ANN Validation via Input Space Partitioning and Synthetic Data

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Trained artificial neural networks (ANNs) often have redundant weights, leading to increased computational requirements for inference and training, necessitating a method to validate ANNs efficiently while minimizing weights.

Innovation Solution

A method and system for validating a trained ANN by partitioning the input space into cells using weight-specific hyperplanes and hypersets, checking for data points in each cell, and generating new data points via a simulation model if necessary to complete the test data set.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If the number of weights in the ANN is increased to improve the network's representational capacity and accuracy, then the quality and performance of the ANN is improved, but the computational requirements and complexity of the network architecture increase

Engineering Contradiction:

Improvequality and performance of ANNVSAvoidcomplexity of network architecture

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The input space is partitioned into multiple cells using hyperplanes derived from the network weights. This segmentation allows the validation process to systematically evaluate different regions of the input space, ensuring comprehensive coverage while maintaining a structured approach to analyzing the network's behavior across various weight configurations

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The method performs preliminary validation by checking whether test data points are present in each cell before final model deployment. This preliminary check ensures that the network has been adequately validated across the entire input space, identifying any gaps in coverage that would indicate insufficient weight configurations or missing validation scenarios

Inventive Principle:

Principle #10Preliminary action

2Reliability

If the number of weights is increased to maintain high accuracy, then the ANN functions more accurately, but more computing power is required for training and inference

Engineering Contradiction:

Improveaccuracy of ANNVSAvoidcomputing power required

Core Design Contradiction:

ReliabilityVSUse of energy by moving object

Solution Approach 1:

The validation method uses a simulation model to generate synthetic test data points in cells where actual data is missing. This partial action approach allows the system to validate the network's behavior in under-sampled regions without requiring additional physical training data, thereby reducing the need for extensive retraining and the associated computational costs while maintaining validation comprehensiveness

Inventive Principle:

Principle #16Partial or excessive action

3Productivity

If the number of weights is reduced to decrease computational requirements, then the efficiency of the ANN is improved, but the quality and performance of the network may deteriorate

Engineering Contradiction:

Improveefficiency of ANNVSAvoidquality and performance of ANN

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The validation process provides feedback by identifying cells in the input space that lack sufficient test data coverage. This feedback mechanism allows developers to iteratively improve the validation process by adding targeted test cases or generating synthetic data for specific regions, ensuring that weight reduction does not compromise the network's ability to handle critical input scenarios

Inventive Principle:

Principle #23Feedback

Data Source

PatentUS20250173580A1Method and system for validating a trained artificial neural network (ANN) on the basis of a test data set

Publication Date: 2025.05.29 ROBERT BOSCH GMBH

AI summary

A method and a system for validating a trained artificial neural network on the basis of a test data set.