AI Server Model Update Merging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing artificial intelligence model updating methods face challenges in managing network traffic and user privacy when transmitting usage logs from AI apparatuses to servers, and they struggle with asynchronous update information management.
Innovation Solution
An AI server and method that selects and merges update information from multiple AI apparatuses, excluding outliers, to update the AI model by adding a sum, average, or weighted sum of the selected information to the model parameters, thereby reducing network traffic and maintaining model reliability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If usage logs are transmitted from AI apparatuses to the server for model updating, then the AI model can be updated based on actual user usage data, but network traffic increases and user privacy may be invaded
Solution Approach 1:
The patent extracts only the essential update information (model parameter changes) from the usage logs at the AI apparatus level, rather than transmitting the complete usage logs. This extraction process removes unnecessary data while preserving the critical information needed for model updates, thereby reducing network traffic while maintaining model accuracy.
Solution Approach 2:
The patent segments the model update process into two parts: local processing at the AI apparatus (computing update information from usage logs) and centralized aggregation at the server (merging updates from multiple apparatuses). This segmentation allows privacy-sensitive data to remain local while only transmitting processed update information, reducing network traffic and privacy risks.
2Reliability
If usage logs are transmitted from AI apparatuses to the server for model updating, then the AI model can be updated based on actual user usage data, but user privacy may be invaded
Solution Approach 1:
The patent extracts only the essential update information (model parameter changes) from the usage logs at the AI apparatus level, rather than transmitting the complete usage logs. This extraction process removes unnecessary data while preserving the critical information needed for model updates, thereby reducing network traffic while maintaining model accuracy.
Solution Approach 2:
The patent introduces an intermediary processing step at the AI apparatus level that transforms raw usage logs into aggregated update information before transmission. This intermediary layer acts as a privacy protector by removing or anonymizing personally identifiable information while preserving the statistical patterns needed for model improvement.
3Reliability
If update information from multiple AI apparatuses is merged to update the AI model, then model reliability improves, but handling asynchronous update information increases system complexity
Solution Approach 1:
The patent changes the parameter representation of update information from raw usage logs to standardized model parameter differences. By transforming diverse asynchronous updates into a common parameter space (model weight changes), the system can merge updates from multiple apparatuses using simple arithmetic operations without complex synchronization protocols.
Solution Approach 2:
The patent merges update information from multiple AI apparatuses by aggregating their model parameter changes using summation or averaging operations. This merging approach handles asynchronous updates naturally by treating each apparatus's contribution as an independent vector that can be combined algebraically, avoiding the need for complex temporal synchronization.
Data Source
AI summary
Disclosed herein is an artificial intelligence server for updating an artificial intelligence model by merging a plurality of pieces of update information including a memory configured to store a first artificial intelligence model, a communication modem configured to communicate with a plurality of artificial intelligence apparatuses, and a processor configured to transmit the first artificial intelligence model to the plurality of artificial intelligence apparatuses, receive, from at least one of the plurality of artificial intelligence apparatuses, first update information of the first artificial intelligence model or second update information of a second artificial intelligence model updated from the first artificial intelligence model, select third update information to be used to update the first artificial intelligence model from the first update information and the second update information, and update the first artificial intelligence model using the third update information.