Neural Network Training Data Manipulation for Autonomous Driving

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Traditional machine learning techniques for training artificial neural networks in autonomous driving often rely on limited and unrealistic training data, leading to potential accidents due to inadequate preparation for real-life scenarios.

Innovation Solution

A method that provides training data in two stages: first, acquiring data in a real or simulated scenario for standard situations, and second, manipulating this data by introducing extreme values to create a more comprehensive and diverse dataset, allowing the neural network to learn correct reactions to critical events.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If traditional machine learning techniques are used to train artificial neural networks with limited training data, then the training process is simpler and faster, but the reliability and safety of the neural network in real-world autonomous driving scenarios deteriorates

Engineering Contradiction:
Improvereliability of neural network in real-world scenariosVSAvoidamount of training data required
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent creates synthetic copies of real-world driving scenarios by manipulating existing training data through parameter transformations. These synthetic scenarios replicate real driving conditions while introducing extreme values and edge cases, thereby expanding the effective training data quantity without requiring additional physical data collection. The synthetic training data mirrors real-world complexity while being generated through controlled data manipulation rather than direct observation.

Inventive Principle:
Principle #26Copying

Solution Approach 2:

The patent systematically changes parameters of training data by introducing extreme values and edge cases into scenario parameters (such as weather conditions, traffic density, road conditions, and vehicle states). This parameter transformation technique generates diverse training scenarios from a limited base dataset, enhancing the neural network's reliability by exposing it to a broader range of conditions including rare and extreme events without requiring proportional increases in data collection effort.

Inventive Principle:
Principle #35Parameter changes

2Adaptability or versatility

If more diverse and comprehensive training data is collected through extensive real-world driving, then the quality and representativeness of training data improve, but the time and effort required for data acquisition increases

Engineering Contradiction:
Improverepresentativeness of training data for real-world situationsVSAvoidtime for data acquisition
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent performs preliminary data manipulation and scenario generation before actual training occurs. By pre-processing existing training data to create synthetic scenarios with extreme values and edge cases, the system prepares comprehensive training material in advance. This preliminary action eliminates the need for time-consuming real-world data collection during the training phase, as the diverse training data has already been generated through controlled data transformation.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The training data generation system serves itself by automatically manipulating existing data to create new training scenarios. Rather than requiring external data collection efforts, the system uses its own computational resources to transform and expand the training dataset. This self-service mechanism generates diverse and representative training data autonomously, significantly reducing the time and human effort required for data acquisition while maintaining high data quality.

Inventive Principle:
Principle #25Self-service

3Reliability

If training data is manipulated by introducing extreme values and edge cases, then the quality and safety of the trained neural network improve, but the complexity of the training process increases

Engineering Contradiction:
Improvesafety of neural network in critical situationsVSAvoidcomplexity of training process
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent segments the training process into distinct stages: first training the neural network with standard training data, then introducing manipulated data with extreme values and edge cases. This segmentation allows the training process to build foundational knowledge before tackling complex edge cases, making the overall process more manageable. The segmentation separates routine training from specialized extreme case training, reducing the cognitive load on the training system while enhancing safety.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent performs preliminary classification and organization of training data before manipulation. By pre-identifying which data points represent normal conditions versus which require extreme value introduction, the system simplifies the subsequent manipulation process. This preliminary action creates a structured approach to data transformation, reducing the complexity of systematically introducing edge cases while ensuring comprehensive coverage of critical situations.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4125007A1Method for providing training data for training an artificial neural network, method for training an artificial neural network, computer program product and data structure
Publication Date: 2023.02.01 VOLKSWAGEN AG
  • EP4125007A1 patent drawingFigure 1~2
  • EP4125007A1 patent drawing

AI summary

The invention relates to a method for providing training data (D1, D2) for training an artificial neural network (ANN), in particular for use in automated and/or autonomous driving, comprising the following steps: - providing training data (D1) in a scenario (S) of an intended use of the trained artificial neural network (ANN) in order to obtain a sequence (D1(t)) of training data (D1) for training the artificial neural network (ANN), - manipulating training data (D1), in particular the same sequence (D1(t)), by imposing at least one extreme value (W) on the scenario (S) in order to obtain a modified sequence (D2(t)) of enhanced training data (D2) for training the artificial neural network (ANN).