Database Reference
In-Depth Information
Fig. 3.1 Automation in an adaptive image/video retrieval system. The pseudo labeling is obtained
by feature space
F 2 , which is used to guide the adaptation of the retrieval system applied to the
feature space
F 1
dynamically generating a model of this topology as it parses the input space. This
results in a representation that tends not to suffer from nodes being trapped in
regions of low density [ 69 , 70 ].
3.2.2
Self-organizing Tree Map
In order to construct a suitable map, the SOTM offers two levels of adaptation:
weight and structure. Weight adaptation is the process of adjusting the weight vector
of the winning nodes. Structure adaptation is the process of adjusting the structure
of the network by changing the number of nodes and the structural relationships
between them. Given a training data set
N
i
T = {
v i }
1 , v i F
2 , the adaptation map
=
using the SOTM algorithm is summarized as follows:
Step 1. Initialization
Choose the root node w j N c
j
￿
1 with a randomly selected training vector from
=
T
, where N c is the total number of nodes currently allocated.
￿
Initialize learning parameters: H
(
0
)
and
ʱ (
0
)
.
Step 2. Similarity matching
￿
Randomly select a new feature vector v , and compute the Euclidean distance,
d to all currently existing nodes w j ,
j
=
1
,
2
,···,
N c :
w j )= v
w j
d
(
v
,
(3.1)
Step 3. Updating
Select the winning node, j , with minimum distance,
￿
d j =
min
j
d
(
v
,
w j )
(3.2)
Search WWH ::




Custom Search