AI Model Retraining via Action Center and Hardware Control
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
AI/ML models in robotic process automation (RPA) become less accurate over time due to data and model drift, requiring retraining, which is expensive and time-consuming, and lacks control over retraining hardware, necessitating an improved approach for model retraining and management.
Innovation Solution
A system that includes an action center to monitor AI/ML model performance, receive requests for review from RPA robots, and provide corrected results for retraining, along with hardware control for efficient retraining using GPUs or CPUs based on optimal resource allocation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If AI/ML models are retrained using traditional methods with labeled training data, then model accuracy is improved, but the process becomes expensive and time-consuming
Solution Approach 1:
The system performs preliminary actions by proactively detecting model drift through continuous performance monitoring before accuracy degrades significantly. When drift is detected, the system automatically initiates retraining workflows in advance, preparing corrected predictions and training data before the model becomes unusable, thus reducing both time loss and maintaining accuracy.
Solution Approach 2:
The system implements feedback mechanisms by continuously monitoring model performance metrics and comparing them against drift thresholds. This feedback loop enables automatic detection of when retraining is needed, triggers corrective actions, and validates improvement after retraining, creating a closed-loop system that reduces time loss through automation while maintaining reliability through continuous validation.
2Reliability
If AI/ML models are retrained frequently to maintain accuracy, then model performance is improved, but computational resources and costs increase
Solution Approach 1:
The system applies partial action by retraining only when necessary, based on drift detection thresholds rather than frequent scheduled retraining. It performs partial retraining using only the specific corrected predictions and drifted data segments that are needed, rather than complete model retraining, thus improving performance only when required while reducing computational resource consumption.
Solution Approach 2:
The system changes parameters by dynamically adjusting retraining frequency and intensity based on detected drift levels. When drift is minimal, no retraining occurs; when drift exceeds thresholds, targeted retraining is triggered. This parameter-based approach optimizes the balance between maintaining model performance and conserving computational resources.
3Measurement precision
If human review is implemented for all AI/ML model predictions, then prediction accuracy is improved, but processing speed decreases
Solution Approach 1:
The system applies partial human review by sending only those predictions to human reviewers that exceed drift thresholds or show signs of inaccuracy. The majority of predictions that fall within acceptable drift ranges are processed automatically without human intervention, maintaining high processing speed while still achieving high accuracy through targeted human review of edge cases.
Solution Approach 2:
The system implements local quality by applying different processing qualities to different predictions based on their drift characteristics. High-drift predictions receive full human review (high quality), while low-drift predictions receive automated processing (standard quality). This localized approach optimizes the balance between accuracy and processing speed by allocating human review resources only where most needed.
4Speed
If more hardware resources are allocated for retraining, then retraining speed is improved, but infrastructure complexity increases
Solution Approach 1:
The system implements universality by creating a unified retraining infrastructure that can dynamically allocate and manage multiple hardware resource types (GPUs, TPUs, CPUs) through a single abstraction layer. This universal platform handles resource provisioning, scheduling, and orchestration, enabling fast retraining through optimized hardware utilization while hiding the underlying complexity from users through standardized interfaces.
Data Source
AI summary
Supplementing artificial intelligence (AI)/machine learning (ML) models via an action center, providing AI/ML model retraining hardware control, and providing AI/ML model settings management are disclosed. AI/ML models may be deployed on hosting infrastructure where the AI/ML models can be called by robotic process automation (RPA) robots. When the performance of an AI/ML model falls below a threshold, the result of the AI/ML model prediction and other data is sent to an action center where a human reviews the data using a suitable application and approves the prediction or provides a correction if the prediction is wrong. This action center-approved result is then sent to the RPA robot to be used instead of the prediction from the AI/ML model.