AI/ML Model Distribution via D2D Scheduling and Local Caching

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for distributing AI/ML models in wireless communication networks are inefficient and do not effectively utilize local D2D servers for timely and localized model distribution.

Innovation Solution

A method involving a Wireless Transmit/Receive Unit (WTRU) that sends a subscription request to a Device to Device (D2D) service manager for downloading an AI/ML model, determines a local D2D server, and receives download information to download the model portion via D2D communication at a scheduled time, using the received information.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If AI/ML models are distributed through traditional network methods, then model distribution can be achieved, but distribution efficiency and timeliness are insufficient

Engineering Contradiction:
Improvemodel distribution efficiencyVSAvoidmodel download time
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The AI/ML model is divided into multiple portions or chunks that can be distributed independently through D2D communication. This segmentation allows parallel downloads from multiple local servers simultaneously, improving distribution efficiency and reducing overall download time.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Local D2D servers pre-cache AI/ML model portions before users need them. This preliminary action enables users to download models faster from local servers rather than waiting for remote server responses, significantly reducing download time and improving distribution efficiency.

Inventive Principle:
Principle #10Preliminary action

2Ease of operation

If centralized server distribution is used, then model management is simplified, but network resources are not optimized and user experience deteriorates

Engineering Contradiction:
Improvenetwork resource utilizationVSAvoiduser experience
Core Design Contradiction:
Ease of operationVSReliability

Solution Approach 1:

Local D2D servers autonomously manage and share AI/ML model portions with nearby users without requiring centralized coordination for each transaction. This self-service approach optimizes network resource utilization by enabling direct peer-to-peer transfers while maintaining reliable user experience through distributed availability.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system merges centralized model management capabilities with decentralized distribution execution. Centralized servers handle model updates and coordination, while local D2D servers handle actual model portions distribution, combining the advantages of both approaches to optimize network resources and improve user experience.

Inventive Principle:
Principle #5Merging (Combining)

3Speed

If local D2D servers are utilized for model distribution, then distribution timeliness improves, but system complexity increases

Engineering Contradiction:
Improvemodel distribution speedVSAvoiddistribution system complexity
Core Design Contradiction:
SpeedVSDevice complexity

Solution Approach 1:

A coordination entity acts as an intermediary between centralized servers and local D2D servers, managing the complexity of distributed model portions distribution. This intermediary handles server discovery, model portion routing, and synchronization, enabling fast local distribution while abstracting away system complexity from individual components.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4381433B1Ai/ML model distribution
Publication Date: 2025.12.31 INTERDIGITAL CE PATENT HOLDINGS SAS
  • EP4381433B1 patent drawingFigure 1A
  • EP4381433B1 patent drawingFigure 1B
  • EP4381433B1 patent drawingFigure 1C

AI summary

Procedures, methods, architectures, apparatuses, systems, devices, and computer program products comprising: sending, by a first Wireless Transmit/Receive Unit (WTRU) to a network entity, a subscription request for downloading an AI/ML model, the AI/ML model comprising a first model portion, and one or more further model portions; determining a second WTRU storing at least the first model portion of the AI/ML model; sending, to the network entity, first information comprising an indication of the second WTRU; receiving, from the network entity, second information indicating a schedule for downloading at least the first model portion of the AI/ML model from the second WTRU; and downloading, from the second WTRU via a device-to-device communication between the first WTRU and the second WTRU, at least the first model portion of the AI/ML model at a scheduled time using the second information.