AI annotation based on signaling interaction

Abstract

An electronic device (400) on an artificial intelligence (AI) platform side, the AI platform side being capable of signaling interaction with an application side to perform at least one AI-informed process having a same application configuration, the electronic device (400) comprising processing circuitry (402) configured to receive a configuration request from the application side for the at least one AI-informed process, the configuration request including information indicative of an application configuration; and in an instance in which the application configuration is satisfied, dispatch the application configuration for use by the at least one AI-informed process at runtime.

Description

TECHNICAL FIELD

[0001] The present disclosure belongs to the field of artificial intelligence, and particularly relates to an artificial intelligence (AI) intelligence injection process. BACKGROUND

[0002] Artificial intelligence (AI) is a new technical science that studies, develops, simulates, extends and expands human intelligence. With the rise of Internet big data, artificial intelligence (AI) has further landed in the fields of security, automobiles, medical treatment, smart homes and intelligent manufacturing, and various open or general AI platforms have emerged as the times require. As a basic and independent computing engine, an AI platform provides intelligent services for application systems. The process in which an AI platform provides intelligent services for application systems is referred to as "intelligence injection" in the industry.

[0003] When an AI platform injects intelligence for an application system, various operations need to be performed. An important link is model training and prediction, and through the training or prediction interaction process, the conversion from "data" to "knowledge" can be realized. At present, the mainstream way to support the training or prediction interaction process of an AI platform is Open API.

[0004] Figure 1 A schematic diagram of an AI platform injecting intelligence for an application system in a model prediction link is shown. Under the Open API process framework, the application system submits to-be-tested data and application scenario information to the AI platform. The AI platform provides the required model and computing resources for the to-be-tested data, performs the corresponding prediction operation, and returns the corresponding recognition result and state.

[0005] In the specific implementation of the intelligence injection process in the model prediction link, the application end initiates a prediction request to the AI platform, the AI platform receives the prediction request from the application end, and performs model validity checking. In the case of a valid model, the AI platform loads the valid model into the memory, and then completes model prediction by using artificially allocated CPU, GPU or other resources to obtain a prediction result. Thereafter, the AI platform returns the prediction result and state to the application end through a prediction response. The application end processes the received prediction result and state, and if the state is normal, the application end analyzes the prediction result and initiates the next prediction. Conversely, if there is an exception, a person will participate in troubleshooting and restore the fault.

[0006] Figure 2 A flow of the current AI platform injecting intelligence for an application system in a model prediction link is shown, which includes the flow of twice face recognition prediction and once number transfer prediction.

[0007] As shown in the figure, in the "face recognition prediction signaling / process (first time)", the application end initiates a face recognition request first; the AI platform receives the face recognition request from the application end, performs model validity verification, model loading operation and resource allocation, then performs face recognition calculation using the model and the allocated resources, and returns the face recognition result and state in the face recognition response.

[0008] The application end parses the recognition state in the received face recognition result, if there is no exception, the next prediction recognition can be initiated in the manner described above, as shown in the figure "face recognition prediction signaling / process (second time)", or "number transfer network prediction signaling / process (first time)". On the contrary, if the application end identifies an abnormal state in the received face recognition result, the artificial participates in troubleshooting and recovers the fault, as shown in the figure "state diagnosis signaling". Then, the next prediction is initiated in the manner described above.

[0009] However, the current AI platform annotation process has a lot of redundant links and manual intervention, resulting in low efficiency. Therefore, how to improve the efficiency of AI platform annotation has become a new challenge for AI platform design, and also a research focus in the AI field.

[0010] Unless otherwise indicated, none of the methods described in this section should be assumed to be prior art merely because of their inclusion in this section. Similarly, unless otherwise indicated, problems recognized in respect of one or more methods should not be assumed to have been recognized in any prior art on the basis of this section. SUMMARY

[0011] The present disclosure proposes improved AI annotation based on signaling transmission, especially AI resource allocation. Through the signaling transmission between the application end and the AI platform, improved application configuration control is realized, the work efficiency is improved, and the manual participation cost in the state diagnosis process is reduced.

[0012] One aspect of the present disclosure relates to an artificial intelligence (AI) platform side electronic device, the AI platform being capable of signaling interaction with an application side to perform at least one AI annotation process having a same application configuration, the AI platform side electronic device comprising processing circuitry configured to receive a configuration request from the application side for the at least one AI annotation process, the configuration request comprising information indicative of the application configuration; and in the event that the application configuration is satisfied, dispatch the application configuration for use by the at least one AI annotation process when running.

[0013] Another aspect of the present disclosure relates to an application-side electronic device capable of signaling interaction with an artificial intelligence (AI) platform-side to perform at least one AI-empowered process with a same application configuration, the application-side electronic device comprising a processing circuitry configured to send a configuration request for the at least one AI-empowered process to the AI platform-side electronic device, the configuration request comprising information indicative of the application configuration; receive a reply information from the AI platform-side electronic device indicating that the application configuration is satisfied, and send an operation request to the AI platform-side electronic device to perform the at least one AI-empowered process without proposing any other configuration request.

[0014] Still another aspect of the present disclosure relates to a method of an artificial intelligence (AI) platform-side capable of signaling interaction with an application-side to perform at least one AI-empowered process with a same application configuration, the method comprising a receiving step of receiving a configuration request for the at least one AI-empowered process from the application-side, the configuration request comprising information indicative of the application configuration; and a dispatching step of dispatching the application configuration for use by the at least one AI-empowered process in running, in case that the application configuration is satisfied.

[0015] Still another aspect of the present disclosure relates to a method of an application-side capable of signaling interaction with an artificial intelligence (AI) platform-side to perform at least one AI-empowered process with a same application configuration, the method comprising a first sending step of sending a configuration request for the at least one AI-empowered process to the AI platform-side electronic device, the configuration request comprising information indicative of the application configuration; a receiving step of receiving a reply information from the AI platform-side electronic device indicating that the application configuration is satisfied, and a second sending step of sending an operation request to the AI platform-side electronic device to perform the at least one AI-empowered process without proposing any other configuration request.

[0016] Still another aspect of the present disclosure relates to a non-transitory computer readable storage medium storing executable instructions that, when executed, implement a method as previously described.

[0017] Yet another aspect of the present disclosure relates to a device. According to one embodiment, the device comprises a processor and a storage device storing executable instructions that, when executed, implement a method as previously described.

[0018] The present disclosure is provided to present a simplified summary of some concepts in order to provide a basic understanding of the disclosure. The disclosure is not intended to identify key or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter. Other aspects and advantages of the technology will become apparent from the detailed description and drawings that follow. BRIEF DESCRIPTION OF DRAWINGS

[0019] The foregoing and other objects and advantages of the present disclosure will be further described and understood by means of the following detailed description and appended claims, in conjunction with the accompanying drawings. In the drawings, like or corresponding elements are denoted by the same or similar reference numbers.

[0020] Figure 1 A basic conceptual diagram of AI platform injecting intelligence for application system is shown.

[0021] Figure 2 The current AI platform intelligence injection process is illustrated by taking the model prediction link as an example.

[0022] Figure 3 The defects in the current AI platform intelligence injection process are illustrated by taking the model prediction link as an example.

[0023] Figure 4A A block diagram of an electronic device on the AI platform side according to an embodiment of the present disclosure is shown, Figure 4B A flowchart of an AI platform side method according to an embodiment of the present disclosure is shown.

[0024] Figure 5 An AI intelligence injection operation process with resource control signaling interaction is shown according to an embodiment of the present disclosure.

[0025] Figure 6 A face recognition process using resource control signaling is shown according to an embodiment of the present disclosure.

[0026] Figure 7 A process of different model prediction scenarios jointly requesting resources is shown according to an embodiment of the present disclosure.

[0027] Figure 8 A schematic diagram of adding state monitoring information to enhance link awareness is shown according to an embodiment of the present disclosure.

[0028] Figure 9A A block diagram of an electronic device on the application side according to an embodiment of the present disclosure is shown, Figure 9B A flowchart of an application side method according to an embodiment of the present disclosure is shown.

[0029] Figure 10 A computer system overview in which embodiments according to the present disclosure can be implemented is shown.

[0030] While the embodiments described in the present disclosure can be susceptible to various modifications and alternative forms, specific embodiments thereof are shown by way of example in the drawings and are herein described in detail. It should be understood, however, that the drawings and detailed description thereto are not intended to limit the embodiments to the particular form disclosed, but on the contrary, the intention is to cover all modifications, equivalents and alternatives falling within the spirit and scope of the claims. DETAILED DESCRIPTION

[0031] Exemplary embodiments of the present disclosure will be described hereinafter with reference to the accompanying drawings. In the description, all specific details of the embodiments are not described in order to provide a clear understanding of the present disclosure. It should be understood that various specific details described in the specification should not be construed as limiting the embodiments, but rather, the intention is to cover all modifications, equivalents and alternatives falling within the spirit and scope of the claims.

[0032] It should also be noted here that, in order to avoid obscuring the present disclosure due to unnecessary details, only the processing steps and / or device structures closely related to the solutions according to the present disclosure are shown in the drawings, and other details not closely related to the present disclosure are omitted.

[0033] As described previously, there are a large number of redundant links and manual interventions in the current AI annotation process, resulting in large operation overhead and low efficiency. The following will refer to Figure 3 to describe the defects existing in the current AI annotation process in detail. Figure 3 The defects existing in the current AI annotation process are illustrated by taking the model prediction link as an example, mainly including the following three defects.

[0034] 1. Large operation overhead and long response time

[0035] In the current AI annotation process, model checking and loading are required before each calculation process, that is, the effectiveness of the model is detected, such as whether optimization update should occur, and the model is loaded, as shown in ① of Figure 3 . Frequent checking and loading operations consume a large amount of time and memory resources, which increases the operation overhead of the system and results in long response time.

[0036] 2. Static resource allocation, lack of adaptation mechanism

[0037] In the current AI-assisted intelligence process, the resources used for computational operations are set and allocated "statically" in the background. That is, the resources used are manually configured and remain fixed. Even if other available hardware resources exist in the resource pool, resource allocation cannot be performed in real-time or flexibly. Figure 3 As shown in ②.

[0038] For example, in face recognition model prediction, theoretically, CPU or GPU resources can be selected. However, if the background configuration is set to CPU, the background configuration will remain unchanged during the face recognition model prediction process, and neither the application system nor the AI ​​platform can freely choose to use GPU for computation.

[0039] 3. Weak perception capabilities and increased labor costs.

[0040] Currently, mainstream centralized AI platforms primarily offer two types of status mechanisms for applications: those using network status codes and those providing custom business status codes. However, each has its own problems:

[0041] 1) Using network status codes (such as 400, 401, etc.) fails to accurately reflect business logic and hinders the localization of code problems;

[0042] 2) Custom business status codes (e.g., 1111 for "insufficient resources", 1112 for "data error"), but lack control over error status indication.

[0043] These issues all lead to excessive human intervention in the problem diagnosis process during AI injection, significantly reducing the efficiency of AI prediction or training. Figure 3 As shown in ③.

[0044] As application scenarios become more complex, the requirements for AI-encoded services also increase, especially in terms of response speed, resource adaptation, and fault diagnosis. Therefore, given the significant performance-impacting defects in current AI platform encoded services, this paper discloses an improved AI platform encoded service based on signaling transmission, particularly an improved application-specific AI-encoded configuration.

[0045] According to this disclosure, AI-enhanced intelligence or platform-enhanced intelligence can refer to providing intelligence to the application that the application system is to execute, that is, intelligence enhancement for the application. This application can include facial recognition, number portability, and other types of applications. For each application, the AI ​​platform-enhanced intelligence process according to this disclosure includes, but is not limited to, the application model training process (e.g., training the application model for the application that the application system is to execute), the application model utilization process (using the trained or pre-defined application model to execute the application, such as the aforementioned model prediction process), and so on.

[0046] The signaling transmission of the AI injection process according to the present disclosure can refer to the signaling sending and / or receiving between the application end and the AI platform, for correlating the application model, the application required resources, the application state information / code, etc. with each other, thereby providing a convenient and efficient injection service. Preferably, the signaling transmission of the AI injection process according to the present disclosure can be the signaling transmission of the information required for the application configuration when the AI injection process is performed for the application. The required configuration can refer to the configuration / setting of the elements required for performing the AI injection process, such as the model, the resources, and other elements necessary for the execution. The information can be referred to as application configuration information, and can include information related to the configuration for the foregoing application model training process, application model utilization process, etc., such as resource configuration, model configuration, etc. Of course, the information can also be other configuration related information required when the AI injection process is performed for the application.

[0047] The signaling transmission of the AI injection process according to the present disclosure can refer to various forms of signaling transmission between the application end and the AI platform. As one example, it can include one-way signaling transmission between the application end and the AI platform. For example, the application end sends a request signal to the AI platform, the AI platform sends a response signal to the application end or a third party, etc. As another example, it can include two-way signaling transmission between the application end and the AI platform, for example, the application end sends a request signal to the AI platform and the AI platform sends corresponding response information to the application end. In this case, the signaling transmission can be referred to as signaling interaction. As an example, the signals mentioned here include application configuration information, application operation related information, system state related information, etc., which will be described in detail below in conjunction with embodiments.

[0048] According to one aspect of the present disclosure, the configuration signaling transmission between the application end and the AI platform can optimize at least one AI injection process with the same application configuration performed for the application, wherein the application configuration can at least include the resource configuration required for performing the AI injection process, and additionally, can also include the model configuration required for performing the AI injection process. In particular, the AI platform pre-meets the configuration required for performing the at least one AI injection process according to the configuration request from the application end, so that the configuration can be fixedly used in each AI injection process, thereby saving the existing redundant operations such as model checking / loading and resource binding required for each AI injection process in the prior art, saving work overhead, and improving efficiency.

[0049] According to another aspect of the present disclosure, configuration signaling transmission between the application side and the AI platform can dynamically adjust the configuration required in the AI intelligent process performed for the application. As an example, the dynamic adjustment can at least involve dynamic adjustment of resources. In particular, the AI platform can dynamically allocate resources for the application side according to specific parameters in the configuration request from the application side, or inform the application side of the available resource status in the current resource pool through the reply information, so that the application side can determine whether to adjust the resource configuration, thereby achieving optimized resource allocation.

[0050] According to still another aspect of the present disclosure, the AI platform can report state monitoring information to the application side through signaling transmission with the application side, the state monitoring information indicating the current state in a coded field, so that the application side can automatically identify the code of the field in the state monitoring information, determine the state type, and perform corresponding processing operations, thereby achieving automatic processing in the state diagnosis process, reducing manual participation costs, and improving the efficiency of the intelligent process.

[0051] Embodiments according to the present disclosure will be described in detail below with reference to the accompanying drawings. It should be noted that the operations performed by the AI platform side according to the embodiments of the present disclosure can be implemented by an electronic device for the AI platform side, which can be separate from the AI platform, can be part of the AI platform, such as components, parts, etc., or can be for the AI platform itself. The operations performed by the application side according to the embodiments of the present disclosure can be implemented by an electronic device for the application side. The electronic device can be separate from the application side, can be part of the application side, such as components, parts, etc., or can be for the application side itself. Therefore, the operations performed by the AI platform side and the application side mentioned in the following description can be executed by the electronic device for the AI platform side and the electronic device for the application side, respectively.

[0052] Figure 4A A block diagram of an electronic device for the AI platform side according to embodiments of the present disclosure is shown. The AI platform can interact with the application side to perform at least one AI intelligent process with the same application configuration. The electronic device 400 includes processing circuitry 420, and the processing circuitry 420 can be configured to receive a configuration request for the at least one AI intelligent process from the application side, the configuration request including information indicating an application configuration; and provide the application configuration for use when the at least one AI intelligent process is run, if the application configuration can be met (e.g., the elements indicated or corresponding by the application configuration are available).

[0053] According to an embodiment, the application configuration can comprise at least resource configuration, thus the information indicating the application configuration can indicate the resource required for performing the at least one AI intelligence process. Thus, in case that the requested resource is available, the resource can be allocated to the at least one AI intelligence process, so that in subsequent operations, the at least one AI intelligence process can be executed fixedly using the allocated resource.

[0054] The information indicating the resource configuration can be provided in various ways. As an example, the information indicating the resource configuration can comprise at least the type of resource required for performing the AI intelligence process, and can further comprise the number of resources and / or the number of resources, which can be in the form of a list or any other appropriate form. For example, the type of resource can comprise physical resources or virtual resources such as CPU, GPU, SPARK, memory, etc. that can be used for application model training and utilization. As an example, determining whether the resource configuration is satisfied can comprise determining whether the resource to be allocated is currently idle in the resource pool, and when the resource is idle, it can be determined that the resource configuration is satisfied. As an example, when the resource configuration indicates that N CPUs are required, or the Nth CPU is required, and there are N CPUs that are idle in the resource pool, or the Nth CPU is idle, it can be considered that the resource to be allocated is available.

[0055] According to an embodiment, the application configuration can further comprise model configuration, thus the information indicating the application configuration can indicate the model adopted for performing the at least one AI intelligence process, and in case that the requested model is available, the model can be loaded into the resource allocated to the at least one AI intelligence process, so that in subsequent operations, the at least one AI intelligence process can be executed fixedly using the loaded model and resource. The resource mentioned here can be the same as or different from the application resource mentioned above, and generally refers to storage resource, such as memory.

[0056] The information indicating the model configuration can be provided in various ways. As an example, the information indicating the model configuration can comprise at least the type of model (such as a model for face recognition, a model for number transfer, etc.), the ID of the model, etc. As another example, in case that the model ID corresponds to the type of model, only the model ID can be included, so that the type of model can be directly derived from the model ID. As an example, determining whether the model is available comprises checking the model, for example, in case that the model is retrieved in the model library and the state of the model is stable, it is considered that the model is valid and available. The state of the model being stable can mean that the model has been the latest, and does not need to be optimized or updated. Of course, the validity of the model can also be indicated by other indicators known in the art, which will not be described in detail here.

[0057] According to another embodiment, additionally or alternatively, the configuration request can comprise at least information indicative of an application scenario of the AI inference process. As an example, the information can comprise at least one of an application scenario name, an application scenario ID, an application scenario parameter, etc. According to an embodiment, an application scenario can be pre-associated with a resource configuration for the AI inference process. For example, a certain application scenario corresponds to a certain resource configuration. Thus, in case the application scenario is indicated in the configuration request, the AI platform can identify the corresponding resource requirement and allocate resources accordingly, it is to be noted that in this case, the information of the application scenario can be used as the information indicative of the resource configuration, thus the information indicative of the resource configuration as described above can be omitted. According to another embodiment, an application scenario can be pre-associated with a certain model configuration. For example, an application scenario can correspond to a certain model. Thus, in case the application scenario is indicated in the configuration request, the AI platform can identify the corresponding model requirement and allocate models (e.g. check, bind, etc.) accordingly, it is to be noted that in this case, the information of the application scenario can be used as the information indicative of the model configuration, thus the information indicative of the model configuration as described above can be omitted.

[0058] According to an embodiment, the same application configuration can refer to the same configuration required in the execution of the at least one AI inference process, such as the same resource configuration, the same model configuration, etc. As an example, the at least one AI inference process with the same application configuration can refer to AI inference processes executed at the same stage for the same application scenario, such as a face recognition model training process, a face recognition model utilization process (i.e. for face recognition, corresponding to the aforementioned face recognition prediction process), a number portability model training process, a number portability model utilization process (corresponding to the aforementioned number portability prediction process), etc. However, it is to be noted that the at least one AI inference process with the same application configuration can also correspond to at least one of different application scenarios or different stages of the same application scenario, as long as the resource and / or model configuration required in the execution of these AI inference processes are the same.

[0059] It is to be noted that the transmission of the application configuration request between the application side and the AI platform as described above can be performed before the start of the at least one AI inference process. Thus, in case the requested configuration can be satisfied, the configuration is allocated to the at least one AI inference process for fixed use of the configuration by each AI inference process. For example, the requested resource can be fixedly used for the at least one AI inference process for application model training, the requested resource and model can be fixedly used for the at least one AI inference process for application model utilization / prediction.

[0060] According to one embodiment, the AI ​​platform can provide the application with a configuration result in response to the application configuration request. As an example, this configuration result can be included in the response information sent by the AI ​​platform to the application, which corresponds to the aforementioned configuration request. In this case, the configuration request and response are part of the configuration signaling interaction. According to an embodiment, the configuration result, in addition to indicating whether the requested configuration is satisfied, can also indicate resource configuration information. According to one embodiment, the processing circuit is further configured to provide the application with information indicating that the requested application configuration is satisfied when the requested application configuration is satisfied. This allows the application to request the at least one AI injection process after receiving this information, without needing to submit a configuration request for each AI injection process. According to another embodiment, if the application does not receive a response from the AI ​​platform for the application configuration request within a specific time period, the application will repeatedly send the configuration request to the AI ​​platform until it receives the application configuration result from the AI ​​platform.

[0061] According to one embodiment, the processing circuit is further configured to, upon receiving a request from the application to execute the at least one AI injection process, directly utilize the application configuration already allocated to the at least one AI injection process to execute the at least one AI injection process, and then feed the execution result back to the application. It should be noted that the request no longer includes requests regarding application configuration (such as resource configuration, model configuration, etc.). For example, after the resource and model configurations for the at least one face recognition process have been allocated, the application sends a request for the at least one face recognition process to the AI ​​platform, and the AI ​​platform can directly apply the allocated resources and the model loaded into the resources to perform face recognition.

[0062] According to one embodiment, the processing circuit is further configured to receive an application resource release request from the application terminal for the at least one AI intelligence injection process, release the resources used for the at least one AI intelligence injection process, and inform the application terminal of the release result as a response. Thus, the application terminal can perform signaling interaction with the AI ​​platform to release currently used resources, thereby freeing up resources for subsequent application operations.

[0063] The following will refer to Figure 5 This section exemplarily illustrates an AI-assisted intelligence injection operation process incorporating resource control signaling interaction according to embodiments of the present disclosure, wherein... Figure 5 The diagram illustrates the facial recognition process in an AI-assisted intelligence-gathering operation. Here, the resource control signaling is an example of the aforementioned configuration signaling and is specifically the resource allocation signaling used in the facial recognition intelligence-gathering process. This resource control signaling includes facial recognition resource activation signaling and facial recognition resource release signaling, and each signaling contains a request and response process.

[0064] exist Figure 5 In this context, "A-1" corresponds to the resource activation signaling interaction preceding the face recognition prediction signaling interaction. This request allocates / activates resources before face recognition, allowing the "face recognition prediction" model to be loaded into and bound to the required resources. Consequently, subsequent face recognition prediction processes can directly utilize the bound resources and model for computation, eliminating the need for model verification and loading at each stage, thus reducing redundant operations. The A-1 resource activation signaling interaction includes a face recognition resource activation request and a face recognition resource activation response.

[0065] In addition, "A-2" corresponds to the face recognition resource release signaling interaction, including a face recognition resource release request and a face recognition resource release response. When face recognition prediction is no longer needed, the "face recognition prediction" model is unbound from the corresponding resource and the resource is returned to the resource pool for other needs to call.

[0066] The following will be referenced Figure 6 Taking the face recognition prediction stage as an example, this paper describes in detail the AI-injected configuration and operation process according to the embodiments of this disclosure.

[0067] First, the A-1 process is performed, involving a "face recognition resource activation signaling" interaction between the application and the AI ​​platform. The application initiates a face recognition resource activation request to the AI ​​platform, which includes AI model information and AI resource configuration information. Upon receiving this request, the AI ​​platform acquires resources based on the AI ​​resource configuration information, verifies and loads the model according to the AI ​​model information, and binds the model and resources. Then, the AI ​​platform sends a response ("face recognition resource activation response") to the application. In this example, the response includes information indicating that both the model and resources are available. Here, the AI ​​model information and AI resource configuration information can be examples of the information indicating model configuration and resource configuration described above, respectively; a more detailed illustrative explanation follows.

[0068] Next, a "face recognition prediction signaling" interaction is performed between the application and the AI ​​platform to enable face recognition (as an example application). This involves n face recognition operations, indicated by B-1, B-2, ..., Bn, where n is an integer greater than or equal to 1. The operations in each face recognition operation are essentially the same.

[0069] First, according to the received indication that both the model and the resource are available, the application initiates a face recognition request. The AI platform receives the request and directly uses the bound resource and model to perform the face recognition operation. The AI platform uses the face recognition response to inform the application of the face recognition prediction result, which includes the recognition result and possible status information. Thus, the first face recognition prediction signaling interaction is completed, i.e., the first face recognition is performed. After the application receives the recognition prediction result, it parses the result. If the result is normal, the application performs the next face recognition prediction signaling interaction in the manner described above, e.g., B-2. If the application finds an abnormality in the parsed result, it goes to a "status diagnosis process" (not shown), in which the problem is diagnosed based on the abnormality information and the fault is recovered, and the prediction signaling interaction can be requested again after the fault is recovered.

[0070] Thus, n times of face recognition prediction signaling interactions are sequentially performed until the face recognition is completed (e.g., the nth face recognition prediction signaling interaction B-n is performed). Then, the A-2 release process is performed, in which "face recognition resource release signaling" interaction is performed between the application and the AI platform. The application initiates a face recognition resource release request to the AI platform. The AI platform receives the request and unbinds the "face recognition prediction" model from the corresponding resource, and puts the resource back into the resource pool for other needs to call. Then, the AI platform informs the application of the release result through a face recognition resource release response.

[0071] As can be seen from the above, by adding the "A-1" configuration control process, the primary configuration allocation / setting for subsequent face recognition operations is performed, e.g., the model and resource are bound for the face recognition process according to the AI model information and the AI resource configuration information, so that in the case where the AI model information and the AI resource configuration information for at least one face recognition prediction operation remain unchanged, the application system can directly send an operation request to the AI platform to initiate AI intelligence for each face recognition prediction without sending any configuration activation request, and the AI platform no longer performs the checking and loading operation, but directly performs the calculation link to complete the training or prediction of the model. It should be pointed out that the above A-1 configuration process and A-2 release process can also be applied to other types of applications, such as number portability applications, and are particularly advantageous for two or more applications with the same configuration requirements.

[0072] Therefore, an AI configuration process based on configuration signaling transmission is proposed, in which a configuration request is sent from the application end to the AI platform before the start of at least one AI annotation process execution with the same application configuration, so that the configuration required by the at least one AI annotation process can be set / assigned in advance before the at least one AI annotation process execution, and the configuration can be fixedly used in each AI annotation process, thereby eliminating the redundant operations such as redundant model checking and loading that need to be performed for each AI annotation process in the prior art, solving the problems of computing overhead, resource allocation and link awareness in the existing AI platform annotation process, and thereby improving the efficiency of AI annotation resources.

[0073] It should be noted that the application configuration process according to the present disclosure is described above in combination with the signaling interaction process. However, this is only exemplary, and embodiments according to the present disclosure can also be implemented in combination with one-way signaling transmission.

[0074] The foregoing mentioned that the AI platform feeds back the configuration result to the application end through a response, but this is not necessary. According to an embodiment, the AI platform can not make any response regarding the configuration status. In the case that the configuration request from the application end is satisfied, the AI platform binds the requested application configuration such as model and resource, and temporarily stores the information of the satisfied configuration request until the resource is released. The application end sends a configuration request in each subsequent application operation, and the AI platform can verify whether the configuration information in the configuration request is consistent with the configuration information in the previously temporarily stored configuration request after receiving the configuration request from the application end, and if consistent, directly utilizes the already bound application configuration such as model and resource for processing, and can also eliminate the operations related to checking and loading. In this way, the work overhead can also be saved to a certain extent, and the efficiency is improved. If not consistent, the AI platform can release the resource, and then perform configuration assignment based on the new application configuration information.

[0075] The foregoing mentions that the resource release signaling interaction is initiated by the application end, but this is not necessary. According to an embodiment, in the case that no application operation request is received from the application end within a certain time interval (which can be referred to as a first time interval), the AI platform can actively release the resource, and inform the application end of the release result. According to another embodiment, in the case that no application operation request is received from the application end within a certain time interval, the AI platform can send inquiry information to the application end to inquire about the subsequent operation, such as continuing similar application operation, performing a new operation, or releasing the resource to wait for a subsequent request, and upon receiving feedback from the application end, perform the corresponding operation according to the feedback message subsequently received from the application end. Further, if no feedback message is received within a second certain time interval after sending the inquiry information, the resource release can be actively performed, and the release result is informed to the application end. The first time interval and the second time interval can be set in various ways, and can be equal to or different from each other. This can help to some extent to efficiently utilize the resource. For example, if information cannot be received due to communication failure, the resource can be temporarily released for use by other application processes, without the resource being occupied for a long time. It should be pointed out that the above operations can be performed by the processing circuit of the electronic device on the AI platform side.

[0076] Dynamic resource allocation

[0077] In the prior art, once the resource is initially set by human, it will basically remain fixed unless changed by human, so if the resource request of the application end cannot be met, the application end can only wait until the available resource in the resource pool exactly meets the demand in the resource request of the application end, which will cause resource vacancy and waste of resources. According to the present disclosure, a dynamic resource allocation mechanism is proposed, which can realize dynamic allocation of resources by means of configuration signaling transmission between the application end and the AI platform, which is particularly effective in the case that the resource indicated by the configuration information sent by the application end is unavailable.

[0078] According to one embodiment, the dynamic allocation of resources can be achieved through the configuration signaling interaction between the application end and the AI platform. In particular, in the case that the resources indicated in the configuration information in the configuration request sent by the application end are unavailable, the AI platform can feed back information indicating the currently available resources in the resource pool to the application end, so that the application end can dynamically adjust its resource configuration. In particular, the application end can decide, according to the feedback information, whether to continue waiting for resources in the resource pool until the configuration request can be satisfied, or to adjust its resource configuration (such as resource type, resource quantity, resource number, etc.) to try to use the available resources in the resource pool. In this way, dynamic configuration of resources can be achieved. As an example, the feedback information can be contained in the resource activation response sent to the application end. In this way, dynamic allocation of AI resources can be achieved to some extent, and the utilization efficiency of resources is improved.

[0079] According to another embodiment, dynamic allocation of resources can also be achieved in the case of one-way configuration signaling transmission between the application end and the AI platform. In particular, the AI platform can achieve dynamic allocation of resources based on specific information related to resource configuration adjustment contained in the resource allocation request from the application end.

[0080] According to one embodiment, the specific information related to resource configuration adjustment can include resource adjustment strategy information, and the processing circuit is configured to make an adjusted resource configuration according to the resource adjustment strategy information in the resource configuration information in the case that the resource configuration indicated in the resource configuration request is unavailable. Here, the resource configuration determined to be unavailable can be referred to as a preliminary resource configuration, which can refer to an initially set resource configuration, or a resource configuration set by the application end to achieve optimal performance, and other configurations.

[0081] According to one embodiment, the resource adjustment strategy information can indicate at least one strategy / way for adjusting resources, such as strategy / way name, ID, etc., represented by any of characters, codes, or numerical values that can be distinguished from each other, etc. As an example, the strategy / way can have a corresponding candidate resource configuration, which can be represented in the same way as the aforementioned resource configuration, for example, candidate resource type, candidate resource quantity, candidate resource number required for AI wisdom process execution, and in this case, additionally or alternatively, the resource adjustment strategy information can also directly indicate the candidate resource configuration. As still another example, the resource adjustment strategy information can also be set to a default value corresponding to no dynamic resource adjustment or not be set to indicate no resource adjustment. Therefore, the AI platform can perform corresponding operations related to resource adjustment according to the setting condition of the resource adjustment strategy information.

[0082] According to one implementation, the resource adjustment policy information can be configured to indicate one resource adjustment policy / way, such that upon confirming that the preliminary resource configuration requirement is not satisfied, the AI platform continues to determine whether a candidate resource configuration corresponding to the resource adjustment policy / way can be satisfied, e.g., whether the resource type, resource quantity, resource number indicated in the candidate resource configuration is available. If it is determined that the candidate resource configuration is satisfied, the resources indicated in the candidate resource configuration are allocated to the AI intelligent process, and the allocation result is informed to the application end. If it is not available, the application end can be informed that the candidate configuration is not available, or the current available resources in the resource pool are informed, so that the application end can determine whether to wait or adjust, as described above. In particular, as an example, one policy / way can correspond to directly using all available resources in the current resource pool, which also belongs to a special candidate resource configuration, and for this policy / way, if there are available resources in the resource pool, the AI platform can directly allocate all available resources in the resource pool, and inform the application end of the successful allocation result.

[0083] According to another implementation, the resource adjustment policy information can be configured to indicate more than one resource adjustment policy / way, such that in this case, the preliminary resource configuration and more than one candidate resource configuration corresponding to the configured resource adjustment policy / way can be implicitly included in the resource configuration request. When it is confirmed that the preliminary configuration resource requirement is not satisfied, the AI platform continues to determine whether any one of the more than one candidate resource configurations is satisfied. The AI platform can perform the determination in various ways. As one example, the more than one resource adjustment policy / way can be sorted in descending order according to the performance of the AI intelligent process resulting therefrom from high to low, and the AI platform can determine in this order, e.g., the higher the resulting performance, the earlier the corresponding candidate resource determination is determined. As another example, the resource adjustment policy / way can be sorted in descending order according to the required resource quantity (e.g., the total resource quantity, the quantity of a specific type of resource, etc.), and the AI platform can determine in this order. If it is determined that the candidate resource configuration is available, the resources indicated in the candidate resource configuration are allocated to the AI intelligent process and the allocation result is informed to the application end. If all candidate resource configurations are not available, the application end can be informed that the candidate configuration is not available, or the current available resources in the resource pool are informed, so that the application end can determine whether to wait or further adjust, as described above. In this way, the resource utilization efficiency can be further appropriately improved, and an effective compromise between execution performance and resource utilization efficiency can be achieved.

[0084] According to an embodiment, the specific information related to resource configuration adjustment can also include resource adjustment permission information, which can be set to a specific value to indicate whether resource adjustment is permitted. As an example, the resource adjustment permission information can be represented by a binary value, where 1 indicates true, i.e. resource adjustment is permitted, and 0 indicates false, i.e. resource adjustment is not permitted, or vice versa. In addition, the resource adjustment permission information can also be represented in other ways, such as a decimal or other numeral value, different characters, etc., as long as it can be distinguished between permitting and not permitting resource adjustment. Thus, in this case, resource adjustment is only performed when the resource adjustment permission information is set to permit resource adjustment. In one implementation, the resource adjustment strategy information can only be set when the resource adjustment permission information is set to permit.

[0085] According to another embodiment, the resource allocation request can include a priority parameter, which indicates the priority of the AI intelligence process or the application for which the request is issued, and the AI platform can also determine resource allocation according to the priority parameter, or even adjust resource allocation. To some extent, the priority parameter can also belong to the specific information related to resource allocation adjustment. In particular, in the case where there is more than one AI intelligence process with different resource configuration requirements, the AI platform can perform resource allocation in descending order of priority. For example, the AI platform first performs resource allocation for the application with the highest priority, that is, first determines whether the resource requirement of the application is met. After the resource requirement of the high-priority application cannot be met, the low-priority application is considered.

[0086] According to an embodiment, when at least two application resource activation requests are received at the same time, the AI platform can perform resource allocation for the applications in descending order of priority, so that the utilization efficiency of resources can be appropriately improved, and the performance of the applications can be improved to some extent. As an example, resources are first allocated to high-priority applications, and then resources are allocated to low-priority applications when the high-priority applications are not met, so that the execution efficiency of the applications and the utilization efficiency of the resources can be improved. It should be pointed out that the high-priority application being met here can refer to the main resource requirement of the high-priority application being met, or can refer to the resource requirement being met after resource adjustment as described above. As another example, the resource requirement of the low-priority application is considered only when the resource requirement of the high-priority application cannot be met.

[0087] According to another embodiment, when a configuration request for another set of one or more application operations is received while a current set of one or more application operations (AI annotation process) is being executed, it can be determined whether the priority indicated in the configuration request for the other set of application operations is higher than the priority of the previously applied, and if the priority indicated in the configuration request for the other set of application operations is higher than the priority of the previously applied, the other set of application operations is prioritized for resource allocation. As an example, even if the current set of application operations is not completed, their corresponding resources are temporarily released, and the resources are allocated for another set of application operations with higher priority. After the other set of application operations with higher priority is completed, the resource allocation is resumed according to the configuration for the current set of application operations, and the remaining application operations are executed. As another example, when the resources are temporarily released, the application end can be informed of the temporary release information. According to another embodiment, when configuration requests for one or more application operations of one or more additional sets are received while a current set of application operations (AI annotation process) is being executed, an application operation queue can be established, in which the configuration requests can be ordered in time sequence, or according to priority, and after the current set of application operations is completed, the individual configuration requests and corresponding subsequent application operations can be executed in the ordered sequence.

[0088] Examples of dynamic configuration of AI resources according to embodiments of the present disclosure will be described below in conjunction with Figure 7 Figure 7 A model prediction is used as an example to illustrate dynamic configuration when different application scenarios request the same resource pool at the same time, in which face recognition and NLP (natural language processing) request CPU resources (the number of CPUs is set to 6) in the resource pool at the same time. Here, the resource activation request can correspond to the aforementioned application configuration request, and the specific information related to resource configuration adjustment is set as a minimum configuration strategy, which in one example can correspond to the minimum resource requirement required for AI annotation process execution, and is included in the AI resource configuration information in the resource activation request. In another example, the minimum configuration strategy can refer to all remaining resources regardless of their number can be used. Therefore, when resources need to be shared, dynamic allocation can be made according to the AI resource configuration information of the activation request information, thereby achieving dynamic resource allocation.

[0089] ​In operation, first, a resource allocation request is initiated for the face recognition application scenario through the configuration signaling transmission between the application end and the AI platform. The application end sends a resource activation request to the AI platform, as "C-1" in the figure, in which the CPU resource is requested, the number is 4, and the minimum configuration strategy is adopted. The AI platform receives the request and gives a successful response, indicating that 4 CPUs have been allocated for face recognition. At this time, there are 2 CPUs left in the resource pool. Then, the application end sends a resource allocation request to the AI platform for the NLP application scenario, as "C-2" in the figure, in which the CPU resource is requested, the number is 4, and the minimum configuration strategy is adopted, and the AI platform receives the request and then performs resource allocation. Specifically, although the number of CPU resources in the resource pool is less than 4 at this time, according to the minimum configuration strategy, a successful response is given under the condition that 2 CPUs can meet the minimum resource requirement, indicating that 2 CPUs are allocated; at this time, the CPU resources in the resource pool are dynamically configured.

[0090] The above is only exemplary. For example, if no specific information related to resource configuration adjustment is set or the minimum resource requirement is still not met, the AI platform can inform the application end of the current state of the resource pool. For example, in the above example, the application end can be informed that only two CPUs are available in the current resource pool, so the application end can determine whether to adopt the minimum configuration strategy, or adopt the allocation strategy that all remaining resources can be used regardless of their number, or wait until 4 CPUs are available.

[0091] It should be pointed out that the order of first performing face recognition and then performing NLP here is exemplary, and other execution modes are also feasible. In particular, the execution order of multiple AI intelligent processes can depend on the types of AI intelligent processes, their causal relationship, priority, and the like.

[0092] By comparing the traditional intelligent process and the signaling interaction of the present application, it can be seen that the CPU / GPU resources in the prior art process are pre-configured by humans based on experience, which belongs to "static" rules. On the contrary, the signaling interaction of the present application is a "dynamic" rule, which can dynamically allocate resources according to the demand, thereby realizing efficient utilization of system resources and facilitating smooth execution of system operations.

[0093] State monitoring

[0094] State monitoring is a means for evaluating whether an application system is stable. The level of detail of state monitoring information not only affects problem positioning, but also affects the subsequent action decisions of the application end or third parties of the system. In addition, the expression method of state monitoring information also affects the efficiency of the system in state monitoring and processing. According to an embodiment of the present disclosure, an improved state monitoring information is proposed, and a state monitoring metric framework is designed to improve the automation in the state diagnosis process, thereby reducing the cost of manual participation and improving the efficiency of the intelligent process. According to an embodiment of the present disclosure, the state monitoring information can particularly include coded information related to the AI intelligent condition when an abnormal state occurs, which can include at least one of an abnormal type, an application scenario where the abnormality occurs, an operation link where the abnormality occurs, etc. According to another embodiment, the state monitoring information can also include information indicating a subsequent operation to be performed. Therefore, when an abnormal state occurs, the problem can be quickly located and the fault can be accurately eliminated, making the resource allocation process more efficient.

[0095] The implementation of state monitoring according to the present disclosure will be described below. The present disclosure designs a binary state metric framework (BSMF), which measures different monitoring objects in the form of "binary bits", reduces the byte length of a single communication, and improves the network transmission efficiency. The state monitoring field is introduced to make up for the lack of indication flags in the traditional state mechanism, which reduces the cost of manual participation to a certain extent and further improves the automatic processing performance. Of course, it should be pointed out that this binary code is only exemplary, and various monitoring fields can also be provided in other ways, such as decimal code, symbol, string, etc. That is, the present disclosure arranges the format of various monitoring fields so that different codes can be used to distinguish different monitoring states, so that different monitoring states can be automatically recognized by machines.

[0096] As an example, the present disclosure divides the 32-bit / 32-bit unsigned integer state monitoring information into at least: type code, reserved code, business code, milestone, control code, and state code. The following table shows an exemplary state metric framework.

[0097] Table 1: State metric framework

[0098]

[0099] As shown in the above table, the basic description of the state metric framework is as follows:

[0100] Type code: 4 bits (bit), maximum 16 bits, used to define the state category. For example: permission category, parameter category, resource category, etc. In the resource control link, this type is fixed as the resource category;

[0101] Reserved code: 3 bits, max 8 bits, for system extension. Default: 000;

[0102] Business code: 6 bits, max 64 bits, for defining application scenario framework. E.g. classification, clustering, image analysis, etc.

[0103] Milestone: 5 bits, max 32 bits, for defining logical flow. E.g. data extraction, pre-processing, training, etc.

[0104] Control code: 4 bits, max 16 bits, for defining status indication flag. E.g. check, retransmission, waiting, etc.

[0105] Status code: 10 bits, max 1024 bits, for defining abnormal status. E.g. password error, resource busy, etc.

[0106] The type code, business code, and milestone described above can be examples of the AI intelligence-related information described above, which aims to clearly and intuitively indicate the abnormal type, the scene where the abnormality occurs, the process stage, etc., and the control code can indicate the operation to be performed subsequently. It should be pointed out that the AI intelligence-related information can also be represented in other ways, as long as it can help to quickly and accurately locate the abnormality when the abnormality occurs and clearly indicate the operation to be performed. Examples of resource class status monitoring feedback information will be described below.

[0107] Table 2: Example of resource class status monitoring feedback information

[0108]

[0109] Wherein, the "definition" row in the table shows the possible values of each field and the corresponding meaning, and the "example" row gives an example of the binary code of the resource class status monitoring feedback information obtained by the combination of the field values, the "decimal" column gives the decimal value corresponding to the binary code of each example respectively, and the "description" column gives the abnormal status information indicated by each binary code example.

[0110] In one example, when an exception occurs, the type field (4 bits) is set to 0011, which indicates that the state is a resource class; the reservation code field (3 bits) is set to 000; the service field (6 bits) is valued as 000101, which indicates that the application scenario is image recognition; the milestone field (5 bits) is set to 00011, which indicates that the logical process is a training process; the control field (4 bits) is set to 0001, which indicates that the application end performs verification; and the state field (10 bits) is set to 0000001000, which indicates that the exception state is a parameter type exception. For this example, the binary code automatically generated by the system, especially the AI platform side, is 00110000001010001100010000001000, and the corresponding decimal number is: 807977992. The code describes a system resource class exception of a parameter type in a picture recognition scenario in a training process, and indicates that the client performs a verification operation. As can be seen from the above, by identifying the coded information indicating the exception state, the type, occurrence scenario, and operation link of the exception can be quickly and accurately judged, and subsequent operations can also be performed according to the information, so that the problem can be quickly located, the fault can be accurately excluded, and the AI resource configuration process and the AI wisdom injection process can be more efficient.

[0111] According to embodiments of the present disclosure, the state monitoring information is provided by the AI platform side to the application end or a third party for diagnosing the application system state. The third party may, for example, refer to a device capable of detecting / diagnosing the application system state. According to embodiments of the present disclosure, the processing of the response information is further optimized by adding the state monitoring information in the response information sent by the AI platform to the application end. The state monitoring information can be included in various response information sent by the AI platform to the application end, such as resource activation response information, prediction recognition response information, resource release response information, and the like.

[0112] The following will be described with reference to Figure 8 Examples of applying state monitoring information in resource configuration signaling interaction according to the present disclosure are described. In the operation process, first, the application end initiates a resource activation request or a resource release request, and the AI platform responds to the request, and the response at least includes state monitoring information and feedback resource state execution effect. As indicated by "D" in the figure; then, the application end receives the response and parses the result and the state, and if there is no exception, the request is initiated again. If there is an exception, the system will automatically perform fault identification based on the state monitoring information, obtain a subsequent operation strategy, perform fault recovery, and initiate the request again, thereby reducing the frequency of manual participation.

[0113] It should be noted that the state monitoring information can also be included in other response information fed back by the AI platform to the application end. According to another embodiment, the sending of the state monitoring information can be triggered by the AI platform side in response to the change of the application system state. As an example, once a state change occurs, the state monitoring information indicating the state change can be sent by the AI platform side to the application end or a third party.

[0114] As can be seen from the above, by designing a state monitoring metric framework to construct the state monitoring information, the application end can quickly locate and exclude the exception based on the state monitoring information, making the resource allocation process more efficient.

[0115] In the structural example of the electronic device, the processing circuit 420 can be in the form of a general-purpose processor or a special-purpose processor such as an ASIC. For example, the processing circuit 120 can be constructed by a circuit (hardware) or a central processing device such as a central processing unit (CPU). In addition, the processing circuit 420 can carry a program (software) for making the circuit (hardware) or the central processing device work. The program can be stored in a memory such as arranged in the memory or an external storage medium connected from the outside, and downloaded via a network such as the Internet.

[0116] According to an embodiment of the present disclosure, the processing circuit 420 can include various units for implementing the above functions, such as a receiving unit 422 that receives a configuration request for at least one AI injection process from an application-side electronic device, the configuration request including information indicating an application configuration; and an assigning unit 424 that assigns the application configuration for use when the at least one AI injection process runs, if the application configuration is available. Preferably, the processing circuit 420 can further include a feedback unit 426 that feeds back information indicating that the requested application configuration is satisfied to the application-side electronic device, if the requested application configuration is satisfied, so that the application-side electronic device can request the AI-side electronic device to perform the at least one AI injection process without making a configuration request again.

[0117] Preferably, the receiving unit 422 can also receive a request from the application-side electronic device to perform the at least one AI injection process, and perform the at least one AI injection process by directly using the application configuration assigned to the at least one AI injection process. Preferably, the receiving unit 422 can also receive a resource release request from the application-side electronic device.

[0118] Preferably, the configuration unit 424 can further perform resource allocation for the at least one AI inference process according to specific information about resource configuration adjustment in the configuration request. Preferably, the dispatch unit 424 can further allocate resources for the at least one AI inference process according to candidate resource configurations corresponding to resource adjustment policy information in the case that the resource configuration indicated in the configuration request cannot be satisfied. Preferably, the dispatch unit 424 can further perform resource dispatch for AI inference processes according to priority information. Preferably, the configuration unit 424 can release the resources allocated for the at least one AI inference process.

[0119] Preferably, the feedback unit 426 can further provide information about currently available resources in the resource pool to the application-side electronic device in the case that the resource configuration included in the requested application configuration cannot be satisfied, so that the application-side electronic device can make resource configuration adjustment. Preferably, the feedback unit 426 can further provide state monitoring information in the response information provided from the AI platform-side electronic device, the state monitoring information containing coded information about AI inference status when an exception occurs. The AI inference status can contain at least one of an exception type, an application scenario where the exception occurs, an operation link where the exception occurs, etc. Preferably, the feedback unit 426 can inform the application-side electronic device of the release result via the response information. The operation of each unit can be performed as described above, and will not be described in detail here. In the drawings, the units are drawn with dashed lines, which is intended to indicate that the unit does not necessarily exist in the processing circuit. As an example, the unit can be in the AI platform-side electronic device but outside the processing circuit, or even outside the AI platform-side electronic device 400. It should be noted that although the units are shown as discrete units in the drawings, one or more of the units can be combined into one unit or split into multiple units. Figure 4A In the drawings, each unit is shown as a discrete unit, but one or more of the units can be combined into one unit or split into multiple units.

[0120] It should be noted that each of the above-mentioned units is only a logical module according to the specific function it implements, and is not intended to limit the specific implementation manner, for example, it can be implemented in software, hardware or a combination of software and hardware. In actual implementation, each of the above-mentioned units can be implemented as an independent physical entity, or can also be implemented by a single entity (for example, a processor (CPU or DSP, etc.), an integrated circuit, etc.). In addition, each of the above-mentioned units is shown with dashed lines in the drawings, indicating that these units can not actually exist, and the operations / functions they implement can be implemented by the processing circuit itself.

[0121] It should be understood that Figure 4AThe AI platform-side electronic device 400 can also include other possible components (e.g., memory, etc.) in addition to the illustrated structure of the AI platform-side electronic device. Optionally, the AI platform-side electronic device 400 can also include other components not shown, such as a memory, a radio frequency link, a baseband processing unit, a network interface, a controller, etc. The processing circuitry can be associated with the memory and / or the antenna. For example, the processing circuitry can be connected to the memory, directly or indirectly (e.g., with other components possibly in between), for access to data. Also for example, the processing circuitry can be connected to the antenna, directly or indirectly (e.g., with other components possibly in between), for transmitting signals via the communication unit and for receiving radio signals via the communication unit.

[0122] The memory can store various pieces of information (e.g., data traffic related information, configured resource information, etc.) generated by the processing circuitry 420, programs and data for AI platform-side electronic device operations, data to be transmitted by the AI platform-side electronic device, etc. The memory can also be located within the AI platform-side electronic device but outside the processing circuitry, or even outside the AI platform-side electronic device. The memory can be a volatile memory and / or a non-volatile memory. For example, the memory can include, but is not limited to, a random access memory (RAM), a dynamic random access memory (DRAM), a static random access memory (SRAM), a read-only memory (ROM), a flash memory.

[0123] Figure 4B A flowchart of an AI platform-side method according to an example embodiment of the present disclosure is shown. The method comprises a receiving step 402 of receiving, from an application-side electronic device, a configuration request for at least one AI injection process, the configuration request comprising information indicative of an application configuration; and a dispatching step 404 of dispatching, in case the application configuration is available, the application configuration for use by the at least one AI injection process at runtime. Preferably, the method can further comprise a feedback step 406 of feeding back, in case the requested application configuration is satisfied, information indicative of the requested application configuration being satisfied to the application-side electronic device, so that the application-side electronic device can request the AI-side electronic device to perform the at least one AI injection process without making a configuration request again. It is noted that the feedback step is drawn with a dashed line, aiming to illustrate that this feedback step does not necessarily have to be included in the method.

[0124] It is noted that the method according to the present disclosure can also comprise operation steps corresponding to the operations performed by the processing circuitry of the AI platform-side electronic device described above, which will not be described in detail here. It is noted that the various operations of the method according to the present disclosure can be performed by the AI platform-side electronic device described above, in particular by the processing circuitry or the corresponding processing unit of the AI platform-side electronic device, which will not be described in detail here.

[0125] The following will be described with reference toFigure 9A and Figure 9B to describe the application side of the exemplary embodiments according to the present disclosure, wherein Figure 9A A block diagram of an application side electronic device according to the exemplary embodiments of the present disclosure is shown. The application side electronic device 900 can include a processing circuit 920, which can be configured to send a configuration request for the at least one AI injection process to the AI platform side electronic device, the configuration request including information indicating the application configuration; receive response information from the AI platform side electronic device indicating that the application configuration is satisfied, and send an operation request to the AI platform side electronic device to perform the at least one AI injection process without proposing any other configuration request.

[0126] Preferably, the processing circuit 920 can be further configured to receive information from the AI platform side electronic device indicating that the resource configuration contained in the application configuration cannot be satisfied and information indicating the current available resources in the resource pool, make resource configuration adjustment, and send the adjusted resource configuration to the AI platform side electronic device via a configuration request.

[0127] Preferably, the processing circuit 920 can be further configured to receive state monitoring information from the AI platform side electronic device, the state monitoring information containing coded information about the AI injection state when an abnormal state occurs, the AI injection state can include at least one of an abnormal type, an application scenario where the abnormality occurs, an operation link where the abnormality occurs, etc., and analyze the state monitoring information to identify and recover the abnormal state.

[0128] Preferably, the processing circuit 920 can be further configured to send a resource release request to the AI platform side electronic device, and receive a release result from the AI platform side electronic device.

[0129] It should be noted that the meanings of various requests, information, etc. in the processes / operations implemented by the processing circuit 920 described above are the same as described above, and will not be described in detail here. In addition, the electronic device 900 and the processing circuit 920 can be implemented in a similar manner as the electronic device 400 and the processing circuit 420 described above, such as processors, units, program modules, etc. For example, the processing circuit 920 can include a receiving unit 922, a sending unit 924, a configuration unit 926, which can respectively implement the functions of the receiving operation, the sending operation, the resource configuration / adjustment operation, etc. described above, and will not be described in detail here. In addition, similar to the processing circuit 420, the processing circuit 920 can also include the additional components described above. Here will not be described in detail.

[0130] Figure 9BA flowchart of an application-side method according to an example embodiment of the present disclosure is shown. The method comprises a first sending step 904 of sending, to an AI platform-side electronic device, a configuration request for the at least one AI injection process, the configuration request comprising information indicative of the application configuration; a receiving step 906 of receiving, from the AI platform-side electronic device, response information indicative of the application configuration being satisfied, and a second sending step 908 of sending, to the AI platform-side electronic device, an operation request for performing the at least one AI injection process without proposing any other configuration request.

[0131] It should be noted that the method according to the present disclosure can further comprise operation steps corresponding to the operations performed by the processing circuitry of the application-side electronic device as described above, which will not be described in detail here. It should be noted that the various operations of the method according to the present disclosure can be performed by the application-side electronic device as described above, in particular by the processing circuitry or the corresponding processing unit of the application-side electronic device, which will not be described in detail here.

[0132] An example design of the configuration request that can be used in the configuration signaling interaction according to the present disclosure will be described exemplarily below.

[0133] An example of the information indicative of the resource configuration contained in the configuration request can be AI resource configuration information, which for example comprises resource type, resource quantity, resource number. In addition, the resource configuration information can further comprise resource configuration policy information. An example of the AI resource configuration information is as follows:

[0134] Table 3: Example of AI resource configuration information parameters

[0135] Parameter name Whether required Parameter type Parameter description resourceType Yes String Resource type, maximum 32 bytes, CPU, GPU, FPGA, SPARK… resourceList Yes String Resource list, maximum 128 bytes, GPU number or SPARK queue name resourceRatio Yes Short Resource ratio, between 0 and 1, default is all ResourceStrategy Yes String Resource allocation strategy, maximum 128 bytes, such as fair allocation, minimum allocation …… …… …… ……

[0136] As shown above, the AI resource configuration information at least contains resource type (resourceType), list (resourceList), ratio (resourceRatio), and configuration policy (resourcePolicy) (such as a fair allocation policy, a minimum configuration policy). Although not shown, the AI resource configuration information can further comprise candidate resource configurations, and each candidate resource configuration can also comprise resourceType and resourceList, as described above. It should be noted that the resource ratio can represent the ratio of the requested resource to the resources listed in the resource list. As an example, if the resource ratio is less than 1, the resource ratio can be considered to correspond to a candidate resource configuration. For example, if the resources listed by resourceType and resourceList are not available, it can be determined whether the resource corresponding to the resource ratio is available.

[0137] The information indicating model configuration included in the configuration request can at least include application scenario model ID, application scenario model name. The configuration request can further include other information indicating application scenario, such as application scenario name, sub-application scenario name, application scenario parameter. In particular, as an example, if the application scenario name / application scenario ID corresponds to resource configuration, model configuration, etc. as described before, the configuration request can only include the application scenario name / application scenario ID, and the resource configuration parameter and the model configuration parameter can be directly derived from the application scenario name / application scenario ID, which further saves the signaling overhead.

[0138] Therefore, corresponding configuration requests can be provided for different AI intelligence annotation processes. For example, one example of the configuration request for the AI training intelligence annotation process is the training activation request, which can include the model training related information in the following table:

[0139] Table 4: Example of model training information parameters

[0140] Parameter name Whether required Parameter type Parameter description appFrame Yes String Application scenario name, maximum 128 bytes, such as face recognition appSubFrame Yes String Sub-application scenario name, maximum 128 bytes, such as feature extraction appParam Yes String Application scenario request parameters, JSON format appModelName No String Application scenario model name, maximum 128 bytes, such as Xinduluo face recognition methodType No Short Calculation method, 1: online resource; 2: offline resource resourceType Yes String Resource type, maximum 32 bytes, CPU, GPU, FPGA, SPARK… resourceList Yes String Resource list, maximum 128 bytes, GPU number or SPARK queue name resourceRatio Yes Short Resource ratio, between 0 and 1, default is all resourceStrategy Yes String Resource allocation strategy, maximum 128 bytes, such as fair allocation, minimum allocation

[0141] appModelName can correspond to the aforementioned information indicating model configuration, and appFrame, appSubFrame, appParam can correspond to the aforementioned information indicating application scenario.

[0142] It should be noted that the model training information parameters given in the above table are only exemplary and do not need to include all parameters. For example, the parameters in the above table can also be represented in other ways. For example, the application scenario name and the sub-application scenario name can be combined into one name parameter, and as another example, the application scenario name and the application scenario model name can be combined into a parameter containing information of both.

[0143] For example, one example of the configuration request for the AI prediction intelligence annotation process can be the prediction activation request, which can include the model prediction information in the following table:

[0144] Table 5: Example of model prediction information parameters

[0145] Parameter name Whether required Parameter type Parameter description appModelName Yes String Application scenario model name, maximum 128 bytes, such as Xinduluo face recognition appModelID Yes String Application scenario model ID, maximum 64 bytes methodType No Short Calculation method, 1: online resource; 2: offline resource resourceType Yes String Resource type, maximum 32 bytes, CPU, GPU, FPGA, SPARK… resourceList Yes String Resource list, maximum 128 bytes, GPU number or SPARK queue name resourceRatio Yes Short Resource ratio, between 0 and 1, default is all resourceStrategy Yes String Resource allocation strategy, maximum 128 bytes, such as fair allocation, minimum allocation

[0146] The appModelName and appModelID included in the model prediction information shown in the above table can belong to the examples of the aforementioned information indicating model configuration. Similarly, the model prediction information parameters given in the above table are only exemplary and do not need to include all parameters. For example, the parameters of the model prediction information can be represented in other ways. For example, the application scenario model name and the application scenario model ID can be combined into a parameter containing information of both.

[0147] The response information for the configuration request according to the present disclosure can at least include state monitoring information. The following shows an example of the response information for the activation request.

[0148] Table 6: Parameter example of activation response information

[0149] Parameter name Whether required Parameter type Parameter description processCode Yes UInt Feedback information, see state metric framework definition resourceUrl is String Resource address, multiple function addresses are allowed resourceID is String Resource ID, max 32 bytes resourceType is String Resource type, max 32 bytes, CPU, GPU, FPGA, SPARK… resourceList is String Resource list, max 128 bytes, GPU number or SPARK queue name resourceRatio is Short Resource ratio, 0-1, default all resourcePolicy is String Resource allocation policy, max 128 bytes, such as fair distribution, minimum allocation …… …… …… ……

[0150] The processCode in the above table can correspond to the state monitoring information described above, and its form can refer to the state metric framework definition described above.

[0151] The AI resource configuration information is also shown in the above table. The resourceUrl and resourceID in the above table are also information related to AI resource configuration, but it should be pointed out that these information are not necessary. In particular, if the resources in the current resource pool cannot meet the configuration requirements of the application, the available resources in the resource pool can be embodied through the AI resource configuration information, so that the application end can actively determine whether resource adjustment can be performed. When the resources in the current resource pool can meet the configuration requirements, the AI resource configuration information is not needed, and only information indicating that the resource configuration is satisfied, such as a specific character, number, code, etc., can be returned.

[0152] The resource release signaling according to the present disclosure can contain release request information and release response information. The release request information at least contains resource ID; the release response information at least contains state monitoring information. The resource ID and state monitoring information can be as described above, which will not be described in detail here.

[0153] The following will exemplarily describe the signaling interaction process for the configuration of the AI annotation process according to the present disclosure. Table 7 shows the signaling interaction process for the activation configuration of the AI model annotation process, mainly including resource allocation.

[0154] Table 7: Resource activation process for model training

[0155]

[0156] Table 8 shows the signaling interaction process for the activation configuration of the AI model annotation process, mainly including model allocation and resource allocation.

[0157] Table 8: Resource activation process for model prediction

[0158]

[0159] Table 9 shows the resource release signaling interaction process when the application configuration in the configuration request is satisfied.

[0160] Table 9: Resource release process

[0161]

[0162] Examples of face recognition model training and prediction in the AI annotation process are described in the present disclosure, but it should be understood that the application scenarios of the present disclosure are not limited thereto. The improvement scheme proposed by the present disclosure can be applied to any AI annotation process, especially a process with at least one application operation of the same configuration, such as object tracking, object detection, etc., which is performed in the model training link or the model application link.

[0163] It should be noted that the above description is merely exemplary. Embodiments of the present disclosure can also be performed in any other appropriate manner, and still achieve the advantageous effects obtained by embodiments of the present disclosure. Moreover, embodiments of the present disclosure are also applicable to other similar application examples, and still achieve the advantageous effects obtained by embodiments of the present disclosure. It should be understood that the machine executable instructions in the machine readable storage medium or program product according to the embodiments of the present disclosure can be configured to perform operations corresponding to the above-mentioned device and method embodiments. When referring to the above-mentioned device and method embodiments, the embodiments of the machine readable storage medium or program product are clear to those skilled in the art, and therefore will not be described again. The machine readable storage medium and program product for carrying or including the above-mentioned machine executable instructions also fall within the scope of the present disclosure. Such storage media can include, but are not limited to, floppy disks, optical disks, magneto-optical disks, memory cards, memory sticks, and the like.

[0164] In addition, it should be understood that the above series of processes and devices can also be implemented by software and / or firmware. In the case of implementation by software and / or firmware, the corresponding programs constituting the corresponding software are stored in the storage medium of the related device, and when the programs are executed, various functions can be performed. As an example, a computer with a special hardware structure, such as Figure 10 The general-purpose computer 1300 shown is installed with programs constituting the software, and the computer can perform various functions and the like when various programs are installed. Figure 10 is a block diagram showing an example structure of a computer of an information processing device that can be employed in the present embodiment. In one example, the computer can correspond to the above-mentioned exemplary AI platform-side electronic device or application-side electronic device according to the present disclosure.

[0165] In Figure 10 the central processing unit (CPU) 1301 performs various processes according to programs stored in the read only memory (ROM) 1302 or programs loaded from the storage section 1308 to the random access memory (RAM) 1303. In the RAM 1303, data required when the CPU 1301 performs various processes and the like is also stored as necessary.

[0166] The CPU 1301, the ROM 1302, and the RAM 1303 are connected to each other via the bus 1304. The input / output interface 1305 is also connected to the bus 1304.

[0167] The following components are connected to the input / output interface 1305: the input part 1306 including a keyboard, a mouse, and the like; the output part 1307 including a display such as a cathode ray tube (CRT), a liquid crystal display (LCD), and the like, and a speaker, and the like; the storage part 1308 including a hard disk and the like; and the communication part 1309 including a network interface card such as a LAN card, a modem, and the like. The communication part 1309 performs a communication process via a network such as the Internet.

[0168] The drive 1310 is also connected to the input / output interface 1305 as necessary. A removable medium 1311 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, and the like is attached to the drive 1310 as necessary, so that a computer program read therefrom is installed in the storage part 1308 as necessary.

[0169] In a case where the above series of processes are implemented by software, a program constituting the software is installed from a network such as the Internet or a storage medium such as the removable medium 1311.

[0170] It is to be understood by those skilled in the art that such a storage medium is not limited to the above-mentioned ones Figure 10 The removable medium 1311 shown therein has a program stored therein, and is distributed separately from an apparatus to provide the program to a user. Examples of the removable medium 1311 include a magnetic disk (including a floppy® disk), a compact disk (including a compact disk read only memory (CD-ROM) and a digital versatile disk (DVD)), a magneto-optical disk (including a mini disk (MD)®), and a semiconductor memory. Alternatively, a storage medium can be the ROM 1302, a hard disk included in the storage part 1308, or the like, in which a program is stored, and which is distributed to a user together with an apparatus including them.

[0171] Further, it is to be understood that a plurality of functions included in one unit in the above-described embodiments can be implemented by a separate apparatus. Alternatively, a plurality of functions implemented by a plurality of units in the above-described embodiments can be implemented by a separate apparatus, respectively. Further, one of the above functions can be implemented by a plurality of units. Needless to say, such a configuration is included in the technical scope of the present disclosure.

[0172] In this specification, steps described in a flowchart describe not only processes performed in time series according to the order described in the flowchart, but also processes performed in parallel or individually rather than necessarily in time series. Further, even in a case of steps performed in time series, needless to say, the order can be changed as appropriate.

[0173] While the disclosure and its advantages have been described in detail, it should be understood that various changes, substitutions and alterations can be made herein without departing from the spirit and scope of the disclosure as defined by the appended claims. Moreover, the specific

[0174] While some embodiments of the disclosure have been described in detail, those skilled in the art will appreciate that the disclosure can be implemented in other specific embodiments without departing from the spirit and essential characteristics thereof. The present embodiments are to be considered in all respects as illustrative and not restrictive. Those skilled in the art will understand that the embodiments described herein can be combined, modified or replaced without departing from the scope and spirit of the disclosure. The scope of the disclosure is defined by the appended claims.

Claims

1. An AI platform-side electronic device, wherein the AI ​​platform side is capable of signaling interaction with an application side to execute at least one AI intelligence injection process having the same application configuration, the AI ​​platform-side electronic device comprising a processing circuit configured to: Receive a configuration request from the application side for the at least one AI injection process, the configuration request including information indicating application configuration; and If the application configuration is satisfied, the application configuration is allocated for use during the execution of the at least one AI injection process. in, The specific information related to resource configuration adjustment in the configuration request includes resource adjustment strategy information, and the processing circuit is further configured to allocate resources for the at least one AI injection process according to the resource configuration corresponding to the resource adjustment strategy information if the resource configuration indicated in the configuration request is not satisfied. The allocation of resources for the at least one AI intelligence injection process according to the resource configuration corresponding to the resource adjustment strategy information includes: The candidate resource configurations corresponding to the resource adjustment strategy information are evaluated in descending order of their resulting AI intelligence injection process performance to determine whether the candidate resource configuration is satisfied. If the candidate resource configuration is satisfied, the resources indicated in the candidate resource configuration are allocated to the AI ​​intelligence injection process.

2. The electronic device according to claim 1, wherein, The application configuration includes at least a resource configuration, and the information in the application configuration indicates the resources required to perform the at least one AI injection process.

3. The electronic device according to claim 1, wherein, The application configuration includes at least a model configuration, and the information in the application configuration indicates the model required to perform the at least one AI injection process.

4. The electronic device according to claim 1, wherein, The configuration request includes at least information indicating the application scenario for the AI ​​intelligence process.

5. The electronic device according to claim 1, wherein, The processing circuit is further configured to, when the requested application configuration is satisfied, feed back information to the application side containing an indication that the requested application configuration is satisfied, so that when the application side receives the information, it can request to execute the at least one AI intelligence injection process, instead of making a configuration request for each AI intelligence injection process.

6. The electronic device according to claim 1, wherein, The processing circuit is further configured to, upon receiving a request from the application side to execute the at least one AI intelligence injection process, directly utilize the application configuration already assigned to the at least one AI intelligence injection process to execute the at least one AI intelligence injection process.

7. The electronic device according to claim 1, wherein, The processing circuit is further configured to provide information indicating the currently available resources in the resource pool to the application side when the resource configuration contained in the configuration information is not satisfied, so that the application side can adjust the resource configuration.

8. The electronic device according to claim 1, wherein, The specific information related to resource allocation adjustment includes priority information for the AI ​​intelligence injection process, and the processing circuit is configured to perform resource allocation for the AI ​​intelligence injection process based on the priority information.

9. The electronic device according to claim 1, wherein, The processing circuit is further configured as follows: Status monitoring information is provided in the response information of the electronic device on the AI ​​platform side. The status monitoring information includes coded information related to the AI ​​intelligence status when an abnormal status occurs. The AI ​​intelligence status can include at least one of the abnormality type, the application scenario in which the abnormality occurred, and the operation step in which the abnormality occurred.

10. The electronic device of claim 1, wherein the processing circuit is further configured to: Received an application resource release request from the application-side electronic device. Release the resources used for the at least one AI injection process, and The release result will be used as a response to inform the application.

11. An application-side electronic device, the application side being capable of signaling interaction with an artificial intelligence (AI) platform to execute at least one AI intelligence injection process having the same application configuration, the application-side electronic device comprising a processing circuit configured to: Send a configuration request to the AI ​​platform-side electronic device for the at least one AI intelligence injection process, the configuration request including information indicating application configuration; Upon receiving a response from the AI ​​platform-side electronic device indicating that the application configuration has been satisfied, and Send an operation request to the AI ​​platform-side electronic device to execute the at least one AI injection process, without making any other configuration requests. in, The specific information related to resource configuration adjustment in the configuration request includes resource adjustment strategy information. Furthermore, if the resource configuration indicated in the configuration request is not satisfied, the resource allocation for the at least one AI intelligence injection process is performed by the AI ​​platform-side electronic device based on the resource configuration corresponding to the resource adjustment strategy information, including: The candidate resource configurations corresponding to the resource adjustment strategy information are evaluated in descending order of their resulting AI intelligence injection process performance to determine whether the candidate resource configuration is satisfied. If the candidate resource configuration is satisfied, the resources indicated in the candidate resource configuration are allocated to the AI ​​intelligence injection process.

12. The electronic device according to claim 11, wherein, The processing circuit is further configured to Receive information from the AI ​​platform-side electronic device indicating that the resource configuration contained in the application configuration is not satisfied, as well as information indicating the currently available resources in the resource pool. Adjust resource allocation, and The adjusted resource configuration is sent to the AI ​​platform-side electronic device via a configuration request.

13. The electronic device according to claim 11, wherein, The processing circuit is further configured as follows: The system receives status monitoring information from an electronic device on the AI ​​platform side. This status monitoring information includes coded information related to the AI's intelligence status when an abnormal state occurs. The AI ​​intelligence status can include at least one of the following: abnormality type, application scenario where the abnormality occurred, and operational step in which the abnormality occurred. The status monitoring information is analyzed to identify and recover from abnormal states.

14. The electronic device according to claim 11, wherein, The processing circuit is further configured as follows: Sending resource release requests to electronic devices on the AI ​​platform side, and Receive the release result from the electronic device on the AI ​​platform side.

15. A method on an artificial intelligence (AI) platform side, wherein the AI ​​platform side is capable of signaling interaction with an application side to execute at least one AI intelligence injection process having the same application configuration, the method comprising: The receiving step is used to receive a configuration request from the application side for the at least one AI injection process, the configuration request including information indicating application configuration; and The assignment step is used to assign the application configuration for use during the execution of the at least one AI injection process, provided that the application configuration is satisfied. The specific information related to resource configuration adjustment in the configuration request includes resource adjustment strategy information, and the method further includes: if the resource configuration indicated in the configuration request is not satisfied, allocating resources for the at least one AI injection process according to the resource configuration corresponding to the resource adjustment strategy information. The allocation of resources for the at least one AI intelligence injection process according to the resource configuration corresponding to the resource adjustment strategy information includes: The candidate resource configurations corresponding to the resource adjustment strategy information are evaluated in descending order of their resulting AI intelligence injection process performance to determine whether the candidate resource configuration is satisfied. If the candidate resource configuration is satisfied, the resources indicated in the candidate resource configuration are allocated to the AI ​​intelligence injection process.

16. An application-side method, wherein the application side is capable of signaling interaction with an artificial intelligence (AI) platform side to execute at least one AI intelligence injection process with the same application configuration, the method being: The first sending step is used to send a configuration request for the at least one AI intelligence injection process to the AI ​​platform-side electronic device, the configuration request including information indicating application configuration; The receiving step is used to receive response information from the AI ​​platform-side electronic device indicating that the application configuration has been satisfied, and The second sending step is used to send an operation request to the AI ​​platform-side electronic device to perform the at least one AI intelligence injection process, without making any other configuration requests. in, The specific information related to resource configuration adjustment in the configuration request includes resource adjustment strategy information. Furthermore, if the resource configuration indicated in the configuration request is not satisfied, the resource allocation for the at least one AI intelligence injection process is performed by the AI ​​platform-side electronic device based on the resource configuration corresponding to the resource adjustment strategy information, including: The candidate resource configurations corresponding to the resource adjustment strategy information are evaluated in descending order of their resulting AI intelligence injection process performance to determine whether the candidate resource configuration is satisfied. If the candidate resource configuration is satisfied, the resources indicated in the candidate resource configuration are allocated to the AI ​​intelligence injection process.

17. An apparatus comprising: One or more processors; as well as One or more storage media, storing instructions that, when executed by the one or more processors, cause the method according to claim 15 or 16 to be performed.

18. A computer-readable storage medium storing instructions that, when executed by one or more processors, cause to perform the method according to any one of claims 15 and 16.

19. An apparatus comprising components for performing the method according to any one of claims 15 and 16.

Images