On-Device Model Orchestration for User Content Quality Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Large foundational machine-learned models require substantial compute resources and often perform worse than tailored models for specific tasks, while modern mobile devices lack sufficient compute power to run these models, leading to inefficient processing and bandwidth utilization in aggregating user-generated content.
Innovation Solution
Orchestrate lightweight, local on-device content-specific machine-learned models on user computing devices to evaluate user-generated content quality, reducing the need for server-side processing and bandwidth usage by selecting and processing user-generated content with models trained for specific tasks using fewer resources.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If large foundational machine-learned models are used to process user-generated content, then comprehensive task coverage is achieved, but substantial compute resources are required and performance deteriorates for specific tasks
Solution Approach 1:
The system segments the model processing task by dividing user-generated content into different categories (e.g., images, text, video) and selecting appropriate specialized models for each category. This segmentation allows the system to avoid using a single large foundational model for all tasks, thereby reducing compute resource requirements while maintaining task coverage.
Solution Approach 2:
The system implements a universal model selection framework that can handle multiple content types and tasks through a single orchestration mechanism. This framework evaluates content characteristics and automatically selects the most appropriate specialized model, providing multi-functional capability without requiring each specialized model to be universally applicable.
2Reliability
If large foundational machine-learned models are deployed on mobile devices, then comprehensive processing capability is achieved, but mobile devices lack sufficient compute power leading to inefficient processing
Solution Approach 1:
The system applies local quality by selecting models with appropriate complexity levels matched to specific content types and device capabilities. Instead of deploying a uniform large model across all scenarios, the system chooses lighter, more efficient models for mobile devices when processing specific content types, optimizing the balance between processing capability and energy consumption.
Solution Approach 2:
The system dynamically changes model selection parameters based on device characteristics, content type, and processing requirements. By adjusting which model is deployed (from lightweight to more complex) based on real-time conditions, the system optimizes processing capability while adapting to the limited compute power of mobile devices.
3Measurement precision
If all user-generated content is transmitted to server computing systems for processing, then centralized processing capability is maintained, but bandwidth utilization increases inefficiently
Solution Approach 1:
The system performs preliminary content evaluation and filtering on the user device before transmission. By pre-processing content locally to identify high-quality submissions and filter out low-quality content, the system reduces the volume of data that needs to be transmitted to the server, thereby maintaining evaluation accuracy while significantly reducing bandwidth consumption.
Solution Approach 2:
The system extracts and processes only the essential evaluation functions locally on user devices, separating these from the centralized server processing. This extraction allows the system to perform initial content quality assessment locally, transmitting only necessary data to the server, thus maintaining centralized processing capability for complex tasks while reducing overall bandwidth requirements.
Data Source
AI summary
User-generated content can be obtained from a user of a mapping application executed by the user computing device, wherein the user-generated content is associated with a particular Point of Interest (POI). A content-specific machine-learned model can be selected based on the user-generated content and/or a set of contextual information. The set of contextual information includes a location of the particular POI, information descriptive of the particular POI, and/or information descriptive of the user-generated content. The user-generated content can be processed with the first content-specific machine-learned model to obtain a quality metric. Based on the quality metric a determination can be made whether to provide the user-generated content to a server computing system associated with the mapping application.