Machine Learning Inference Activation

Description

ML inference activation is a management and control procedure defined in 3GPP Release 19 (specifically in 21.905) that enables the inference capability of an ML inference function. In the context of 3GPP networks, machine learning (ML) is increasingly used to optimize various aspects of network operation, such as radio resource management, mobility prediction, and network slicing. An ML inference function is a logical entity that executes a trained ML model to generate outputs (inferences) based on input data. Activation is the process that transitions this function from a disabled or idle state to an enabled state where it can perform inference. This is distinct from ML model activation, which enables a specific model for a feature; ML inference activation is about enabling the inference capability itself, independent of which model is used.

The activation process typically involves coordination between management functions (e.g., OAM) and the network entity hosting the ML inference function. It may include steps such as verifying resource availability, loading the necessary model and runtime environment, and configuring input/output interfaces. Once activated, the inference function can receive input data (e.g., measurements, context information) and produce outputs that are consumed by other network functions or features. Activation can be triggered by management commands, policy decisions, or autonomous network actions. Deactivation is the reverse process, disabling the inference capability to free resources or due to operational changes.

In 3GPP architecture, ML inference activation is relevant to both the core network and radio access network, as ML can be deployed in various locations (e.g., gNB, AMF, SMF, NWDAF). The activation state is managed per inference function instance. The procedure ensures that ML capabilities are only active when needed, allowing efficient use of computational resources and energy. It also supports dynamic reconfiguration, where inference functions can be activated or deactivated based on network conditions or service requirements. The concept is part of the broader 3GPP effort to integrate AI/ML into network operations, as seen in specifications like TS 23.288 (NWDAF) and TS 38.300 (NR).

Key technical aspects include the signaling and management protocols used to activate inference, the state model (e.g., activated/deactivated), and the interaction with other management functions such as fault management and performance management. Activation may be subject to authorization and authentication to ensure only legitimate entities can enable inference. The process also considers the lifecycle of ML models: inference activation is independent of model activation, meaning an inference function can be activated but no model may be active, or vice versa. This separation allows flexible management of ML capabilities.

Overall, ML inference activation is a foundational concept for AI/ML-enabled 3GPP networks, enabling on-demand use of inference capabilities while maintaining control and efficiency. It is expected to evolve as more ML use cases are standardized in future releases.

Purpose & Motivation

ML inference activation exists to provide a standardized, controllable mechanism for enabling the execution of ML models within 3GPP network functions. Before this concept, ML deployment in networks was often ad-hoc, with no clear separation between model management and inference execution. This led to inefficiencies: models could be loaded but not actively inferring, or inference could run without proper resource allocation. The activation process solves this by introducing an explicit state transition that network operators can manage, ensuring that inference only occurs when intended and that resources are allocated appropriately.

Another motivation is the need for energy efficiency and resource optimization. ML inference can be computationally intensive, so being able to activate and deactivate it on demand allows networks to save power and processing capacity when inference is not needed. This is particularly important in multi-vendor, multi-domain environments where different ML functions may compete for resources. The activation concept also supports dynamic network behavior, where inference can be turned on in response to changing conditions (e.g., high load, specific service requests) and turned off otherwise.

Furthermore, the separation of inference activation from model activation provides flexibility. Operators can prepare multiple models but only activate inference when a specific model is selected and activated. This decoupling simplifies management and enables scenarios like model switching without re-activating the entire inference function. The concept aligns with 3GPP's overall AI/ML management framework, which aims to make ML integration seamless and interoperable across network elements.