Machine Learning Model Adaptation via Statistical Feature Comparison
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning systems face challenges in developing accurate and cost-effective learned models for new domains, as the narrow scope of domains often results in limited information for model development, and obtaining training data can be expensive, especially when human experts are required.
Innovation Solution
The technique involves identifying existing learned models with similar statistical characteristics to new domain input data, using metrics like the Hellinger metric for comparison, and leveraging these models to retrain a new domain model, potentially combining or subset models to improve initial accuracy and reduce deployment delays.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the domain scope is restricted to improve model accuracy, then the learned model accuracy improves, but the available information for model development decreases
Solution Approach 1:
The patent combines multiple existing learned models from different domains to form a new model for a target domain. By merging statistical characteristics from multiple source models, the system accumulates more information than would be available from a single restricted domain, thereby resolving the contradiction between domain restriction and information availability.
Solution Approach 2:
The patent creates a universal framework that can adapt existing learned models across different domains to a new target domain. The system evaluates statistical characteristics of multiple source models and selectively combines them, making the model development process universally applicable regardless of the specific domain restrictions, thus maintaining both accuracy and information availability.
2Measurement precision
If human experts are used to develop training data to improve model accuracy, then the learned model accuracy improves, but the cost increases
Solution Approach 1:
The patent copies statistical characteristics from existing learned models that were previously developed with expert input, rather than creating new training data from scratch. By copying and adapting the statistical patterns from source models, the system preserves accuracy while avoiding the high cost of human expert involvement in new model development.
Solution Approach 2:
The patent performs preliminary model development in source domains where expert input may have already been invested, then reuses these pre-developed models for new domains. This preliminary action in source domains eliminates the need for repeated expert involvement in every new domain, thereby reducing costs while maintaining accuracy through statistical characteristic transfer.
3Measurement precision
If new training data is collected for a new domain to improve model accuracy, then the learned model accuracy improves, but the time required for model development increases
Solution Approach 1:
The patent performs preliminary model development in source domains, creating learned models that can be later adapted to new domains. This preliminary action eliminates the need for time-consuming data collection and model training in every new domain, thereby reducing development time while maintaining accuracy through statistical characteristic evaluation and combination.
Solution Approach 2:
The patent copies statistical characteristics from existing learned models instead of collecting new training data for each domain. This copying approach dramatically reduces the time required for model development while preserving accuracy by leveraging previously learned statistical patterns from source domains that are similar to the target domain.
Data Source
AI summary
In a machine learning system in which a plurality of learned models, each corresponding to a unique domain, already exist, new domain input for training a new domain model may be provided. Statistical characteristics of features in the new domain input are first determined. The resulting new domain statistical characteristics are then compared with statistical characteristics of features in prior input previously provided for training at least some of the plurality of learned models. Thereafter, at least one learned model of the plurality of learned models is identified as the basis for the new domain model when the new domain input statistical characteristics compare favorably with the statistical characteristics of the features in the prior input corresponding to the at least one learned model.