Generator Neural Network Training With Localized Discriminator Feedback

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Training generator neural networks to generate synthesized sensor data is hindered by misleading global feedback from discriminators, leading to slow training and suboptimal solutions, as the discriminator's global decision can be noisy and incomplete.

Innovation Solution

Configuring the discriminator to provide localized distinguishing information, allowing the generator to receive more detailed feedback on which parts of the input data appear real or fake, and training the discriminator on composed sensor data to improve consistency regularization and focus on semantic changes.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Ease of operation

If the discriminator provides global feedback decisions, then the training process is simpler to implement, but the training signal becomes noisy and incomplete leading to suboptimal generator performance

Engineering Contradiction:
Improvesimplicity of training implementationVSAvoidcompleteness of training signal
Core Design Contradiction:
Ease of operationVSLoss of information

Solution Approach 1:

The patent segments the discriminator's feedback into localized distinguishing information for different parts of the input data. Instead of providing a single global decision, the discriminator outputs localized feedback indicating which specific regions or features appear real or fake, thereby preserving complete training information while maintaining implementation simplicity through modular architecture.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The discriminator is configured to provide different types of feedback for different parts of the input data. Localized distinguishing information is provided for specific regions where the generator fails, while other regions receive appropriate feedback. This local quality approach ensures that the training signal is both complete and targeted, improving generator performance without complicating the overall training process.

Inventive Principle:
Principle #3Local quality

2Loss of information

If the discriminator provides localized distinguishing information, then the training signal becomes more informative, but the discriminator complexity increases

Engineering Contradiction:
Improveinformation completeness of feedbackVSAvoiddiscriminator network complexity
Core Design Contradiction:
Loss of informationVSDevice complexity

Solution Approach 1:

The discriminator is segmented into multiple components: a feature extractor that processes input data, a classifier that makes global decisions, and a localization module that identifies specific regions. This segmentation allows the discriminator to provide comprehensive localized feedback without requiring a completely redesign of the network architecture, thereby managing complexity while maintaining information completeness.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The discriminator output is extended from a single global dimension to multiple localized dimensions corresponding to different regions of the input data. By adding this spatial dimension to the feedback, the discriminator provides more informative outputs without significantly increasing computational complexity, as the additional outputs can be generated through standard neural network operations applied across different spatial locations.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Reliability

If training data is obtained from real sensor measurements, then the training data reflects real-world conditions, but obtaining sufficient and diverse training data becomes difficult and costly

Engineering Contradiction:
Improverealism of training dataVSAvoidefficiency of data acquisition
Core Design Contradiction:
ReliabilityVSProductivity

Solution Approach 1:

The discriminator provides feedback to the generator about which features of generated data appear fake. This feedback loop enables the generator to learn from its mistakes and progressively improve the realism of synthesized training data. The localized distinguishing information guides the generator to focus on specific features that need improvement, making the training process more efficient and reducing reliance on large amounts of real sensor data.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The generator creates copies of real sensor data through synthesis, transforming real training data into additional diverse training samples. By copying and transforming existing real data into various domains and conditions, the system generates sufficient and diverse training data while maintaining the realism and reliability characteristics of the original sensor measurements.

Inventive Principle:
Principle #26Copying

4Manufacturing precision

If the generator is trained to produce indistinguishable synthesized data, then the quality of generated samples improves, but the training time increases due to noisy global feedback

Engineering Contradiction:
Improvequality of synthesized dataVSAvoidtraining duration
Core Design Contradiction:
Manufacturing precisionVSLoss of time

Solution Approach 1:

The localized feedback enables the generator to focus its optimization efforts on specific regions where it produces poor results. Instead of uniformly improving all parts of the generated data, the generator can target specific localized features that the discriminator identifies as fake. This targeted approach accelerates training convergence while maintaining high overall quality, reducing the time required to achieve indistinguishable synthesized data.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent replaces the mechanical process of iterative global feedback evaluation with a more efficient information-based system. The localized distinguishing information provides direct guidance to the generator about what needs improvement, eliminating the need for multiple rounds of global evaluation and trial-and-error learning. This substitution of the feedback mechanism significantly reduces training time while achieving the same quality improvements.

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentEP3859599B1Training a generator neural network using a discriminator with localized distinguishing information
Publication Date: 2023.12.27 ROBERT BOSCH GMBH
  • EP3859599B1 patent drawingFigure 1a~1b
  • EP3859599B1 patent drawingFigure 1c~2b
  • EP3859599B1 patent drawingFigure 2c~3

AI summary

Some embodiments are directed to a training method (300) for training a generator neural network (110; 120, 130, 140) configured to generate synthesized sensor data (141). A discriminator network is configured to receive (360) discriminator input data comprising synthesized sensor data and/or measured sensor data, and to produce (370) as output localized distinguishing information, the localized distinguishing information indicating for a plurality of sub-sets of the discriminator input data if said subset corresponds to measured sensor data or to synthesized sensor data.