Condensed SVM
The present invent ion provides a condensed SVM for high-speed learning using a large amount of training data. A first stage WS selector samples a plurality of training data from a training data DB, selects an optimal training vector x t among the plurality of training data, and outputs it to the WS manager. After the first stage finishes, a second stage WS selector extracts training data one by one from the training data DB and selects training data x t satisfying optimality and outputs it to the WS manager. An SVM optimizer extracts training data closest to the training data x t selected by the first and second stage WS selectors from the WS being managed by the WS manager, and condenses the two first and second training data to one training data when the distance between these is smaller than a predetermined value.
1. A condensed SVM comprising:
a training database having large training data;
selecting training data means for repeatedly selecting a plurality of training data from the training database and obtaining one optimal training vector among the plurality of training data in a first stage;
extracting training data means for extracting training data one by one from the training database and selecting training data satisfying optimality after the first stage finishes;
managing training data means for managing the training data selected by said selecting training data means and said extracting training data means;
optimizing means for extracting a second training data closest to a first training data selected by said selecting training data means and said extracting training data means from a working set (WS) managed by said managing training data means, and condensing the first and second training data to one training data when the distance between the first and second training data is smaller than a predetermined value; and
condensing means for condensing the two first and second training data to one training data that obtains a condensed vector z, a coefficient β, and a parameter D from the following formula: when the first and second training data are x i and x j , coefficients are ∝ i and ∝ i , and parameters are C i and C j ,
z
=
C
i
x
i
+
C
j
x
j
C
i
+
C
j
,
β
=
∝
i
K
(
z
,
x
i
)
+
∝
j
K
(
z
,
x
j
)
K
(
z
,
z
)
,
D
=
C
i
K
(
z
,
x
i
)
+
C
j
K
(
z
,
x
j
)
K
(
z
,
z
)
,
where K is a kernel function.
2. A condensed SVM comprising:
a training database having large training data;
selecting training data means for repeatedly selecting a plurality of training data from the training database and obtaining one optimal training vector among the plurality of training data in a first stage;
extracting training data means for extracting training data one by one from the training database and selecting training data satisfying optimality after the first stage finishes;
managing training data means for managing the training data selected by said selecting training data means and said extracting training data means;
optimizing means for extracting a second training data closest to a first training data selected by said selecting training data means and said extracting training data means from a working set (WS) managed by said managing training data means, and condensing the first and second training data to one training data when the distance between the first and second training data is smaller than a predetermined value; and
condensing means for condensing the two first and second training data to one training data that obtains a condensed vector z, a coefficient β, and a parameter D from the following formula: when the first and second training data are x i and x j , coefficients are ∝ i and ∝ i , and parameters are C i and C j ,
z
=
C
i
x
i
+
C
j
x
j
C
i
+
C
j
,
β
=
∝
i
K
(
z
,
x
i
)
+
∝
j
K
(
z
,
x
j
)
K
(
z
,
z
)
,
D
=
C
i
K
(
z
,
x
i
)
+
C
j
K
(
z
,
x
j
)
K
(
z
,
z
)
,
where K is a kernel function,
wherein said optimizing means further inspects whether non-support data exists in the WS when the distance between the first and second training data is larger than the predetermined value, and when non-support data exists, deletes the non-support data.