JP7699791B2 - AI processor - Google Patents

Description

The present invention relates to an AI processor.

Recently, COVID-19 (hereinafter also referred to as novel coronavirus infection) caused by the SARS-CoV2 virus (hereinafter referred to as novel coronavirus) has become widespread. A standard method for testing whether or not a patient is infected with the novel coronavirus is, for example, reverse transcription polymerase chain reaction (RT-PCR) using a sample collected from the patient, which has a sensitivity of about 60% to 97%. Another method for testing whether or not a patient is infected with the novel coronavirus is, for example, analysis of X-ray images of the patient's lungs, which has an accuracy of about 80% to 90%.

When analyzing X-ray images as described above, doctors must manually diagnose each X-ray image of the patient, which leads to inefficient diagnostic procedures.

Therefore, hospitals may use machine learning models (hereinafter also referred to as diagnostic models) to diagnose whether patients are infected with the new coronavirus or whether they have pneumonia, for example, in order to carry out efficient diagnostic procedures for many patients.

US Patent Application Publication No. 20200026992A1

Kun-Chih (Jimmy) Chen, Ting-Yi Wang, “NN-Noxim: High-Level Cycle-Accurate NoC-based Neural Networks Simulator”, 2018 11th International Workshop on Network on Chip Architectures (NoCArc). Xiaoxiao Liu, Wei Wen, Xuehai Qian, Hai Li, Yiran Chen, “Neu-NoC: A High-efficient Interconnection Network for Accelerated Neuromorphic Systems”, 2018 23rd Asia and South Pacific Design Automation Conference (ASP-DAC) Kun-Chih (Jimmy) Chen, Masoumeh Ebrahimi, Ting-Yi Wang, Yuch-Chi Yang, “NoC-based DNN Accelerator: A Future Design Paradigm”, NOCS ’19, October 17-18, 2019, New York, NY, USA.

However, the efficiency of diagnosis using such diagnostic models varies greatly depending on the performance of the computer that executes the diagnostic model. For this reason, in medical settings such as hospitals, there is a demand for the development of AI (Artificial Intelligence) processors that can perform diagnosis using such diagnostic models more efficiently.

The object of the present invention is to provide an AI processor that can perform diagnosis using a diagnostic model more efficiently.

In one aspect of the present invention, an AI processor has multiple computation cores, and at least one of the multiple computation cores executes a mapping process in which a computation program corresponding to each of multiple neurons included in a machine learning model of a CNN (Convolutional Neural Network) having a convolutional layer and a fully connected layer is divided and assigned to each of the multiple computation cores, and each of the multiple computation cores executes the computation program assigned by the mapping process, and in the mapping process, the computation program is assigned to the multiple computation cores by a genetic algorithm so that the communication cost between the multiple computation cores is equal to or less than a predetermined threshold.

According to one aspect of the present invention, it becomes possible to perform diagnosis using a diagnostic model more efficiently.

FIG. 1 is a diagram illustrating a configuration of a conventional information processing system 200. As shown in FIG. FIG. 2 is a diagram illustrating a configuration of the information processing system 100 according to the present embodiment. FIG. 3 is a diagram illustrating an example of a user interface in this embodiment. FIG. 4 is a diagram illustrating an example of a user interface in this embodiment. FIG. 5 is a diagram illustrating an example of a user interface in this embodiment. FIG. 6 is a diagram for explaining a specific example of the diagnostic model MD. 7 is a diagram for explaining the update process in the diagnostic model MD. FIG. 8 is a diagram illustrating a specific example of the AI processor PR in this embodiment. FIG. 9 is a diagram illustrating the mapping process in this embodiment. FIG. 10 is a diagram illustrating the mapping process in this embodiment. FIG. 11 is a diagram illustrating the mapping process in this embodiment. FIG. 12 is a diagram illustrating the mapping process in this embodiment. FIG. 13 is a diagram illustrating the mapping process in this embodiment. FIG. 14 is a diagram illustrating the mapping process in this embodiment. FIG. 15 is a diagram illustrating the mapping process in this embodiment. FIG. 16 is a diagram illustrating the mapping process in this embodiment. FIG. 17 is a diagram illustrating the mapping process in this embodiment. FIG. 18 is a diagram illustrating the mapping process in this embodiment. FIG. 19 is a diagram illustrating the mapping process in this embodiment. FIG. 20 is a diagram illustrating the mapping process in this embodiment. FIG. 21 is a diagram illustrating the mapping process in this embodiment.

The following describes embodiments of the present invention with reference to the drawings. Each embodiment is prepared for a better understanding of the present invention. However, such embodiments do not limit the technical scope of the present invention. The scope of the present invention also includes the claims and their equivalents.

[Conventional Information Processing System]
First, a conventional information processing system 200 will be described. Fig. 1 is a diagram for explaining the configuration of the conventional information processing system 200. Note that, although the following description will be given for a case where there are three hospitals (hospital 11, hospital 12, and hospital 1n), the number of hospitals may be other than three.

In the information processing system 200 shown in FIG. 1, a doctor at a hospital 11 performs a diagnosis by analyzing an X-ray image 21 of a patient. Then, for example, by using an information processing device (not shown) set up at the hospital 11, the doctor periodically generates patient information 31 indicating the diagnosis result as to whether or not each patient is infected with COVID-19 (e.g., the number of infected people, the severity of each infected person, and the number of deaths, etc.), and transmits the generated patient information 31 to a government information processing device 4.

Similarly, in the example shown in FIG. 1, at hospital 12, a doctor generates patient information 32 indicating the diagnosis result for the patient's X-ray image 22, and transmits the generated patient information 32 to the government information processing device 4. Also, at hospital 1n, a doctor generates patient information 3n indicating the diagnosis result for the patient's X-ray image 2n, and transmits the generated patient information 3n to the government information processing device 4.

Here, in the information processing system 200, each of hospitals 11, 12, and 1n cannot cooperate with the other hospitals. In this case, each doctor cannot efficiently diagnose the X-ray images.

[Information Processing System in the Present Embodiment]
Next, the information processing system 100 according to the present embodiment will be described. Fig. 2 is a diagram for explaining the configuration of the information processing system 100 according to the present embodiment. Figs. 3 to 5 are diagrams for explaining examples of a user interface according to the present embodiment.

In the information processing system 100 shown in FIG. 2, in the hospital 11, an X-ray image transmitted from a patient's mobile terminal 51 is input to a diagnostic system 61 installed in the hospital 11. The detection system 61 is, for example, an information processing device that executes a diagnostic model to determine whether or not the X-ray image transmitted from the mobile terminal 51 is an image of a patient infected with the new coronavirus. The diagnostic system 61 then transmits a diagnosis result for the X-ray image transmitted from the mobile terminal 51 to the mobile terminal 51.

Similarly, in the example shown in FIG. 2, a diagnostic system 62 installed in hospital 12 diagnoses the X-ray image transmitted from mobile terminal 52 and transmits the diagnostic result to mobile terminal 52. Also, a diagnostic system 6n installed in hospital 1n diagnoses the X-ray image transmitted from mobile terminal 5n and transmits the diagnostic result to mobile terminal 5n.

This makes it possible for hospitals 11, 12, and 1n to reduce the burden on doctors associated with diagnosing X-ray images. It also makes it possible for hospitals 11, 12, and 1n to prevent erroneous diagnoses from occurring on X-ray images. It also makes it possible for each doctor to check the diagnosis status of each patient in real time, for example, by viewing the user interface U1 shown in FIG. 3.

In addition, in hospitals 11, 12, and 1n, diagnosis of X-ray images can be performed quickly by performing diagnosis using the diagnostic model, and the diagnosis results can be quickly notified to mobile terminals 51, 52, and 5n. Therefore, each patient can quickly check the diagnosis results of the X-ray images by viewing, for example, user interface U2 shown in FIG. 4.

Then, the diagnostic system 61 generates patient information in real time indicating the diagnosis results based on the diagnostic model (e.g., the number of infected people, the severity of each infected person, and the number of deaths, etc.), and transmits the generated patient information to the government's information processing device 4.

Similarly, in the example shown in FIG. 2, each of diagnostic system 62 and diagnostic system 6n generates patient information indicating the diagnosis results based on the diagnostic model in real time and transmits it to the government information processing device 4.

Specifically, each of diagnostic systems 61, 62, and 6n transmits patient information to cloud server 7, for example. Then, cloud server 7 generates real-time information 8 from diagnostic systems 61, 62, and 6n, for example, and transmits the generated real-time information 8 to government information processing device 4.

This allows Hospital 11, Hospital 12, and Hospital 1n to quickly transmit the latest patient information to the government's information processing device 4. This allows government personnel to quickly check the diagnosis results for each patient by viewing, for example, the user interface U3 shown in FIG. 5.

[Specific example of diagnostic model]
Next, a specific example of the diagnostic model MD executed in the diagnostic system 61, the diagnostic system 62, and the diagnostic system 6n will be described. Fig. 6 is a diagram for explaining a specific example of the diagnostic model MD.

In the diagnostic model MD shown in Figure 6, when an X-ray image is input from the input layer, it outputs either "Normal," a category indicating that the X-ray image is not infected with the new coronavirus (no pneumonia), or "Suspect," a category indicating that the X-ray image is suspected of being infected with the new coronavirus (suspect pneumonia).

Specifically, the diagnostic model MD shown in FIG. 6 classifies X-ray images by using, for example, a softmax function. In addition, the diagnostic model MD shown in FIG. 6 uses, for example, the cross-entropy loss expressed by the following formula (1) as the loss function E.

In formula (1), w _o and b _o are parameters in the diagnostic model MD, and y and y' indicate the actual label (correct label) and the predicted label, respectively. In the diagnostic model MD shown in FIG. 6, a loss function E between the actual label and the predicted label is calculated by using formula (1). Then, in the diagnostic model MD, the loss function E is minimized using a stochastic gradient reduction algorithm or a backpropagation algorithm, and parameters are further optimized.

[Update diagnostic model]
Next, the process of updating the diagnostic model MD (hereinafter also referred to as the update process) will be described. Fig. 7 is a diagram for explaining the update process in the diagnostic model MD. Specifically, Fig. 7(A) is a flow chart for explaining the update process. Fig. 7(B) is a diagram for explaining the transmission and reception of information when the update process is executed. Note that the update process performed between the diagnostic system 61 and the cloud server 7 described in Fig. 2 will be described below.

The diagnostic system 61 generates a diagnostic model MD in advance for the hospital 11 by using a data set including X-ray images transmitted from patients in the hospital 11. The diagnostic system 61 then transmits parameters for the generated diagnostic model MD (e.g., the gradient ∇gL shown in the following equation (2)) to the cloud server 7 (S1).

Next, the cloud server 7 calculates a global parameter (e.g., the global gradient ∇gG shown in the following formula (3)) from each of the received parameters, for example, in response to receiving parameters from each of the diagnostic systems 61, 62, and 6n (S2).

Furthermore, the cloud server 7 transmits the calculated global parameters to each of the diagnostic systems 61, 62, and 6n (S3).

Then, in response to receiving the global parameters transmitted from the cloud server 7, the diagnostic system 61 updates the parameters of the diagnostic model MD in the hospital 11 as shown in the following equations (4) and (5) (S4).

In equations (4) and (5), Wr and br respectively represent the weight and bias in the rth process of S1 to S4 (the rth training round). Also, in equations (4) and (5), η represents the learning rate.

That is, the information processing system 100 generates diagnostic models MD (parameters of the diagnostic models MD) for each hospital using federated model learning (FML).

This enables the information processing system 100 to improve the accuracy of each hospital's diagnostic model MD without transmitting a data set containing personal information about patients at each hospital to other hospitals or other external parties. As a result, each hospital's diagnostic model MD can accurately diagnose the new coronavirus while protecting the privacy of each patient.

The processes from S1 to S4 may be repeated until the judgment accuracy of each of the diagnostic models MD of each hospital exceeds the required condition.

In addition, the diagnostic model MD of each hospital may be generated, for example, in a computer other than diagnostic system 61, diagnostic system 62, and diagnostic system 6n (for example, a host computer HC shown in FIG. 8 described below).

[Diagnostic model mapping]
Next, a process of mapping a diagnostic model MD to a conventional AI processor (hereinafter also referred to as an AI chip) PR (hereinafter also referred to as a mapping process) will be described.

The AI processor PR is, for example, a processor mounted in diagnostic system 61, diagnostic system 62, and diagnostic system 6n. When mapping of the diagnostic model MD to the AI processor PR is performed, the AI processor PR, for example, clusters the neurons that make up the diagnostic model MD into multiple groups, and then maps each group to each of the multiple arithmetic cores included in the AI processor PR.

However, conventional mapping processes may not take into account the communication costs between multiple computing cores. As a result, conventional mapping processes may not be completed within the required time.

Furthermore, both clustering neurons and mapping them to multiple computing cores are generally NP-hard problems and may not be optimally solvable in polynomial time.

Therefore, in this embodiment, the AI processor PR uses a genetic algorithm to map the neurons that make up each layer of the diagnostic model MD so as to reduce communication costs between multiple computing cores.

[Specific example of AI processor in this embodiment]
Next, a specific example of the AI processor PR in this embodiment will be described. Fig. 8 is a diagram for explaining a specific example of the AI processor PR in this embodiment. In the following, the description will be given on the assumption that the diagnostic model MD is generated in the host computer HC.

The AI processor PR shown in FIG. 8 has 15 arithmetic cores. Specifically, the AI processor PR shown in FIG. 8 has 10 arithmetic cores C associated with the convolutional layer and 3 arithmetic cores F associated with the fully connected layer. Each arithmetic core is also connected to a router R. Note that, although the following description will be given of the AI processor PR having 15 arithmetic cores, the AI processor PR may have a number of arithmetic cores other than this.

The AI processor PR shown in FIG. 8 has one calculation core U associated with the pooling layer and activation function, and one calculation core I/O that transmits weighting coefficients to each calculation core. The calculation core I/O may, for example, communicate with other AI processors PR (inter-chip communication) or communicate with a host computer HC that controls each AI processor PR.

The AI processor PR shown in FIG. 8 also has an on-chip memory M (hereinafter also simply referred to as memory M) that stores inputs to the AI processor PR. Each calculation core, such as the calculation core C, loads inputs from the memory M and executes processing. Then, each calculation core stores the output of each calculation core associated with the execution of processing in the memory M to enable the execution of processing corresponding to the next layer.

The AI processor PR shown in FIG. 8 further includes a router R corresponding to each of the arithmetic cores and the memory M, and an external DRAM (Dynamic Random Access Memory).

[Mapping Process in the Present Embodiment]
Next, the mapping process in this embodiment will be described. Fig. 9 is a flow chart for explaining the mapping process in this embodiment. Figs. 10 to 21 are diagrams for explaining the mapping process in this embodiment.

The computational core U of the AI processor PR randomly determines N mapping solutions for the neurons that make up the diagnostic model MD (neural network) (S11).

Specifically, if the AI processor PR has four calculation cores that can each correspond to eight neurons, and the number of neurons that make up the diagnostic model MD is 30, the combinations of mapping solutions will be 496, which are combinations in which 30 neurons are arranged among 32 neurons. Therefore, in this case, the calculation core U randomly determines N mapping solutions that correspond to N of the 496 combinations.

The following description will be given assuming that the AI processor PR has four computing cores that can each correspond to eight neurons, and that the number of neurons constituting the diagnostic model MD is 30. Furthermore, the description will be given assuming that the diagnostic model MD is a neural network having layers L1, L2, and L3, each of which is made up of 10 neurons.

[Example of mapping results]
Next, a specific example of the mapping result in this embodiment will be described. Fig. 10 is a diagram showing the number of neurons mapped to each processor core. Fig. 11 is a diagram showing identification information of the neurons mapped to each processor core.

The examples shown in Figures 10 and 11 show that the neurons associated with the computation core in the first row and first column are three neurons (neurons 2, 3, and 5) corresponding to layer L1, two neurons (neurons 11 and 12) corresponding to layer L2, and three neurons (neurons 22, 23, and 28) corresponding to layer L3.

The examples shown in Figures 10 and 11 also show that the neurons associated with the arithmetic core in the first row, second column are three neurons (neurons 6, 7, and 8) corresponding to layer L1, one neuron (neuron 13) corresponding to layer L2, and three neurons (neurons 21, 29, and 30) corresponding to layer L3. The examples shown in Figures 10 and 11 also show that the number of neurons (F) that can be further supported by the arithmetic core in the first row, second column is 1.

The examples shown in Figures 10 and 11 also show that the neurons associated with the computation core in the second row, first column are two neurons (neurons 1 and 4) corresponding to layer L1, three neurons (neurons 15, 17, and 19) corresponding to layer L2, and three neurons (neurons 24, 25, and 26) corresponding to layer L3.

The examples shown in Figures 10 and 11 also show that the neurons associated with the arithmetic core in the second row, second column are two neurons (neurons 9 and 10) corresponding to layer L1, four neurons (neurons 14, 16, 18, and 20) corresponding to layer L2, and one neuron (neuron 27) corresponding to layer L3. The examples shown in Figures 10 and 11 also show that the number of neurons (F) that can be further supported by the arithmetic core in the second row, second column is 1.

Returning to FIG. 9, the computation core U deletes inappropriate mapping solutions from the N mapping solutions determined in the process of S11 (S12).

Specifically, the computation core U deletes a mapping solution in which there is a computation core (the computation core in the first row and first column) associated with nine neurons, as shown in FIG. 12. In addition, the computation core U deletes a mapping solution in which the same neuron (neuron 8) is associated with multiple computation cores, as shown in FIG. 13.

Then, the computation core U calculates the communication cost corresponding to each mapping solution that was not deleted in the process of S12 (S13).

Specifically, the communication cost of each mapping solution is expressed by the following equation (6).

In equation (6), d _i,j indicates the distance between neuron i and neuron j, and c _i,j indicates the connection status between neuron i and neuron j.

Specifically, d _i,j is, for example, a value obtained by adding 1 to the number of routers R existing between the processing core associated with neuron i and the processing core associated with neuron j. Also, c _i,j is, for example, 1 when neuron i and neuron j are directly connected, and is 0 when neuron i and neuron j are not directly connected.

More specifically, in the example described in Fig. 11, neuron 1 and neuron 2 are associated with layer L1. Therefore, in this case, _d1,2 are 1, and _c1,2 are 0. Also, in the example described in Fig. 11, neuron 1 is associated with layer L1, and neuron 14 is associated with layer L2. Therefore, in this case, _d1,14 are 4, and _c1,2 are 1.

Next, the computation core U identifies M mapping solutions that satisfy the conditions for the communication cost calculated in the process of S13 from among the mapping solutions that were not deleted in the process of S12 (S14).

Specifically, the computation core U identifies M mapping solutions in descending order of communication cost calculated in the process of S13 from among the mapping solutions that were not deleted in the process of S12, for example.

Then, the computation core U determines N-M new mapping solutions by crossing over the M mapping solutions identified in the process of S14 (S15).

Specifically, for example, if parent 1 and parent 2 shown in Figures 14 and 15 are included in the M mapping solutions identified in the processing of S14, the computation core U creates a new descendant with the proportions of parent 1 and parent 2 set to 50 (%), as shown in Figure 16.

More specifically, in the example shown in FIG. 14, the arithmetic core in the first row and first column for parent 1 is associated with three neurons corresponding to layer L1, two neurons corresponding to layer L2, and three neurons corresponding to layer L3. Also, in the example shown in FIG. 15, the arithmetic core in the first row and first column for parent 2 is associated with one neuron corresponding to layer L1, four neurons corresponding to layer L2, and three neurons corresponding to layer L3. Therefore, in this case, the arithmetic core in the first row and first column for the new descendant is associated with two (3*0.5+1*0.5) neurons corresponding to layer L1, three (4*0.5+2*0.5) neurons corresponding to layer L2, and three (3*0.5+3*0.5) neurons corresponding to layer L3, as shown in FIG. 16.

For example, when the number of neurons associated with each computation core includes a decimal point as shown in FIG. 17, computation core U may adjust the number of neurons associated with each computation core so that each number is an integer as shown in FIG. 18.

Returning to FIG. 9, the computational core U generates mutations in the N mapping solutions (the sum of the M mapping solutions identified in the process of S14 and the N-M new mapping solutions determined in the process of S15) (S16).

Specifically, as shown in FIG. 19, the computing core U identifies, for example, the computing core in the first row, first column and the computing core in the second row, second column from the multiple computing cores shown in FIG. 18, and further identifies layer L1 and layer L2 from the multiple layers constituting the diagnostic model MD. Then, the computing core U identifies, for example, 2, which is the minimum value of 2, which is the number of neurons associated with layer L1 in the identified computing core in the first row, first column, and 3, which is the number of neurons associated with layer L2 in the computing core in the second row, second column. After that, the computing core U subtracts 2, which is the value identified as the minimum value, from 2, which is the number of neurons associated with layer L1 in the computing core in the first row, first column, and further adds 2, which is the value identified as the minimum value, to 3, which is the number of neurons associated with layer L2 in the computing core in the first row, first column.

Similarly, the computation core U subtracts 2 from the number of neurons associated with layer L2 in the computation core in the second row, second column, which is 3, and then adds 2 to the number of neurons associated with layer L1 in the computation core in the second row, second column, which is 4.

Then, the computation core U determines whether or not the N mapping solutions after the processing of S16 have been performed satisfy the constraints, for example, in the same manner as in the processing of S12. If the result is that there is a mapping solution that does not satisfy the constraints, the computation core U deletes the existing mapping solution. Furthermore, the computation core U calculates the communication costs of the mapping solutions that have not been deleted (S17), in the same manner as in the processing of S13.

Then, it is determined whether the optimal cost (smallest cost) among the communication costs calculated in the process of S17 satisfies a predetermined condition (S18).

As a result, if it is determined that the optimal communication cost calculated in the processing of S17 satisfies the predetermined condition, the calculation core U ends the mapping process (normal end).

On the other hand, if it is determined that the optimal communication cost calculated in the process of S17 does not satisfy the predetermined condition, the calculation core U determines, for example, whether the number of times the process from S12 onwards (i.e., the number of generations) has been executed has reached a predetermined number (S19).

As a result, if it is determined that the number of times the processing from S12 onwards has been executed has not reached a predetermined number of times, the calculation core U, for example, performs the processing from S12 onwards again.

On the other hand, if it is determined that the number of times the processes from S12 onwards have been executed has reached a predetermined number, the computation core U, for example, terminates (abnormally terminates) the mapping of the neuron.

In this way, the AI processor PR in this embodiment has multiple calculation cores, and at least one of the multiple calculation cores executes a mapping process that divides a calculation program associated with each of the multiple neurons included in the diagnostic model MD having a convolutional layer and a fully connected layer and assigns it to each of the multiple calculation cores.

In the AI processor of this embodiment, each of the multiple computing cores executes the computation program assigned by the mapping process.

Specifically, in the mapping process in this embodiment, a computation program is assigned to multiple computing cores using a genetic algorithm so that the communication cost between multiple cores is below a predetermined threshold.

More specifically, when the mapping process is completed successfully, the computation core U generates a mapping table (not shown) that indicates the results of the mapping process. Next, the computation core U downloads the parameters of the diagnostic model MD from the host computer HC. Furthermore, the computation core U refers to the mapping table and transmits the parameters of the diagnostic model MD to each of the computation cores C and F. The computation core U also transmits the mapping table to each router R.

After that, when input data (e.g., an X-ray image of a patient) is input, the computation core U refers to the mapping table and transmits the input data to each of the computation cores corresponding to the neurons included in the first layer. Then, upon completion of the processing corresponding to the first layer, each router R transmits the output data from the first layer to the computation core corresponding to the neuron included in the next layer. Furthermore, each router R repeatedly transmits data to the computation core corresponding to the neuron included in the layer next to the layer being processed until the processing corresponding to the last layer is completed. Then, each router R stores the output data from the last layer (output data of the diagnostic model MD) in DRAM.

In other words, the AI processor PR in this embodiment performs the mapping process by using a genetic algorithm, which is a simpler algorithm than conventional methods and is capable of obtaining results within a predictable time.

As a result, the AI processor PR in this embodiment can reduce communication costs when mapping each neuron to an arithmetic core. Therefore, the AI processor PR in this embodiment can perform diagnosis using a diagnostic model more efficiently.

4: Information processing device 7: Cloud server 8: Real-time information 11: Hospital 12: Hospital 1n: Hospital 51: Mobile terminal 52: Mobile terminal 5n: Mobile terminal 61: Diagnostic system 62: Diagnostic system 6n: Diagnostic system 100: Information processing system MD: Diagnostic model PR: AI processor

Claims (5)

It has multiple computing cores,
At least one of the plurality of arithmetic cores executes a mapping process of dividing a calculation program corresponding to each of a plurality of neurons included in a machine learning model of a CNN (Convolutional Neural Network) having a convolutional layer and a fully connected layer, and assigning the divided calculation program to each of the plurality of arithmetic cores;
Each of the plurality of computing cores executes the computing program assigned by the mapping process;
In the mapping process, the computing programs are assigned to the plurality of computing cores by a genetic algorithm so that a communication cost between the plurality of computing cores is equal to or less than a predetermined threshold value; and
updating parameters in the machine learning model by using global parameters generated from parameters in the machine learning model and parameters in other machine learning models;
the machine learning model is a machine learning model that outputs, when image data including personal information is input, information as to whether or not a person corresponding to the personal information is in a predetermined state;
In the mapping process, a first processor core included in the plurality of processor cores generates a mapping table indicating a result of the mapping process;
In the process of updating the parameters in the machine learning model, the first processor core refers to the mapping table and transmits the parameters in the machine learning model to a plurality of other processor cores among the plurality of processor cores, thereby updating the parameters in the machine learning model .
An AI processor characterized by: In claim 1,
The image data is image data showing a patient,
The machine learning model is a machine learning model that outputs information on whether the patient is in the predetermined state or not in response to input of the image data.
An AI processor characterized by: In claim 1,
The CNN further comprises a pooling layer.
An AI processor characterized by: In claim 1,
In the mapping process,
Randomly generating N mapping solutions for the computing program to the plurality of computing cores;
Calculating the communication cost in a case where each of the N mapping solutions is adopted;
Identifying M mapping solutions in ascending order of the calculated communication costs from the N mapping solutions;
generating N-M new mapping solutions by intersecting the M mapping solutions;
generating mutations in N new mapping solutions including the M mapping solutions and the N-M new mapping solutions;
recalculating the communication cost when each of the N new mapping solutions resulting from the mutation is adopted;
Identifying a particular mapping solution among the N new mapping solutions for which the re-computed communication cost is the smallest;
determining whether the communication cost of the particular mapping solution is less than or equal to the predetermined threshold;
when it is determined that the communication cost of the specific mapping solution is equal to or less than the predetermined threshold, the computing program is allocated to the plurality of computing cores according to the specific mapping solution.
An AI processor characterized by: In claim 4,
In the mapping process,
if it is determined that the communication cost of the specific mapping solution is not equal to or less than the predetermined threshold, the process of calculating the communication cost for the N new mapping solutions, the process of identifying the M mapping solutions, the process of generating the N-M new mapping solutions, the process of causing the mutation, the process of calculating the communication cost again, the process of identifying the specific mapping solution, and the process of determining whether the communication cost is equal to or less than the predetermined threshold are performed again.
An AI processor characterized by:

Images