Neural Network Domain Transfer via Synthetic Data Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing neural networks trained for specific sensor configurations struggle to process data from new or improved configurations without requiring extensive manual labeling and retraining, limiting their adaptability in driving assistance and automated driving systems.
Innovation Solution
A method involving a generator network that transforms data records from one domain to another, using saliency records to create synthetic data that mimics the effects of real data, allowing neural networks to adapt without needing new labeling or data acquisition, and adversarial training with a discriminator network to refine the synthetic data's indistinguishability from real data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If neural networks are trained with training data from a specific sensor configuration, then the neural network can process measurement data from the same sensor configuration effectively, but it cannot supply usable results for measurement data from new or improved sensor configurations
Solution Approach 1:
The patent creates synthetic training data that copies the essential characteristics and statistical properties of real sensor data from the target configuration. By generating artificial training samples that mimic the distribution, noise patterns, and feature structures of actual sensor outputs, the neural network learns to process data from new sensor configurations without requiring physical samples from those specific sensors.
Solution Approach 2:
The patent transforms training data by applying parameter changes that simulate different sensor configurations. This includes modifying statistical parameters such as noise levels, signal intensity distributions, and temporal correlation structures to match the characteristics of new sensors. By adjusting these parameters, existing training data adapts to represent data from unseen sensor configurations.
2Reliability
If training data is manually labeled with target outputs, then supervised training can be performed effectively, but the greatest expense comes from the manual labeling work required
Solution Approach 1:
The patent implements self-service labeling by using the neural network's own predictions and intermediate representations to automatically generate labels for synthetic training data. The system leverages pre-trained model outputs, attention mechanisms, or feature activations to create supervisory signals without human intervention, thereby eliminating the time-consuming manual labeling process while maintaining training quality.
Solution Approach 2:
The patent performs preliminary labeling actions during the data generation phase rather than requiring separate manual labeling steps. By pre-computing target outputs from available information, intermediate model states, or simplified proxy tasks during synthetic data creation, the system prepares labeled training samples in advance, eliminating the need for subsequent manual annotation work.
3Measurement precision
If training data is recorded using a specific sensor configuration, then the neural network learns to process that specific configuration, but extensive manual labeling and retraining is required for new configurations
Solution Approach 1:
The patent creates a universal training data generation system that can produce training samples for multiple sensor configurations using a single framework. The synthetic data generation pipeline is designed to be configuration-agnostic, accepting different sensor parameters and outputting appropriate training data. This multi-functional approach eliminates the need for separate retraining pipelines for each sensor type, reducing overall system complexity.
Solution Approach 2:
The patent introduces synthetic training data as an intermediary between existing labeled data from one sensor configuration and the requirements of new sensor configurations. This intermediate representation layer translates characteristics from source domain data to target domain requirements, serving as a mediator that bridges the gap without requiring direct retraining on physical samples from each new sensor.
Data Source
AI summary
A method for training a generator network. In the method: training data records of the first domain and training data records of the second domain are provided; the training data records of the first domain are transformed into synthetic data records of the second domain using the generator network; the training data records and synthetic data records of the second domain are mapped by a task network to outputs relating to a predefined task; a saliency record is created comprising the saliencies with which portions of the training data record and of the synthetic data record respectively have contributed to the respective output of the task network; saliency records sampled from the pool of saliency records are classified by a discriminator network according to whether they belong to a training data record or a synthetic data record; the accuracy achieved in this classification is evaluated.