AI Model Distribution via Network Manifest
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The existing methods for distributing AI/ML models over wireless networks face challenges in efficiently managing network congestion and ensuring quality of service, particularly during events where numerous users simultaneously download large AI/ML models, leading to limited throughput and potential delays in user experience.
Innovation Solution
A method involving a wireless transmit/receive unit (WTRU) that receives information on available network communication paths, determines AI/ML model chunks, selects an appropriate path, downloads and builds the model chunks, and performs inference, while a server selects an AI/ML model based on subscription information and generates information on available paths for downloading, optimizing network efficiency and reducing congestion.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If numerous users simultaneously download large AI/ML models during events, then model availability is improved, but network congestion increases and throughput decreases
Solution Approach 1:
The patent divides the AI/ML model into multiple smaller chunks that can be downloaded independently through different network paths. This segmentation allows users to receive different portions of the model simultaneously without blocking each other, reducing congestion on individual paths while maintaining overall model availability.
Solution Approach 2:
The patent introduces a multi-path dimension for model distribution, moving from single-path sequential downloads to parallel multi-path downloads. By utilizing multiple network communication paths simultaneously, the system increases throughput capacity without increasing the number of users, effectively handling the congestion problem.
2Ease of operation
If AI/ML models are distributed over wireless networks, then user access to models is improved, but spectrum resources are consumed and may become limited
Solution Approach 1:
By segmenting the model into chunks and distributing them across multiple paths, the patent reduces the data load on any single spectrum resource, allowing more efficient utilization of available spectrum while maintaining broad model access for users.
Solution Approach 2:
The system dynamically selects and switches between different network paths based on real-time spectrum availability and congestion conditions. This dynamic adaptation allows the system to optimize spectrum usage by utilizing available resources flexibly rather than following a fixed distribution pattern.
3Speed
If multiple network communication paths are provided for model download, then download speed is improved, but system complexity increases
Solution Approach 1:
The patent implements self-service mechanisms where the WTRU autonomously selects optimal paths and manages chunk downloads without requiring complex centralized control. The device automatically handles path selection, chunk assembly, and error recovery, reducing the operational complexity burden on the network side while maintaining high download speeds.
Data Source
AI summary
In one implementation, a manifest application server provides to a UE a description text file, called “manifest.” which indicates several network communication paths to download or update a particular AI/ML model adapted to the target UE capabilities. The AI/ML model is supposed to be split into chunks. The manifest file is centralized and controlled by the manifest application server for delivering the best overall network efficiency with respect to different types of UEs in the system. The manifest file includes device to device communication paths and relevant information (bandwidth, chunk IDs, etc.) provided by the UE themselves. The manifest application server publishes a set of different manifests describing different network communication paths and related expected network limitations for downloading particular model chunks.