JP6643905B2 - Machine learning method and machine learning device - Google Patents

Description

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

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

  The steepest descent method is used not only for deep learning but also as a basic algorithm for a back propagation learning method used in neural network training and the like, and as various learning and optimization techniques. This method is a deterministic search. However, since this method is almost certainly limited to a local optimum value that is not global optimum, a stochastic gradient descent method, which is a stochastic search, is usually used. In neural network learning, there is a learning rate as a parameter for controlling the learning. The initial value of the learning rate is determined by the experimenter (human), and is a constant during the learning process or changes according to a predetermined schedule. In Patent Literature 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. As described above, other methods for adaptively determining the learning rate have been proposed, but all of them have limited applicable problems and networks.

  As a method for stochastic optimization, there is a genetic algorithm (GA). GA is an optimization method originally developed independently of a neural network, but it is sometimes used in combination with a machine learning method. In particular, in a neural network, a number of methods for using back propagation learning and GA in combination have been developed. The most frequent method is the method of optimizing the structure and weight of a neural network by using 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 operation of the GA, that is, the mutation or crossover, is repeatedly performed after the entire back propagation learning process is performed. In addition, as a combination with other stochastic search methods, a method combining with particle swarm optimization methods, which is a stochastic search method that has been developed relatively recently and has been successful, has been studied. I have. Also, an algorithm in which the stochastic gradient descent method is parallelized has been developed.

U.S. 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 developments in research on deep learning and back propagation learning of neural networks, there are some difficult issues in back propagation learning of neural networks. The first problem is how to determine the learning rate. That is, the optimum learning rate differs depending on the structure of the neural network, and also depends on the problem. In addition, a method of keeping the learning rate constant in the learning process is relatively widely used, but it is generally better to decrease the learning rate as the learning progresses. is there. There is a great deal of literature 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. A variety of other adaptive methods have been devised these days, but many of these methods are subtle but may not work. A simpler and more powerful method is needed.

  The second challenge in backpropagation learning is to escape from the local minima. That is, depending on the initial value of the weight of the neural network, the value of the function to be minimized (such as the error rate) may not be as small as the minimum value even if the back propagation learning is performed. Also, once a relatively good value is determined, it may not return to the minimum value after further learning. The challenge is to escape from such a blind alley and determine the optimal value.

  The reason that backpropagation learning is easily caught in a dead end is that backpropagation learning is a local search method. As described above, a stochastic gradient descent method is usually used as a method for back propagation. However, if a stochastic search is to be performed, it can be improved by using it in combination with one of the various established stochastic search methods. The purpose of combining conventional backpropagation learning and GA is that, but in the conventional method, GA does not act on the process of one backpropagation learning, so the learning rate is reduced in one learning. It cannot be optimized.

  According to one embodiment of the present invention, there is provided a machine learning method executed by a computer having a processor and a storage device connected to the processor, wherein the storage device performs a predetermined process. Holding a plurality of programs for realizing a plurality of systems for executing, 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, The learning parameter is a parameter that specifies a change in the structural parameter in the learning performed by each system, and the machine learning method is such that the processor uses the learning parameter corresponding to each system to perform the processing on each system. A first step of learning a predetermined data set, and a second step in which the processor evaluates each system 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 and the plurality of structural parameters corresponding to the second system. A program for realizing a copy of the second system and a plurality of structural parameters corresponding to the program are generated, and a copy of the learning parameter corresponding to the second system is generated as a learning parameter corresponding to the copy of the second system. At least the learning parameter corresponding to the second system and the learning parameter corresponding to the copy of the second system such that the learning parameter corresponding to the second system and the learning parameter corresponding to the copy of the second system are different from each other. And a third procedure for changing one of the first and second systems. For serial multiple systems, wherein the third instructions from said first procedure is executed again.

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

FIG. 4 is an explanatory diagram of encoding of parameters of a multilayer neural network into chromosomes according to the embodiment of the present invention. FIG. 7 is an explanatory diagram of a learning method in which back propagation learning and GA are combined in the embodiment of the present invention. It is an explanatory view of an example of a change of a learning rate in an embodiment of the present invention. It is explanatory drawing of the example of the Euclidean distance between the best individual and each other individual in embodiment of this invention. It is an explanatory view of the outline of the composition and operation of the neural network design tool for data identification of the embodiment of the present invention. 1 is a block diagram illustrating a hardware configuration of a data identification neural network design tool according to an embodiment of the present invention.

  An embodiment of the present invention will be described. In this embodiment, the conventional back-propagation learning method and the GA are combined by performing selection and mutation in the GA for each step (1 epoch) of the back-propagation learning parallelized by the method in this embodiment. In the same way as the method described above, the neural network is optimized as a learning result, and the learning rate in the back propagation learning process, which cannot be achieved by the conventional method, is optimized.

(overall structure)
The overall configuration of this embodiment, that is, the configuration and operation of the neural network design tool 500 for data identification 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 identifying image data and the like, and includes original data 504 as learning data and teacher information 505 for explaining the original data 504. Is entered. A video or a still image can be input as the original data 504, and the teacher information 505 includes information to be identified at which position of the video or the still image, for example, a plurality of rectangular areas (for example, indicating a vehicle or a pedestrian). bounding box) information. The designed neural network is extracted from the learning neural net group 507 in the form of a bias and a 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 inputs a user's input to the learning control program 502 and outputs an output from the learning control program to the user. Further, the learning control computer 501 can operate the learning control program 502, the learning data generation program 503, the learning neural net group 507 composed of a plurality of neural nets, and the identification neural net 508 thereon. Original data 504, teacher information 505, learning data with teacher information 506, data 509 to be identified, and an identification result output 510, which are input / output of a program, can be stored. However, the learning data generation program 503, the neural network group for learning 507, the neural network for identification 508, the original data 504, the teacher information 505, the learning data with teacher information 506, the data to be identified 509, and the identification result output 510 are used for 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 net group 507 and the identification neural net 508. Enter The parameters for the learning neural network group 507 include the number of neural nets, that is, the number of individuals, seeds of random numbers for randomly determining weights and biases of the neural nets, weights and biases of a plurality of neural nets, as described later. , A distribution function such as a normal distribution that determines the distribution of, the average value and the standard deviation, and the average value and the standard deviation of the initial value of the learning rate. However, when using default values for these values, the user does not need to enter them. This input can include the upper limit of the number of steps (the number of epochs), the target value of the error, and the target value of the learning rate as the conditions for stopping the neural network. The user can also input data 509 to be identified at the time of this input.

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

  The learning control program 502 operates the learning neural net group 507 according to a user's instruction to perform back propagation learning. That is, the 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 of learning a plurality of neural nets will be described later.

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

  The data identification neural network design tool 500 of the present embodiment can be realized by, for example, computers 600, 610, and 620 interconnected 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 which are interconnected. 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 and data to be processed. The memory 602 of the present embodiment stores the learning control program 502. In the present embodiment, the processing executed by the learning control program 502 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 programs 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 needed.

  The computer 610 has a CPU 611, a memory 617, an I / F 616, and an HDD 615 connected to each other, and further has a GPU (Graphics Processing Unit) 612 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 programs executed by the CPU 611, data to be processed, and the like. The memory 602 of the present embodiment stores a learning data generation program 503. In the present embodiment, the processing executed by the learning data generation program 503 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. In the memory 614 of this embodiment, a learning neural network group 507 and learning data with teacher information 506 are stored. The operation of the learning neural network group 507 of the present embodiment is executed by the GPU 612.

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

  Each of the learning neural nets included in the learning neural net group 507 includes a set of structural parameters such as a weight and a bias between neurons included in each of the learning neural nets stored in the memory 614, and the structural parameters and input. And a program that calculates an output based on the obtained learning data and performs learning by a predetermined learning method (for example, back propagation learning) based on the evaluation of the output. That is, each neural network for learning is a system realized by the GPU 612 executing the corresponding program using the structural parameters.

  The computer 620 has a CPU 621, a memory 626, an I / F 625, and an HDD 627 connected to each other, and further has a GPU 622 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 programs executed by the CPU 621, data to be processed, and the like. 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. In the memory 624 of this embodiment, a neural network 508 for identification, data 509 to be identified, and an identification result output 510 are stored. The operation of the neural network for identification 508 of the present 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 may be copied to the memory 624 by the CPU 621. Further, 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 identification neural network design tool 500, and there may be various modifications in practice. For example, the computer 610 may have a plurality of GPUs 612. In this case, each learning neural net included in the learning neural net group 507 and the learning data with teacher information 506 are stored in the memory 614 of each GPU 612, and each GPU 612 learns one learning neural net. May go. Thereby, learning of a plurality of neural nets is performed in parallel, so that the time required for learning is reduced.

  Alternatively, any two or all of the 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 and the like may be replaced by a storage device other than the HDD such as a flash memory.

(Chromosome expression)
In this embodiment, as shown in FIG. 1, the parameters of the multilayer neural network are encoded on the chromosome of the genetic algorithm (GA). 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 that the joint 101 is encoded on the chromosome is the same as the encoding method in the method of combining the conventional neural network and GA, but in the present embodiment, the learning rate 102 is also encoded. . In FIG. 1 (a), only the connection parameter among the neurons is described, but the constant term, that is, the bias can also be encoded in the chromosome. Here, the structure of the neural network is not represented on the chromosome because the structure of the neural network is fixed, but by representing the structure, the neurons and the connections between the neurons are changed according to a predetermined mutation rule in the learning process (for example, A structural optimization such as deleting) can also be realized using GA. That is, the chromosome is made variable length, and the parameters of each neuron are described (when deleting a neuron, the whole is deleted), or the parameters related to connections between neurons are described (when deleting a connection, Remove it).

  FIG. 1B illustrates encoding of a convolutional neural network (CNN) often used in image recognition and the like. Compared to FIG. 1 (a), the number of parameters increases and the scale of the chromosome increases, but the same is true in that parameters are encoded.

  The structure of the chromosome need not be the same for all individuals. That is, the calculations can be performed using neural nets of different structures (eg, different connections between neurons or different numbers of neurons and connections between neurons). In this case, the mutation can be similarly performed. 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, it is possible to divide each of the two neural nets between any of the layers, or to divide them into two in a specific layer and to intersect them. In this case, the structure of the neural network can be maintained by simply reconnecting if the number of bonds to be cut is reduced to one, but when the number of bonds is insufficient, a bond of zero is introduced. When the connection has a remainder, the structure of the neural network can be reconstructed by deleting the connection.

  The coding on the chromosome is not limited to the multilayer neural network, but can be applied to other types of systems for learning or optimization / search. In other words, coding the structural parameters of the system (corresponding to the joints of the neural network) and the learning parameters or the parameters that control the optimization / search process (corresponding to the learning rate), and applying the mutation and crossover operations Can be.

(Learning method)
Hereinafter, a learning method combining back propagation learning and GA will be described with reference to FIG. This learning method is called LOG-BP learning method (learning-rate-optimizing genetic back-propapation learning method). In this learning method, by preparing a plurality of chromosomes described in the previous section and performing backpropagation learning in parallel, the weight on those chromosomes changes autonomously. In addition, the learning rate is stochastically varied.

  First, the computer (the computer 610 in the example of FIG. 6) into which the program of FIG. 2 (that is, the program included in the learning neural network group 507) is incorporated performs chromosome initialization (201). The number of individuals can be variable, but here is a fixed number (for example, 20). The weight and learning rate of those chromosomes are determined by random numbers. Initialization of the weight may be performed in the same manner as in normal backpropagation learning. For example, weight and bias may be determined by normally distributed random numbers. The learning rate is also appropriately distributed using random numbers, but may be determined by, for example, a normal distribution. It seems that the learning rate should be set to a relatively large value so that the divergence frequency does not become too high. The parameters for determining these initial values can be externally input via the learning control program 502. That is, the learning rate, the average value of the rice, the standard deviation, the shape of the distribution, and the seed of the random number can be designated before the learning is started.

  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, the computer prepares as many image data as possible as training data in advance, and specially uses a part of the training data as verification data. Also, evaluation data composed of images in the same format is prepared as needed. Then, the computer trains all training data once (learning using image data will be described later). Stochastic gradient descent in units of mini-batch (that is, steepest descent method in which all training data is learned at once, and basic stochastic gradient descent method in which training data is learned one by one; When using (method), the training data stored in the array is divided for each mini-batch and learned by backpropagating once. At this time, a value encoded in each chromosome is used as the learning rate.

  Next, the computer updates the weight on the chromosome with the weight changed by learning (204). That is, in this method, the value of the weight on the chromosome is not changed externally, but is updated autonomously based on random numbers and learning of each individual. That is, Lamarck-like inheritance in which acquired traits are inherited as they are is realized, instead of Darwin-like inheritance. However, this process is consistent with the basics of GA, assuming that the updating of fir is a mutation.

  Next, the computer evaluates each updated individual (eg, an asexually regenerated egg 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 lower the value, the better, and the computer makes a selection based on the value. That is, the one with the highest value, i.e., the worst evaluated (i.e., chromosome) is killed (i.e., deleted), and the smallest, i.e., the individual with the best evaluation is copied (to produce a dizygotic twin) (207). ). This makes the population unchanged. However, by making the generation (copy) 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 equation for the individual generated by copying.

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

  That is, which equation is applied is determined by random numbers so as to have equal probability. f is a value of, for example, about 1.2, which is similar to a rule used in an adaptive back propagation learning method (this rule is originally irrelevant to GA). However, the change of the learning rate by the above equation is an example, and as long as the learning rate of the individual whose evaluation is the best and the learning rate of the individual generated by copying it are determined to be different, for example, The learning rate may be determined by a method other than the above, such as changing the learning rate of. Further, in the above example, the worst individual is deleted and a copy of the individual with the best evaluation is generated.However, as long as the evaluation of the individual whose copy is generated is better than the evaluation of the individual to be deleted, deletion is performed as long as it is better. It is not necessary to limit the objects to be copied to the worst individual and the best individual.

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

  The computer repeats the above steps (epoch) until a proper solution is obtained or little change occurs (209). The condition for stopping the calculation 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 repetition.

  In the above description, selection and mutation are performed for each step. However, in practice, it is considered better to select and control the average value of the number of selections and mutations for each step. That is, if the selection / mutation is performed exactly once for each step, the number of individuals becomes excessive when the number of individuals is small, and the number becomes excessively small when the number of individuals is large. Therefore, the average value of the number of times of performing 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 too large, the search range may be too fast, and if too small, the search range may be too wide. By appropriately controlling the number of selections, a wide area can be searched at the start of calculation, and the search range can be gradually narrowed, so that a solution can be obtained well.

  When another learning system or optimization / search system is used instead of the neural network, evaluation is performed for each step of the learning or optimization / search that is repeatedly executed, and the selection is made based on the result. Then, the value of the learning parameter for controlling the learning process or the value of the optimization / search parameter for controlling the optimization / search process is varied. With regard to this variation, when a distance (difference in scalar value) can be defined between a plurality of values of these parameters, a parameter value close to the distance may be generated using random numbers. When the distance cannot be defined, any value may be selected by using a random number.

(Supplementary information on learning methods and application range)
Hereinafter, six points related to the variation of the LOG-BP learning method and the expansion of its application range will be described. First, each individual can refer to the evaluation value based on the verification data and reflect that in learning. In the above algorithm, the evaluation value is also used for selection, but since the selection is considered to be external, the evaluation value for selection can be different from that. For example, it is possible to provide evaluation data in addition to verification data and use it for selection. That is, a different criterion can be used as a selection criterion by each individual (a criterion in backpropagation learning) and an external selection criterion (a criterion in GA).

  Second, in the above method, the structure and parameters of the neural network are not changed by selection / mutation. No extensions will be made in the following evaluations, but it is possible to extend them for the purpose of optimizing structure and parameters and to use mutations and crossovers. For example, each chromosome includes information indicating the number of neurons of each neural network and the presence / absence of connection between neurons, and the computer 610 mutates the chromosomes based on a predetermined mutation rule in the process 207, thereby obtaining the connection between neurons. Can be cut or neurons can be killed. The same applies to the addition of neurons described later. When a system other than a neural network is used, a part thereof can be changed or deleted by the same method.

  Third, it is possible to add neurons by mutation. When adding a neuron, it is necessary to reduce the value of the fir tree so that it does not invalidate the learning that has already been performed (it is the same as not adding a fir tree if it is 0). However, this may not activate the added neurons. If neurons are added without increasing the training data, overfitting occurs, and apparently the evaluation value is likely to improve. Therefore, it is considered that it is better to determine the evaluation function so that the evaluation value is improved when the size of the neural network is small. For example, it is expected 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 explained, the LOG-BP learning method can be extended not only in the application to the neural network as described above, but also in other learning methods. That is, it is assumed that a system that performs classification and detection (corresponding to a neural network) exists, and a learning method for training the system exists. 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 process of iteration is given, the present learning method (LOG learning method) can be applied in the same manner as in the back propagation learning of the neural network. That is, parameters that determine the structure of the system are expressed as chromosomes, learning is started by initializing a plurality of chromosomes, and learning, evaluation, mutation / selection are repeated. The learning control parameter to be mutated does not need to be a real value, and a method of simply mutating it to another value may be provided instead of the two expressions for the mutation.

  Fifth, in the above embodiment, all individuals are of the same type of neural network. However, even for individuals using different types of neural nets or other learning methods for each individual, the evaluation parameters are aligned and the learning parameters and If the method of the mutation is designated, it can be evaluated, selected and mutated by the above method. That is, the 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 nets having different connections between neurons, or include a plurality of neural nets having different numbers of neurons and different connections between neurons. May be. 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 specify an optimal one among different types of neural nets or different types of learning methods.

  Sixth, in the above-described embodiment, learning on one computer (for example, the computer 610 in FIG. 6) is basically performed. However, a plurality of computers (for example, a plurality of computers 610) are prepared, and By allocating one chromosome, parallel calculations can be performed by a processor (for example, CPU 611 or GPU 612 of each computer 610) of these computers. The evaluation value of each chromosome can be collected by one computer and selected. The next epoch requires that one or more of the chromosomes be replaced, but this can be done by exchanging small amounts of data between the computers. Performing normal backpropagation learning using different parameters on multiple computers can be realized by conventional techniques, but in comparison with this, the method described above allows faster and more optimal values to be obtained. This has the advantage that the probability of finding it increases. Alternatively, as described with reference to FIG. 6, for example, the computer 610 has a plurality of GPUs 612, and the same processing and execution as described above can be performed using each GPU 612.

(Example of image data set learning)
In this section, a method of learning a pedestrian image data set according to the learning method described above will be described. An example of a pedestrian image dataset is the Caltech pedestrian dataset. The same method can be applied when a face image, an object image, a character image, or the like is used instead of the pedestrian image. The dataset used in this training contains multiple videos. In addition to the video, annotation data is attached, and there is also bounding box data that shows 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, which are generated as follows. A 100,000 image was prepared by cutting out the bounding box specified in the above data set and normalizing it to a size of 24 × 48, and was combined with the 100,000 obtained by inverting left and right. Two hundred thousand images are used twice as positive examples.

  Further, negative examples of the remaining half of the training data are generated as follows. Use 200,000 images of size 24x48 extracted from the Caltech pedestrian dataset other than the bounding box. In generating this initial negative example, its position is determined by a random number. Although an image of a size different from 24 × 48 can be cut out and resized, it is also possible to cut out only an image of exactly 24 × 48 size. Then, 200,000 positive examples and 200,000 negative examples are applied, and the CNN of one stage of the convolutional layer trained is applied to the original data set to generate 200,000 negative examples from the erroneously recognized portion. That is, an image in which the CNN determines that a pedestrian is included but is out of the bounding box is set as a new negative example. A total of 400,000 images are prepared including these negative examples and 800,000 images including the positive examples.

  The numerical values included in these data are 0 to 255 in the original data such as the Caltech pedestrian data set, but these are converted to floating-point numbers in the range of approximately -1 to 1, and further corrected so that the average becomes 0.

  Hereinafter, the structure of the CNN to be used and the hyper parameters will be described. The example is as follows, assuming that there are two convolutional layers.

・ First stage of convolutional layer: filter size 5 × 5, number of filters 16, nonlinear (activation) function: ReLU
・ First stage of pooling layer: Maximum pooling. Size 2 × 2
• Second stage of convolutional layer: filter size 3 × 3, number of filters 26, 28, or 32, nonlinear function: ReLU
-2nd pooling layer: maximum pooling. Size 2 × 2
・ Hidden layer (1 stage): 50 neurons
-Output layer: Logistic regression. 2 neurons (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 broken line function of f (x) = if x <0 then 0 else x.

(Change in learning rate)
FIG. 3 shows an example of a change in the learning rate in the learning process of the present embodiment. Although the average value and the standard deviation of the learning rate can be measured for each epoch, they are plotted every 5 epochs in this figure. When the initial value 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 is recognized that the initial value has been 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 performing machine learning or optimization / search for other systems instead of neural network learning, other learning parameters that control the learning process or parameters that control the optimization / search process instead of the learning rate are autonomous. Is adjusted.

  The value of the learning rate (or the learning parameter or the optimization / search parameter) 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.

(Supplementary information on learning performance, etc.)
First, the number of individuals (the number of chromosomes) will be described. The larger the number, the more stochastically the closest solution can be obtained. However, unless all of them can be calculated in parallel, the larger the number of individuals, the longer the calculation time. As an example of a value that can balance the calculation time and the search range, it is possible to reduce the number of individuals to about 12. It takes, for example, about 8 hours using a GPU to perform an experiment up to 100 epoch with the number of individuals being 12.

  Second, a method for maintaining diversity, that is, a method for controlling the globality of search, will be described. If the frequency of selection / mutation is high, all individuals are likely to be copied from one individual. Therefore, if the individual is located in a portion having only the local optimal value deviating from the global optimal value, a satisfactory solution cannot be reached. It seems that the probability of selection / mutation in one epoch needs to be 5% or less (0.6 or less 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, global search performance will be described. About 12 chromosomes is not enough for global search. In this case, initially, 12 locations are searched, but since individuals generated by duplication gradually increase, even if the above-mentioned method for maintaining diversity is applied, the search location is immediately reduced to around 3 points. Because it is spilled. The number of search ranges can be estimated by calculating and displaying the Euclidean distance between the best individual determined so far and each individual at present. FIG. 4 shows an example of such a 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 digit of 9) is the best individual, and at least two neighborhoods are searched in parallel.

  Fourth, the deletion of divergent individuals will be described. In the simple back propagation learning algorithm used in this experiment, the weight often diverges (becomes "nan") due to the back propagation. Such individuals are considered to be zombies, that is, individuals that are unlikely to be solved even if calculations are continued, and should be deleted. Although it is possible to incorporate logic for deletion, it is necessary to incorporate special logic because such an individual has an extremely deteriorated evaluation value and is therefore preferentially deleted in this algorithm. However, in that case, the frequency of selection / mutation must be higher than the frequency of zombies. In addition, when a large number of zombies are generated, the calculation efficiency is improved by incorporating the logic to eliminate 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 may generate the same number of chromosome copies as the deleted chromosomes. For example, when a plurality of chromosomes are deleted, the computer 610 generates the same number of copies of the chromosome with the best evaluation as the number of the deleted chromosomes, and the learning rate of the chromosome with the best evaluation and the learning rates of the plurality of chromosomes that duplicated them are all equal. The learning rate of those chromosomes may be changed differently. 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 the selected chromosomes one by one, and copies them to the source chromosome. At least one of them may be changed so that the learning rates of the chromosomes differ.

(Summary of Embodiment of the Present Invention)
As described above, the present embodiment relates to a new learning method in which the GA method is used 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 hyper-parameter of the neural network, that is, the connection between neurons, etc. is stored in each individual chromosome. Is encoded (represented). In addition, 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 parallelized back propagation learning. That is, individuals with poor grades are deleted and replaced with mutated learning rates of individuals with good grades.

  Further, the method of the present embodiment can be applied not only to learning of a neural network but also to other machine learning. In other words, iterative learning of data such as images, sounds, documents, etc. is performed, and when the results can be evaluated numerically, learning parameters that control the machine learning are coded on chromosomes, and parallelized learning is performed. Selection and mutation in GA are performed for each of the following steps.

  Furthermore, the method of the present invention can be applied to optimization and search. In other words, when performing optimization or search by repeating movement in the search space, if the current point in the search space can be numerically evaluated, optimization / search to control the optimization or search The control parameters are coded on the chromosome, and the 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 parameters, optimization / search control parameters) is determined autonomously. That is, by performing selection and mutation at each step of back propagation learning (or learning, optimization, search), the conventional back propagation learning method (or learning method, optimization method, search method) and the GA As well as optimizing the neural network (or system) as a learning result in the same way as the method of combining 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 approximated to the optimum value in that step by the selection and mutation performed for each step, and the learning (or optimization, search) Follow the 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 the learning (or optimization, search) proceeds and the optimal schedule , But when it is better not to lower it, it becomes like that.

  At the same time, in the present embodiment, there is an effect that the search range in learning (or optimization, search) can be appropriately controlled. At the beginning of learning (or optimization, search), a global search can be performed, and the search range can be narrowed as the learning (or optimization, search) progresses. Performing a global search in the early stage lowers the probability of falling to the local optimum, and in the latter stage, it is possible to efficiently search in a narrow range efficiently. However, for proper control, it is necessary to appropriately control the frequency of selection and mutation.

  Note that the present invention is not limited to the above-described embodiment, and includes various modifications. 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, with respect to a part of the configuration of the above-described embodiment, it is possible to add, delete, or replace another configuration.

  In addition, each of the above configurations, functions, processing units, processing means, and the like may be partially or entirely realized by hardware, for example, by designing an integrated circuit. In addition, the above-described configurations, functions, and the like may be implemented by software by a processor interpreting and executing a program that implements each function. Information such as a program, a table, and a file 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 non-readable computer 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 the information lines are shown as necessary for the explanation, and not all the control lines and the information lines are necessarily shown on the product. In fact, almost all components may be considered to be interconnected.

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 net group 508 Identification neural net 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 that execute a predetermined process, 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 the
Each of the learning parameters is a parameter that specifies a change in the structural parameter in learning performed by each of the systems,
The machine learning method includes:
A first procedure in which the processor causes each system to learn a predetermined data set by using a learning parameter corresponding to each system;
A second procedure in which the processor evaluates each of the systems according to 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 and the plurality of structural parameters corresponding to the second system to the second system. Generating a program for realizing a copy of the system and a plurality of structural parameters corresponding thereto, generating a copy of the learning parameter corresponding to the second system as a learning parameter corresponding to the copy of the second system, At least one of the learning parameter corresponding to the second system and the learning parameter corresponding to the copy of the second system is set such that the learning parameter corresponding to the second system and the learning parameter corresponding to the copy of the second system are different from each other. A third step of changing
A machine learning method, wherein the first to third steps are executed again for the plurality of systems other than the first system. The machine learning method according to claim 1, wherein
The combination of the plurality of structural parameters corresponding to each system and the learning parameters corresponding to each system is held as one chromosome in the genetic algorithm,
The machine learning method further includes a step in which the processor uses random numbers to determine initial values of the plurality of structural parameters corresponding to the respective systems and initial values of learning parameters corresponding to the respective systems,
In the third step, the processor generates a copy of a chromosome corresponding to the second system, and uses a random number to generate a learning parameter corresponding to the second system and a learning parameter corresponding to the copy of the second system. Determining a value of at least one of the following. The machine learning method according to claim 1, wherein
Each of the systems is a neural network,
The plurality of structural parameters corresponding to each system include weights of connections between neurons in the neural network,
In the first procedure, the processor causes the respective systems to learn the predetermined data set by back propagation learning,
The machine learning method according to claim 1, 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, wherein
The combination of the plurality of structural parameters corresponding to each system and the learning parameters corresponding to each system is held as one chromosome in the genetic algorithm,
Each chromosome further includes information indicating the presence or absence of connection between the neurons,
In the third step, 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, wherein
Each of the chromosomes further includes information indicating the number of neurons included in each of the systems,
The machine learning method according to claim 3, wherein in the third step, the processor changes the number of neurons based on a predetermined mutation rule. The machine learning method according to claim 4, wherein
Each chromosome further includes information indicating the number of stages of the neural network included in each system,
In the third step, the processor changes the number of stages of the neural network based on a predetermined mutation rule. The machine learning method according to claim 3, wherein
The plurality of systems include a plurality of neural networks having different connections between the neurons, or a plurality of neural networks having different numbers of the neurons and different connections between the neurons. Learning method. The machine learning method according to claim 1, wherein
In the third step, when the evaluation of two or more systems is lower than a predetermined value, the processor is configured to realize the same number of systems as the two or more systems, other than the two or more systems. A copy of a program, a copy of a plurality of structural parameters corresponding to a system other than the two or more systems, and a copy of a learning parameter corresponding to a system other than the two or more systems,
A machine learning method, wherein the first to third steps are executed again for the plurality of systems other than the two or more systems. The machine learning method according to claim 1, wherein
The computer has a plurality of the processors,
Each of the plurality of systems is assigned to each of the plurality of processors;
The machine learning method according to claim 1, wherein, in the first procedure, each of the processors causes one of the systems assigned to each of the processors 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 a learning parameter corresponding to each system,
Each of the learning parameters is a parameter that specifies a change in the structural parameter in learning performed by each of the systems,
The processor comprises:
A first procedure for causing each system to learn a predetermined data set by using a learning parameter corresponding to each system;
A second procedure for evaluating each of the systems according to a predetermined evaluation method;
Selecting a first system and a second system having a higher evaluation than the first system from the plurality of systems, generating a copy of the second system, and copying the learning parameter corresponding to the second system to the second system; A learning parameter corresponding to the second system and a learning parameter corresponding to the second system such that the learning parameter corresponding to the second system and the learning parameter corresponding to the replication of the second system are different from each other. And 3) changing at least one of the learning parameters corresponding to the duplication of the two systems.
A machine learning apparatus, wherein the first to third procedures are executed again for the plurality of systems other than the first system.

Images