Neural Network Verification Device for Safety-Critical Applications
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The reliability of artificial neural networks in safety-critical applications is limited due to unpredictability in their output, making it challenging to ensure operational safety in systems that rely on these models for approximating physical and chemical relations.
Innovation Solution
A method for statistically evaluating artificial neural networks by determining a measure of susceptibility to error within the input space, using test points and reference values to assess the network's reliability, and controlling systems based on this measure to ensure operational safety.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If artificial neural networks are used to approximate functions in safety-critical applications, then the modeling capability and flexibility are improved, but the reliability and predictability of output are worsened
Solution Approach 1:
The patent introduces an intermediary verification system that acts as a mediator between the neural network and the safety-critical application. This verification device evaluates the neural network's predictions by comparing them against reference values and calculating deviation measures, providing a safety layer that enables reliable use of flexible neural network models without directly compromising output predictability
Solution Approach 2:
The patent implements a feedback mechanism where the verification system continuously evaluates neural network outputs and provides information about prediction reliability. The deviation measure and susceptibility to error calculations feed back into the system, allowing dynamic adjustment of trust in neural network predictions based on their performance characteristics
2Measurement precision
If statistical evaluation with multiple test points is performed to improve reliability assessment, then the measurement precision is improved, but the computational complexity and time are worsened
Solution Approach 1:
The patent applies preliminary action by pre-calculating and storing reference values for multiple test points before actual verification occurs. This preparation phase allows the verification system to quickly compare neural network outputs against pre-established references during runtime, reducing the time penalty of statistical evaluation while maintaining high measurement precision
Data Source
AI summary
A device and a method, in particular a computer-implemented method, for the verification of an artificial neural network that is trained to map an input point from an input space of a function, in particular a limited or Lipschitz-constant function, as accurately as possible onto a functional value of the function. A test point is specified, the test point including a pair of a test input point from the input space of the function and a test functional value, the input point being determined from the input space, the input point being mapped by the artificial neural network onto the functional value, a reference for the functional value being determined using the test input point, a deviation of the functional value from the reference being determined, and a measure of a susceptibility to error of the artificial neural network being determined as a function of the deviation.