Patent · US Active

System and method for layer-wise training of deep neural networks

US10163197B2 · kind B2 · utility

1Cited by
2References
12Claims
0Family size

Assignee

Inventors

Key dates

Filing dateMar 30, 2017
Grant dateDec 25, 2018
Priority date
Expiry dateAug 3, 2037

Classification

  • Technology area (CPC G)Physics
  • CPC primaryG06T2207/20084
  • WIPO fieldComputer technology
  • WIPO sectorElectrical engineering

Abstract

System and method for layer-wise training of deep neural networks (DNNs) are disclosed. In an embodiment, multiple labelled images are received at a layer of multiple layers of a DNN. Further, the labelled images are pre-processed. The pre-processed images are then transformed based on a predetermined weight matrix to obtain feature representation of the pre-processed images at the layer, the feature representation comprise feature vectors and associated labels. Furthermore, kernel similarity between the feature vectors is determined based on a predefined kernel function. Moreover, a Gaussian kernel matrix is determined based on the kernel similarity. In addition, an error function is computed based on the predetermined weight matrix and the Gaussian kernel matrix. Also, a weight matrix associated with the layer is computed based on the error function and predetermined weight matrix, thereby training the layer of the multiple layers.

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