Patent · US Active

Training artificial neural networks with reduced computational complexity

US11620514B2 · kind B2 · utility

0Cited by
2References
20Claims
0Family size

Assignee

Inventors

Key dates

Filing dateJun 28, 2018
Grant dateApr 4, 2023
Priority date
Expiry dateDec 14, 2039

Classification

  • Technology area (CPC G)Physics
  • CPC primaryG10L15/16
  • WIPO fieldComputer technology
  • WIPO sectorElectrical engineering

Abstract

At least some embodiments of the present disclosure relate to a method of training an artificial neural network (ANN) for an artificial intelligence recognition. The method includes producing, by an ANN, outputs by feeding inputs of a training data set to the ANN; determining errors of the generated outputs from target outputs of the training data set; generating a first-order derivative matrix including first-order derivatives of the errors and a second-order derivative matrix including second-order derivatives of the errors; obtaining an approximation of the first-order derivative matrix or an approximation of the second-order derivative matrix by compressing the first-order derivative matrix or the second-order derivative matrix; and updating weights of the ANN based on the approximation of the first-order derivative matrix or the approximation of the second-order derivative matrix.

Source: USPTO / EPO open patent data. Objective bibliographic and citation counts.