18313254. DOMAIN TRANSFER OF TRAINING DATA FOR NEURAL NETWORKS simplified abstract (Robert Bosch GmbH)

From WikiPatents
Jump to navigation Jump to search

DOMAIN TRANSFER OF TRAINING DATA FOR NEURAL NETWORKS

Organization Name

Robert Bosch GmbH

Inventor(s)

Jens Eric Markus Mehnert of Malmsheim (DE)

DOMAIN TRANSFER OF TRAINING DATA FOR NEURAL NETWORKS - A simplified explanation of the abstract

This abstract first appeared for US patent application 18313254 titled 'DOMAIN TRANSFER OF TRAINING DATA FOR NEURAL NETWORKS

Simplified Explanation

The abstract describes a method for training a generator network that can transform training data records from one domain into synthetic data records of another domain. These synthetic data records are then mapped by a task network to generate outputs related to a predefined task. A saliency record is created to determine the contribution of different portions of the training data and synthetic data to the task network's output. The saliency records are classified by a discriminator network to evaluate the accuracy of distinguishing between training data records and synthetic data records.

  • The method involves training a generator network to transform data records from one domain to another.
  • The transformed data records are then used by a task network to generate outputs for a specific task.
  • A saliency record is created to determine the contribution of different portions of the data records to the task network's output.
  • The saliency records are classified by a discriminator network to evaluate the accuracy of distinguishing between real and synthetic data records.

Potential Applications

  • This method can be applied in various domains such as image generation, text-to-speech synthesis, or data augmentation for machine learning tasks.
  • It can be used to generate synthetic data records for training models when real data is limited or expensive to obtain.
  • The method can also be used to improve the performance of models by generating additional training data that covers a wider range of scenarios.

Problems Solved

  • Limited availability of real data for training models can be addressed by generating synthetic data records.
  • The method helps in understanding the contribution of different portions of the data records to the task network's output.
  • It provides a way to evaluate the accuracy of distinguishing between real and synthetic data records.

Benefits

  • The method allows for the generation of synthetic data records that can be used to train models in various domains.
  • It provides insights into the importance of different portions of the data records for the task network's output.
  • The accuracy of distinguishing between real and synthetic data records can be evaluated, ensuring the reliability of the generated data.


Original Abstract Submitted

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.

Retrieved from "http://wikipatents.org/index.php?title=18313254._DOMAIN_TRANSFER_OF_TRAINING_DATA_FOR_NEURAL_NETWORKS_simplified_abstract_(Robert_Bosch_GmbH)&oldid=18520"