Component-Based Neural Networks for Explainable Sensor Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep neural networks in driver assistance systems and automated driving are opaque, uninterpretable, susceptible to adversarial perturbations, and lack robustness and explainability, making them challenging for systematic testing, formal verification, and integration of new concepts.
Innovation Solution
A method involving expert-based selection of components as the basis for a differentiable neural network structure, using feature distributions to train the network with primitives representing typical input variables, allowing positive, negative, and indefinite reasoning to enhance explainability and robustness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If deep neural networks are used for processing sensor data in driver assistance systems, then the processing capability and pattern recognition performance are improved, but the system becomes opaque and uninterpretable, making systematic testing and formal verification difficult
Solution Approach 1:
The patent segments the neural network into modular components with clearly defined interfaces and data flows. Each module processes specific aspects of sensor data independently, allowing the complex system to be understood through its constituent parts. This modular architecture maintains high processing capability while enabling systematic verification of individual components.
Solution Approach 2:
The patent introduces intermediate representation layers that bridge the gap between raw sensor data and final decisions. These intermediate layers provide interpretable feature representations that can be analyzed and verified, serving as mediators between the opaque neural network processing and the need for explainable results.
2Measurement precision
If deep neural networks are trained with large volumes of data, then the accuracy and functional quality are improved, but the training time and computational resources required increase significantly
Solution Approach 1:
The patent performs preliminary processing and feature extraction before main neural network training. By pre-processing sensor data to extract relevant features and reduce dimensionality, the system achieves high accuracy with reduced training time, as the neural network receives pre-curated input rather than raw high-volume data.
Solution Approach 2:
The patent extracts and removes redundant or irrelevant features from the training data before feeding it to the neural network. This extraction process maintains the essential information needed for accurate processing while eliminating unnecessary data that would increase training time and computational burden.
3Speed
If conventional neural networks are used, then processing speed is maintained, but robustness against adversarial perturbations and noise is insufficient
Solution Approach 1:
The patent implements preprocessing steps that cushion against adversarial perturbations before data enters the neural network. This includes noise filtering, normalization, and anomaly detection that protect the system from harmful inputs while maintaining processing speed, providing a defensive layer that preserves both speed and robustness.
Solution Approach 2:
The patent dynamically adjusts processing parameters based on input characteristics to maintain robustness without sacrificing speed. By changing parameters such as processing depth, filter strength, or activation thresholds based on the detected input type, the system adapts to maintain high robustness against perturbations while preserving fast processing for normal inputs.
Data Source
Figure 1~2
Figure 3~4
Figure 5~6
AI summary
The present invention relates to a method, a computer program containing instructions and an apparatus for training a neural network, and to a neural network trained using a method according to the invention. The invention additionally relates to a method, a computer program containing instructions and an apparatus for processing input variables, and to a motor vehicle that uses such a method or such an apparatus. Provided in a first step (10) are components that represent typical input variables for an application of the neural network. The neural network is then trained using at least the provided components (11). Additionally, feature distributions for the components can be determined (12) and parameters for conclusions can be learned (13).