JP2020144849A - Device and method for training meta learning network - Google Patents

Abstract

To provide a meta learning network training device and method.SOLUTION: A device acquires a plurality of network examples of a task neural network; acquires loss of each task network example; samples the loss; calculates generalization loss reflecting a total loss; calculates an inclination of a weight parameter of the task network example corresponding to each loss of a plurality of losses after sampling; inputs the inclination to one meta learning network and acquires update value of the weight parameter; updates the weight parameter of the task network example on the basis of the update value; includes a unit training the meta learning network on the basis of the generalization loss when a prescribed condition is satisfied, and repeatedly performs processing of the unit by the updated meta learning network until a first repetition termination condition is satisfied.SELECTED DRAWING: Figure 1

Description

The present invention relates to the field of information processing, and more particularly to an apparatus and method for training a meta-learning network, and an apparatus and method for training a neural network using the meta-learning network obtained by the training.

Neural networks are widely applied in academia and industry. Attention is being paid to how to effectively learn the weight parameters of neural networks in application scenarios that are aging and have few samples.

An object of the present invention is to train a neural network using a device and a method for training a meta-learning network that can solve one or more drawbacks in the prior art, and a meta-learning network obtained by the training. The purpose is to provide equipment and methods.

According to one aspect of the invention, a device for training a meta-learning network is provided.
A task network example acquisition unit that obtains a plurality of task network instances of the task neural network for each task neural network among a plurality of task neural networks for different tasks, among which the different tasks are obtained. Task network example acquisition unit with similar tasks;
A loss acquisition unit that acquires the loss of each task network example of each task neural network using training data;
A sampling unit that performs sampling on the loss and acquires a plurality of losses after sampling;
A generalization loss calculation unit that calculates a generalization loss that reflects the total loss of the plurality of losses after sampling based on the plurality of losses after sampling;
A gradient calculation unit that calculates the gradient corresponding to the weight parameter of the task network example corresponding to each loss of the generalized loss among the plurality of losses after sampling;
The gradient is input to one meta-learning network of at least one meta-learning network, and the updated value of the weight parameter of the task network example corresponding to each loss of the plurality of losses after sampling is acquired. A learning unit that is a neural network in which the meta-learning network learns with respect to a gradient;
The first update unit that updates the weight parameter of the task network example corresponding to each loss of the plurality of losses after the sampling based on the update value; and the loss acquisition unit, the sampling unit, and the generalized loss. The processing of the calculation unit, the gradient calculation unit, the learning unit, and the first update unit is repeated, and when the predetermined conditions are satisfied, based on the generalization loss when the predetermined conditions are satisfied. A second update unit is included that trains the at least one meta-learning network in a direction in which the loss of the task network example is smaller and obtains at least one meta-learning network after the update.
Based on at least one meta-learning network after the update of the second update unit, the loss acquisition unit, the sampling unit, the generalized loss calculation unit, the gradient calculation unit, the learning unit, the first update unit, And the processing of the second update unit is repeated until the first iteration end condition is satisfied.

According to another aspect of the invention, a method of training a meta-learning network is provided.
For each task neural network among a plurality of task neural networks for different tasks, a task network example acquisition step for obtaining a plurality of task network examples of the task neural network, wherein the different tasks have similarities. Task network example acquisition step;
Loss acquisition step to acquire the loss of each task network example of each task neural network using the training data;
A sampling step in which sampling is performed on the loss to obtain a plurality of losses after sampling;
A generalization loss calculation step that calculates a generalization loss that reflects the total loss of the plurality of losses after sampling based on the plurality of losses after sampling;
Gradient calculation step to calculate the gradient corresponding to the weight parameter of the task network example corresponding to each loss of the generalized loss among the plurality of losses after sampling;
Learning to input the gradient into one meta-learning network of at least one meta-learning network and obtain the updated value of the weight parameter of the task network example corresponding to each loss of the plurality of losses after sampling. A learning step, which is a neural network in which the meta-learning network learns with respect to a gradient;
The first update step of updating the weight parameter of the task network example corresponding to each loss of the plurality of losses after the sampling based on the update value; and the loss acquisition step, the sampling step, and the generalization loss. The calculation step, the gradient calculation step, the learning step, and the first update step are repeatedly performed, and when the predetermined condition is satisfied, based on the generalization loss when the predetermined condition is satisfied. Including a second update step of training the at least one meta-learning network in a direction in which the loss of the profitable task network example is smaller and obtaining at least one meta-learning network after the update.
The loss acquisition step, the sampling step, the generalized loss calculation step, the gradient calculation step, the learning step, the first update step, and the second update step based on at least one meta-learning network after the update. Is repeated until the first iteration end condition is satisfied.

According to another aspect of the present invention, there is provided an apparatus for training a neural network using the meta-learning network obtained by training the apparatus for training the meta-learning network described above.
A second loss acquisition unit that uses training data to obtain the loss of a neural network awaiting training;
A second gradient calculation unit that calculates the gradient of the loss with respect to the weight parameter of the neural network awaiting training;
The gradient is input to one of the trained at least one meta-learning networks acquired when the first iteration end condition is satisfied, and the weight parameter of the neural network awaiting training is input. A second learning unit that obtains the update value of; and a third update unit that updates the weight parameter of the neural network awaiting training based on the update value.
The processing of the second loss acquisition unit, the second gradient calculation unit, the second learning unit, and the third update unit is repeatedly executed until the second iteration end condition is satisfied.

According to another aspect of the present invention, computer program code and a computer program product for realizing the above-mentioned method according to the present invention are further provided.

It is a block diagram of the composition of the meta-learning network training apparatus in the Example of this invention. It is a block diagram of the task neural network in the Example of this invention. It is an exemplary architecture diagram of the process performed by the meta-learning network training apparatus in the embodiment of the present invention. It is a figure which shows an example of the process executed by the meta-learning network in the Example of this invention. It is a top view of the two-dimensional loss curved surface for a single task of a meta-learning network based on gradient information and a task network based on traditional data in an example of a multitasking network in the embodiment of the present invention. It is an exemplary flowchart of the meta-learning network training method in the embodiment of the present invention. It is a block diagram of a device which trains a neural network using a trained meta-learning network. It is an exemplary flowchart of a method of training a neural network using a trained meta-learning network. It is an exemplary block diagram of a general-purpose computer that can be adopted in the examples of the present invention.

Hereinafter, preferred embodiments for carrying out the present invention will be described in detail with reference to the attached drawings. It should be noted that such an embodiment is merely an example and does not limit the present invention.

First, a functional block diagram of the meta-learning network training device 100 according to the embodiment of the present invention will be described with reference to FIG. FIG. 1 is a block diagram of the configuration of the meta-learning network training device 100 according to the embodiment of the present invention. As shown in FIG. 1, the meta-learning network training device 100 in the embodiment of the present invention includes a task network example acquisition unit 102, a loss acquisition unit 104, a sampling unit 106, a generalized loss calculation unit 108, a gradient calculation unit 110, and learning. Includes unit 112, first update unit 114, and second update unit 116.

The task network example acquisition unit 102 acquires a plurality of task network examples of the task neural network for each task neural network among the plurality of task neural networks for different tasks, among which different tasks have similarities. Have.

As an example, one task neural network of a plurality of task neural networks may be a neural network for an English recognition task, and another task neural network of a plurality of task neural networks recognizes French. It may be a neural network for the task, and the English recognition task and the French recognition task have similarities.

As an example, each task neural network among a plurality of task neural networks is a neural network system for classifying emotions.

As an example, each task neural network among a plurality of task neural networks is a neural network recommender system for recommending a list of goods on a website to a user. For example, the goods list may include a goods list.

As an example, each task neural network among a plurality of task neural networks is a neural network recommender system for recommending a program list of a movie / television website to a user.

Also, as an example, each task neural network among a plurality of task neural networks is a neural network recommender system for recommending a list of goods on a shopping website to a user.

FIG. 2 is a block diagram of a task neural network according to an embodiment of the present invention.

In Figure 2, Q _{1: n-1} = [Q ₁ , Q ₂ ,…, Q _n-1 ] is the contextual basic unit index sequence, and each Q _i (i = 1,2,…, n-1). Represents the index number of the basic unit, of which the context basic unit is the name of the item viewed, where the task neural network is the recommended neural network for recommending the user a list of items on the shopping website. The inset layer transforms the base unit index number into a low-dimensional continuous non-sparse word vector. A recurrent neural network (for example, the recurrent neural network may be a long / short-term memory network LSTM) takes a word vector sequence corresponding to the context basic unit index sequence as an input, and loop unrolls to Q _n-1 . The context feature h _n-1 at the corresponding n-1 position is output, of which h ₀ is the initial feature. The affine layer transforms h _n-1 into a vector O _n-1 of the same dimensions as the basic unit dictionary. Of these, On _-1 under the weight parameter of the current task neural network corresponds to the recommended item, and the one-hot vector vector of the next one basic unit (unit corresponding to Q _n ) becomes a realistic browsing item. Correspond.

As an example, the similarity of different tasks means that the business scenes of the tasks are similar, for example, the inputs of the task neural network of different tasks are similar and / or the tasks of different tasks. The output of the neural network is similar, or the configuration of the task neural network of different tasks is similar.

FIG. 3 is an exemplary architectural diagram of the processing performed by the meta-learning network training apparatus 100 in the embodiment of the present invention.

As shown in FIG. 3, it is assumed that there are N tasks, that is, task 1, task 2, ..., Task N. Of these, task 1, task 2, ..., and task N have similarities.

By obtaining a plurality of task network examples for each task neural network, the problem that the number of neural network samples is small can be effectively solved, and more input samples can be provided to the meta-learning network described later.

Preferably, the task network example acquisition unit 102 is configured to acquire a plurality of task network examples of each task neural network by performing random initialization for each task neural network among the plurality of task neural networks. Can be done.

As an example, when the task neural network is a recommended neural network system for recommending a list of goods on a shopping website to a user, a plurality of task network examples of each task neural network are examples of the recommended neural network system.

As shown in Figure 3, by performing random initialization on the network parameters of the neural network for task 1, it is possible to obtain multiple task network examples of the neural network for task 1; task 2 By performing random initialization on the network parameters of the neural network for, we can obtain multiple task network examples of the neural network for task 2; and the network of the neural network for task N. By performing random initialization on the parameters, it is possible to obtain multiple task network examples of the neural network for task N. Figure 3 shows that, for convenience, each neural network for each task has three task network examples. It should be noted that those skilled in the art should understand that the neural network for each task may have a plurality of other task network examples.

The loss acquisition unit 104 may be configured to acquire the loss of each task network example of each task neural network, respectively, using the training data.

As an example, if the task neural network is a recommended neural network system for recommending a list of goods on a shopping website to the user, the loss acquisition unit 104 is configured to acquire the actual loss of each recommended neural network system. obtain.

As an example, in the loss acquisition unit 104, each task network example of each task neural network is used by a normal loss calculation function in this field, for example, softmax, using the training data in the training set of each task neural network. You can get the loss of.

If the task neural network for each task has the same number of task network examples, then the loss of all task network examples can form one loss matrix, which is the number of columns in the number of tasks. And the number of rows is the number of examples of task neural networks.

In the example shown in FIG. 3, each cube (square block) with "loss" represents the loss of each task network example, these losses constitute one loss matrix, and the number of columns in the matrix is the number of tasks. N, the number of rows is the number of examples of each task neural network, i.e. 3.

The sampling unit 106 may be configured to acquire a plurality of losses after sampling by sampling the losses.

As an example, if the task neural network is a recommendation neural network system for recommending a list of goods on a shopping website to a user, the sampling unit 106 is configured to sample for the loss of an example of the recommendation neural network system. Can be done.

As will be described later, the gradient of the task network example corresponding to a plurality of losses after sampling is used as the input of the meta-learning network described later, and in this case, sampling for the loss is used as meta information of the meta-learning network. It is equivalent to performing selective filtering on the gradient, which can greatly improve the robustness of the meta-learning network.

Preferably, the sampling unit 106 may be configured to acquire a plurality of losses after the sampling by sampling the losses by a random selection operation. Of these, the number of losses after sampling is controlled by the retention probability parameter of the selection operation.

As an example, the random selection operation may be a dropout operation well known to those skilled in the art, and the number of multiple losses after sampling is controlled by the retention probability parameter of the dropout operation.

Sampling the loss by a random selection operation, for example a dropout operation, is equivalent to performing random selectivity filtering on the gradient as meta information of the meta-learning network, thereby causing the meta-learning network. Robustness can be further improved.

In FIG. 3, in the illustration of the loss corresponding to the “random selection operation”, the gray cube shows the plurality of losses after sampling, and the white cube shows the loss not sampled (that is, not selected).

The generalization loss calculation unit 108 may be configured to calculate the generalization loss that reflects the total loss of the plurality of losses after sampling, based on the plurality of losses after sampling.

As an example, if the task neural network is a recommendation neural network system for recommending a list of goods on a shopping website to a user, the generalization loss calculation unit 108 generalizes the loss after sampling of an example of the recommendation neural network system. It can be configured to calculate the loss.

Preferably, the generalization loss calculation unit 108 can be further configured to perform averaging on a plurality of losses after sampling and to make the loss after averaging the generalization loss. There is a difference in the time length to calculate the loss of multiple task network examples. As an example, the loss of the task network example with the shortest time to calculate the loss can be further referred to as the generalized loss. It should be noted that those skilled in the art should understand that the result after performing other processing on a plurality of losses after sampling may be regarded as a generalization loss, but detailed description thereof will be omitted here.

The gradient calculation unit 110 may be configured to calculate the gradient of the generalized loss with respect to the weight parameter of the task network example corresponding to each of the multiple losses after sampling.

Specifically, the gradient is the result of obtaining the partial derivative of the weight parameter of the task network example corresponding to each loss of the plurality of losses after sampling of the generalized loss.

The task network example learning process is essentially a process that minimizes generalization loss by continuously updating the weight parameters of the task network example. The above-mentioned gradient may be used to control the direction of updating the weight parameter of the task network example, and the above-mentioned gradient can also control the width of updating the weight parameter of the task network example. ..

Preferably, the gradient calculation unit 110 can be configured to calculate the gradient of the generalized loss for the weight parameter of each network layer in the task network example corresponding to each loss of the plurality of post-sampling losses.

As an example, the task neural network may include a plurality of network layers (for example, when the task neural network is a convolutional neural network, the convolutional neural network may include a plurality of convolutional layers, a plurality of fully connected layers, and the like). As a result, the task network example also includes a plurality of network layers accordingly. Learning the weight parameter of the task network example is to learn the weight parameter of each network layer in the task network example. Therefore, the gradient calculation unit 110 needs to calculate the gradient of the generalization loss with respect to the weight parameter of each network layer in the task network example corresponding to each loss among the plurality of losses after sampling.

Preferably, the gradient calculation unit 110 is used for each of the task network examples other than the task network example corresponding to each loss of the plurality of losses after sampling among all the task network examples acquired for the plurality of task neural networks. It can be configured to set the gradient of the network layer weight parameter to zero.

In FIG. 3, among the cubes corresponding to the “gradient”, the gray cube shows the gradient of the generalization loss with respect to the weight parameter of a certain layer in the task network example corresponding to the loss after sampling. The gradient of the weight parameter of each network layer of the task network example other than the task network example corresponding to each loss of the loss is set to zero, and the gradient having a value of zero is shown by a white cube.

The learning unit 112 updates the weight parameters of the task network example corresponding to each loss of the plurality of losses after sampling by inputting the gradient to each of the meta learning networks of at least one meta learning network. It can be configured to get each value. Among them, the meta-learning network is a neural network that learns with respect to the gradient.

As described above, the task network example learning process can be embodied as a process that minimizes generalization loss by continuously updating the weight parameters of each network layer of the task network example.

The update of the weight parameter of one network layer in the task network example can be written as W _{i + 1} = W _i + Δ _W. Among them, W _i is a weight parameter before the update of the network layer, W _{i + 1} is a weight parameter after the update of the network layer, and ΔW is an update value of the weight parameter of the network layer. ..

The update value ΔW of the weight parameter can be expressed as ΔW = λ * grad. Among them, λ is the learning rate, which is used to control the update width of the weight parameter, and is, for example, an empirical value or a parameter determined by an experiment, and grad is the network of generalization loss. It is a gradient with respect to the weight parameter of the layer and is used to control the direction of updating the weight parameter, and grad can also control the width of updating the weight parameter. As can be seen from the above formula, the update value ΔW directly affects the learning speed of the task network example.

The meta-learning network in the embodiment of the present invention is a neural network that learns with respect to the gradient. The gradient grad of each network layer of the task network example corresponding to the plurality of losses after sampling is input to the meta-learning network, and the meta-learning network can output the updated value ΔW of the weight parameter of the layer.

Preferably, the meta-learning network is based on the state of the last update of the weight parameter of each network layer of the task network example corresponding to each loss of the multiple losses after sampling, and the update value of the weight parameter of that layer. Can be output.

Specifically, the meta-learning network is generated by the gradient of the weight parameter of each network layer of the task network example corresponding to each loss of the plurality of losses after sampling, and the previous update of the weight parameter of the layer. Based on the obtained state vector (the initial state vector is a zero vector), the updated value of the weight parameter of the layer and the state vector generated by this update of the weight parameter of the layer are output. As can be seen, the update of the weight parameter has a post-effect because the update result of the previous weight parameter may affect the update result of the later weight parameter.

FIG. 4 is a diagram showing an example of processing executed by the meta-learning network in the embodiment of the present invention.

In FIG. 4, the gradient grad of the weight parameter of one network layer in the task network example corresponding to each loss among the plurality of losses after sampling is input to the meta-learning network, of which grad is one two-dimensional matrix. Is.

In the meta-learning network, first, the grad of the two-dimensional matrix is expanded in the row direction into the preprocessing gradient vector of the one-dimensional vector by the preprocessing operation.

Recurrent neural networks can be used to represent the post-effect of updating weight parameters, and as shown in FIG. 4, multiple recurrent neural networks, such as "1", "2", "3", "4". Suppose there is a recurrent neural network with, "5" and "6". Although FIG. 4 shows that the recurrent neural network is a long / short-term memory network LSTM, those skilled in the art should understand that the recurrent neural network is a network other than the long / short-term memory network. It is okay to have it. The preprocessed gradient vector and the previous state vector of the recurrent neural network (which corresponds to the state vector generated by the previous update of the weight parameter of the task network example described above, and is in the “last updated state” in FIG. (The last updated state vector) is the input of the recurrent neural network layer, and the output vector of a series of recurrent neural networks (the one with the “LSTM output vector” in Fig. 4) is output. , The state vector of the recurrent neural network is unified and initialized with the zero vector. Among them, each element of the preprocessing gradient vector shares the weight parameter of the recurrent neural network layer. Note that the input weights inside the recurrent neural network may not be shared due to the different sizes of the network layers of the same type, and in order to solve the problem of different sizes, the recurrent neural network Has adopted the policy of loop unrolling and weight sharing of each scalar element in the multi-vector. It is not necessary to save the intermediate state of the expansion (for example, a small block (small block) including the diagonal line in Fig. 4), and only the output of the state at the last one position of the expansion is the state generated by this update. It needs to be saved as a vector (that is, the "currently updated state vector" in the "currently updated state" in FIG. 4).

Subsequently, the output vector of each recurrent neural network is converted into a scalar element by one linear transformation layer, and all the scalar elements are merged into a final output vector.

Finally, the post-processing converts the final output vector into a matrix that is the same size as the grad and gives it ΔW.

In addition, in FIG. 3, the cubes with the “last updated state” represent the state generated by the previous update of each task network example, and in FIG. 3, the “last updated state of the task network example after sampling” is shown. Of the cubes marked with ", the gray cube indicates the state generated by the previous update of each task network example after sampling, and the white cube indicates that the task network example is not sampled.

The first update unit 114 may be configured to update the weight parameter of the task network example corresponding to each loss of the plurality of losses after sampling based on the update value.

Preferably, the first update unit 114 may be configured to update the weight parameters of each network layer in the task network example corresponding to each loss of the plurality of losses after sampling based on the update value.

The second update unit 116 repeatedly executes the processes of the loss acquisition unit 104, the sampling unit 106, the generalized loss calculation unit 108, the gradient calculation unit 110, the learning unit 112, and the first update unit 114, and the predetermined conditions are satisfied. After the update, by training the at least one meta-learning network in a direction in which the loss of the task network example is smaller, based on the generalized loss when the predetermined condition is satisfied, when satisfied. It may be configured to obtain at least one meta-learning network.

Specifically, before performing the above-mentioned iterative processing, first, the weight parameters of the at least one meta-learning network are initialized, and as an example, the weight parameters of the at least one meta-learning network are initialized based on experience. Alternatively, the weight parameters of the at least one meta-learning network may be randomly initialized, or the weight parameters of the at least one meta-learning network may be initialized by other methods conceivable by those skilled in the art. In the process of iteratively updating the weight parameters of the task network example, the weight parameters of at least one meta-learning network are fixed.

As an example, the predetermined condition described above may reach a predetermined number of repetitions. As an example, the weight parameters of the task network example are repeatedly updated, and when the predetermined number of iterations is reached, the weight parameters of the plurality of task network examples are fixed so that the loss of the task network example is reduced. By training the at least one meta-learning network, the weight parameters of the at least one meta-learning network can be updated.

As described above, the meta-learning network in the embodiment of the present invention is a neural network that learns with respect to the gradient. As mentioned above, by obtaining multiple task network examples for each task neural network, the problem of few neural network samples is effectively solved, and more task network examples can be added to the meta-learning network. A gradient can be provided. The gradients of these many task network examples are used as sample data of the meta-learning network to train the meta-learning network, and the trained meta-learning network estimates the learning update rules of the task neural network and is more efficient. By predicting the optimization path of the weight parameter of the task neural network and calculating the update value ΔW of the better weight parameter, the speed of learning the task neural network can be accelerated. In addition, sampling the loss is equivalent to performing selective filtering on the gradient of the meta-learning network as meta-information, which can greatly improve the robustness of the meta-learning network. ..

Preferably, the number of the at least one meta-learning network is the same as the number of network layer types in the plurality of task neural networks, and the task network example corresponding to each loss among the plurality of losses after sampling. The gradient of each network layer is input to the meta-learning network corresponding to the layer type. In other words, different meta-learning networks are used to learn the gradient information of different types of network layers in the task network example, and the same meta-learning network is used to learn the gradient information of the same type of network layer in the task network example. Learn information. As an example, when the task neural network is a convolutional neural network and the convolutional neural network includes a plurality of convolutional layers and a plurality of fully connected layers, the first meta-learning network is adopted to adopt the first meta-learning network and the plurality of convolutions in the convolutional neural network example. The gradient information about the layer is learned, and the gradient information about the plurality of fully connected layers in the convolutional neural network example is learned by adopting the second meta learning network different from the first meta learning network.

FIG. 5 is a top view of a two-dimensional loss curved surface for a single task of a meta-learning network based on gradient information and a task network based on traditional data in an example of a multitasking network in an embodiment of the present invention.

In the loss curved surface shown in FIG. 5, the deeper the color of the curved surface, the smaller the loss value at that position, and each line segment with an arrow represents the direction and step length of the loss drop after one network layer weight parameter update. Among them, the route composed of the white arrowed line segment is the loss reduction method of the task network based on the traditional data, and the route composed of the black arrowed line segment included in the dotted frame is the multi in the embodiment of the present invention. It is a loss drop method of the meta-learning network based on the gradient information of the task network example. In traditional data-based task networks, for example, task networks using the mini-batch random gradient descent method, the rate of convergence to relatively small losses is often slow. The main reason for this is that it is biased towards the update gradient obtained based on the loss on the local batch data, so the direction of each parameter update is not always optimal (it may bypass). Once the deviation of the previous update is too large, it can slow down the convergence and cause it to not converge or to converge where the local minimum loss is not very good. As can be seen from FIG. 5, the number of updates required for the black line segment to reach the point where the loss is relatively low is much smaller than the number of updates of the task network based on the data, and thus in the embodiment of the present invention. A meta-learning network based on the gradient information of the multitasking network example can predict the optimization path of the weight parameter of the task neural network with higher efficiency, so that the speed of learning the task neural network can be accelerated.

In the meta-learning network training device 100 in the embodiment of the present invention, the loss acquisition unit 104, the sampling unit 106, the generalized loss calculation unit 108, and the gradient are based on at least one meta-learning network after the update of the second update unit 116. The processing of the calculation unit 110, the learning unit 112, the first update unit 114, and the second update unit 116 is repeatedly executed until the first iteration end condition is satisfied. As described above, the updated meta-learning network can predict the optimization path of the weight parameter of the task neural network with higher efficiency and calculate the update value ΔW of the better weight parameter. Based on at least one meta-learning network after the update of unit 116, the weight parameter of the task network example can be iteratively updated again to accelerate the learning speed of the task network example. The above process is repeated, that is, the weight parameters of the plurality of task network examples and the parameters of the at least one meta-learning network are alternately updated until the first iteration end condition is satisfied (as described above). , The weight parameters of the plurality of task network examples are repeatedly updated, and when the predetermined conditions are satisfied, the parameters of the at least one meta-learning network are updated). As an example, the first iteration end condition is at least one of reaching a predetermined number of iterations, the generalization loss converging, the generalization loss decrease being very small, and the generalization loss increasing. You may. The meta-learning network obtained when the first iteration end condition is satisfied can estimate the learning update rules of the new task neural network more quickly and accurately, thereby rapidly learning the new task neural network. Can be used to teach.

In the meta-learning network estimated based on the conventional parameter update value, one task network example is initialized under each task, and then the parameters of the meta-learning network are learned according to the parameter update rule of each task example. However, a single network example under each task can lead to inadequate task rule extraction, which causes the weight parameters of the task network example of the meta-learning network trained under multitasking. It can be induced that there is a guidance deviation in the update value, that is, it can cause the meta-learning network trained under multitasking to be unable to accurately predict the update value of a better task network example parameter. For example, if multiple task neural networks are a neural network recommender system for recommending a shopping website item list to a user, it may cause the user to be unable to accurately recommend the current item list.

In the meta-learning network training device 100 in the embodiment of the present invention, the problem that the number of neural network samples is small is effectively solved by acquiring a plurality of task network examples for each task neural network. It can provide gradients for many task network examples and significantly increase the robustness of the meta-learning network by sampling the loss of the task network examples to selectively filter the gradient as meta information. Also, the meta-learning network predicts the optimization path of the weight parameter of the task neural network with higher efficiency, and calculates the update value of the weight parameter of the current better task network example. You can also accelerate the training speed of new task neural networks. For example, if multiple task neural networks are a neural network recommender system for recommending a user a list of items on a shopping website, the user is quickly and accurately recommended the current item list based on the user's historical browsing records. can do.

Corresponding to the above-mentioned embodiment of the meta-learning network training apparatus, the present invention further provides an embodiment of the following meta-learning network training method.

FIG. 6 is an exemplary flowchart of the meta-learning network training method 500 in the embodiment of the present invention.

As shown in FIG. 6, the meta-learning network training method 500 in the embodiment of the present invention includes a task network example acquisition step S502, a loss acquisition step S504, a sampling step S506, a generalized loss calculation step S508, and a gradient calculation step S510. Includes step S512, first update step S514, and second update step S516.

In the task network example acquisition step S502, for each task neural network among the plurality of task neural networks for different tasks, a plurality of task network examples of the task neural network are acquired, and among them, different tasks have similarities. Have.

For a detailed description of the task neural network and the task network example, the description of the task network example acquisition unit 102 in the embodiment of the apparatus can be referred to, and thus the duplicate description is omitted here.

In the loss acquisition step S504, the loss of each task network example of each task neural network is acquired by using the training data.

As an example, using the training data in the training set of each task neural network, the loss of each task network example of each task neural network can be obtained by a normal loss calculation function in this field, for example, softmax. Can be done.

In the sampling step S506, a plurality of losses after sampling are obtained by sampling the losses.

Entering the gradient of the task network example corresponding to multiple losses after sampling in the meta-learning network, in this case sampling for the loss is selective filtering for the gradient as meta-information in the meta-learning network. This is equivalent to doing this, which can greatly improve the robustness of the meta-learning network.

Preferably, in sampling step S506, a plurality of losses after sampling are acquired by sampling the losses by a random selection operation, and the number of the plurality of losses after sampling is the retention of the selection operation. Controlled by stochastic parameters.

As an example, the random selection operation is a dropout operation well known to those skilled in the art, and the number of multiple losses after sampling is controlled by the retention probability parameter of the dropout operation.

Sampling the loss by a random selection operation, for example a dropout operation, is equivalent to performing random selective filtering on the gradient as meta information of the meta-learning network, thereby causing the meta-learning network to perform random selective filtering. Robustness can be further improved.

In the generalization loss calculation step S508, the generalization loss that reflects the total loss of the plurality of losses after sampling is calculated based on the plurality of losses after sampling.

Preferably, in the generalization loss calculation step S508, the plurality of losses after sampling are averaged, and the loss after averaging is defined as the generalization loss. There is a difference in the time length of loss calculation for multiple task network examples. As an example, let the loss of the task network example with the shortest loss calculation time be the generalized loss. It should be noted that those skilled in the art should understand that the result after performing other processing on a plurality of losses after sampling may be regarded as a generalization loss, but detailed description thereof will be omitted here.

In the gradient calculation step S510, the gradient of the generalized loss with respect to the weight parameter of the task network example corresponding to each loss among the plurality of losses after sampling is calculated.

The task network example learning process is essentially a process that minimizes generalization loss by continuously updating the weight parameters of the task network example. The gradient described above may be used to control the direction of updating the weight parameters of the task network example, and the gradient described above may also control the width of updating the weight parameters of the task network example. it can.

Preferably, in the gradient calculation step S510, the gradient of the generalization loss is calculated for the weight parameter of each network layer in the task network example corresponding to each loss among the plurality of losses after sampling.

As an example, the task neural network may include a plurality of network layers (for example, when the task neural network is a convolutional neural network, the convolutional neural network may include a plurality of convolutional layers, a plurality of fully connected layers, and the like. ), Thereby, the task network example also includes a plurality of network layers accordingly. Learning the weight parameter of the task network example is to learn the weight parameter of each network layer in the task network example. Therefore, in the gradient calculation step S510, it is necessary to calculate the gradient of the generalization loss with respect to the weight parameter of each network layer in the task network example corresponding to each loss among the plurality of losses after sampling.

Preferably, in the gradient calculation step S510, among all the task network examples acquired for the plurality of task neural networks, each of the task network examples other than the task network example corresponding to each loss of the plurality of losses after sampling. Set the gradient of the network layer weight parameter to zero.

In learning step S512, the gradient is input to each of the meta-learning networks of at least one meta-learning network, and the updated value of the weight parameter of the task network example corresponding to each loss of the plurality of losses after sampling is set. The meta-learning network is a neural network that learns the gradient.

Preferably, the meta-learning network is based on the state of the last update of the weight parameter of each network layer of the task network example corresponding to each loss of the multiple losses after sampling, and the update value of the weight parameter of that layer. Is output.

For a detailed description of the meta-learning network, the description of the learning unit 112 in the embodiment of the apparatus can be referred to, and thus duplicate description will be omitted here.

In the first update step S514, the weight parameter of the task network example corresponding to each loss among the plurality of losses after sampling is updated based on the update value.

Preferably, in the first update step S514, the weight parameter of each network layer in the task network example corresponding to each loss among the plurality of losses after sampling is updated based on the update value.

In the second update step S516, the processes of the loss acquisition step S504, the sampling step S506, the generalized loss calculation step S508, the gradient calculation step S510, the learning step S512, and the first update step S514 are repeated, and the predetermined conditions are satisfied. Occasionally, at least one after update, by training the at least one meta-learning network in a direction in which the loss of the task network example is smaller, based on the generalization loss when the predetermined condition is satisfied. Get a meta-learning network.

Specifically, before performing the above-mentioned iterative processing, first, the weight parameters of the at least one meta-learning network are initialized, and as an example, the weight parameters of the at least one meta-learning network are initialized based on experience. Alternatively, the weight parameters of the at least one meta-learning network may be randomly initialized, or the weight parameters of the at least one meta-learning network may be initialized by other methods conceivable by those skilled in the art. In the process of iteratively updating the weight parameters of the task network example, the weight parameters of at least one meta-learning network are fixed.

As an example, the predetermined condition described above may reach a predetermined number of repetitions. As an example, the weight parameter of the task network example is repeatedly updated, and when the predetermined number of iterations is reached, the parameters of the plurality of task network examples are fixed so that the loss of the task network example is smaller. By training the at least one meta-learning network, the weight parameters of at least one meta-learning network can be updated.

As described above, the meta-learning network in the embodiment of the present invention is a neural network that learns with respect to the gradient. By obtaining a plurality of task network examples for each task neural network, it is possible to effectively solve the problem of a small number of neural network samples, and to provide the meta-learning network with a gradient of more task network examples. The gradients of these many task network examples are used as sample data for the meta-learning network to train the meta-learning network, and the trained meta-learning network estimates the learning update rules of the task neural network and is more efficient. By predicting the optimization path of the weight parameter of the task neural network and calculating the updated value of the better weight parameter, the speed of learning the task neural network can be accelerated. In addition, sampling the loss is equivalent to performing selective filtering on the gradient of the meta-learning network as meta-information, which can greatly improve the robustness of the meta-learning network. ..

Preferably, the number of at least one meta-learning network is the same as the number of layer types in the plurality of task neural networks, and the task network example corresponding to each loss of the plurality of losses after sampling. The gradient of each network layer is input to the meta-learning network corresponding to the layer type. In other words, different meta-learning networks are adopted to learn the gradient information of different types of network layers in the task network example, and the same meta-learning network is adopted to learn the gradient information of the same type of network layer in the task network example. Learn information.

In the meta-learning network training method 500 in the embodiment of the present invention, the loss acquisition step S504, the sampling step S506, and the generalized loss calculation step S508 are based on at least one meta-learning network after being updated in the second update step S516. , Gradient calculation step S510, learning step S512, first update step S514, and second update step S516 are repeatedly executed until the first iteration end condition is satisfied. As described above, the updated meta-learning network can predict the optimization path of the weight parameter of the task neural network with higher efficiency and calculate the update value of the better weight parameter. Therefore, the second update step. By iteratively updating the weight parameters of the task network example again based on at least one meta-learning network updated in S516, the speed of learning the learning task network example can be accelerated. The above process is repeated, that is, the weight parameters of the plurality of task network examples and the parameters of the at least one meta-learning network are alternately updated until the first iteration end condition is satisfied (as described above). , The weight parameters of the plurality of task network examples are repeatedly updated, and when the predetermined conditions are satisfied, the parameters of the at least one meta-learning network are updated). As an example, the condition for ending the first iteration is that at least one of reaching a predetermined number of iterations, the generalization loss converges, the generalization loss decrease is very small, and the generalization loss increases. May be included. The meta-learning network obtained when the first iteration end condition is satisfied can estimate the learning update rules of the new task neural network more quickly and accurately, thereby rapidly learning the new task neural network. Can be instructed.

In the meta-learning network training method 500 in the embodiment of the present invention, the problem of a small number of neural network samples is effectively solved by obtaining a plurality of task network examples for each task neural network. The gradient of many task network examples can be provided, and by sampling the loss of the task network example, the robustness of the meta-learning network can be improved by selectively filtering the gradient as meta information. It can be significantly improved, and the meta-learning network predicts the optimization path of the weight parameter of the more efficient task neural network and calculates the update value of the weight parameter of the current better task network example. This can also accelerate the training speed of new task neural networks. For example, if multiple task neural networks are a neural network recommender system for recommending a shopping website item list to a user, the current item list is quickly and accurately recommended to the user based on the user's history browsing record. Can be recommended.

The present invention further provides an apparatus for training a neural network using the meta-learning network obtained by training the meta-learning network training apparatus 100 or the meta-learning network training method 500 described above. FIG. 7 is a block diagram of a device 600 that trains a neural network using the meta-learning network obtained by training the meta-learning network training device 100 or the meta-learning network training method 500. As shown in FIG. 7, the apparatus 600 that trains the neural network using the trained meta-learning network in the embodiment of the present invention includes the second loss acquisition unit 602, the second gradient calculation unit 604, and the second. Includes learning unit 606 and third update unit 608.

The second loss acquisition unit 602 may be configured to acquire the loss of the neural network awaiting training using the training data.

As an example, the second loss acquisition unit 602 acquires the loss of the neural network awaiting training by using a normal loss calculation function in this field, for example, softmax.

The second gradient calculation unit 604 may be configured to calculate the gradient of the loss relative to the weight parameter of the neural network awaiting training.

The training-waiting neural network learning process is essentially the process of continuously updating the weight parameters of the training-waiting neural network to minimize losses. The gradients described above may be used to control the direction of updates of the weight parameters of the neural network awaiting training, and the gradients described above further reduce the range of updates of the weight parameters of the neural network awaiting training. Can be controlled.

Preferably, the second gradient calculation unit 604 can be configured to calculate the gradient of the loss with respect to the weight parameter of each network layer in the awaiting neural network.

As an example, a training-waiting neural network may include a plurality of network layers (for example, when the training-waiting neural network is a convolutional neural network, the convolutional neural network may include a plurality of convolutional layers, a plurality of fully connected layers, and the like. May be included). Learning the weight parameter of the neural network waiting for training is to learn the weight parameter of each network layer in the neural network waiting for training. Therefore, the second gradient calculation unit 604 needs to calculate the gradient of the loss with respect to the weight parameter of each network layer in the awaiting neural network.

The second learning unit 606 inputs the gradient into one of the trained at least one meta-learning networks acquired when the first iteration end condition is satisfied, and waits for training. It can be configured to obtain updated values for the weight parameters of the neural network of.

With reference to the description of the meta-learning network training device 100 in the embodiment of the present invention, the meta-learning network in the embodiment of the present invention is a neural network that learns with respect to the gradient. The gradient of each network layer of the neural network awaiting training is input to the meta-learning network, and the meta-learning network can output the updated value of the weight parameter of the layer. Preferably, the second learning unit 606 inputs the gradient of each network layer in the neural network waiting for training into one meta learning network corresponding to the type of the layer among the at least one meta learning network. It may be configured to obtain updated values for the weight parameters of the layer.

With reference to the description of the meta-learning network training device 100 in the examples of the present invention, the number of meta-learning networks is the same as the number of layer types in the neural network. As an example, when the neural network awaiting training is a convolutional neural network, and the convolutional neural network includes a plurality of convolutional layers and a plurality of fully connected layers, the plurality of said above in the convolutional neural network using the first meta-learning network. The gradient information of the convolutional layer is learned, and the gradient information of the plurality of fully connected layers in the convolutional neural network is learned by using a second meta-learning network different from the first meta-learning network.

The third update unit 608 may be configured to update the weight parameters of the neural network awaiting training based on the update values.

Preferably, the third update unit 608 may be configured to update the weight parameters of each network layer in the neural network awaiting training based on the update values.

Preferably, the meta-learning network outputs the updated value of the weight parameter of each network layer of the neural network awaiting training based on the state of the previous update of the weight parameter of the layer.

With reference to the description of the meta-learning network training device 100 in the examples of the present invention, the meta-learning network was generated by the gradient of the weight parameter of the neural network awaiting training and the previous update of the weight parameter of the layer. Based on the state vector, the updated value of the weight parameter of the layer and the state vector generated by this update of the weight parameter of the layer are output.

In the device 600 that trains the neural network using the trained meta-learning network, the processing of the second loss acquisition unit 602, the second gradient calculation unit 604, the second learning unit 606, and the third update unit 608 is performed. , The second iteration is repeated until the end condition is satisfied. As an example, the conditions for ending the second iteration are that the predetermined number of iterations is reached, the loss of the neural network awaiting training converges, the loss drop of the neural network awaiting training is very small, and the loss of the neural network awaiting training is lost. May include at least one of the rises. When the second iteration end condition is satisfied, the training for the weight parameter of the neural network waiting for training can be completed, and the final weight parameter of the neural network waiting for training can be obtained.

The meta-learning network obtained by training the above-mentioned meta-learning network training device 100 or meta-learning network training method 500 has robustness, and the meta-learning network allows the optimization of the weight parameter of the task neural network to be more efficient. Since it is possible to predict the transformation path and calculate the updated value of the weight parameter of the current better task network example, the trained meta-learning network in the embodiment of the present invention is used to train the neural network. In the device 600 to perform, a neural network with excellent performance can be trained quickly and accurately. For example, if a neural network awaiting training is a neural network recommender system for recommending a user a list of items on a shopping website, the current item list is recommended to the user quickly and accurately based on the user's historical browsing record. can do.

The present invention further corresponds to an embodiment of a device 600 that trains a neural network using the meta-learning network obtained by training the meta-learning network training device 100 or the meta-learning network training method 500 described above. , An example of a method of training a neural network using the meta-learning network obtained by training of the meta-learning network training device 100 or the meta-learning network training method 500 is provided.

FIG. 8 is an exemplary flowchart of Method 700 training a neural network using a trained meta-learning network.

As shown in FIG. 8, the method 700 for training the neural network using the trained meta-learning network in the embodiment of the present invention includes a second loss acquisition step S702, a second gradient calculation step S704, and a second. Includes learning step S706 and third update step S708.

In the second loss acquisition step S702, the training data is used to acquire the loss of the neural network waiting for training.

As an example, in the second loss acquisition step S702, the loss of the neural network awaiting training can be obtained by using a normal loss calculation function in this field, for example, softmax.

In the second gradient calculation step S704, the gradient of the loss with respect to the weight parameter of the neural network awaiting training is calculated.

The training-waiting neural network learning process is essentially the process of continuously updating the weight parameters of the training-waiting neural network to minimize losses. The gradients described above may be used to control the direction of updates of the weight parameters of the neural network awaiting training, and the gradients described above further reduce the range of updates of the weight parameters of the neural network awaiting training. Can be controlled.

Preferably, in the second gradient calculation step S704, the gradient of the loss for each network layer weight parameter in the awaiting training neural network can be calculated.

As an example, a training-waiting neural network may include a plurality of network layers (for example, when the training-waiting neural network is a convolutional neural network, the convolutional neural network may include a plurality of convolutional layers, a plurality of fully connected layers, and the like. May be included). Learning the weight parameter of the neural network waiting for training is to learn the weight parameter of each network layer in the neural network waiting for training. Therefore, in the second gradient calculation step S704, it is necessary to calculate the gradient of the loss for the weight parameter of each network layer in the training-waiting neural network.

In the second learning step S706, the gradient is input to each one of the trained at least one meta-learning networks acquired when the first iteration end condition is satisfied, and awaiting training. Get the updated value of the weight parameter of the neural network of.

With reference to the description of the meta-learning network training device 100 in the embodiment of the present invention, the meta-learning network in the embodiment of the present invention is a neural network that learns with respect to the gradient. The gradient of each network layer of the neural network awaiting training is input to the meta-learning network, and the meta-learning network can output the updated value of the weight parameter of the layer.

Preferably, in the second learning step S706, the gradient of each network layer in the neural network waiting for training is input to one meta learning network corresponding to the type of the layer among the at least one meta learning network. , The updated value of the weight parameter of the layer can be obtained.

With reference to the description of the meta-learning network training device 100 in the embodiment of the present invention, the number of meta-learning networks is the same as the number of layer types in the neural network. As an example, when the neural network awaiting training is a convolutional neural network, and the convolutional neural network includes a plurality of convolutional layers and a plurality of fully connected layers, the plurality of said above in the convolutional neural network using the first meta-learning network. The gradient information of the convolutional layer is learned, and the gradient information of the plurality of fully connected layers in the convolutional neural network is learned using a second meta-learning network different from the first meta-learning network.

In the third update step S708, the weight parameter of the neural network waiting for training is updated based on the update value.

Preferably, the meta-learning network can output the updated value of the weight parameter of each network layer of the neural network awaiting training based on the state of the previous update of the weight parameter of the layer.

With reference to the description of the meta-learning network training device 100 in the embodiment of the present invention, the meta-learning network is a state generated by the gradient of the weight parameter of the neural network waiting for training and the previous update of the weight parameter of the layer. Based on the vector, the updated value of the weight parameter of the layer and the state vector generated by this update of the weight parameter of the layer can be output.

In the method 700 of training the neural network using the trained meta-learning network, the processing of the second loss acquisition step S702, the second gradient calculation step S704, the second learning step S706, and the third update step S708 is performed. , The second iteration is repeated until the end condition is satisfied. As an example, the conditions for ending the second iteration are that the predetermined number of iterations is reached, the loss of the neural network awaiting training converges, the loss drop of the neural network awaiting training is very small, and the loss of the neural network awaiting training is lost. Is at least one of the rises. When the second iteration end condition is satisfied, the training for the weight parameter of the neural network waiting for training can be completed, and the final weight parameter of the neural network waiting for training can be obtained.

The meta-learning network obtained by training the meta-learning network training device 100 or the meta-learning network training method 500 described above has robustness, and the meta-learning network allows the optimization path of the weight parameter of the task neural network to be more efficient. The method of training a neural network using a trained meta-learning network in the embodiment of the present invention because it is possible to predict and calculate the updated value of the weight parameter of the current better task network example. With the 700, you can train high-performance neural networks quickly and accurately. For example, if a neural network awaiting training is a neural network recommender system for recommending a user a list of items on a shopping website, the current item list is provided to the user quickly and accurately based on the user's historical browsing record. can do.

Further, the above-mentioned series of processes and the like may be realized by software and / or firmware. When implemented by software and / or firmware, install the programs that make up the software from a storage medium or network on a computer with a dedicated hardware structure, such as the general-purpose machine 800 (eg, computer) shown in FIG. , The computer can perform various functions and the like when various programs are installed.

FIG. 9 is a block diagram of a general-purpose machine 800 that can realize the method and apparatus according to the embodiment of the present invention. The general-purpose machine 800 may be, for example, a computer system. The general-purpose machine 800 is merely an example, and does not limit the application range or function of the method and device according to the present invention. Further, the general-purpose machine 800 does not depend on any module, assembly, or a combination thereof in the above-mentioned method and device.

In FIG. 9, the central processing unit (CPU) 801 performs various processes based on the program stored in the ROM 802 or the program rodged from the storage unit 808 to the RAM 803. The RAM 803 can also store data required when the CPU 801 performs various processes according to needs. CPU 801 and ROM 802 and RAM 803 are connected to each other via bus 804. The input / output interface 805 is also connected to the bus 804.

Further, the following components are connected to the input / output interface 805, that is, an input unit 806 including a keyboard and the like, an output unit including a display such as a liquid crystal display (LCD), and a speaker. The 807 is a storage unit 808 including a hard disk and the like, and a communication unit 809 including a network interface card such as a LAN card and a modem. The communication unit 809 performs communication processing via a network such as the Internet or LAN.

Drive 810 may be connected to input / output interface 805, if desired. The removable medium 811 such as a semiconductor memory can be set in the drive 810 as needed, and the computer program read from the medium can be installed in the storage unit 908.

The present invention also provides a program product that includes a machine-readable command code. When such a command code is read and executed by the machine, the method according to the embodiment of the present invention described above can be executed. Correspondingly, carry such program products, such as magnetic disks (including floppy disks (registered trademarks)), optical disks (including CD-ROMs and DVDs), magneto-optical disks (MD (registered trademarks)). ), And various storage media such as semiconductor storage devices are also included in the present invention.

The above-mentioned storage medium may include, but is not limited to, for example, a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor storage device, and the like.

Further, each operation (process) in the above method can be realized by a method of a computer-executable program stored in various machine-readable storage media.

In addition, the above examples and the like will be further disclosed as additional notes as follows.

(Appendix 1)
A meta-learning network training device
A task network example acquisition unit for obtaining a plurality of task network examples of the task neural network for each task neural network among a plurality of task neural networks for different tasks, and the task network in which the different tasks have similarities. Example acquisition unit;
A loss acquisition unit that acquires the loss of each task network example of each task neural network using training data;
A sampling unit that performs sampling on the loss and acquires a plurality of losses after sampling;
A generalization loss calculation unit that calculates a generalization loss that reflects the total loss of the plurality of losses after sampling based on the plurality of losses after sampling;
A gradient calculation unit that calculates the gradient of the generalized loss corresponding to the weight parameter of the task network example corresponding to each loss of the plurality of losses after sampling;
Each of the gradients is input to one of the at least one meta-learning networks, and the updated value of the weight parameter of the task network example corresponding to each loss of the plurality of losses after sampling. A learning unit in which the meta-learning network is a neural network that learns with respect to a gradient;
The first update unit that updates the weight parameter of the task network example corresponding to each loss of the plurality of losses after the sampling based on the update value; and the loss acquisition unit, the sampling unit, and the generalized loss. The processing of the calculation unit, the gradient calculation unit, the learning unit, and the first update unit is repeatedly executed, and when the predetermined conditions are satisfied, based on the generalization loss when the predetermined conditions are satisfied. Including a second update unit that obtains at least one meta-learning network after the update by training the at least one meta-learning network in a direction in which the loss of the task network example is smaller.
Based on at least one meta-learning network after the update of the second update unit, the loss acquisition unit, the sampling unit, the generalized loss calculation unit, the gradient calculation unit, the learning unit, the first update unit, And an apparatus that repeatedly executes the processing of the second update unit until the first iteration end condition is satisfied.

(Appendix 2)
The device described in Appendix 1
The task neural network is a device that is a neural network recommender system for recommending a list of goods on a shopping website to a user.

(Appendix 3)
The device described in Appendix 1
The task network example acquisition unit is a device that acquires the plurality of task network examples of each task neural network by performing random initialization on each task neural network among the plurality of task neural networks.

(Appendix 4)
The device described in Appendix 1
The sampling unit acquires a plurality of losses after the sampling by sampling the loss by a random selection operation.
A device in which the number of losses after the sampling is controlled by the retention probability parameter of the selection operation.

(Appendix 5)
The device described in Appendix 1
The generalization loss calculation unit is an apparatus that further averages a plurality of losses after sampling, and sets the loss after averaging as the generalization loss.

(Appendix 6)
The device described in Appendix 1
The gradient calculation unit calculates the gradient of the generalized loss with respect to the weight parameter of each network layer in the task network example corresponding to each loss among the plurality of losses after sampling.

(Appendix 7)
The device described in Appendix 6
The gradient calculation unit is used for each network of the task network example other than the task network example corresponding to each loss of the plurality of losses after sampling among all the task network examples acquired for the plurality of task neural networks. A device that sets the gradient of a layer weight parameter to zero.

(Appendix 8)
The device described in Appendix 7
The meta-learning network sets the updated value of the weight parameter of each network layer of the task network example corresponding to each loss of the plurality of losses after sampling based on the state of the previous update of the weight parameter of the layer. A device to output.

(Appendix 9)
The device described in Appendix 1
The number of at least one meta-learning network is the same as the number of layer types in the plurality of task neural networks.
A device in which the gradient of each network layer of a task network example corresponding to each loss of the plurality of losses after sampling is input to a meta-learning network corresponding to the type of the layer.

(Appendix 10)
A meta-learning network training method
For each task neural network among a plurality of task neural networks for different tasks, a task network example acquisition step for obtaining a plurality of task network examples of the task neural network, wherein the different tasks have similarities. Example acquisition step;
Loss acquisition step to acquire the loss of each task network example of each task neural network using the training data;
A sampling step in which sampling is performed on the loss to obtain a plurality of losses after sampling;
A generalization loss calculation step that calculates a generalization loss that reflects the total loss of the plurality of losses after sampling based on the plurality of losses after sampling;
Gradient calculation step to calculate the gradient corresponding to the weight parameter of the task network example corresponding to each loss of the generalized loss among the plurality of losses after sampling;
Each of the gradients is input to one of the at least one meta-learning networks, and the updated value of the weight parameter of the task network example corresponding to each loss of the plurality of losses after sampling. A learning step in which the meta-learning network is a neural network that learns with respect to a gradient;
The first update step of updating the weight parameter of the task network example corresponding to each loss of the plurality of losses after the sampling based on the update value; and the loss acquisition step, the sampling step, and the generalization loss. The processing of the calculation step, the gradient calculation step, the learning step, and the first update step is repeatedly executed, and when the predetermined condition is satisfied, based on the generalization loss when the predetermined condition is satisfied. Including a second update step to obtain at least one updated meta-learning network by training the at least one meta-learning network in a direction in which the loss of the task network example is smaller.
The loss acquisition step, the sampling step, the generalized loss calculation step, the gradient calculation step, the learning step, the first update step, and the second update step based on at least one meta-learning network after the update. A method of iterating through the process until the first iteration end condition is satisfied.

(Appendix 11)
The method described in Appendix 10
The task neural network is a method, which is a neural network recommender system for recommending a list of goods on a shopping website to a user.

(Appendix 12)
The method described in Appendix 10
In the task network example acquisition step, a method of obtaining the plurality of task network examples of each task neural network by performing random initialization for each task neural network among the plurality of task neural networks.

(Appendix 13)
The method described in Appendix 10
In the sampling unit, a plurality of losses after the sampling are acquired by sampling the loss by a random selection operation.
A method in which the number of losses after the sampling is controlled by the retention probability parameter of the selection operation.

(Appendix 14)
The method described in Appendix 10
In the generalization loss calculation step, a method in which a plurality of losses after sampling are further averaged, and the loss after averaging is defined as the generalization loss.

(Appendix 15)
The method described in Appendix 10
In the gradient calculation step, a method of calculating the gradient of the generalized loss with respect to the weight parameter of each network layer in the task network example corresponding to each loss of the plurality of losses after sampling.

(Appendix 16)
The method described in Appendix 15
In the gradient calculation step, among all the task network examples acquired for the plurality of task neural networks, each network of the task network example other than the task network example corresponding to each loss of the plurality of losses after sampling. A method of setting the gradient of a layer weight parameter to zero.

(Appendix 17)
The method described in Appendix 16
The meta-learning network sets the updated value of the weight parameter of each network layer of the task network example corresponding to each loss of the plurality of losses after sampling based on the state of the previous update of the weight parameter of the layer. How to output.

(Appendix 18)
The method described in Appendix 10
The number of at least one meta-learning network is the same as the number of layer types in the plurality of task neural networks.
A method in which the gradient of each network layer of a task network example corresponding to each loss of the plurality of losses after sampling is input to the meta-learning network corresponding to the layer type.

(Appendix 19)
A device that trains a neural network using the meta-learning network obtained by training the meta-learning network training device according to any one of Appendix 1 to 9.
A second loss acquisition unit that uses training data to obtain the loss of a neural network awaiting training;
A second gradient calculation unit that calculates the gradient of the loss with respect to the weight parameter of the neural network awaiting training;
Each of the gradients is input to the meta-learning network of one of the trained at least one meta-learning networks acquired when the first iteration end condition is satisfied, and the neural network awaiting training A second learning unit that obtains an updated value of the weight parameter; and a third update unit that updates the weight parameter of the neural network awaiting training based on the updated value.
An apparatus that repeatedly executes the processes of the second loss acquisition unit, the second gradient calculation unit, the second learning unit, and the third update unit until the second iteration end condition is satisfied.

(Appendix 20)
The device described in Appendix 19
A device in which the gradient of each network layer in the neural network awaiting training is input to the meta-learning network corresponding to the type of the layer among the at least one meta-learning network.

Although the preferred embodiment of the present invention has been described above, the present invention is not limited to this embodiment, and any modification to the present invention belongs to the technical scope of the present invention unless the gist of the present invention is abandoned.

Claims (10)

A meta-learning network training device
A task network example acquisition unit for obtaining a plurality of task network examples of the task neural network for each task neural network among a plurality of task neural networks for different tasks, and the task network in which the different tasks have similarities. Example acquisition unit;
A loss acquisition unit that acquires the loss of each task network example of each task neural network using training data;
A sampling unit that performs sampling on the loss and acquires a plurality of losses after sampling;
A generalization loss calculation unit that calculates a generalization loss that reflects the total loss of the plurality of losses after sampling based on the plurality of losses after sampling;
A gradient calculation unit that calculates the gradient of the generalized loss corresponding to the weight parameter of the task network example corresponding to each loss of the plurality of losses after sampling;
Each of the gradients is input to one of the at least one meta-learning networks, and the updated value of the weight parameter of the task network example corresponding to each loss of the plurality of losses after sampling. A learning unit in which the meta-learning network is a neural network that learns with respect to a gradient;
The first update unit that updates the weight parameter of the task network example corresponding to each loss of the plurality of losses after the sampling based on the update value; and the loss acquisition unit, the sampling unit, and the generalized loss. The processing of the calculation unit, the gradient calculation unit, the learning unit, and the first update unit is repeatedly executed, and when the predetermined condition is satisfied, the generalization loss when the predetermined condition is satisfied is obtained. Based on this, the task network example includes a second update unit that obtains at least one updated meta-learning network by training the at least one meta-learning network in a direction in which the loss is smaller.
Based on at least one meta-learning network after the update of the second update unit, the loss acquisition unit, the sampling unit, the generalized loss calculation unit, the gradient calculation unit, the learning unit, the first update unit, And an apparatus that repeatedly executes the processing of the second update unit until the first iteration end condition is satisfied. The device according to claim 1.
The task neural network is a device that is a neural network recommender system for recommending a list of goods on a shopping website to a user. The device according to claim 1.
The task network example acquisition unit is a device that acquires the plurality of task network examples of each task neural network by performing random initialization on each task neural network among the plurality of task neural networks. The device according to claim 1.
The sampling unit acquires a plurality of losses after the sampling by sampling the loss by a random selection operation.
A device in which the number of losses after the sampling is controlled by the retention probability parameter of the selection operation. The device according to claim 1.
The generalization loss calculation unit is an apparatus that further averages a plurality of losses after sampling, and sets the loss after averaging as the generalization loss. The device according to claim 1.
The gradient calculation unit calculates the gradient of the generalized loss with respect to the weight parameter of each network layer in the task network example corresponding to each loss among the plurality of losses after sampling. The device according to claim 6.
The gradient calculation unit is used for each network of the task network example other than the task network example corresponding to each loss of the plurality of losses after sampling among all the task network examples acquired for the plurality of task neural networks. A device that sets the gradient of a layer weight parameter to zero. The device according to claim 7.
The meta-learning network sets the updated value of the weight parameter of each network layer of the task network example corresponding to each loss of the plurality of losses after sampling based on the state of the previous update of the weight parameter of the layer. A device to output. A meta-learning network training method
For each task neural network among a plurality of task neural networks for different tasks, a task network example acquisition step for obtaining a plurality of task network examples of the task neural network, wherein the different tasks have similarities. Example acquisition step;
Loss acquisition step to acquire the loss of each task network example of each task neural network using the training data;
A sampling step in which sampling is performed on the loss to obtain a plurality of losses after sampling;
A generalization loss calculation step that calculates a generalization loss that reflects the total loss of the plurality of losses after sampling based on the plurality of losses after sampling;
Gradient calculation step to calculate the gradient corresponding to the weight parameter of the task network example corresponding to each loss of the generalized loss among the plurality of losses after sampling;
Each of the gradients is input to one of the at least one meta-learning networks, and the updated value of the weight parameter of the task network example corresponding to each loss of the plurality of losses after sampling. A learning step in which the meta-learning network is a neural network that learns with respect to a gradient;
The first update step of updating the weight parameter of the task network example corresponding to each loss of the plurality of losses after the sampling based on the update value; and the loss acquisition step, the sampling step, and the generalization loss. The processing of the calculation step, the gradient calculation step, the learning step, and the first update step is repeatedly executed, and when the predetermined condition is satisfied, based on the generalization loss when the predetermined condition is satisfied. Including a second update step to obtain at least one updated meta-learning network by training the at least one meta-learning network in a direction in which the loss of the task network example is smaller.
The loss acquisition step, the sampling step, the generalized loss calculation step, the gradient calculation step, the learning step, the first update step, and the second update step based on at least one meta-learning network after the update. A method of iterating through the process until the first iteration end condition is satisfied. A device that trains a neural network using the meta-learning network obtained by training the meta-learning network training device according to any one of claims 1 to 8.
A second loss acquisition unit that uses training data to obtain the loss of a neural network awaiting training;
A second gradient calculation unit that calculates the gradient of the loss with respect to the weight parameter of the neural network awaiting training;
Each of the gradients is input to the meta-learning network of one of the trained at least one meta-learning networks acquired when the first iteration end condition is satisfied, and the neural network awaiting training A second learning unit that obtains an updated value of the weight parameter; and a third update unit that updates the weight parameter of the neural network awaiting training based on the updated value.
An apparatus that repeatedly executes the processes of the second loss acquisition unit, the second gradient calculation unit, the second learning unit, and the third update unit until the second iteration end condition is satisfied.

Images