17549006. Reducing Exposure Bias in Machine Learning Training of Sequence-to-Sequence Transducers simplified abstract (INTERNATIONAL BUSINESS MACHINES CORPORATION)

From WikiPatents
Jump to navigation Jump to search

Reducing Exposure Bias in Machine Learning Training of Sequence-to-Sequence Transducers

Organization Name

INTERNATIONAL BUSINESS MACHINES CORPORATION

Inventor(s)

Xiaodong Cui of Chappaqua NY (US)

Brian E. D. Kingsbury of Cortlandt Manor NY (US)

George Andrei Saon of Stamford CT (US)

David Haws of New York City NY (US)

Zoltan Tueske of White Plains NY (US)

Reducing Exposure Bias in Machine Learning Training of Sequence-to-Sequence Transducers - A simplified explanation of the abstract

This abstract first appeared for US patent application 17549006 titled 'Reducing Exposure Bias in Machine Learning Training of Sequence-to-Sequence Transducers

Simplified Explanation

The patent application describes a mechanism for training a computer model using machine learning techniques. Here is a simplified explanation of the abstract:

  • A perturbation generator is used to create modified training data by injecting perturbations into the original training data.
  • These perturbations cause data corruption in the original training data.
  • The modified training data is then fed into a prediction network of the computer model.
  • The prediction network processes the modified training data and generates a prediction output.
  • The machine learning training is performed on the prediction network using the prediction output and the original training data.
  • This training process results in a trained prediction network, which is a trained computer model.
  • The trained computer model can be deployed to an artificial intelligence computing system to perform inference operations.

Potential applications of this technology:

  • This technology can be used in various fields where machine learning is applied, such as image recognition, natural language processing, and data analysis.
  • It can improve the accuracy and performance of computer models by training them with modified data that simulates real-world data corruption.

Problems solved by this technology:

  • Traditional machine learning training methods may not adequately prepare computer models for real-world scenarios where data corruption or perturbations can occur.
  • This technology addresses this problem by introducing perturbations into the training data, allowing the computer model to learn and adapt to such scenarios.

Benefits of this technology:

  • By training computer models with perturbed data, the models can become more robust and accurate in handling real-world data corruption.
  • The trained computer models can be deployed to artificial intelligence systems, enabling them to perform inference operations with improved accuracy and reliability.


Original Abstract Submitted

Mechanisms are provided for performing machine learning training of a computer model. A perturbation generator generates a modified training data comprising perturbations injected into original training data, where the perturbations cause a data corruption of the original training data. The modified training data is input into a prediction network of the computer model and processing the modified training data through the prediction network to generate a prediction output. Machine learning training is executed of the prediction network based on the prediction output and the original training data to generate a trained prediction network of a trained computer model. The trained computer model is deployed to an artificial intelligence computing system for performance of an inference operation.

Retrieved from "http://wikipatents.org/index.php?title=17549006._Reducing_Exposure_Bias_in_Machine_Learning_Training_of_Sequence-to-Sequence_Transducers_simplified_abstract_(INTERNATIONAL_BUSINESS_MACHINES_CORPORATION)&oldid=26292"