AI Model Perturbation for Uncertainty and Reliability Measurement

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing AI models lack effective methods to quantify uncertainty and reliability, leading to inconsistent performance and potential errors in classification tasks.

Innovation Solution

The method involves adding perturbations based on class importances to representative vectors of an AI model, using noise distributions to measure inference uncertainty and determine when to terminate training, thereby enhancing model robustness and accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If AI models are used for classification tasks, then classification capability is achieved, but uncertainty and reliability cannot be quantified

Engineering Contradiction:
Improvemodel reliabilityVSAvoiduncertainty quantification
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

The patent applies preliminary action by adding perturbation noises to representative vectors before performing classification. This pre-processing step allows the model to evaluate multiple possible outcomes and quantify uncertainty in advance, rather than assessing reliability only after classification. The perturbation-based approach enables upfront assessment of model confidence and reliability metrics.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If standard AI classification is used, then classification speed is maintained, but performance measurement requires separate test sets

Engineering Contradiction:
Improveperformance measurementVSAvoidtesting infrastructure
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent implements universality by designing a system where the same AI model performs both classification and self-evaluation functions. By integrating perturbation-based uncertainty quantification into the classification process, the model simultaneously produces class predictions and reliability metrics using the same computational infrastructure, eliminating the need for separate test sets while maintaining measurement precision.

Inventive Principle:
Principle #6Universality (Multi-functionality)

3Reliability

If perturbation noises are added to representative vectors, then inference uncertainty can be measured, but computational complexity increases

Engineering Contradiction:
Improveinference uncertainty measurementVSAvoidcomputational process
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent applies parameter changes by systematically varying perturbation noise parameters (magnitude, distribution type, application method) to optimize the balance between uncertainty measurement accuracy and computational complexity. Different perturbation strategies are evaluated to find parameters that provide sufficient reliability information without excessive computational overhead, adapting the complexity level to specific application requirements.

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4583006A1Method and apparatus with ai model performance measuring using perturbation
Publication Date: 2025.07.09 SAMSUNG ELECTRONICS CO LTD
  • EP4583006A1 patent drawingFigure 1
  • EP4583006A1 patent drawingFigure 2
  • EP4583006A1 patent drawingFigure 3

AI summary

A device for measuring performance of an artificial intelligence (AI) model includes: one or more processors and a memory; and the memory storing instructions configured to cause the one or more processors to perform a process including: determining perturbations for respective classes based on respective class importances and adding noises determined based on the respective perturbations to respective representative vectors of the respective classes; and generating an inference uncertainty of the AI model from inference results outputted by the AI model using a weight matrix including the noise-added representative vectors.