Long-Running Workflows for AI Model Retraining

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current AI tools lack dynamic input capabilities for managing the training and retraining lifecycle of AI/ML models, which can lead to inefficiencies and inaccuracies, particularly when human validation is required for model confidence thresholds.

Innovation Solution

Implementing long-running workflows with AI flows that suspend execution when model confidence is below a human validation threshold, allowing for dynamic input collection from human validation, and resuming execution once valid input is received, thereby enabling continuous training and retraining of AI/ML models.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If AI/ML models operate autonomously without human validation, then productivity is improved, but reliability deteriorates when model confidence is low

Engineering Contradiction:
Improveautomation efficiencyVSAvoidmodel prediction accuracy
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent introduces an RPA robot as an intermediary component that mediates between the AI/ML model and the workflow execution. The robot monitors model confidence scores and dynamically routes decisions: when confidence exceeds the threshold, the workflow continues autonomously; when confidence falls below the threshold, the robot suspends execution and triggers human validation. This intermediary mechanism enables the system to maintain high productivity through automation while ensuring reliability by activating human oversight only when necessary, thus resolving the contradiction between autonomous operation and validated accuracy.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If human validation is required for low-confidence predictions, then reliability is improved, but loss of time increases due to workflow suspension

Engineering Contradiction:
Improvemodel prediction accuracyVSAvoidworkflow execution time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent implements partial human validation by applying human-in-the-loop intervention only partially - specifically when the AI/ML model's confidence score falls below a predetermined threshold. For high-confidence predictions, the workflow executes fully autonomously without human intervention. This selective application of human validation minimizes time loss while maintaining reliability for uncertain predictions, resolving the contradiction between ensuring accuracy and maintaining workflow efficiency.

Inventive Principle:
Principle #16Partial or excessive action

3Adaptability or versatility

If dynamic input collection from human validation is implemented, then adaptability is improved, but device complexity increases

Engineering Contradiction:
Improvemodel retraining capabilityVSAvoidworkflow management complexity
Core Design Contradiction:
Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The patent establishes a feedback loop where the RPA robot collects dynamic input data from human validators and uses this feedback to retrains the AI/ML model. The system continuously monitors workflow decisions, collects correction data from humans when needed, and feeds this information back into the model training process. This feedback mechanism enhances adaptability by enabling the model to learn from real-world corrections while the RPA robot automates the complexity of data collection and model retraining, thus managing system complexity.

Inventive Principle:
Principle #23Feedback

4Manufacturing precision

If AI flows call current AI/ML models in long-running workflows, then manufacturing precision is improved, but ease of operation deteriorates due to suspension and resumption logic

Engineering Contradiction:
Improvetraining data accuracyVSAvoidworkflow execution simplicity
Core Design Contradiction:
Manufacturing precisionVSEase of operation

Solution Approach 1:

The patent implements self-service through the RPA robot, which autonomously manages the complex suspension and resumption logic without requiring manual intervention. The robot automatically monitors model confidence scores, suspends workflow execution when needed, collects human validation input, and resumes the workflow with corrected data. This self-managing approach maintains high manufacturing precision in training data while hiding the operational complexity from users, as they simply interact with the standardized RPA interface rather than managing the underlying suspension/resumption mechanics.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP4276709A1Building and managing artificial intelligence flows using long-running workflows for robotic process automation
Publication Date: 2023.11.15 UIPATH INC
  • EP4276709A1 patent drawingFigure 1
  • EP4276709A1 patent drawingFigure 2
  • EP4276709A1 patent drawingFigure 3

AI summary

Using long-running workflows with artificial intelligence flows to manage the training/retraining lifecycle of artificial intelligence (AI) / machine learning (ML) models is disclosed. Validation may be desired when an AI/ML model is called by a robotic process automation (RPA) robot executing the long-running workflow. This validation includes dynamic input from users. The RPA robot receives the dynamic input from the users and uses this data for training a replacement AI/ML model or retraining the called AI/ML model. The state of the long-running workflow may be preserved, both in training and serving. Long-running workflows may be used to keep track of where the current execution is in the ML model lifecycle.