Categorizer with user-controllable calibration
US8189930B2 · kind B2 · utility
Assignee
Inventors
Key dates
| Filing date | Jul 17, 2008 |
| Grant date | May 29, 2012 |
| Priority date | — |
| Expiry date | Mar 29, 2031 |
Classification
- Technology area (CPC G)Physics
- CPC primaryG06F18/2415
- WIPO fieldComputer technology
- WIPO sectorElectrical engineering
Abstract
A calibrated categorizer comprises: a multi-class categorizer configured to output class probabilities for an input object corresponding to a set of classes; a class probabilities rescaler configured to rescale class probabilities to generate rescaled class probabilities; and a resealing model learner configured to learn calibration parameters for the class probabilities rescaler based on (i) class probabilities output by the multi-class categorizer for a calibration set of class-labeled objects, (ii) confidence measures output by the multi-class categorizer for the calibration set of class-labeled objects, and (iii) class labels of the calibration set of class-labeled objects, the class probabilities rescaler calibrated by the learned calibration parameters defining a calibrated class probabilities rescaler. In a method embodiment, class probabilities are generated for an input object corresponding to a set of classes using a classifier trained on a first set of objects, and are rescaled to form rescaled class probabilities using a resealing algorithm calibrated using a second set of objects different from the first set of objects. The method may further entail thresholding the re…
Source: USPTO / EPO open patent data. Objective bibliographic and citation counts.