Back to results
US 8,943,005 B2Granted
Metric learning apparatus
AssigneeKabushiki Kaisha Toshiba
InventorTomoyuki Shibata
Life of this patent
Estimating the term…
Loading patent…
Loading patent…
AssigneeKabushiki Kaisha Toshiba
InventorTomoyuki Shibata
Estimating the term…
A metric learning apparatus memorizes a learning pattern in a feature space and a category which the learning pattern belongs to, performs variable transformation of the learning pattern to a metric space by a transformation matrix, calculates a transformation matrix having a minimum loss value of a loss function in which the loss value is increased when there is a learning pattern belonging to a different category but closer than learning patterns up to kth closest learning pattern in an aggregate of the learning pattern belonging to the same category on the basis of the k-nearest neighbor algorithm and belonging to the same category, and determines that the metric learning is terminated and outputs the transformation matrix corresponding to the minimum loss value when the minimum loss value is at least equal to or smaller than a threshold value.
A metric learning apparatus comprising:
Dependencies are inferred from the claim text.
Drawings are temporarily unavailable. Open the PDF at USPTO
From the US filing history. Foreign family members aren’t listed.
No references of this kind are recorded.
Machine learning
Using statistics or function optimisation, e.g. modelling of probability density functions: with fixed number of clusters, e.g. K-means clustering
Based on distances to training or reference patterns: distances to closest patterns, e.g. nearest neighbour classification
Classification techniques: relating to the decision surface
About 4 years 1 month of term left · ends 12 Nov 2030 (estimated)
The present invention relates to a metric learning technology using a k-nearest neighbor algorithm.
As disclosed in Patent Documents 1 to 4 and Non-patent Document 1, a metric learning apparatus of the related art using a classification on the basis of k-nearest neighbor algorithm improves accuracy of the k-nearest neighbor classification by learning a metric in which the distance between patterns in the same category is small and the distance between patterns in different categories is large.
Patent Document 1: U.S. Unexamined Patent Application Publication No. 2008/0199075
FIG. 1(a) is a drawing showing a search range in a method of the related art, and
FIG. 1(b) is a drawing showing a search range according to a first embodiment.
FIG. 2 is a block diagram showing a configuration of a metric space learning apparatus according to the first embodiment of the present invention.
Referring now to FIG. 1 to FIG. 3, a metric learning apparatus 1 according to a first embodiment of the present invention will be described.
First of all, the theory of the first embodiment will be described in comparison with a method of the related art.
“K-nearest neighbor algorithm” is a method of classification of statistics on the basis of a nearest example of learning in a feature space, and is used in pattern recognition. The classification of a certain pattern is determined by the votes of pattern groups nearby. In other words, the most common category in k-nearest pattern groups is assigned to the corresponding pattern. The symbol “k” is a positive integer, which is generally a small integer. For example, if k=1, the corresponding pattern is simply classified in the same category as the nearest pattern.