ML Model Validation via Safety Penalty Assignment

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current machine learning models used in safety-critical applications, such as autonomous vehicles and industrial plants, face challenges in preventing misclassifications that can lead to severe accidents or downtimes due to insufficient consideration of safety-critical consequences, and existing validation methods are time-consuming and expensive.

Innovation Solution

A computer-implemented method for validating trained machine learning models by using a representative validation data set to determine output probabilities, assigning penalties for misclassifications based on potential safety-critical consequences, and verifying that the quality criteria meet pre-defined thresholds to ensure reliable operation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If machine learning models are used in safety-critical applications, then productivity and efficiency are improved, but reliability deteriorates due to potential misclassifications causing accidents or downtimes

Engineering Contradiction:
ImproveefficiencyVSAvoidsafety
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The patent applies preliminary action by performing validation of the machine learning model before deployment in safety-critical applications. The validation process includes determining output probabilities for classification, assigning penalties based on safety-critical consequences, and verifying quality criteria against pre-defined thresholds. This preliminary validation ensures that the model meets reliability requirements before being used in production, preventing misclassifications that could cause accidents or downtimes.

Inventive Principle:
Principle #10Preliminary action

2Reliability

If conventional algorithmic means are used in addition to machine learning models for validation, then reliability is improved by detecting deviations, but device complexity and cost increase

Engineering Contradiction:
Improvedetection accuracyVSAvoidsystem complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent introduces an intermediary validation layer that sits between the machine learning model and the safety-critical application. This validation layer includes components for determining output probabilities, assigning penalties based on safety consequences, and verifying quality criteria. Rather than running parallel conventional algorithms, this intermediary validation mechanism enhances the ML model's reliability by systematically evaluating its outputs against pre-defined safety thresholds, thereby improving detection accuracy without proportionally increasing system complexity.

Inventive Principle:
Principle #24Intermediary (Mediator)

3Reliability

If comprehensive validation of machine learning models is performed to ensure safety, then reliability is improved, but loss of time increases due to extensive validation processes

Engineering Contradiction:
Improvesafety assuranceVSAvoidvalidation time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent applies parameter changes by dynamically adjusting validation parameters such as penalty assignments and quality criteria thresholds based on the specific safety-critical application context. The validation process determines output probabilities and assigns penalties according to the severity of potential misclassifications. By parameterizing the validation process, the system can adapt the comprehensiveness of validation to the risk level, ensuring adequate safety assurance while minimizing unnecessary validation time for lower-risk applications.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4375889A1Computer-implemented method for validating a trained machine learning model with regard to any potential misclassification and associated safety-critical consequence
Publication Date: 2024.05.29 SIEMENS MOBILITY GMBH
  • EP4375889A1 patent drawingFigure 1
  • EP4375889A1 patent drawingFigure 2
  • EP4375889A1 patent drawing

AI summary

The invention relates to a computer-implemented method for validating a trained machine learning model with regard to any potential misclassification and associated safety-critical consequence, comprising the steps comprising the steps a. providing the trained machine learning model (S1); b. providing a validation data set comprising a plurality of data elements to be classified (S2); wherein the validation data set is a representative validation data set for a safety-critical application ; c. determining an output vector with a plurality of probabilities using the trained machine learning model for classifying each data element based on the validation data set (S3); d. assigning a plurality of penalties to the classified data elements using at least one rule based on the output vector (S4); wherein the penalty is assigned to each classified data element of the plurality of classified data elements; wherein the penalty indicates a severity and/or an impact of a misclassification considering at least one associated safety-critical consequence; e. determining a quality criterium for the trained machine learning model based on the assigned penalties (S5); f. verifying that the determined quality criterium lies within a pre-defined range and/or meets a pre-defined threshold (S6); and g. providing the verified quality criterium and/or the validated machine learning model (S7) . Further, the invention relates to a corresponding computer program product and system.