Edge Multitask Model Tuning Using Encoded Vector Retrieval
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional systems face inefficiencies due to the need to choose between data security and processing capabilities, as they are unable to handle intense processing and data storage constraints for AI model development, especially at edge locations, leading to operational and performance bottlenecks.
Innovation Solution
A method involving multitask model tuning at edge locations using a vector database to select source datasets that provide context during training, allowing for relevant transfer learning, while maintaining privacy and security by encoding and combining vectors without exposing raw data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If data is sent to a centralized cloud location for processing, then processing capabilities are improved, but data security and privacy are compromised
Solution Approach 1:
The patent introduces an encoding model as an intermediary between the edge location and cloud location. The encoding model transforms raw data into encoded vectors that preserve the essential information needed for processing while removing sensitive details. This intermediary transformation allows the cloud to perform computations on the encoded representations without directly accessing or exposing the original sensitive data, thus resolving the contradiction between processing power and data security.
Solution Approach 2:
The patent extracts only the necessary features from the raw data by using an encoding model to generate encoded vectors. Instead of transmitting or storing the complete raw datasets at cloud locations, only the extracted encoded representations are sent to the cloud. This extraction principle allows centralized processing to occur on simplified data representations while the sensitive original data remains secured at edge locations.
2Measurement precision
If complex machine learning models are used to process data, then data analysis capability is improved, but processing overhead and computational strain increase
Solution Approach 1:
The patent segments the machine learning processing into two distinct parts: an encoding model that runs at the edge location to transform data into encoded vectors, and a larger language model that runs at the cloud location to perform complex analysis. This segmentation allows the computationally intensive encoding to occur locally with minimal overhead, while the complex analytical work occurs on already-transformed data in the cloud, distributing the processing burden effectively.
Solution Approach 2:
The encoding model performs preliminary transformation of the raw data into encoded vectors before the data is sent to the cloud. This preliminary action of encoding the data in advance simplifies the subsequent processing tasks that the large language model must perform, reducing the overall computational overhead while maintaining the ability to perform complex analysis on the encoded representations.
3Manufacturing precision
If all data is transmitted to centralized locations for model training, then model accuracy is improved, but data transfer overhead and processing time increase
Solution Approach 1:
The patent extracts only the essential information from the raw data by transforming it into encoded vectors through the encoding model. Instead of transmitting all the original data to the cloud for model training, only these compressed encoded representations are sent. This extraction maintains the critical information needed for accurate model training while dramatically reducing the volume of data that needs to be transferred, thus reducing transfer time and overhead.
Data Source
AI summary
A computer-implemented method, according to one approach, includes: receiving a user dataset having textual entries, and sampling a number of the textual entries in the received user dataset. The textual entries sampled from the user dataset are applied to an encoding model that corresponds to a given large language model architecture. The vectors that are produced by the encoding model are combined. The combined vectors are further sent to a vector database. In response, one or more public datasets are received from the vector database, the one or more public datasets being a closest match to the combined vectors.