7.3 AI/ML Training Model Management
7.3.1 Overview
oneM2M serves as a fundamental Internet of Things (IoT) platform for collecting and managing diverse data. Artificial Intelligence (AI) and Machine Learning (ML) applications extensively utilise data gathered within IoT platforms for model training. The performance of AI models is directly influenced by the quality and quantity of the collected dataset used for training. In order to facilitate the effective deployment of AI/ML models, it is crucial that the IoT platform provides robust capabilities for managing the entire AI/ML lifecycle, particularly concerning data and model management. This includes the ability to organise datasets into distinct categories such as training, validation, and testing sets, which are essential for rigorous model development and evaluation. Furthermore, effective AI/ML execution requires the management of selected algorithms, their corresponding parameters, and the resulting trained models directly within the oneM2M ecosystem.
Given the increasing integration of AI/ML technologies across various network systems, including telecommunication core networks, smart factory platforms, and IoT platforms, it is imperative to incorporate necessary AI enablement features directly into IoT platforms. By leveraging oneM2M platforms that support standardized AI/ML data and model management, AI applications can more readily develop and deploy intelligent services by utilizing platform-provided functionalities rather than implementing complex, bespoke solutions.
This proposal introduces a new resource, <mlModelBuiler>, within the oneM2M platform to facilitate comprehensive data and model management for AI/ML applications. This resource provides the necessary functionalities for AI/ML applications to construct predictive models.
Key features of the <mlModelBuilder> resource include:
- Training Dataset Classification: Enables the classification and management of datasets into training, validation, and testing sets, vital for the iterative process of model development and evaluation.
- AI/ML Model Selection and Parameters: Provides mechanisms to specify the AI/ML algorithm an AI application intends to use and to define the required hyperparameters for the selected model.
- Trained Model Management: Facilitates the management and storage of the resulting model subsequent to the training and validation processes, making the learned model accessible for inference.
- Model Building Control: Provides a controlling mechanism to initiate the model building process, assuming that all prerequisites are met.
Preconditions and Assumptions for this Proposal:
This proposed mechanism is based on the following key preconditions and assumptions:
- Data Availability on oneM2M Platform: It is assumed that the oneM2M platform either already holds or can reliably acquire all necessary data for both model training and subsequent prediction tasks. This implies that relevant data sources (e.g., from sensors, devices) are accessible via oneM2M resources.
- AI/ML Algorithm Awareness: It is assumed that the oneM2M platform, potentially through an "AI-enabled Common Service Function (CSF)" or similar entity, possesses knowledge of a predefined list of available ML algorithms that can be utilized for model building.
- AI/ML Application Interaction: AI/ML applications are expected to interact with the oneM2M platform by creating and configuring <mlModelBuilder> resources, controlling model builds, and subsequently retrieving the trained models for their specific inference tasks.
By providing these capabilities within oneM2M, AI/ML applications can streamline the model development workflow, leveraging the platform's existing data management strengths and reducing the complexity associated with off-platform data handling and model lifecycle management.
7.3.2 Resource types
7.3.2.1 Resource Type mlModelBuilder
The <mlModelBuilder> resource represents.
Table 7.3.2.1-1: Child resources of <mlModelBuilder> resource
| Child Resources of <mlModelBuilder> | Child Resource Type | Multiplicity | Description | <mlModelBuilder> Child Resource Types |
|---|---|---|---|---|
| [variable] | <semanticDescriptor> | 0..n | See clause 9.6.30 [1] | <semanticDescriptor>, <semanticDescriptorAnnc> |
| [variable] | <subscription> | 0..n | See clause 9.6.8 [1] | <subscription> |
| [variable] | <transaction> | 0..n | See clause 9.6.48 [1] | <transaction> |
| [variable] | <action> | 0..n | See clause 9.6.61 [1] | <actionAnnc> |
The <mlModelBuilder> resource shall contain the attributes specified in table 7.3.2.1-2.
Table 7.3.2.1-2: Attributes of <mlModelBuilder> resource
| Attributes of <mlModelBuilder> |
Multiplicity | RW/ RO/ WO |
Description | <mlModelBuilderAnnc> Attributes |
|---|---|---|---|---|
| resourceType | 1 | RO | See clause 9.6.1.3 [1] | NA |
| resourceID | 1 | RO | See clause 9.6.1.3 [1] | NA |
| resourceName | 1 | WO | See clause 9.6.1.3 [1] | NA |
| parentID | 1 | RO | See clause 9.6.1.3 [1] | NA |
| expirationTime | 1 | RW | See clause 9.6.1.3 [1] | MA |
| accessControlPolicyIDs | 0..1 (L) | RW | See clause 9.6.1.3 [1] | MA |
| labels | 0..1 (L) | RW | See clause 9.6.1.3 [1] | MA |
| creationTime | 1 | RO | See clause 9.6.1.3 [1] | NA |
| lastModifiedTime | 1 | RO | See clause 9.6.1.3 [1] | NA |
| announceTo | 0..1 (L) | RW | See clause 9.6.1.3 [1] | NA |
| announcedAttribute | 0..1 (L) | RW | See clause 9.6.1.3 [1] | NA |
| announceSyncType | 0..1 | RW | See clause 9.6.1.3 [1] | MA |
| dynamicAuthorizationConsultationIDs | 0..1 (L) | RW | See clause 9.6.1.3 [1] | OA |
| creator | 0..1 | RO | See clause 9.6.1.3 [1] | NA |
| custodian | 0..1 | RW | See clause 9.6.1.3 [1] | NA |
| datasetTrain | 0..1 (L) | RW | A list of resources storing training data. This attribute can contain multiple resource references, allowing for distributed or segmented training datasets. | OA |
| datasetValidation | 0..1 (L) | RW | A list of resources storing validation data. Similar to datasetTrain, it can hold multiple references. | OA |
| datasetTest | 0..1 (L) | RW | A list of resources for testing a model. This attribute can also contain multiple resource references. | OA |
| selectedModel | 0..1 | RW | An ML algorithm that represents the model to be performed. This specifies which type of ML model (e.g., Linear Regression, Neural Network, Decision Tree) the platform should use. | OA |
| modelParameters | 0..1 | RW | The parameters (hyperparameters) utilized by the selected algorithm. This could be a complex data structure (e.g., JSON or XML) defining learning rates, epochs, layer configurations, etc. | OA |
| trainedModel | 0..1 | RO | The resulting model (e.g., executable software, a model file, or a reference to a model registry) after the successful completion of training and validation. | OA |
| controlBuildModel | 0..1 | WO | This is a write-only attribute used to control the model building process. For example, this attribute can indicate various control actions such as start, pause, stop, and restart. When updated, the platform performs the corresponding operation. The assumption is that proper values for datasetTrain, datasetValidation, datasetTest, and selectedModel must be pre-configured before triggering. | OA |
| modelBuildStatus | 0..1 | RO | This attribute indicates the current status of the model build process initiated by controlBuildModel. This attribute allows applications to monitor the build progression and completion. Applications can subscribe to changes in this attribute to detect when the model build has been successfully completed or altered its state. Possible status this attribute could indicate include: Pending, In progress, Successful and Failed. | OA |
Editor’s note: It is FFS how and where to manage an actual trained model build out of this resource including architectural options (e.g., dedicated model repository resource, 3rd party location) and overall lifecycle management (e.g., versioning, deployment).
Editor’s note: It is FFS how to manage and operate the mlModelBuilder resource, especially for recurrent model build processes or when external control is required (e.g., via a semantic mash-up profile or other resources supporting process management).