Model Update

Description

Model Update is a concept introduced in 3GPP Release 19 to describe the process of modifying a trained machine learning (ML) model used within the 5G system. As specified in TS 26.927 and TS 38.843, a model update can be partial or full, and may involve changes to the model's internal structure (e.g., adding or removing layers) and/or its related parameters (e.g., weights, biases). This concept is part of the broader framework for AI/ML management in 5G, which aims to enable the integration of AI/ML capabilities into network functions and services.

In the context of 5G, AI/ML models can be deployed in various network elements, such as the RAN (e.g., for beam management, positioning), the core network (e.g., for traffic prediction, anomaly detection), and the UE (e.g., for application-specific tasks). The model update procedure is essential for maintaining model accuracy and adapting to changing conditions. It can be triggered by various events, such as performance degradation, new data availability, or network policy changes. The update can be performed by the model owner (e.g., the network operator or a third party) and may involve transferring the updated model or its components to the inference entity.

Architecturally, the model update process involves several functional entities. The AI/ML Model Management Function (e.g., in the core network or OAM) is responsible for managing the lifecycle of ML models, including training, testing, deployment, and updates. The model update may be delivered to the inference function (e.g., in the gNB or UE) via signaling protocols. The specifications define the procedures for model update, including the exchange of model identifiers, versioning, and the actual model data. The update can be incremental (partial) to reduce signaling overhead, or full to ensure consistency.

Key components include the model itself, the model identifier, the update trigger, and the transport mechanism. The model update may be performed over the air (OTA) for UEs, or via network interfaces for RAN and core network functions. Security and integrity protection are crucial to prevent tampering. The specifications also address the need for fallback mechanisms if an update fails. The model update concept is closely related to other AI/ML management aspects, such as model training, inference, and performance monitoring.

In summary, model update is a fundamental operation for maintaining and improving AI/ML models in 5G networks. It enables the network to adapt to dynamic environments and ensures that ML-based functionalities remain effective over time. As AI/ML becomes more pervasive in 5G-Advanced and 6G, the model update procedures will continue to evolve, focusing on efficiency, automation, and interoperability.

Purpose & Motivation

The model update concept exists to address the need for maintaining and evolving AI/ML models deployed in 5G networks. Traditional software updates are not sufficient for ML models because their performance depends on the data distribution and environment, which can change over time. Without updates, models may become stale, leading to degraded performance and potentially negative impacts on network operations. The model update mechanism provides a standardized way to modify models, ensuring that they remain accurate and effective.

Prior to Rel-19, AI/ML management in 3GPP was less formalized, with proprietary solutions for model updates. The introduction of a standardized concept enables interoperability and multi-vendor deployments. It solves problems such as model drift, where a model's performance degrades due to changes in the underlying data patterns. By allowing partial updates, it reduces the overhead of transmitting large models, which is particularly important for UEs with limited bandwidth. Full updates ensure that the model is completely refreshed when necessary.

The motivation for model update also includes the need for lifecycle management, where models can be updated to incorporate new features or improvements from continuous training. It supports the vision of a self-optimizing network that leverages AI/ML to automate operations. The specifications in TS 26.927 and TS 38.843 define the requirements and procedures, providing a foundation for future enhancements. As networks become more AI-driven, the model update concept will be critical for ensuring reliability, security, and performance.