Edge Node ML Model Ranking via Data Sketching
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning model training approaches in distributed computer systems are inefficient due to the manual and trial-and-error method of selecting seed models, which fails to support continual learning and updates at edge nodes, and requires significant CPU, storage, and network resources.
Innovation Solution
A computer-implemented method that ranks and updates machine learning models by using a hash function and sketch of training datasets to select a seed model based on similarity with new input data, reducing the need for user input and resource usage through automated re-training and continuous adaptation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of time
If manual trial-and-error approach is used to select seed models, then model selection flexibility is maintained, but time consumption and resource usage increase significantly
Solution Approach 1:
The system enables self-service by automatically selecting seed models based on data similarity metrics. The automated process compares new input data with historical training datasets using hash functions and similarity calculations, eliminating the need for manual trial-and-error selection while reducing time consumption and resource usage.
Solution Approach 2:
The patent replaces the mechanical manual selection process with an automated computational system. Hash functions, similarity metrics, and algorithmic comparisons substitute for human judgment and manual experimentation, enabling rapid automated model selection that significantly reduces time and computational resources.
2Adaptability or versatility
If cloud-based model training is used, then centralized model management is achieved, but ability to track edge node data changes and support continual learning is lost
Solution Approach 1:
The system implements feedback mechanisms by continuously monitoring data changes at edge nodes and using this information to automatically select and retrain seed models. The feedback loop compares new input data characteristics with historical training data, enabling the system to adapt to evolving data distributions and support continual learning in distributed environments.
Solution Approach 2:
The patent applies preliminary action by pre-computing hash functions and similarity metrics for historical training datasets. This preparation enables rapid comparison and selection of appropriate seed models when new data arrives at edge nodes, facilitating quick adaptation and continual learning without requiring complex real-time analysis.
3Reliability
If re-training with complete training datasets is performed, then model accuracy is maximized, but CPU, storage, and network resource requirements increase
Solution Approach 1:
The system extracts only the essential elements needed for model selection by using hash functions and similarity metrics instead of comparing complete training datasets. This extraction approach identifies the most relevant seed models based on data characteristics without requiring the full datasets, significantly reducing CPU, storage, and network resource consumption while maintaining the ability to achieve model accuracy through targeted re-training.
Data Source
AI summary
An input dataset for training a new machine learning model is received by a processor. For each of a plurality of trained machine learning models, a hash function and a sketch of a training dataset used to train the machine learning model is retrieved. A sketch of the input dataset is computed based on the hash function and the input dataset, along with a distance between the sketch of the training dataset and the sketch of the input dataset. The computed distances of the trained machine learning models are ranked from smallest to largest, and a seed machine learning model for the input dataset is selected from the trained machine learning models based at least in part on the ranking. A training process of the new machine learning model using the selected seed machine learning model and the input dataset is initiated.