the loaded weight matrix W to the calculation unit 240.

The operation unit 240 has the same function as the operation unit 140 of the first embodiment. Note that the configuration of the calculation unit 240 of the present embodiment may be the configuration shown in FIG.

The weight store selection unit 250 has a function of updating the weight matrix W stored in the first weight memory 210 or the second weight memory 220 with the input weight matrix W.

The weight store selection unit 250 holds a memory address of a weight memory in which a weight matrix W for inference processing is stored and a memory address of a weight memory in which a weight matrix W for learning processing is stored.

The weight store selection unit 250 updates the weight matrix W stored in the first weight memory 210 or the second weight memory 220 to the input weight matrix W according to the stored memory address.

When the learning is completed, the arithmetic unit 240 notifies the weight update control unit 260 that the learning has been completed. Upon receiving the notification, the weight update control unit 260 instructs the weight load selection unit 230 and the weight store selection unit 250 to exchange the role of the held memory address.

Specifically, the weight update control unit 260 treats the memory address of the weight memory in which the weight matrix W for inference processing is stored as the memory address of the weight memory in which the weight matrix W for learning processing is stored. To instruct.

At the same time, the weight update control unit 260 instructs to treat the memory address of the weight memory storing the weight matrix W for the learning process as the memory address of the weight memory storing the weight matrix W for the inference process. I do.

The case where the learning is completed is a case where a learning process including a weight update process using a predetermined number of learning data is repeatedly executed a predetermined number of times. In other words, when the learning is completed, the weight update control unit 260 instructs the weight load selection unit 230 to load the weight used in the inference process from the weight memory in which the weight matrix W for the learning process is stored. I have.

Similarly, when the learning is completed, the weight update control unit 260 instructs the weight load selection unit 230 to load the update target weight in the learning process from the weight memory storing the weight matrix W for the inference process. are doing. Therefore, the weight copy unit 170 of the first embodiment is unnecessary in the learning device 200 of the present embodiment.

Note that the weight load selection unit 230 and the weight store selection unit 250 receive from the weight update control unit 260 an instruction to switch the role of the memory address in advance before the learning is completed, and at a predetermined timing such as after the learning is completed, Roles may be switched.

When the weight load selection unit 230 and the weight store selection unit 250 receive an instruction in advance, the weight update control unit 260 instructs the weight load selection unit 230 and the weight store selection unit 250 to change the memory address of the role exchange destination before learning is completed. May be notified in advance.

The control unit 270 has the same function as the control unit 180 of the first embodiment.

FIG. 6 is a block diagram showing another configuration example of the learning apparatus according to the second embodiment of the present invention. The learning device 201 illustrated in FIG. 6 includes the components included in the learning device 200 illustrated in FIG. 5 and a fixed-layer weight memory 280.

学習 The learning device 201 of the present embodiment also does not need to update all parameters (weights in this example) used for calculating the output of each unit 71 in the parameter update processing (learning processing). The fixed layer weight memory 280 stores a weight matrix related to fixed weights. Further, the first weight memory 210 and the second weight memory 220 store weight matrices for weights that are not fixed.

[Description of operation]
Hereinafter, the operation of the inference processing and the learning processing of the learning device 200 of the present embodiment will be described with reference to FIG. FIG. 7 is a flowchart illustrating an operation of an inference process and a learning process performed by the learning device 200 according to the second embodiment.

First, an image is input to the control unit 270 of the learning device 200 (Step S201). The control unit 270 determines the type of the input image (Step S202).

When determining that the input image is a learning image (“learning image” in step S202), the control unit 270 determines that the process to be executed is a learning process by the weight load selection unit 230 and the arithmetic unit 240. Is notified (step S203).

In step S203, the control unit 270 inputs the input learning image to the calculation unit 240. Alternatively, the control unit 270 inputs predetermined input data obtained from the learning image, such as the feature amount of the learning image extracted by the unit 71 of the previous layer, to the arithmetic unit 240.

Next, the weight load selection unit 230 loads the weight matrix W from the weight memory in which the weight matrix W for the learning process is stored (Step S204). The weight load selection unit 230 inputs the loaded weight matrix W to the calculation unit 240.

Next, the arithmetic unit 240 performs a learning process using the loaded weight matrix W (step S205). The calculation unit 240 inputs the weight matrix W updated in the learning process to the weight store selection unit 250.

Next, the weight store selection unit 250 updates the weight matrix W for the learning process stored in the weight memory with the weight matrix W updated in the learning process (step S206).

Next, the control unit 270 determines whether or not the learning has been completed (step S207). If it is determined that the learning has not been completed (No in step S207), the learning device 200 ends the learning process on the input learning image.

When it is determined that the learning has been completed (Yes in step S207), the control unit 270 notifies the weight update control unit 260 via the arithmetic unit 240 that the learning has been completed.

Next, the weight update control unit 260 instructs the weight load selection unit 230 and the weight store selection unit 250 to exchange the role of the stored memory address (step S208).

Next, the weight load selector 230 and the weight store selector 250 swap the roles of the stored memory addresses (step S209). After the replacement, the learning device 200 ends the learning process on the input learning image.

When it is determined that the input image is the image for determination (“image for determination” in step S202), the control unit 270 determines that the process to be executed is the inference process by the weight load selection unit 230 and the arithmetic unit 240. (Step S210).

In step S210, the control unit 270 inputs the input determination image to the calculation unit 240. Alternatively, the control unit 270 inputs predetermined input data obtained from the discrimination image, such as the feature amount of the discrimination image extracted by the unit 71 of the previous layer, to the calculation unit 240.

Next, the weight load selection unit 230 loads the weight matrix W from the weight memory storing the weight matrix W for inference processing (step S211). The weight load selection unit 230 inputs the loaded weight matrix W to the calculation unit 240.

The weight load selector 230 of the learning device 201 shown in FIG. 6 also loads the weight matrix from the fixed-layer weight memory 280 in the process of step S211.

Next, the calculation unit 240 performs an inference process using the loaded weight matrix W (step S212). After executing the inference processing, the learning device 200 ends the inference processing on the input discriminating image. Note that the learning device 200 may collectively execute inference processing on a plurality of determination images.

The learning device 200 can execute the inference process and the learning process illustrated in FIG. 7 in parallel. Therefore, the weight load selection unit 230 may load the weight matrix W obtained from the weight memory when the process to be executed is switched. The timing at which the weight load selector 230 loads the weight matrix W is not particularly limited.

[Explanation of effects]
The learning device 200 of the present embodiment includes a first weight memory 210 and a second weight memory 220. Therefore, for example, even while the weighting matrix W for learning processing stored in the first weight memory 210 is being updated, the arithmetic unit 240 performs the processing for inference processing stored in the second weight memory 220. Inference processing can be performed using the weight matrix W. That is, the learning device 200 can execute the inference process and the learning process in parallel.

The learning device 200 of the present embodiment can update the weight matrix W for inference processing to the learned weight matrix W without using the weight copy unit 170 of the first embodiment. That is, the learning device 200 can use the learned weight matrix W more quickly than in the first embodiment.

Hereinafter, specific examples of the hardware configuration of the learning device of each embodiment will be described. FIG. 8 is an explanatory diagram showing an example of a hardware configuration of the learning device according to the present invention.

The learning device illustrated in FIG. 8 includes a processor 108, a main storage device 102, an auxiliary storage device 103, an interface 104, an output device 105, and an input device 106. Further, the processor 108 may include various arithmetic and processing devices such as the CPU 109 and the GPU 107.

When implemented as shown in FIG. 8, the operation of the learning device may be stored in the auxiliary storage device 103 in the form of a program. When the program is stored in the auxiliary storage device 103, the CPU 109 reads the program from the auxiliary storage device 103, expands the program in the main storage device 102, and executes a predetermined process in the learning device according to the expanded program.

The CPU 109 is an example of an information processing device that operates according to a program. The learning device may include, for example, an MPU (Micro Processing Unit), an MCU (Memory Control Unit), or a GPU (Graphics Processing Unit) in addition to the CPU (Central Processing Unit). FIG. 8 illustrates an example in which the learning apparatus further includes a GPU 107 in addition to the CPU 109.

The auxiliary storage device 103 is an example of a non-transitory tangible medium. Other examples of non-transitory tangible media include a magnetic disk, a magneto-optical disk, a CD-ROM (Compact Disk Read Only Memory), a DVD-ROM (Digital Versatile Disk Read Only Memory) connected via the interface 104, A semiconductor memory and the like are included.

When the target program stored in the auxiliary storage device 103 is distributed to the learning device through a communication line instead of being stored in the auxiliary storage device 103, the distributed learning device stores the distributed program in the main storage. The processing may be developed on the device 102 and a predetermined process may be executed.

The program may be for realizing a part of a predetermined process in the learning device. Further, the program may be a difference program for implementing a predetermined process in the learning device, which is used in combination with another program already stored in the auxiliary storage device 103.

The interface 104 transmits and receives information to and from other devices. The output device 105 presents information to a user. Further, the input device 106 receives input of information from a user.

Also, some elements shown in FIG. 8 can be omitted depending on the processing content in the learning device. For example, if the learning device does not present information to the user, the output device 105 can be omitted. Also, for example, if the learning device does not accept information input from the user, the input device 106 can be omitted.

{Some or all of the above components are realized by general-purpose or special-purpose circuits (Circuitry), processors, and the like, and combinations thereof. These may be constituted by a single chip, or may be constituted by a plurality of chips connected via a bus. In addition, some or all of the above-described components may be realized by a combination of the above-described circuit and the like and a program.

When some or all of the above-described components are realized by a plurality of information processing devices or circuits, the plurality of information processing devices or circuits may be centrally arranged or distributed. Good. For example, the information processing device, the circuit, and the like may be realized as a form in which each is connected via a communication network, such as a client and server system or a cloud computing system.

Next, the outline of the present invention will be described. FIG. 9 is a block diagram showing an outline of a learning device according to the present invention. The learning apparatus 10 according to the present invention is used in an inference process of calculating an output for each unit of discrimination data of a discrimination model in which a plurality of layers each composed of one or more units are combined in a layered manner in a predetermined order. A first storage unit 11 (for example, an inference weight memory 110) that stores parameters of each unit, and an update in a learning process of updating at least a part of parameters of each unit based on an output of each unit with respect to learning data. A second storage unit 12 (for example, an updating weight memory 120) for storing target parameters;

に よ り With such a configuration, the learning device can execute inference processing and learning processing in parallel.

Further, the learning device 10 executes a reasoning process or a learning process using the loaded parameters (for example, the weight load selecting unit 130) for loading parameters from the first storage unit 11 or the second storage unit 12. (For example, the operation unit 140).

に よ り With such a configuration, the learning device can execute inference processing and learning processing in parallel.

Further, the learning device 10 includes a plurality of execution units corresponding to each unit, and the execution unit is used in the inference processing and stores a parameter of the unit corresponding to the execution unit in a third storage unit (for example, an inference weight). A register 141) and a fourth storage unit (for example, the updating weight register 142) that stores parameters of a unit to be updated in the learning process and corresponding to the execution unit.

With such a configuration, the learning device can reduce the time for switching between execution of the calculation for the inference process and execution of the calculation for the learning process.

After completing the learning process using the predetermined number of learning data, the learning device 10 stores the parameters stored in the first storage unit 11 in the second storage unit 12 corresponding to the parameters. An update unit (for example, a weight copy unit 170) for updating the parameter to be updated may be provided.

に よ り With such a configuration, the learning device can execute the inference process using the learned model.

After completing the learning process using the predetermined number of learning data, the learning device 10 stores the parameters used in the inference process from the second storage unit 12 and the parameters to be updated in the learning process in the first storage unit. An instruction unit (for example, a weight update control unit 260) for instructing the load unit to load each from the unit 11 may be provided.

With such a configuration, the learning device can omit the process of copying the learned model to the inference weight memory.

Further, the learning device 10 includes a fifth storage unit (for example, a fixed-layer weight memory 190) that stores the fixed parameters of the parameters of each unit, and the first storage unit 11 and the second storage unit 12 May store unfixed parameters among the parameters of each unit.

に よ り With such a configuration, the learning device can support a discrimination model in which the weight of each unit in some layers is fixed.

判別 The discrimination model may be a neural network.

学習 With such a configuration, the learning device can execute deep learning.

Although the present invention has been described with reference to the exemplary embodiments and examples, the present invention is not limited to the exemplary embodiments and examples. Various changes that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention.

10, 100, 101, 200, 201 Learning device 11 First storage unit 12 Second storage unit 70 Large-scale learning circuit 71 Unit 80 Inference device 81, 91 Weight memory 82, 92 Weight load unit 83, 93, 140, 240 Operation Units 94 and 150 Weight storage unit 102 Main storage device 103 Auxiliary storage device 104 Interface 105 Output device 106 Input device 107 GPU
108 processor 109 CPU
110 Inference weight memory 120 Updating weight memory 130, 230 Weight load selection unit 141 Inference weight register 142 Updating weight register 143 Operation unit weight load selection unit 144 Operation unit 145 Operation unit weight storage unit 160, 260 Weight update control unit 170 Weight copy unit 180, 270 Control unit 190, 280 Fixed layer weight memory 210 First weight memory 220 Second weight memory 250 Weight store selection unit

Claims (10)

The parameters of the respective units used in the inference processing for calculating the output of each unit of the discrimination model of the discrimination model composed of one or more units in layers in a predetermined order are calculated. A first storage unit,
A learning device, comprising: a second storage unit that stores a parameter to be updated in a learning process of updating at least a part of a parameter of each unit based on an output of the unit for learning data. A loading unit that loads a parameter from the first storage unit or the second storage unit;
The learning device according to claim 1, further comprising: an execution unit configured to execute an inference process or a learning process using the loaded parameters. It has multiple execution units corresponding to each unit,
The execution unit is a third storage unit that is used in the inference process and stores a parameter of a unit corresponding to the execution unit, and a fourth storage unit that is an update target in the learning process and stores a parameter of the unit corresponding to the execution unit. The learning device according to claim 2, comprising: After the learning process using the predetermined number of learning data is completed, the parameter stored in the first storage unit is updated to the parameter to be updated stored in the second storage unit corresponding to the parameter. The learning device according to any one of claims 1 to 3, further comprising an updating unit that performs the updating. After the learning process using a predetermined number of learning data is completed, the parameters used in the inference process are loaded from the second storage unit, and the parameters to be updated in the learning process are loaded from the first storage unit. The learning device according to claim 2, further comprising an instruction unit that instructs the unit. A fifth storage unit for storing a fixed parameter among the parameters of each unit,
The learning device according to any one of claims 1 to 5, wherein the first storage unit and the second storage unit store parameters that are not fixed among the parameters of each unit. The parameters of each unit used in the inference processing for calculating the output for the discrimination data of each unit of the discrimination model of the discrimination model in which a plurality of layers each composed of one or more units are combined in a layered manner in a predetermined order are respectively described as follows. 1 Store in the storage unit,
A learning method, wherein a parameter to be updated in a learning process of updating at least a part of a parameter of each unit based on an output of each unit for learning data is stored in a second storage unit. Loading parameters from the first storage unit or the second storage unit,
The learning method according to claim 7, wherein inference processing or learning processing is performed using the loaded parameters. On the computer,
The parameters of each unit used in the inference processing for calculating the output for the discrimination data of each unit of the discrimination model of the discrimination model in which a plurality of layers each composed of one or more units are combined in a layered manner in a predetermined order are respectively described as follows. A first storage process to be stored in the first storage unit; and a parameter to be updated in the learning process of updating at least a part of the parameter of each unit based on the output of the unit for learning data, to the second storage unit. A learning program for executing a second storage process for storing. On the computer,
The learning program according to claim 9, wherein the learning program is configured to execute a loading process of loading a parameter from the first storage unit or the second storage unit, and an inference process or a learning process using the loaded parameter.

Images