Training neural networks to assign scores
US11657268B1 · kind B1 · utility
Assignee
Inventors
Key dates
| Filing date | Sep 27, 2019 |
| Grant date | May 23, 2023 |
| Priority date | — |
| Expiry date | Nov 26, 2041 |
Classification
- Technology area (CPC G)Physics
- CPC primaryG05D1/0221
- WIPO fieldComputer technology
- WIPO sectorElectrical engineering
Abstract
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a neural network configured to receive a network input and to assign a respective score to each of a plurality of locations in the network input. In one aspect, a method includes obtaining a training input and a corresponding ground truth output; processing the training input to generate a training output; computing a loss for the training input, comprising: selecting a plurality of candidate locations; setting to zero the training scores for any location in the selected candidate locations that has a ground truth score below a threshold value; for each of a plurality of pairs of locations in the selected candidate locations: computing a pair-wise loss for the pair; and combining the pair-wise losses to compute the loss for the training input; and determining an update to the current values of the parameters.
Source: USPTO / EPO open patent data. Objective bibliographic and citation counts.