JP2017126260A - Machine learning method and machine learning device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a method for enabling the learning rate to be autonomously optimized in the learning process such as a neural network and preventing learning from being led to a dead end.SOLUTION: A storage device holds programs, a plurality of systems including a plurality of structure parameters corresponding to the programs, and a learning parameter corresponding to each system. A machine learning method includes: the first procedure for causing each system to learn a prescribed data set by using a learning parameter corresponding to each system; the second procedure for evaluating each system by a prescribed evaluation method; and the third procedure for selecting the first system having the lower evaluation and the second system having the higher evaluation than the first system out of the plurality of systems, generating the copy of the second system and changing at least one of the learning parameter corresponding to the second system and the learning parameter corresponding to the copy of the second system. The machine learning method executes the first procedure to the third procedure again for the plurality of systems other than the first system.SELECTED DRAWING: Figure 5

Description

  The present invention relates to machine learning using a neural network or the like.

  In recent years, research on the recognition of speech, images, etc. by multilayer neural networks, so-called deep learning research, has been activated. This activation is based on the development of a method to learn a multi-layer (deep) neural network of 4 or more layers that was difficult to learn by using a mechanism called auto-encoder. This is because the recognition rate of voice and images by the convolutional neural network has been greatly improved.

  The steepest descent method is used as a basic algorithm for the back propagation learning method used in neural network training and the like as well as deep learning, and as various learning and optimization methods. This method is a deterministic search. However, since this method is almost certainly caught by a local optimum value that is not a global optimum, a stochastic gradient descent method that is usually a stochastic search is used. In learning of a neural network, there is a learning rate as a parameter for controlling learning. The initial value of the learning rate is determined by an experimenter (human) and is a constant in the learning process or changes according to a predetermined schedule. In Patent Document 1, the learning rate is increased when the evaluation of the solution obtained in the learning process is high, and the learning rate is decreased when the evaluation is poor. Other methods for adaptively determining the learning rate in this way have been proposed, but all have problems and networks that can be adapted.

  There is a genetic algorithm (GA) as a method for probabilistic optimization. GA is an optimization method originally developed independently of neural networks, but is sometimes used in conjunction with machine learning methods. In particular, many methods for using back propagation learning and GA in combination have been developed in neural networks. The most common method is to optimize the structure and weight of a neural network by GA as in Patent Document 2 and Non-Patent Document 1, but GA is also used to optimize the learning method. In these methods, the GA operation, that is, mutation or crossover is repeated after the entire process of backpropagation learning is performed. In addition, as a combination with other probabilistic search methods, a method combined with particle swarm optimization methods, which is a probabilistic search method that has been developed relatively recently, has been studied. Yes. An algorithm that parallels the stochastic gradient descent method has also been developed.

US Pat. No. 6,269,351 US Pat. No. 6,601,053

Marshall, S. J. and Harrison, R. F., “Optimization and Training of Feedforward Neural Networks by Genetic Algorithms”, 2nd International Conference on Artificial Neural Networks, pp. 39-43, November 1991.

  Despite recent advances in research related to deep learning and back propagation learning of neural networks, there are some difficult problems in back propagation learning of neural networks. The first problem is how to determine the learning rate. In other words, the optimal learning rate depends on the structure of the neural network and also on the problem. Furthermore, the method of keeping the learning rate constant in the learning process is relatively widely used, but usually it is better to decrease as learning progresses, so it is necessary to include a method to change it in the learning process. is there. There are many references on how to determine the learning rate. A method has also been proposed in which the learning rate is automatically determined as a function of time in the learning process. Recently, various other adaptive methods have been devised, but most of these methods are clever but may not work. A simpler and more powerful method is sought.

  The second problem in back propagation learning is to escape from the local minima. That is, depending on the initial value of the weight of the neural network, only a value that is as low as the minimum value can be obtained as a value of a function (such as an error rate) that should be minimized even by back propagation learning. In addition, even if a relatively good value is obtained once, it may not return from the minimum value if learning is further promoted. It is a problem to escape from such a narrow path and find the optimum threshold value.

  The reason why back propagation learning is easily caught by a dead path is that back propagation learning is a method of local search. As described above, the stochastic gradient descent method is usually used as a method for back propagation. However, if a probabilistic search is to be performed, improvement can be expected by using it in combination with any of a variety of established probabilistic search methods. The purpose of combining conventional back propagation learning and GA is to achieve this, but in the conventional method, GA does not act in the process of one back propagation learning. It cannot be optimized.

  In order to solve the above problems, an embodiment of the present invention is a machine learning method executed by a computer including a processor and a storage device connected to the processor, wherein the storage device is a predetermined process. Holding a plurality of programs for realizing a plurality of systems for executing a plurality of, a plurality of structural parameters corresponding to each of the plurality of programs, a plurality of learning parameters corresponding to the plurality of systems, The learning parameter is a parameter for designating the change of the structural parameter in the learning executed by each system, and the machine learning method uses the learning parameter corresponding to each system to cause each processor to A first procedure for learning a predetermined data set; and a second procedure in which the processor evaluates each of the systems by a predetermined evaluation method. And the processor selects a first system and a second system having a higher evaluation than the first system from the plurality of systems, and copies the program corresponding to the second system and the plurality of structure parameters. A program for realizing replication of the second system and a plurality of structural parameters corresponding thereto are generated, and a copy of the learning parameter corresponding to the second system is generated as a learning parameter corresponding to the replication of the second system. At least of the learning parameter corresponding to the second system and the learning parameter corresponding to the duplication of the second system so that the learning parameter corresponding to the second system and the learning parameter corresponding to the duplication of the second system are different from each other A third procedure for changing one of the other systems, For serial multiple systems, wherein the third instructions from said first procedure is executed again.

  According to one aspect of the present invention, optimization can be performed by autonomously determining the learning rate (or learning parameter, optimization / search control parameter).

It is explanatory drawing of encoding to the chromosome of the parameter of the multilayer neural network in embodiment of this invention. It is explanatory drawing of the learning method which combined back propagation learning and GA in embodiment of this invention. It is explanatory drawing of the example of the change of the learning rate in embodiment of this invention. It is explanatory drawing of the example of the Euclidean distance of the best individual | organism | solid and each other individual | organism | solid in embodiment of this invention. It is explanatory drawing of the outline | summary of a structure and operation | movement of the neural network design tool for data identification of embodiment of this invention. It is a block diagram explaining the hardware constitutions of the neural network design tool for data identification of embodiment of this invention.

  An embodiment of the present invention will be described. In this embodiment, the method in this embodiment combines the conventional back propagation learning method and the GA by performing selection and mutation in the GA for each step of parallel back propagation learning (1 epoch). Similar to the method described above, the neural network is optimized as a learning result, and the learning rate in the back propagation learning process, which was not possible with the conventional method, is optimized.

(overall structure)
An overall configuration of this embodiment, that is, an overview of the configuration and operation of the data identification neural network design tool 500 will be described with reference to FIG. The data identification neural network design tool 500 is a tool for the user to design a neural network for image data identification and the like. The original data 504 as learning data and teacher information 505 for explaining the original data 504 are used. Is entered. A video or a still image can be input as the original data 504, and the teacher information 505 is a plurality of rectangular areas (indicating where to identify information such as vehicles or pedestrians in the video or still image). Include bounding box) information. The designed neural network is taken out from the learning neural network group 507 in the form of weight and bias for the identifying neural network 508. This design result can be tested by giving data 509 to be identified as test data, and as a result, an identification result output 510 is output.

  The learning control computer 501 is a terminal that gives a user input to the learning control program 502 and gives an output from the learning control program to the user. The learning control computer 501 can also operate a learning control program 502, a learning data generation program 503, a learning neural network group 507 constituted by a plurality of neural networks, and an identifying neural network 508. It is possible to store original data 504 which is input / output of a program, teacher information 505, learning data 506 with teacher information, data 509 to be identified, and identification result output 510. However, the learning data generation program 503, the learning neural network group 507, the identification neural network 508, the original data 504, the teacher information 505, the learning data with teacher information 506, the data 509 to be identified, and the identification result output 510 are learning control. It can also be stored and executed on another computer designated by the program 502 (see FIG. 6).

  The user instructs the learning control computer 501 which data to use as the original data 504 and the teacher information 505, the learning data generation program 503, and the parameters for the learning neural network group 507 and the identifying neural network 508. Enter. Parameters for the learning neural network group 507 include, as will be described later, the number of neural networks, that is, the number of individuals, the seeds of random numbers for randomly determining the weights and biases of the neural network, the weights and biases of a plurality of neural networks. The distribution function such as a normal distribution that determines the distribution of, the average value and the standard deviation, and the average value and standard deviation of the initial value of the learning rate are included. However, when using default values as these values, the user does not need to input them. This input can include the upper limit of the number of steps (epoch number), the target value of error, and the target value of learning rate as stop conditions of the neural network. The user can also input data 509 to be identified at the time of this input.

  In response to a user instruction, the learning control program 502 operates the learning data generation program 503 to generate learning data 506 with teacher information. The original data 504 is a frame image having a relatively large size (for example, 640 × 480), and the learning neural network group 507 cannot be used as it is, so that a single frame image is used for a relatively small size (for example, 24). A large number of × 48) patch images are generated and used as learning data 506 with teacher information. These images are classified into positive examples, that is, images to be detected and negative examples, ie, images that should not be detected, by the learning data generation program 503 using the teacher information 505. The classification is stored in a one-to-one correspondence with the image. In the teacher information 505, the images to be detected may be divided into classes. In this case, the class is stored in the learning data with teacher information 506. That is, a pair of patch image and class is stored.

  In response to a user instruction, the learning control program 502 operates the learning neural network group 507 to perform back propagation learning. That is, an image including the learning data with teacher information 506 is input to the neural network, and the weight and bias of the learning neural network group 507 are updated based on the difference between the output and the teacher information. A method for learning a plurality of neural networks will be described later.

  FIG. 6 is a block diagram illustrating the hardware configuration of the data identification neural network design tool 500 according to the embodiment of this invention.

  The neural network design tool 500 for data identification according to the present embodiment can be realized by, for example, computers 600, 610, and 620 connected to each other by a network 630.

  The computer 600 corresponds to the learning control computer 501 in FIG. 5, and includes a CPU (Central Processing Unit) 601, a memory 602, an I / F (Interface) 603, and an HDD (Hard Disk Drive) 604 that are connected to one another. The CPU 601 is a processor that executes a program stored in the memory 602. The memory 602 is a so-called main storage device that stores programs executed by the CPU 601, data to be processed, and the like. The memory 602 of this embodiment stores a learning control program 502. The processing executed by the learning control program 502 in the present embodiment is actually executed by the CPU 601 according to the learning control program 502. The I / F 603 transmits and receives data to and from the computers 610 and 620 via the network 630. The HDD 604 is a so-called auxiliary storage device that stores a program executed by the CPU 601 and data to be processed. For example, the learning control program 502 may be stored in the HDD 604 and copied to the memory 602 as necessary.

  The computer 610 includes a CPU 611, a memory 617, an I / F 616, and an HDD 615 that are connected to each other, and further includes a GPU (Graphics Processing Unit) 612 that is connected to the CPU 611. The CPU 611 is a processor that executes a program stored in the memory 617. The memory 617 is a so-called main storage device that stores a program executed by the CPU 611 and data to be processed. The memory 602 of the present embodiment stores a learning data generation program 503. The processing executed by the learning data generation program 503 in the present embodiment is actually executed by the CPU 611 according to the learning data generation program 503. The I / F 616 transmits and receives data to and from the computers 600 and 620 via the network 630. The HDD 615 is a so-called auxiliary storage device that stores programs executed by the CPU 611 and the like, data to be processed, and the like. The HDD 615 of this embodiment stores original data 504 and teacher information 505.

  The GPU 612 is a processor having a plurality of processor cores 613 and a memory 614. The memory 614 of this embodiment stores a learning neural network group 507 and learning data 506 with teacher information. The operation of the learning neural network group 507 in this embodiment is executed by the GPU 612.

  Note that the learning data with teacher information 506 may be stored in the HDD 615 after being generated from the original data 504 and the teacher information 505 by the learning data generation program 503, and then copied to the memory 614 by the CPU 611. Similarly, the learning neural network group 507 may be stored in the HDD 615 or the memory 617 and copied to the memory 614 by the CPU 611.

  Each learning neural network included in the learning neural network group 507 includes a set of structural parameters such as weights and biases between neurons included in each learning neural network stored in the memory 614, and the structural parameters and inputs. And a program that calculates an output based on the learned data and performs learning by a predetermined learning method (for example, reverse propagation learning) based on the evaluation of the output. That is, each learning neural network is a system realized by the GPU 612 executing a corresponding program using the structure parameter.

  The computer 620 includes a CPU 621, a memory 626, an I / F 625, and an HDD 627 that are connected to each other, and further includes a GPU 622 that is connected to the CPU 621. The CPU 621 is a processor that executes a program stored in the memory 626. The memory 626 is a so-called main storage device that stores a program executed by the CPU 621 and data to be processed. The I / F 625 transmits and receives data to and from the computers 600 and 610 via the network 630. The HDD 627 is a so-called auxiliary storage device that stores programs executed by the CPU 621 and the like, data to be processed, and the like.

  The GPU 622 is a processor having a plurality of processor cores 623 and a memory 624. The memory 624 of this embodiment stores an identification neural network 508, data 509 to be identified, and an identification result output 510. The operation of the identification neural network 508 of this embodiment is executed by the GPU 622.

  Note that the identification neural network 508 and the data 509 to be identified may be stored in the HDD 627 or the memory 626 and copied to the memory 624 by the CPU 621. The identification result output 510 may be copied from the memory 624 to the memory 626 or the HDD 627 by the CPU 621 and further transmitted to the computer 600 or the like via the I / F 625 and the network 630 as necessary.

  FIG. 6 shows an example of the hardware configuration of the data identifying neural network design tool 500, and there may actually be various modifications. For example, the computer 610 may have a plurality of GPUs 612. In this case, the memory 614 of each GPU 612 stores each learning neural network included in the learning neural network group 507 and learning data 506 with teacher information, and each GPU 612 learns one learning neural network. You may go. As a result, learning of a plurality of neural networks is executed in parallel, and the time required for learning is shortened.

  Alternatively, any two or all functions of the computers 600, 610, and 620 may be realized by one computer. Alternatively, the processing executed by the GPU 612 or the like in the above example may be executed by the CPU 611 or the like. Alternatively, the HDD 604 or the like may be replaced with a storage device other than the HDD such as a flash memory.

(Chromosome expression)
In this embodiment, the parameters of the multilayer neural network are encoded in the chromosome of the genetic algorithm (GA) as shown in FIG. Since one individual has only one chromosome, chromosome and individual are synonymous here. FIG. 1A shows an example of a three-layer perceptron. The point of encoding the binding weight 101 on the chromosome is the same as the encoding method in the conventional method combining the neural network and the GA, but in this embodiment, the learning rate 102 is also encoded. . In FIG. 1 (a), only the weight is described among the connection parameters between neurons. However, a constant term, that is, a bias can also be encoded into the chromosome. Here, since the structure of the neural network is fixed, the structure is not expressed on the chromosome. However, by expressing the structure as well, neurons and connections between neurons are changed according to a predetermined mutation rule in the learning process (for example, Structural optimization such as (deletion) can also be realized using GA. That is, make the chromosome variable length and describe the parameters of each neuron (delete the entire neuron when deleting a neuron), or describe the parameters related to the connection between neurons (when deleting a connection) Delete it).

  FIG. 1B shows the encoding of a convolutional neural network (CNN) often used in image recognition or the like. Compared to FIG. 1 (a), the number of parameters increases and the size of the chromosome increases, but the same is true in terms of encoding parameters.

  The chromosome structure need not be the same for all individuals. That is, the calculation can be performed using a neural network having a different structure (for example, different connections between neurons, or different numbers of neurons and connections between neurons). Even in this case, the mutation can be carried out in the same manner. In GA, an operation called crossover is used in addition to mutation, but crossover can be performed not only between chromosomes having the same structure but also between chromosomes having different structures. For example, each of the two neural nets can be divided between any layers, or divided into two in a specific layer and crossed together. In this case, the structure of the neural network can be maintained by simply reconnecting if the number of connections to be cut is made small. When there is a remainder in the connection, the structure of the neural network can be reconstructed by deleting the connection.

  Chromosome coding is not limited to multilayer neural networks, and can be applied to other types of systems for learning or optimization / search. That is, coding system structural parameters (corresponding to the binding weight in a neural network) and learning parameters or parameters that control the optimization / search process (corresponding to the learning rate) and applying mutation and crossover operations Can do.

(Learning method)
Hereinafter, a learning method in which back propagation learning and GA are combined will be described with reference to FIG. This learning method is called a LOG-BP learning method (learning-rate-optimizing genetic back-propapation learning method). In this learning method, by preparing a plurality of chromosomes shown in the previous section and performing reverse propagation learning in parallel, the masses on those chromosomes are autonomously mutated. The learning rate is stochastically varied.

  First, a computer (computer 610 in the example of FIG. 6) in which the program of FIG. 2 (that is, the program included in the learning neural network group 507) is embedded initializes the chromosome (201). The number of individuals can be variable, but here it is a fixed number (for example, 20). The masses and learning rates of these chromosomes are determined by random numbers. The initialization of the omigami may be performed in the same manner as in the normal back propagation learning. However, the omigami and the bias may be determined by a normally distributed random number, for example. The learning rate is also appropriately distributed using random numbers, but may be determined by, for example, a normal distribution. It can be said that the learning rate should be relatively large so that the frequency of divergence is not too high. Parameters for determining these initial values can be input from the outside via the learning control program 502. That is, it is possible to specify the learning rate, the average value of the weight, the standard deviation, the shape of the distribution, and the seed of the random number before starting the learning.

  Next, the computer performs one step (1 epoch) of back propagation learning for each individual (203). This step corresponds to one generation in GA. For example, when learning image data, a computer prepares as many image data as possible as training data in advance, and uses a part of the data as verification data. Also, evaluation data consisting of the same image format is prepared as necessary. Then, the computer learns all the training data once (learning using image data will be described later). Probabilistic gradient descent in units of mini-batch (i.e., the steepest descent method that learns all of the training data at once and the basic probabilistic gradient descent method that learns one by one) When the method is used, the training data stored in the array is divided for each mini-batch and back-propagated once to learn. At this time, the value encoded in each chromosome is used as the learning rate.

  Next, the computer updates the mass on the chromosome with the mass changed by learning (204). That is, in this method, the value of the worm on the chromosome is not changed externally, but is autonomously updated based on the random number and learning of each individual. In other words, instead of Darwinian inheritance, Lamarck-like inheritance in which acquired traits are inherited as they are is realized. However, this process is consistent with the basics of GA if the renewal of rice cake is considered a mutation.

  Next, the computer evaluates each updated individual (eg, an egg by asexual reproduction of the original individual) using the verification data (205). If there is an individual having a sufficient evaluation value, the calculation may be terminated here (206). When there is no individual having a sufficient evaluation value, if the evaluation result is an error rate, the value is better. Therefore, the computer selects based on the value. That is, the individual with the highest value, i.e. the worst evaluation (i.e. the chromosome), is killed (i.e. deleted) and the one with the lowest value, i.e. the best evaluation, is copied (creating a dizygotic twin) (207 ). This makes the population unchanged. However, by making the generation (copying) probability and the death probability not the same, the number of individuals can be gradually increased or decreased. The computer mutates the learning rate η on the chromosome according to the following formula for the individual generated by copying.

η '= fη (probability 0.5)
η '= η / f (probability 0.5)

  That is, which expression is applied is determined so as to have an equal probability by a random number. f is a value of about 1.2, for example, and is related to a rule used in the adaptive back propagation learning method (this rule is originally independent of GA). However, the change in the learning rate according to the above formula is an example, and as long as it is determined that the learning rate of the individual with the best evaluation differs from the learning rate of the individual generated by copying it, for example, both The learning rate may be determined by a method other than the above, such as changing the learning rate. In the above example, the individual with the worst evaluation is deleted and a copy of the individual with the best evaluation is generated. However, as long as the evaluation of the individual with the copy is better than the evaluation of the individual to be deleted, it is deleted. It is not necessary to limit the target of copying to the worst individual and the best individual.

  Deletion of the chromosome in the process 207 is an example of a process for excluding the chromosome from the subsequent machine learning process, and the chromosome may be actually deleted from the memory 614, or the chromosome may be stored in the memory. For example, the learning control program 502 may exclude the chromosome from the machine learning process by controlling the learning so that the learning control program 502 does not perform the machine learning on the chromosome in the subsequent epoch. The same applies to the deletion of chromosomes in the following description.

  The computer repeatedly calculates the above steps (epoch) until an appropriate solution is obtained or until almost no change occurs (209). The calculation stop condition may be determined in the same manner as in the normal back propagation learning method. Processes 202 and 208 are processes for initializing and updating parameters relating to this iteration.

  In the above description, selection and mutation are performed for each step. However, in practice, it is better to select and control the average value of the selection and the number of mutations for each step. That is, if selection / mutation is performed exactly once in each step, the number of times becomes excessive when the number of individuals is not large, and the number becomes too small when the number of individuals is large. Therefore, an average value of the number of times of selection / mutation for each step is determined in advance, and the actual number of times may be determined stochastically. If the number of selections is excessive, the search range is too fast. If it is too small, the search range is too wide. By appropriately controlling the number of selections, it is possible to search a wide area at the start of calculation, gradually narrow the search range, and obtain a solution well.

  When using another learning system or optimization / search system instead of the neural network, evaluation is performed for each step of the repeated learning or optimization / search, and selection is made based on the result. The learning parameter that controls the learning process or the value of the optimization / search parameter that controls the optimization / search process is mutated. Regarding this variation, if a distance (difference when it is a scalar value) can be defined between a plurality of values of these parameters, a random parameter value may be generated using a random number. If the distance cannot be defined, any value can be selected by a random number.

(Supplement about learning method and application range)
The following describes six points related to the variation of the LOG-BP learning method and its application range expansion. First, each individual can refer to an evaluation value based on the verification data and reflect it in learning. In the above algorithm, the evaluation value is also used for selection. However, since the selection is considered to be external, the evaluation value for selection can be assigned separately. For example, it is possible to give evaluation data in addition to the verification data and use it for selection. That is, different criteria can be used as selection criteria for each individual (standard in back propagation learning) and external selection criteria (standard in GA).

  Second, in the above method, the structure and parameters of the neural network are not changed by selection / mutation. There is no expansion in the following evaluation, but it is possible to extend it to optimize structure and parameters and use mutations and crossings. For example, each chromosome includes information indicating the number of neurons in each neural network and the presence / absence of connections between neurons, and the computer 610 mutates the chromosomes based on a predetermined mutation rule in process 207, thereby connecting the neurons. Can be cut off or neurons can be extinguished. The same applies to the addition of neurons described later. Even when a system other than a neural network is used, a part of the system can be changed or deleted by the same method.

  Third, it is also possible to add neurons by mutation. In order to avoid invalidating the learning that has already been done by adding a neuron, it can be thought that the value of the peach should be reduced (same as not adding if the peach is 0). However, the added neuron may not be activated. If neurons are added without increasing the training data, overfitting occurs, and it seems that the evaluation value is likely to improve. For this reason, it can be said that the evaluation function should be determined so that the evaluation value is improved when the scale of the neural network is small. For example, it can be considered that a minimum description length (MDL) is added as a part of the evaluation value (in other words, a description length penalty is given). That is, the description length of the neural network model is set as a part of the evaluation value. When a chromosome describes the structure of a neural network, it can be said that it describes a model, so the length of the chromosome can be used as the description length. When using a system other than a neural network, a part of the system can be added.

  Fourth, as already described in supplementary explanation, the LOG-BP learning method can be extended not only to the application to the neural network as described above but also to other learning methods. That is, it is assumed that there is a system that performs classification / detection (corresponding to a neural network) and a learning method for training it. The learning is performed iteratively, and there are parameters that control the learning. At this time, if a method for evaluating the effect of learning in the iteration process is provided, the present learning method (LOG learning method) can be applied in the same manner as in the back propagation learning of a neural network. That is, the parameters that determine the structure of the system are expressed as chromosomes, a plurality of chromosomes are initialized, learning is started, and learning, evaluation, mutation and selection are repeated. The learning control parameter to be mutated does not need to be a real value, and a method of mutating it to another value only needs to be provided instead of the two expressions for the mutation.

  Fifth, in the above embodiment, all the individuals are the same type of neural network, but even if the individuals use different types of neural networks or other learning methods for each individual, the evaluation functions are arranged and the learning parameters If the mutation method is designated, it can be evaluated, selected and mutated by the above method. That is, a neural network and other learning methods can be mixed and applied. Specifically, for example, the learning neural network group 507 may include a plurality of neural networks having different connections between neurons, or a plurality of neural networks having different numbers of neurons and connections between neurons. But you can. Back propagation learning may be used for all of them, or a different learning method may be used for each neural network. As a result, it is possible to identify an optimal one among different types of neural networks or different types of learning methods.

  Sixth, in the above embodiment, learning is basically performed on one computer (for example, the computer 610 in FIG. 6). However, a plurality of computers (for example, a plurality of computers 610) are prepared, and each computer is provided. By assigning one chromosome, parallel calculation can be performed by a processor (for example, CPU 611 or GPU 612 of each computer 610) included in these computers. The evaluation value of each chromosome can be selected by collecting in one computer. When proceeding to the next epoch, it is necessary to replace one or more chromosomes of those computers. This operation can be performed by exchanging a small amount of data between computers. Conventional back-propagation learning using different parameters can be performed on multiple computers by the conventional technology, but compared to that, the above method can be used to achieve faster and more optimal values. There is an advantage that the probability of being obtained increases. Alternatively, as described with reference to FIG. 6, for example, the computer 610 includes a plurality of GPUs 612, and each GPU 612 can be used to perform the same processing as described above.

(Learning example of image data set)
This section describes a method for learning a pedestrian image data set according to the learning method described above. An example of a pedestrian image data set is the Caltech pedestrian data set. The same method can be applied when a face image, an object image, a character image, or the like is used instead of a pedestrian image. The data set used in this learning includes multiple videos. Annotation data is attached to each video, and there is also bounding box data that indicates the position and size of the pedestrian. The video is composed of one or more videos for training and one or more videos for testing.

  Half of the training data are positive examples, but they are generated as follows. The bounding box specified in the above data set was cut out and normalized to a size of 24 × 48 to prepare 100,000 images, which were combined with the 100,000 obtained by reversing left and right. 200,000 images are used twice as positive examples.

  A negative example of the remaining half of the training data is generated as follows. Use 200,000 images with a size of 24 x 48 that are cut out from parts other than the bounding box of the Caltech pedestrian dataset. In generating this initial negative example, the position is determined by a random number. An image with a size different from 24 × 48 can be extracted and resized, but only an image with a size of 24 × 48 can be extracted. Further, 200,000 negative examples are generated from the misrecognized portion by applying the CNN of one stage of the convolution layer trained using 200,000 positive examples and 200,000 negative examples to the original data set. That is, the CNN determines that a pedestrian is included, but an image deviating from the bounding box is a new negative example. These negative examples are combined to 400,000, and 800,000 images are prepared including positive examples.

  The numerical value included in these data is 0 to 255 in the original data such as the Caltech pedestrian data set, but this is made a floating point number in the range of about −1 to 1 and further corrected so that the average becomes 0.

  The following describes the CNN structure and hyper parameters to be used. The example is as follows, assuming that there are two layers of convolution layers.

First stage of convolution layer: filter size 5 × 5, number of filters 16, non-linear (activation) function: ReLU
・ Pooling layer first stage: Maximum pooling. Size 2x2
2nd stage of convolution layer: filter size 3 × 3, number of filters 26, 28 or 32, nonlinear function: ReLU
・ Pooling layer 2nd stage: Maximum pooling. Size 2x2
-Hide layer (1 stage): 50 neurons
-Output layer: Logistic regression. Number of neurons 2 (expected output is [1, 0] or [0, 1])
・ Mini-batch size: 250 (smaller than the original deep learning tutorial, but may be too large)
Here, ReLU means a line function of f (x) = if x <0 then 0 else x.

(Change in learning rate)
An example of a change in the learning rate in the learning process of this embodiment is shown in FIG. The average value and standard deviation of the learning rate can be measured every epoch, but in this figure they are plotted every 5 epochs. When the initial value of the learning rate is too low (FIG. 3 (b)), the average value initially increases and then decreases, but does not increase in FIG. 3 (a). In FIG. 3B, the initial value is slightly too small, but it can be considered that it was adjusted autonomously. The standard deviation of the learning rate is relatively large in the initial state, but usually decreases as learning progresses. However, it may increase. In any case, the learning rate is adjusted autonomously. When learning or optimizing / searching another system instead of learning a neural network, other learning parameters controlling the learning process or parameters controlling the optimization / searching process are autonomous instead of the learning rate. Adjusted.

  The learning rate (or learning parameter, optimization / search parameter) value and its change can be displayed by means such as a graph or table as shown in FIG. 3 when learning is performed or at the end of learning.

(Supplement about learning performance)
First, the number of individuals (number of chromosomes) will be described. The more it can be, the stochastic solution can be obtained more probabilistically. However, if not all can be calculated in parallel, the larger the number of individuals, the longer it takes. As an example of a value that can balance the calculation time and the search range, it can be considered that the number of individuals is about 12. It takes about 8 hours using a GPU to experiment to 100 epoch with 12 individuals.

  Second, a method for maintaining diversity, that is, a method for controlling the globality of the search will be described. If the frequency of selection / mutation is high, all individuals tend to be copies from one individual. Therefore, a satisfactory solution cannot be reached if the individual is located in a portion having only local optimum values that are separated from the global optimum values. It can be considered that the probability of selection / mutation in 1 epoch needs to be 5% or less (or 0.6 if the number of individuals is 12). If the probability of selection / mutation is reduced, a more global search is performed even if the learning process proceeds. Conversely, if the probability of selection / mutation is increased, the search becomes local relatively early.

  Third, the global search performance will be described. A chromosome number of about 12 is not sufficient for global search. In this case, the search is initially made at 12 locations. However, since individuals that are generated by copying gradually increase, even if the above-described method for maintaining diversity is applied, the search location is immediately set to around 3 points. Because it is drowned. The number of search ranges can be estimated by calculating and displaying the Euclidean distance between the best individual so far and the current individual. FIG. 4 shows an example of such display. The four numerical values on the right of each row are the Euclidean distances of the weights in each row. It can be seen that the individual 9 (the row with the leftmost number being 9) is the best individual and is searching for at least two neighborhoods in parallel.

  Fourth, the removal of the divergent individual will be described. In the simple back propagation learning algorithm used in this experiment, the weight often diverges (becomes “nan”) due to back propagation. Such individuals are considered zombies, that is, individuals that are unlikely to be solved by continuing calculations, and should be deleted. Although the logic for deletion can be included, such an individual has an extremely deteriorated evaluation value. Therefore, since this algorithm is deleted preferentially, it is not always necessary to include special logic. In this case, however, the selection / mutation frequency must be greater than the zombie frequency. In addition, when many zombies are generated, calculation efficiency is improved by incorporating logic to remove them.

  For example, in the process 207, the computer 610 may delete all chromosomes whose evaluation satisfies a predetermined condition (for example, worse than a predetermined value), and generate the same number of duplicated chromosomes as the deleted chromosomes. For example, when a plurality of chromosomes are deleted, the computer 610 generates the same number of duplicates of the best evaluated chromosome as the deleted chromosome, and the learning rate of the best evaluated chromosome and the plurality of chromosomes replicating them is all. You may change the learning rate of those chromosomes so that it may differ. Alternatively, when a plurality of chromosomes are deleted, the computer 610 selects the same number of chromosomes with the highest evaluation as the deleted chromosomes, generates one copy of each selected chromosome, and is replicated with the original chromosome. At least one of them may be changed so that the learning rate of the chromosomes is different.

(Summary of embodiments of the present invention)
As described above, the present embodiment relates to a new learning method in which the GA method is adopted in the back propagation learning process. In this method, a neural network is represented on a computer (as a program and data) as an individual having one chromosome (data), and the hyperparameters of the neural network, that is, the connection between neurons, etc. Is coded. At the same time, the learning rate of the neural network is coded for each chromosome. A plurality of individuals are prepared and calculated in parallel, and selection and mutation in GA are performed for each step (1 epoch) of parallel back propagation learning. In other words, individuals with poor grades are deleted and replaced with those obtained by mutating the learning rate of individuals with good grades.

  Further, the method of the present embodiment is not limited to neural network learning but can be applied to other machine learning. In other words, when iterative learning of data such as images, voices, documents, etc. is performed and the results can be evaluated numerically, the learning parameters that control the machine learning are coded on the chromosome and parallelized learning In each step, selection and mutation in GA are performed.

  Furthermore, the method of the present invention can also be applied to optimization and searching. In other words, when performing optimization and search by repeatedly moving in the search space, when the current point in the search space can be evaluated numerically, optimization / search that controls the optimization and search A control parameter is coded on a chromosome, and selection and mutation in GA are performed for each step of parallel optimization and search.

  The greatest effect of this embodiment is that the learning rate (or learning parameter, optimization / search control parameter) is determined autonomously. That is, by performing selection and mutation for each step of backpropagation learning (or learning, optimization, search), the conventional backpropagation learning method (or learning method, optimization method, search method), and GA and In the same way as the method that combines, the neural network (or system) is optimized as a learning result, and at the same time, in the back propagation learning process (learning process, optimization process, or search process) that was not possible with the conventional method The learning rate (or learning parameter, optimization / search control parameter) can be optimized. That is, the average value of the learning rate (or learning parameter, optimization / search control parameter) is assigned to the optimum value in that step by the selection and mutation performed at each step, and the learning (or optimization, search) Follow optimal values that change with progress. The learning rate (or learning parameter, optimization / search control parameter) usually takes a relatively large value at the beginning of learning (or optimization, search), and learning (or optimization, search) proceeds as the schedule is optimized. It can be lowered with, but when it is better not to lower it, it will be like that.

  At the same time, this embodiment has an effect that the search range in learning (or optimization, search) can be appropriately controlled. A global search can be performed at the initial stage of learning (or optimization, search), and the search range can be narrowed as learning (or optimization, search) progresses. By performing a global search in the initial stage, the probability of falling to the local optimum value decreases, and in the latter stage, a narrow range can be efficiently searched in parallel. However, it is necessary to appropriately control the frequency of selection and mutation for proper control.

  In addition, this invention is not limited to an above-described Example, Various modifications are included. For example, the above-described embodiments have been described in detail for better understanding of the present invention, and are not necessarily limited to those having all the configurations described. For example, it is possible to add, delete, or replace another configuration with respect to a part of the configuration of the above-described embodiment.

  Each of the above-described configurations, functions, processing units, processing means, and the like may be realized by hardware by designing a part or all of them with, for example, an integrated circuit. Further, each of the above-described configurations, functions, and the like may be realized by software by interpreting and executing a program that realizes each function by the processor. Information such as programs, tables, and files for realizing each function is stored in a non-volatile semiconductor memory, a hard disk drive, a storage device such as an SSD (Solid State Drive), or a computer-readable non-readable information such as an IC card, an SD card, or a DVD. It can be stored on a temporary data storage medium.

  Further, the control lines and information lines indicate what is considered necessary for the explanation, and not all the control lines and information lines on the product are necessarily shown. Actually, it may be considered that almost all the components are connected to each other.

501 Learning control computer 502 Learning control program 503 Learning data generation program 504 Original data 505 Teacher information 506 Learning data with teacher information 507 Learning neural network group 508 Identification neural network 509 Data to be identified 510 Identification result output

Claims (10)

A machine learning method executed by a computer having a processor and a storage device connected to the processor,
The storage device includes a plurality of programs for realizing a plurality of systems for executing predetermined processing, a plurality of structural parameters corresponding to each of the plurality of programs, and a plurality of learning parameters corresponding to the plurality of systems. And hold
Each learning parameter is a parameter that specifies a change in the structural parameter in learning performed by each system,
The machine learning method includes:
A first procedure in which the processor causes each system to learn a predetermined data set using a learning parameter corresponding to each system;
A second procedure in which the processor evaluates the systems by a predetermined evaluation method;
The processor selects a first system and a second system having a higher evaluation than the first system from the plurality of systems, and copies the program corresponding to the second system and the plurality of structural parameters to the second system. A program for realizing replication of the system and a plurality of structural parameters corresponding thereto are generated, a replica of the learning parameter corresponding to the second system is generated as a learning parameter corresponding to the replication of the second system, At least one of a learning parameter corresponding to the second system and a learning parameter corresponding to the duplication of the second system so that the learning parameter corresponding to the second system and the learning parameter corresponding to the duplication of the second system are different from each other A third procedure to be changed,
The machine learning method, wherein the third procedure is executed again from the first procedure for the plurality of systems other than the first system. The machine learning method according to claim 1,
A combination of the plurality of structural parameters corresponding to each system and a learning parameter corresponding to each system is retained as one chromosome in the genetic algorithm,
The machine learning method further includes a step in which the processor determines initial values of the plurality of structural parameters corresponding to the systems and initial values of learning parameters corresponding to the systems using random numbers,
In the third procedure, the processor generates a replica of a chromosome corresponding to the second system, uses a random number to learn a parameter corresponding to the second system, and a learning parameter corresponding to the replica of the second system A machine learning method characterized by determining at least one of the values. The machine learning method according to claim 1,
Each of the systems is a neural network,
The plurality of structural parameters corresponding to each of the systems includes connection weights between neurons in the neural network,
In the first procedure, the processor causes each system to learn the predetermined data set by back propagation learning,
The machine learning method, wherein the learning parameter is a learning rate indicating a magnitude of a change amount of the weight in the back propagation learning. The machine learning method according to claim 3,
A combination of the plurality of structural parameters corresponding to each system and a learning parameter corresponding to each system is retained as one chromosome in the genetic algorithm,
Each chromosome further includes information indicating the presence or absence of connections between the neurons,
In the third procedure, the processor changes the presence or absence of the connection between the neurons based on a predetermined mutation rule. The machine learning method according to claim 4,
Each chromosome further includes information indicating the number of neurons included in each system,
In the third procedure, the processor changes the number of neurons based on a predetermined mutation rule. The machine learning method according to claim 4,
Each chromosome further includes information indicating the number of stages of the neural network included in each system,
In the third procedure, the processor changes the number of stages of the neural network based on a predetermined variation rule. The machine learning method according to claim 3,
The plurality of systems includes a plurality of neural networks having different connections between the neurons, or a plurality of neural networks having different numbers of neurons and connections between the neurons. Learning method. The machine learning method according to claim 1,
In the third procedure, when the evaluation of two or more of the systems is lower than a predetermined value, the processor is configured to realize the same number of systems other than the two or more systems as many as the two or more systems. Generating a copy of a program, a plurality of structural parameters corresponding to a system other than the two or more systems, and a learning parameter corresponding to a system other than the two or more systems;
The machine learning method, wherein the first procedure to the third procedure are executed again for the plurality of systems other than the two or more systems. The machine learning method according to claim 1,
The computer has a plurality of the processors,
Each of the plurality of systems is assigned to each of the plurality of processors;
In the first procedure, each processor causes the one system assigned to each processor to learn a predetermined data set using the learning parameter. A machine learning device having a processor and a storage device connected to the processor,
The storage device holds a plurality of systems each including a program executed by the processor and a plurality of structural parameters for the program, and learning parameters corresponding to each system,
Each learning parameter is a parameter that specifies a change in the structural parameter in learning performed by each system,
The processor is
A first procedure for causing each system to learn a predetermined data set using a learning parameter corresponding to each system;
A second procedure for evaluating each of the systems by a predetermined evaluation method;
A first system and a second system having a higher evaluation than the first system are selected from the plurality of systems, a duplicate of the second system is generated, and a duplicate of the learning parameter corresponding to the second system is generated in the second A learning parameter corresponding to the second system so that a learning parameter corresponding to the second system and a learning parameter corresponding to the second system replication are different from each other. Executing at least one of the learning parameters corresponding to the duplication of the two systems;
The machine learning device, wherein the third procedure is executed again from the first procedure for the plurality of systems other than the first system.

Images