3GPP TR22.876 Version 19.10: AIML Model Transfer Phase 2 for Release 19

Jul 10, 2024

3GPP TR22.876 Version 19.10: AIML Model Transfer Phase 2 for Release 19

Overview

  • Focus: Application and optimization of AIML within the 5G system
  • Main topic: Transfer of AIML models between different entities in the network

Phases of Study

  • Phase 1: Initial exploration, use cases, basic requirements, gap analysis
    • Outcomes: Visibility confirmation, initial recommendations
  • Phase 2: Advanced exploration, detailed use cases, enhanced requirements
    • Outcomes: Advanced recommendations, detailed technical solutions
    • Note: Knowledge cut-off date: October 2021

Agenda Items

  1. Scope and Objectives
  2. AIML Operations Overview
  3. Proximity Based Work Task Off-loading
  4. Local AIML Model Split on Factory Robots
  5. AIML Model Transfer Management
  6. 5G System Assisted Transfer Learning
  7. Direct Device Connection Assisted Federated Learning
  8. Asynchronous Federated Learning
  9. Distributed Joint Interface for 3D Object Detection
  10. Consolidated Potential Requirements and KPIs
  11. Conclusion and Recommendations

Scope and Objectives

  • Study use cases and identify functional and performance requirements for AIML operations using direct device connections.
  • Focus on:
    • Distributed AI inference
    • Decentralized model training
    • Gap analysis of existing 5G system mechanisms
  • Ultimate goal: Ensure the 5G system can effectively support AIML operations for applications like autonomous driving, robot remote control, and video recognition.

AIML Operations Overview

  • Types of AIML operations:
    • Splitting operations between AIML end-points
    • Distributing/sharing AIML models and data over the 5G system
    • Utilizing distributed or federated learning
  • Phase 2: Enhanced by direct device connections for efficient AIML processes

Proximity Based Work Task Off-loading

  • Concept: Offloading work from one device to another via direct device connections
  • Process:
    • Discover nearby device
    • Establish direct connection
    • Transfer intermediate data
  • Example: AlexNet model for image recognition
    • Mobile device offloads computation to a network server
    • Evaluation of layer-level computation and communication resource
    • Selection of split points for effective load management

Local AIML Model Split on Factory Robots

  • Use case: Autonomous robots assist human operators
    • Offload AIML model to another robot or service hosting environment when battery is low
  • Service Flows:
    • Sharing AIML models and sensors
    • Maintain efficiency through task offloading
  • Conditions: Proximity service support, continuous operation despite low battery

AIML Model Transfer Management

  • Goal: Relay or store AIML models for stable and reliable transfers using direct device connections
  • Components:
    • Operator MEC stores AIML models
    • Volunteer UEs assist in relaying models
    • Third party sets policies and requirements
  • Process:
    • Model transfer paths (direct communication)
    • Control mechanisms (manage transfer)
    • Volunteer UE selection (assist in transfer)

5G System Assisted Transfer Learning

  • Concept: Leverage pre-trained models for better training efficiency
  • Process:
    • Transfer model from UE1 to UE2 via 5G network
    • Fine-tune based on local data from UE2
  • Result: Reduced training time and enhanced performance

Direct Device Connection Assisted Federated Learning

  • Concept: Collaborative model training via direct device connections
  • Process:
    • Sharing training results locally
    • Decentralized aggregation of local model updates
  • Comparison:
    • Original FL vs. Lacy Metropolis algorithm
    • Enhanced efficiency and reduced communication overhead with Lacy Metropolis

Asynchronous Federated Learning

  • Concept: Devices report results when ready, without waiting for synchronization
  • Advantages: Resilient to varying Network conditions
  • Process:
    • Relay UEs determine QoS for member UEs
    • Efficient model training and data aggregation

Distributed Joint Interface for 3D Object Detection

  • Scenario: Multiple vehicles collaborate for 3D object detection
    • Share and process sensor data for enhanced accuracy
  • Process:
    • Local maps from individual vehicles aggregated
    • Creation of a more accurate global map
  • Applications: Autonomous driving, accident detection

Consolidated Potential Requirements and KPIs

  • Authorization: User consent and operator policy
  • QoS Control: Configure and monitor QoS
  • Information Exposure: Share info with third parties
  • Charging Mechanisms: Support for AIML applications
  • Forming the baseline for future normative work

Conclusion and Recommendations

  • Key Findings:
    • Direct device connections enhance AIML operations in 5G
    • Significant improvements in performance and efficiency
  • Recommendations:
    • Proceed with normative work based on identified requirements
    • Enhance AIML applications within the 5G ecosystem for robust solutions