Systems and methods for discriminative density model selection
AssigneeMicrosoft Corporation
InventorsBo Thiesson, Christopher A. Meek
Life of this patent
Estimating the term…
Abstract
The present invention utilizes a discriminative density model selection method to provide an optimized density model subset employable in constructing a classifier. By allowing multiple alternative density models to be considered for each class in a multi-class classification system and then developing an optimal configuration comprised of a single density model for each class, the classifier can be tuned to exhibit a desired characteristic such as, for example, high classification accuracy, low cost, and/or a balance of both. In one instance of the present invention, error graph, junction tree, and min-sum propagation algorithms are utilized to obtain an optimization from discriminatively selected density models.
Claims37
A recognition system, comprising a memory having stored therein computer executable components and a processor that executes the following computer executable components:
- a density model identifier component that identifies a plurality of density models for each of at least two classes that are employed in connection with recognition;
- and
- an optimization component that at least:
- identifies a subset of the density models for employment in a recognition application;
- selects a model of the subset of the density models for each class to achieve an optimization based on, at least in part, multi-class classification accuracy;
- and
- determines the optimization based on, at least in part:
- forming an error graph that captures confusability between the at least two classes;
- triangulating the error graph to form a junction tree that includes cliques;
- and
- associating each clique of the junction tree with a potential, wherein the potential is based on a value for each of the possible configurations of variables in the clique.
A method of selecting density models for a classifier, wherein the method is performed by a processor that executes at least the following acts:
- constructing a set of at least two alternative density models for at least one class in a classification system having at least two classes;
- identifying a subset of the density models for employment in a recognition application;
- selecting a density model from the set of at least two alternative density models for the at least one class based on, at least in part, constructing an error graph to capture confusability between classes, constructing a junction tree, and employing at least one desired characteristic;
- and
- triangulating an error graph to form cliques of the junction tree based on, at least in part, a plurality of potentials associated with at least one clique of the junction tree, wherein each potential of the plurality of potentials is based on values indexed by a model configuration for classes that appear in a clique.
A classifier optimization system comprising a memory having stored therein computer executable components and a processor that executes the following computer executable components:
- means for discriminatively selecting a density model from a plurality of density models for each class of a classification system having at least two classes;
- and
- means for optimizing a configuration of the density model by utilizing, at least in part, potentials associated with at least one clique of a junction tree, wherein the junction tree is formed by triangulating an error graph that captures confusability between at least two classes, and wherein each potential is based on a value for each of the possible configurations of variables of a clique associated with the potential.
Dependencies are inferred from the claim text.