Sensor Feature Selection for Resource-Efficient Classification Models

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Autonomous vehicles require efficient classification models that utilize a reduced set of features from sensor data to minimize computing resources without compromising accuracy, as conventional models consume excessive memory and bandwidth due to large input vectors.

Innovation Solution

A method to identify a subset of features by computing influence scores using backpropagation, excluding less influential features like noise, and training a second model on this reduced set, thereby reducing resource requirements.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If a large number of features are used to train the computer-implemented model, then classification accuracy is improved, but computing resources and memory consumption increase

Engineering Contradiction:
Improveclassification accuracyVSAvoidcomputing resources
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts and removes less influential features from the feature set through automated feature selection. The system evaluates feature importance and excludes features that contribute minimally to classification accuracy, thereby reducing the input vector size and computing resource requirements while preserving the essential information needed for accurate object classification in autonomous vehicle environments

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent changes the parameter of feature dimensionality by selecting a reduced subset of features from the original large set. Through automated feature selection processes, the system transforms the input from a high-dimensional feature space to a lower-dimensional space that maintains classification performance while reducing computational complexity and memory consumption

Inventive Principle:
Principle #35Parameter changes

2Measurement precision

If a large number of features are used to train the computer-implemented model, then classification accuracy is improved, but memory consumption increases

Engineering Contradiction:
Improveclassification accuracyVSAvoidcomputer-readable memory
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent extracts and removes less influential features from the feature set through automated feature selection. The system evaluates feature importance and excludes features that contribute minimally to classification accuracy, thereby reducing the input vector size and computing resource requirements while preserving the essential information needed for accurate object classification in autonomous vehicle environments

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent changes the parameter of feature dimensionality by selecting a reduced subset of features from the original large set. Through automated feature selection processes, the system transforms the input from a high-dimensional feature space to a lower-dimensional space that maintains classification performance while reducing computational complexity and memory consumption

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If a large number of features are transmitted from sensor system to computing system, then classification accuracy is improved, but bandwidth consumption increases

Engineering Contradiction:
Improveclassification accuracyVSAvoidbandwidth
Core Design Contradiction:
Measurement precisionVSQuantity of substance

Solution Approach 1:

The patent extracts and removes less influential features from the feature set through automated feature selection. The system evaluates feature importance and excludes features that contribute minimally to classification accuracy, thereby reducing the input vector size and computing resource requirements while preserving the essential information needed for accurate object classification in autonomous vehicle environments

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent changes the parameter of feature dimensionality by selecting a reduced subset of features from the original large set. Through automated feature selection processes, the system transforms the input from a high-dimensional feature space to a lower-dimensional space that maintains classification performance while reducing computational complexity and memory consumption

Inventive Principle:
Principle #35Parameter changes

Applied Scientific Principles

This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.

Function Achieved in This Case

The approach maintains classification accuracy while significantly reducing computing resources, memory, and bandwidth needs, enabling more efficient execution and potentially cheaper sensor deployment.

Implementation Method 1

computing gradients corresponding to the input values by backpropagating an output score through the trained first computer-implemented neural network

Methodology Applied
Scientific EffectBackpropagation:

Data Source

PatentEP4586113A1Technologies for identifying features to employ when training a computer-implemented model
Publication Date: 2025.07.16 GM CRUISE HOLDINGS LLC
  • EP4586113A1 patent drawingFigure 1
  • EP4586113A1 patent drawingFigure 2
  • EP4586113A1 patent drawingFigure 3

AI summary

Technologies pertaining to identifying a set of features to use when training a computer-implemented model are described. A first computer-implemented model is trained to perform a classification task based upon values for features in a set of features, where the features include features of sensor data and a noise feature. An importance score is computed for each feature in the set of features, and a subset of the set of features is identified based upon the importance scores computed for the features. A second computer-implemented model is then trained to perform the classification task based upon values for the features in the subset of features.