Information processing device, information processing method, and program

Abstract

The information processing device 100 includes a group setting unit 10 that groups candidate operations into one or more groups for each layer in a machine learning model network in which numerous candidate operations are arranged in each layer; a learning processing unit 20 that performs machine learning on the machine learning model network and selects candidate operations for each group using the parameter values ​​assigned to the candidate operations by the execution of machine learning; and a candidate operation output unit 30 that outputs a machine learning model composed of the selected candidate operations.

Description

Technical Field

[0001] The present disclosure relates to an information processing apparatus, an image processing unit, an information processing method, and a computer-readable recording medium storing a program for realizing these.

Background Art

[0002] In the era of artificial intelligence, research and development of convolutional neural network (CNN) models has grown significantly in the field of computer vision. An important reason behind the growth of CNN models is that CNN models can achieve high accuracy for various computer vision tasks such as image classification, object detection, and segmentation.

[0003] However, designing the architecture of a CNN model requires specialized knowledge from the designer and is time-consuming. The main reason for the time consumption is that there are several options for convolution operation types. Furthermore, adjusting the number of parameters used in the selected convolution operation type also requires specialized knowledge from the designer and takes additional time. For real-time applications, the architecture of a CNN model is not designed solely based on achieving high accuracy, but is also designed with the goal of achieving high speed.

[0004] One prominent solution to address the problem of requiring specialized knowledge from designers and the problem of time consumption is Neural Architecture Search (NAS). NAS is defined as a method for automatically searching several options of convolution operation types, and automatically searching for the number of parameters to be used in the selected convolution operation type. There are several variations of NAS schemes, such as reinforcement learning-based NAS, Bayesian optimization-based NAS, one-shot NAS, few-shot NAS, zero-shot NAS, evolutionary NAS, and the like. Among various NAS schemes, one-shot NAS is one of the faster schemes with respect to design space search.

[0005] In a one-shot NAS, several candidate convolution operation types with different numbers of parameters (where applicable), such as standard convolution, extended convolution, group convolution, pooling, and Eltwise operations, are arranged in parallel within a layer. Several such layers are stacked sequentially to form a large CNN architecture, commonly called a SuperNet®. For object detection tasks, convolution-based layers, commonly called the neck and head, exist within the SuperNet. A special architectural parameter (α) is added to the output of each candidate operation.

[0006] During the training of the supernet using the training dataset, the parameters of candidate operations are trained in the same way as the parameters (α) of the special architecture. At the end of training, at each layer, the candidate operations with the highest values ​​of the special architecture parameters (α) are selected, forming the final neural network, commonly called the subnet or AI model in the current real-world system.

[0007] Ideally, the AI ​​model derived from a supernet should consist of good candidate operations that can achieve high accuracy and speed, where applicable. However, if the supernet is not trained efficiently, training it may lead to the selection of poor candidate operations.

[0008] In other words, the parameter (α) value of the special architecture for poor candidate operations tends to be higher at the start of supernet training for faster convergence. Conversely, the parameter (α) value of the special architecture for good candidate operations tends to be lower at the start of supernet training for slower convergence.

[0009] To overcome such inefficient supernet training scenarios, it becomes essential to train until good candidate operations converge, which leads to higher parameter (α) values ​​for each special architecture. However, the amount of time required to train large-scale neural networks like supernets becomes significantly larger.

[0010] Conversely, if training time is kept shorter, there is a potential risk of unfair evaluation of both good and bad candidate calculations. Therefore, there is a trade-off between the efficiency of supernet training and the amount of time spent training the supernet.

[0011] One method for solving such problems is proposed in Non-Patent Document 1. Non-Patent Document 1 proposes a novel method for removing candidate operations. Here, during the training of a supernet, one or more candidate operations are removed at each training interval based on the value of the special architecture parameter (α).

[0012] At each interval, the candidate operation with the minimum value of the special architecture parameter (α) is eliminated, and the number of candidate operations gradually decreases. At the end of training the supernet, the last remaining candidate operation is selected for the final model. [Prior art documents] [Non-patent literature]

[0013] [Non-Patent Document 1] Progressive Differentiable Architecture Search Bridging the Depth Gap between Search and Evaluation; Xin Chen, Lingxi Xie, Jun Wu, Qi Tian;<https: / / arxiv.org / abs / 1904.12760> [Overview of the project] [Problems that the invention aims to solve]

[0014] However, the problem with Non-Patent Document 1 is that slow-converging but good candidate operations may tend to obtain smaller special architecture parameter (α) values ​​at the start of supernet training.

[0015] Such a situation can lead to the elimination of good candidate operations in the early stages of supernet training. If good candidate operations are eliminated first, there is a risk that poor candidate operations will be selected last.

[0016] Thus, the main problem with one-shot NAS is that poor candidate operations are selected instead of good candidate operations. The reason this problem occurs is that some poor candidate operations tend to converge faster than good candidate operations, and a higher value for the special architecture parameter (α) is derived, which in turn leads to the selection of poor candidate operations.

[0017] One example of the objectives of the present invention is to suppress the occurrence of situations in which poor candidate operations are selected during neural architecture searching. [Means for solving the problem]

[0018] To achieve the above objective, the information processing device in one aspect of this disclosure is: In a machine learning model network in which numerous candidate operations are placed in each layer, a group setting means is provided for grouping the candidate operations in each layer to set up one or more groups. A learning processing means that performs machine learning on the machine learning model network and selects candidate operations for each group using the parameter values ​​assigned to the candidate operations by the machine learning execution, candidate operation output means that outputs a machine learning model composed of the selected candidate operations, comprising: the feature of which is as follows.

[0019] Moreover, to achieve the above object, an information processing method according to one aspect of the present disclosure is: in a machine learning model network in which a large number of candidate operations are arranged in each layer, grouping the candidate operations for each of the layers to set one or more groups; executing machine learning on the machine learning model network, and selecting a candidate operation for each of the groups using the values of parameters assigned to the candidate operations by the execution of the machine learning; outputting a machine learning model composed of the selected candidate operations; the feature of which is as follows.

[0020] Further, to achieve the above object, a computer-readable recording medium according to one aspect of the present disclosure: causes a computer to: in a machine learning model network in which a large number of candidate operations are arranged in each layer, group the candidate operations for each of the layers to set one or more groups; execute machine learning on the machine learning model network, and select a candidate operation for each of the groups using the values of parameters assigned to the candidate operations by the execution of the machine learning; output a machine learning model composed of the selected candidate operations; records a program including instructions for performing the above processes; the feature of which is as follows. [Advantageous Effects of the Invention]

[0021] As described above, according to the present disclosure, it is possible to suppress the occurrence of a situation where a defective candidate operation is selected in neural architecture search. [Brief Description of the Drawings]

[0022] [Figure 1] Figure 1 is a block diagram that schematically shows the configuration of an example of an information processing device. [Figure 2] Figure 2 is a diagram illustrating the configuration of an example of an information processing device in more detail. [Figure 3] Figure 3 shows an example of a supernet. [Figure 4] Figure 4 shows an example of grouping of candidate operations performed in the embodiment. [Figure 5] Figure 5 shows an example of a supernet after candidate operations have been selected. [Figure 6] Figure 6 shows a supernet where one group is set up for each layer. [Figure 7] Figure 7 shows an example of a supernet that will be output. [Figure 8] Figure 8 is a flowchart illustrating an example of the operation of an information processing device. [Figure 9] Figure 9 is a block diagram showing an example of a computer that implements an information processing device. [Modes for carrying out the invention]

[0023] (Embodiment) Embodiments of this disclosure will be described in detail below with reference to the drawings.

[0024] [Device configuration] First, the configuration of the information processing device 100 in the embodiment will be explained using Figure 1. Figure 1 is a block diagram that schematically shows the configuration of an example of the information processing device.

[0025] The information processing device 100 shown in Figure 1 is a device for generating machine learning models. As shown in Figure 1, the information processing device 100 comprises a group setting unit 10, a learning processing unit 20, and a candidate operation output unit 30.

[0026] The group setting unit 10 groups candidate operations for each layer in a machine learning model network where numerous candidate operations are placed in each layer, setting up one or more groups. The learning processing unit 20 performs machine learning on the machine learning model network. The learning processing unit 20 then selects candidate operations for each group using the parameter values ​​assigned to the candidate operations by the machine learning execution. The candidate operation output unit 30 outputs a machine learning model composed of the selected candidate operations.

[0027] In this way, the information processing device 100 groups candidate operations that make up each layer during machine learning model network training and selects candidate operations for each group. Therefore, convergence is fast, and bad candidate operations can be efficiently eliminated. In other words, the information processing device 100 can suppress the occurrence of situations where bad candidate operations are selected during neural architecture searching.

[0028] Next, in addition to Figure 1, Figures 2 to 7 will be used to specifically explain the configuration and functions of the information processing device 10. Figure 2 is a configuration diagram that more specifically explains the configuration of an example of an information processing device.

[0029] As shown in Figure 2, the information processing device 100 includes a network construction unit 40 in addition to the group setting unit 10, learning processing unit 20, and candidate operation output unit 30 described above. Also in Figure 2, 31 indicates the dataset used as training data in the machine learning of the machine learning model network. 50 is the machine learning model (AI model) output from the candidate operation output unit 30.

[0030] Dataset 31 includes images and labels that describe the content of the images. In AI model 50, the images are the explanatory variables and the labels are the dependent variables. When an image is input to the trained AI model, it outputs a predicted label.

[0031] The network construction unit 40 constructs a machine learning model network. In this embodiment, the network construction unit 40 constructs a supernet 60 as the machine learning model network, as shown in Figure 3. Figure 3 is a diagram showing an example of a supernet.

[0032] The supernet 60 consists of numerous layers with candidate operations arranged in parallel, and many such layers are arranged in a row. The number of candidate operations and the selection of candidate operations in each layer are not particularly limited. In all layers, a special architecture parameter (α) is assigned to the output of each candidate operation. For object detection tasks, as shown in Figure 3, the neck 80 and head 90 are generally added as additional blocks. The neck 80 and head 90 are convolutional layers.

[0033] The sample example of Supernet 60 shown in Figure 3 consists of various candidate operations at each layer. Figure 3 also shows the following candidate operations: A standard convolution operation with a kernel size of 3x3 and an expansion ratio of 1, called k3e1 candidate operation 7010. A standard convolution operation with a kernel size of 3x3 and an expansion ratio of 3, called k3e3 candidate operation 7011. A standard convolution operation with a kernel size of 3x3 and an expansion ratio of 6, called k3e6 candidate operation 7012. A standard convolution operation with a kernel size of 5x5 and an expansion ratio of 1, called k5e1 candidate operation 7013. A standard convolution operation with a kernel size of 5x5 and an expansion ratio of 3, called k3e3 candidate operation 7014. A standard convolution operation with a kernel size of 5x5 and an expansion ratio of 6, called k5e6 candidate operation 7015. A 3x3 average pooling operation called Avg.k3 candidate operation 7016. A 5x5 average pooling operation called Avg.k5 candidate operation 7017. Max.k3 candidate operation 7018, a maximum pooling operation with a kernel size of 3x3. Max.k5 candidate operation 7019, a maximum pooling operation with a kernel size of 5x5. Furthermore, as shown in Figure 3, the supernet 60 also includes a neck 80 and a head 90, which are convolution-based layers, for object detection tasks.

[0034] The group configuration unit 10 receives the constructed supernet 60 as input from the network construction unit 40. The group configuration unit 10 performs grouping of candidate operations within each layer. Each layer has multiple candidate operations, and all candidate operations in each layer are grouped based on a predetermined strategy. The number of groups formed in the candidate operations may differ from layer to layer. Also, the number of groups may differ from supernet to supernet.

[0035] The grouping strategy may differ from layer to layer and from supernet to supernet. The group setting unit 10 groups the candidate operations according to a rule that uses at least one of the following as a criterion: operation type, number of parameters, kernel size, number of multiply-accumulate operations, similar latency, dissimilar operation type, and heuristics. The rule is not limited to those mentioned above.

[0036] Grouping based on operation type means that candidate operations with similar operation types are grouped into one group. For example, candidate operations with the convolution type are grouped into one group, pooling operation types into another group, and eltwise operation types into yet another group.

[0037] Grouping based on the number of parameters means that candidate operations with similar parameter ranges are grouped into one group. For example, candidate operations with a large number of parameters, such as the standard convolution operation, are grouped into one group, while candidate operations with a very small number of parameters, such as the group convolution operation or the extended convolution operation, are grouped into another group. In this case, candidate operations without parameters, such as the skip operation, the Eltwise operation, and the channel swapping operation, are also grouped into another group.

[0038] Kernel size-based grouping means grouping candidate operations with similar kernel array sizes into a single group. For example, candidate operations with a kernel size of 3x3, such as a standard convolution operation and a maximum pooling operation, would be in one group, while candidate operations with a kernel size of 1x1, such as a pointwise convolution operation and a scale operation, would be in a separate group.

[0039] Grouping based on the number of multiply-accumulate operations means grouping candidate operations with similar multiply-accumulate operations into a single group. For example, candidate operations with many multiply-accumulate operations, such as standard convolution, group convolution, and extended convolution, will be in one group, while candidate operations with fewer multiply-accumulate operations, such as Eltwise and batch norm operations, will be in a separate group. In this case, candidate operations with few multiply-accumulate operations, such as skip operations and channel shuffle operations, will also be in a separate group.

[0040] Grouping based on similar latency means grouping candidate operations based on the amount of execution time they consume on specific hardware. For example, candidate operations with high latency, such as standard convolution and Eltwise operations, would belong to one group, while those with low latency, such as shuffle-channel operations or scale operations, would belong to a different group. In this case, candidate operations with almost negligible latency, such as skip operations and concatenation operations, would also belong to a different group.

[0041] Grouping based on dissimilar operation types means grouping candidate operations with dissimilar operation types into a single group. For example, candidate operations with convolution type operations and skip operations would be in one group, while pooling operations and Eltwise operations would be in a separate group. In this case, shuffle channel operations and batch norm operations would also be in a separate group.

[0042] Grouping may be based on certain expert heuristics. Furthermore, grouping may be a combination of two or more of the aforementioned grouping strategies (hybrid-based).

[0043] Figure 4 shows an example of grouping of candidate operations performed in the embodiment. In the example in Figure 4, a hybrid-based grouping strategy is adopted. In the example in Figure 4, in each layer, one group is formed by candidate operations with a standard convolution operation of kernel size 3x3, another group is formed by candidate operations with a standard convolution operation of kernel size 5x5, and yet another group is formed by candidate operations with a pooling operation.

[0044] A special architecture parameter (α) is assigned to the output of all candidate operations. Specifically, as shown in Figure 4, for example, the output of candidate operation 7010 k3e1 in the first layer is assigned the special architecture parameter (α) of the first candidate in the first group of the first layer. 1 L1g1 This is assigned. Additionally, the output of the Max.k5 candidate operation 7019 in the first layer includes α, which represents the special architecture parameter (α) of the fourth candidate in the third group of the first layer. 4 L1g3 This is assigned. Furthermore, the output of the Max.k5 candidate operation 7039 of the nth layer includes α, which represents the special architecture parameter (α) of the fourth candidate of the third group in the nth layer. 4 Lng3 It will be assigned.

[0045] The group setting unit 10 uses one of the strategies described above to group the candidate operations of each layer in the supernet 60. The group setting unit 10 then outputs the supernet 60 with the grouped candidate operations to the learning processing unit 20.

[0046] The learning processing unit 20 receives a supernet 60, in which candidate operations are grouped, as input. The learning processing unit 20 performs machine learning on the received supernet 60 using the dataset 31. Through machine learning on the supernet 60, the weight parameters and special architecture parameters (α) of all candidate operations are trained and updated.

[0047] Specifically, the learning processing unit 20 partially trains the supernet 60. After the partial training of the supernet 60, the learning processing unit 20 selects the candidate operation with the largest value of the special architecture parameter (α) in each group of each layer.

[0048] Poor candidate operations with high convergence rates, even if they have high values ​​for their special architecture parameter (α), only affect other candidate operations within the group to which they belong. Therefore, good candidate operations with slower convergence rates are not excluded, even if they do not have higher special architecture parameter (α) compared to poor candidate operations belonging to other groups. Such scenarios introduce fairness into the evaluation of candidate operations with different convergence rates, and with fairness evaluation and sufficient convergence time, the probability of selecting good candidates increases. This improves the training efficiency of supernet60.

[0049] Once the selection of candidate operations is complete, a corresponding number of candidate operations remain in each layer, depending on the number of groups within that layer. Figure 5 shows an example of a supernet after the selection of candidate operations. As shown in Figure 5, in the first layer, candidate k3e1 operation 7010, candidate k5e3 operation 7014, and candidate Max.k5 operation 7019 have been selected. In the second layer, candidate k3e3 operation 7021, candidate k5e1 operation 7023, and candidate Avg.k3 operation 7026 have been selected.

[0050] The learning processing unit 20 outputs the reduced supernet 60, in which the number of candidate operations in each layer has been reduced through partial training, to the candidate operation output unit 30. However, in this embodiment, as described below, grouping by the group setting unit 10 and machine learning by the learning processing unit 20 are performed in all layers until there is only one selected candidate operation.

[0051] If two or more candidate operations remain in any of the layers of the supernet 60, the partially trained and reduced supernet 60 is output to the group setting unit 10. The group setting unit 10 performs grouping again based on one of the grouping strategies. All remaining candidate operations in each layer are grouped according to a certain strategy. The number of groups into which the remaining candidate operations are grouped may differ from layer to layer and from supernet to supernet. The strategies employed for grouping may also differ from layer to layer and from supernet to supernet. Furthermore, if there are not many candidate operations remaining, all candidate operations may be grouped into one group per layer, as shown in Figure 6. Figure 6 shows a state in the supernet where one group is set per layer. The group setting unit 10 outputs the supernet 60, which has been grouped again, to the learning processing unit 20.

[0052] The learning processing unit 20 receives the regrouped supernet 60 as input. The learning processing unit 20 performs machine learning on the received supernet 60. Using the dataset, the learning processing unit 20 trains and updates the weight parameters and special architecture parameters (α) of all candidate operations in this machine learning of supernet 60 as well. In this case, training may be performed only on layers where the number of candidate operations is not single, and the supernet 60 is partially trained.

[0053] After training the supernet 60, the learning processing unit 20 again selects the candidate operation with the largest value of the special architecture parameter (α) in each group of each layer. At this time, if there is only one group within a layer, only one candidate operation remains. The learning processing unit 20 shrinks the supernet 60 until there is only one candidate operation in all layers, and then passes the shrunk supernet 60 to the candidate operation output unit 30.

[0054] The candidate operation output unit 30 outputs a reduced supernet 60, as shown in Figure 7. Figure 7 shows an example of the output supernet. Such a reduced supernet 60 is the AI ​​model 50. Generally, a reduced supernet 60 is also called a subnet.

[0055] [Device operation] Next, the operation of the information processing device 100 in the embodiment will be explained using Figure 8. Figure 8 is a flowchart showing an example of the operation of the information processing device. In the following explanation, Figures 1 to 7 will be referred to as appropriate. In this embodiment, the information processing method is implemented by operating the information processing device 100. Therefore, the explanation of the information processing method in this embodiment will be replaced by the following explanation of the operation of the information processing device 100.

[0056] As shown in Figure 8, first, the network construction unit 40 constructs a supernet 60 as a machine learning model network (step S1). An example of a supernet 60 is shown in Figure 3.

[0057] Next, the group setting unit 10 receives the supernet 60 constructed in step S1 and groups the candidate operations for each layer that makes up the supernet 60 (step S2).

[0058] Next, the learning processing unit 20 uses the dataset 31 to perform machine learning on the supernet 60, which was grouped in step S2 (step S3).

[0059] Next, after training in step S4, the learning processing unit 20 selects the candidate operation with the largest value of the special architecture parameter (α) in each group of each layer and eliminates the remaining candidate operations (step S4).

[0060] Next, the learning processing unit 20 determines whether there is a single selected candidate operation in all layers of the supernet 60 (step S5).

[0061] If, as a result of the determination in step S5, the selected candidate operation is not unique in any of the layers of the supernet 60, the learning processing unit 20 passes the supernet 60 after the execution of step S4 to the group setting unit 10 and has it execute step S2 again. After that, steps S3 to S5 are also executed again.

[0062] On the other hand, if the result of the determination in step S5 is that there is only one selected candidate operation in all layers of the supernet 60, the learning processing unit 20 passes the supernet 60 after the execution of step S4 to the candidate operation output unit 30. As a result, the candidate operation output unit 30 outputs the reduced supernet 60, i.e., the AI ​​model 50, as shown in Figure 7 (step S6).

[0063] Thus, if any of the layers of Supernet 60 have two or more candidate operations, steps S2 to S5 are repeated until finally the subnet is output.

[0064] Thus, according to this embodiment, a good candidate operation with a slow convergence rate in one group will not be prematurely eliminated due to the presence of a candidate operation with a fast convergence rate in another group. Therefore, since candidate operations in the supernet 60 are evaluated on a group basis, fairness is ensured in the evaluation, thereby increasing the probability that a good candidate operation will be selected in the final AI model 50. Furthermore, in this embodiment, the supernet 60 can be trained in a partially iterative manner, allowing more time to be spent training good candidate operations with slow convergence rates. As a result, the training efficiency of the supernet 60 is increased.

[0065] [program] In this embodiment, the program is one that causes a computer to execute steps S1 to S6 shown in Figure 8. By installing and executing this program on a computer, the information processing device 100 and the information processing method in this embodiment can be realized. In this case, the computer's processor functions as a group setting unit 10, a learning processing unit 20, a candidate operation output unit 30, and a network construction unit 40, and performs the processing. Examples of computers include general-purpose PCs, smartphones, and tablet terminal devices.

[0066] Furthermore, the program in this embodiment may be executed by a computer system composed of multiple computers. In this case, for example, each computer may function as one of the following: the group setting unit 10, the learning processing unit 20, the candidate operation output unit 30, and the network construction unit 40.

[0067] [Physical configuration] Here, a computer that implements the information processing device 10 by executing the program in the embodiment will be described with reference to Figure 9. Figure 9 is a block diagram showing an example of a computer that implements the information processing device.

[0068] As shown in Figure 9, the computer 110 comprises a CPU (Central Processing Unit) 111, main memory 112, storage device 113, input interface 114, display controller 115, data reader / writer 116, and communication interface 117. Each of these components is connected to the others via a bus 121, enabling data communication.

[0069] Furthermore, the computer 110 may include a GPU (Graphics Processing Unit) or an FPGA (Field-Programmable Gate Array) in addition to, or instead of, the CPU 111. In this embodiment, the GPU or FPGA can execute the program in the embodiment.

[0070] The CPU 111 loads the program in the embodiment, which consists of a set of codes stored in the storage device 113, into the main memory 112, and performs various calculations by executing each code in a predetermined order. The main memory 112 is typically a volatile storage device such as DRAM (Dynamic Random Access Memory).

[0071] Furthermore, the program in this embodiment is provided stored on a computer-readable recording medium 120. The program in this embodiment may also be distributed over the internet via a communication interface 117.

[0072] Specific examples of the storage device 113 include hard disk drives and semiconductor storage devices such as flash memory. The input interface 114 mediates data transmission between the CPU 111 and input devices 118 such as a keyboard and mouse. The display controller 115 is connected to the display device 119 and controls the display on the display device 119.

[0073] The data reader / writer 116 mediates data transmission between the CPU 111 and the recording medium 120, reads programs from the recording medium 120, and writes processing results from the computer 110 to the recording medium 120. The communication interface 117 mediates data transmission between the CPU 111 and other computers.

[0074] Furthermore, specific examples of the recording medium 120 include general-purpose semiconductor memory devices such as CF (Compact Flash®) and SD (Secure Digital), magnetic recording media such as Flexible Disks, or optical recording media such as CD-ROMs (Compact Disk Read Only Memory).

[0075] Furthermore, the information processing device 100 can be implemented not only by a computer with a program installed, but also by using hardware corresponding to each part, such as electronic circuits. Moreover, the information processing device 100 may be partially implemented by a program and the remaining parts by hardware. The computer is not limited to the one shown in Figure 9.

[0076] Some or all of the embodiments described above can be expressed by (Appendix 1) to (Appendix 15) described below, but are not limited to the following descriptions.

[0077] (Note 1) In a machine learning model network in which numerous candidate operations are placed in each layer, a group setting means is provided for grouping the candidate operations in each layer to set up one or more groups. A learning processing means that performs machine learning on the machine learning model network and selects candidate operations for each group using the parameter values ​​assigned to the candidate operations by the machine learning execution, A candidate operation output means that outputs a machine learning model composed of selected candidate operations, It is equipped with An information processing device characterized by the following:

[0078] (Note 2) In any of the above-mentioned layers, if there are two or more selected candidate operations, The group setting means performs grouping of the selected candidate operations in a layer where there are two or more selected candidate operations. The learning processing means performs machine learning on the machine learning model network composed of the selected candidate operations, and selects the candidate operations again in the layer where there are two or more candidate operations. The candidate operation output means outputs a machine learning model composed of the selected candidate operation when there is only one selected candidate operation in all layers. The information processing device described in Appendix 1.

[0079] (Note 3) The group setting means groups the candidate operations according to a rule that uses at least one of the following as a criterion: operation type, number of parameters, kernel size, number of multiply-accumulate operations, similar latency, dissimilar operation type, and heuristic. The information processing device described in Appendix 1.

[0080] (Note 4) The network construction means for constructing the aforementioned machine learning model network is Furthermore, it is equipped with The information processing device described in Appendix 1.

[0081] (Note 5) The network construction means constructs a supernet as the machine learning model network. The information processing device described in Appendix 1.

[0082] (Note 6) In a machine learning model network in which numerous candidate operations are placed in each layer, the candidate operations are grouped into one or more groups for each layer, Machine learning is performed on the machine learning model network, and a candidate operation is selected for each group using the parameter values ​​assigned to the candidate operations by the machine learning execution. Outputs a machine learning model composed of the selected candidate operations. An information processing method characterized by the following:

[0083] (Note 7) In any of the above-mentioned layers, if there are two or more selected candidate operations, In layers where there are two or more selected candidate operations, grouping is performed on the selected candidate operations. Machine learning is performed on the machine learning model network composed of the selected candidate operations, and the candidate operations are selected again in the layer where there are two or more candidate operations. In all of the aforementioned layers, if there is only one selected candidate operation, the machine learning model composed of the selected candidate operation is output. The information processing method described in Appendix 6.

[0084] (Note 8) In the group setting described above, candidate operations are grouped according to a rule that uses at least one of the following as a criterion: operation type, number of parameters, kernel size, number of multiply-accumulate operations, similar latency, dissimilar operation type, and heuristics. The information processing method described in Appendix 6.

[0085] (Note 9) Furthermore, to construct the aforementioned machine learning model network, The information processing method described in Appendix 6.

[0086] (Note 10) In constructing the aforementioned network, a supernet is constructed as the machine learning model network. The information processing method described in Appendix 6.

[0087] (Note 11) On the computer, In a machine learning model network in which numerous candidate operations are placed in each layer, the candidate operations are grouped into one or more groups for each layer. Machine learning is performed on the machine learning model network, and the candidate operations are selected for each group using the parameter values ​​assigned to the candidate operations by the execution of the machine learning. Output a machine learning model composed of the selected candidate operations. A computer-readable recording medium that stores programs containing instructions.

[0088] (Note 12) The program further provides the computer with: In any of the above-mentioned layers, if there are two or more selected candidate operations, In layers where there are two or more selected candidate operations, grouping is performed on the selected candidate operations. Machine learning is performed on the machine learning model network composed of the selected candidate operations, and the candidate operations are selected again in the layer where there are two or more candidate operations. In all of the aforementioned layers, if there is only one selected candidate operation, a machine learning model composed of the selected candidate operation is output. A computer-readable recording medium as described in Appendix 11.

[0089] (Note 13) In the group setting described above, the candidate operations are grouped according to a rule that uses at least one of the following as a criterion: operation type, number of parameters, kernel size, number of multiply-accumulate operations, similar latency, dissimilar operation type, and heuristic. A computer-readable recording medium as described in Appendix 11.

[0090] (Note 14) The program further provides the computer with: To construct the aforementioned machine learning model network, A computer-readable recording medium as described in Appendix 11.

[0091] (Note 15) The program further provides the computer with: In constructing the aforementioned network, a supernet is constructed as the machine learning model network. A computer-readable recording medium as described in Appendix 11.

[0092] Although the present invention has been described above with reference to embodiments, the present invention is not limited to the above embodiments. Various modifications to the structure and details of the present invention can be made, as can be understood by those skilled in the art within the scope of the present invention. [Industrial applicability]

[0093] As described above, this disclosure makes it possible to suppress the occurrence of situations where poor candidate operations are selected during neural architecture searching. This disclosure is useful for machine learning systems. [Explanation of Symbols]

[0094] 10 Group Setting Section 20 Learning Processing Unit 30 Candidate Operation Output Unit 31 datasets 40 Network Construction Department 50 Machine Learning Models (AI Models) 60 Supernet 70 Supernet Backbone 80 Neck 90 heads 100 Information Processing Devices 110 Computer 111 CPU 112 Main Memory 113 Storage device 114 Input Interface 115 Display Controller 116 Data Readers / Writers 117 Communication Interface 118 Input devices 119 Display device 120 recording media 121 Bus 501 AI Model Backbone 700 Fixed blocks within the Supernet backbone 7010 Standard convolution operation with kernel size 1x1 and expansion ratio 1 in the first layer 7011 Standard convolution operation with kernel size 3x3 and expansion ratio 1 in the first layer 7020 Standard convolution operation with kernel size 1x1 and expansion ratio 1 in the second layer 7021 Standard convolution operation with kernel size 3x3 and expansion ratio 1 in the second layer

Claims

1. In a machine learning model network in which numerous candidate operations are placed in each layer, a group setting means is provided for grouping the candidate operations in each layer to set up one or more groups. A learning processing means that performs machine learning on the machine learning model network and selects candidate operations for each group using the parameter values ​​assigned to the candidate operations by the machine learning execution, A candidate operation output means that outputs a machine learning model composed of selected candidate operations, It is equipped with An information processing device characterized by the following:

2. In any of the above-mentioned layers, if there are two or more selected candidate operations, The group setting means performs grouping of the selected candidate operations in a layer where there are two or more selected candidate operations. The learning processing means performs machine learning on the machine learning model network composed of the selected candidate operations, and in the layer where there are two or more candidate operations, select the candidate operations again. The candidate operation output means outputs a machine learning model composed of the selected candidate operation when there is only one selected candidate operation in all layers. The information processing apparatus according to claim 1.

3. The group setting means groups the candidate operations according to a rule that uses at least one of the following as a criterion: operation type, number of parameters, kernel size, number of multiply-accumulate operations, similar latency, dissimilar operation type, and heuristic. The information processing apparatus according to claim 1.

4. The network construction means for constructing the aforementioned machine learning model network is Furthermore, it is equipped with The information processing apparatus according to claim 1.

5. The network construction means constructs a supernet as the machine learning model network. The information processing apparatus according to claim 1.

6. In a machine learning model network in which numerous candidate operations are placed in each layer, the candidate operations are grouped into one or more groups for each layer, Machine learning is performed on the machine learning model network, and a candidate operation is selected for each group using the parameter values ​​assigned to the candidate operations by the machine learning execution. Outputs a machine learning model composed of the selected candidate operations. An information processing method characterized by the following:

7. In any of the above-mentioned layers, if there are two or more selected candidate operations, In layers where there are two or more selected candidate operations, grouping is performed on the selected candidate operations. Machine learning is performed on the machine learning model network composed of the selected candidate operations, and in the layer where there are two or more candidate operations, the candidate operations are selected again. In all of the aforementioned layers, if there is only one selected candidate operation, the machine learning model composed of the selected candidate operation is output. The information processing method according to claim 6.

8. In the group setting described above, candidate operations are grouped according to a rule that uses at least one of the following as a criterion: operation type, number of parameters, kernel size, number of multiply-accumulate operations, similar latency, dissimilar operation type, and heuristics. The information processing method according to claim 6.

9. Furthermore, to construct the aforementioned machine learning model network, The information processing method according to claim 6.

10. In constructing the aforementioned network, a supernet is constructed as the machine learning model network. The information processing method according to claim 6.

11. On the computer, In a machine learning model network in which numerous candidate operations are placed in each layer, the candidate operations are grouped into one or more groups for each layer. Machine learning is performed on the machine learning model network, and the candidate operations are selected for each group using the parameter values ​​assigned to the candidate operations by the execution of the machine learning. Output a machine learning model composed of the selected candidate operations. A computer-readable recording medium that stores programs containing instructions.

12. The program further provides the computer with: In any of the above-mentioned layers, if there are two or more selected candidate operations, In layers where there are two or more selected candidate operations, grouping is performed on the selected candidate operations. Machine learning is performed on the machine learning model network composed of the selected candidate operations, and the candidate operations are selected again in the layer where there are two or more candidate operations. In all of the aforementioned layers, if there is only one selected candidate operation, a machine learning model composed of the selected candidate operation is output. The computer-readable recording medium according to claim 11.

13. In the group setting described above, the candidate operations are grouped according to a rule that uses at least one of the following as a criterion: operation type, number of parameters, kernel size, number of multiply-accumulate operations, similar latency, dissimilar operation type, and heuristic. The computer-readable recording medium according to claim 11.

14. The program further provides the computer with: To construct the aforementioned machine learning model network, The computer-readable recording medium according to claim 11.

15. The program further provides the computer with: In constructing the aforementioned network, a supernet is constructed as the machine learning model network. The computer-readable recording medium according to claim 11.

Images