Multi-Task AI Model Training via Graph-Aware Error Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing multi-task learning technologies struggle to effectively predict future values of multiple target variables connected via bipartite or semantic graphs, especially when dealing with partial relatedness between tasks and additional data sources, and when historical and live data are available in varying forms.
Innovation Solution
A computer-implemented method for training ML-AI models that inputs data for multiple tasks and additional data sources through a common trainable task representation function, followed by individual trainable linear functions to generate predictions and adversarial predictions. The method calculates prediction errors, adversarial errors, auxiliary errors, and graph errors, then trains parameters based on a weighted sum of these errors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a single multi-task model is trained to predict multiple target variables, then prediction accuracy improves through task synergies, but model complexity increases
Solution Approach 1:
The patent segments the multi-task model into separate task-specific models, each predicting a single target variable. This segmentation allows each model to be trained independently on its specific task data, reducing the complexity burden on any single model while maintaining the ability to handle multiple tasks through an ensemble approach.
Solution Approach 2:
The patent creates a universal prediction framework where multiple independent models can be trained and deployed to handle different prediction tasks. The system provides universal functionality by accepting various input types and graph structures while maintaining specialized prediction capabilities for each target variable through separate models.
2Device complexity
If multiple independent models are trained for each task, then model complexity is reduced, but computational resources and training time increase
Solution Approach 1:
The patent merges multiple independent task-specific models into a unified prediction system that shares common infrastructure. The models are trained separately but can be efficiently deployed together, and the system provides mechanisms for coordinating predictions across tasks to improve overall computational efficiency.
3Measurement precision
If graph-based relationships between tasks are utilized, then prediction accuracy improves through task relatedness, but computational overhead increases
Solution Approach 1:
The patent applies partial action by selectively using graph-based relationships only when and where they provide benefit. The system can operate with or without graph information depending on the specific task and available data, allowing it to maintain accuracy improvements from graph utilization while avoiding unnecessary computational overhead when graphs are not applicable.
Data Source
AI summary
A computer-implemented method for training a machine learning—artificial intelligence model for multiple prediction tasks includes inputting data for tasks and additional data sources through a common trainable task representation function to obtain a data representation for each. Each resulting data representation is input through two individual trainable linear functions to obtain a corresponding prediction and adversarial prediction. A prediction error for the tasks, an adversarial error across edges of a graph, an auxiliary error for the additional data sources, and a graph error are determined. Parameters of the common trainable task representation function and the trainable linear functions are trained based on a comparison against a weighted sum of the errors. The present invention can be used in a variety of applications including, but not limited to, several anticipated use cases in drug development, material synthesis, and medical/healthcare.