Databases Reference
In-Depth Information
FIGURE 4.11: Results for a 1-bit distortion pattern of character “A.” The
first HGN subnet shows that a new subpattern has been found (with assigned
index 0), whereas other compositions correctly recall this as the pattern as-
sociated with index 1 (bitmap pattern of “A” ). (With kind permission from
John Wiley & Sons, Inc.: Mobile Intelligence, “An Online Scheme for Threat
Detection Within Mobile Ad Hoc Networks,” pp. 380-411, 2010, Khan, A.
I. and Muhamad Amin, A. H. and Raja Mahmood, R. A., Figure 17.15,
http://dx.doi.org/10.1002/9780470579398.ch17.)
identified, the active top neuron outputs the index value 0. Otherwise, the
recalled index of the subpattern will be output. Figure 4.11 shows the result
of a 1-bit distorted character pattern “A” introduced to the network after the
character patterns “A,”“I,”“J,”“S,”“X,” and “Z” have been stored.
Figure 4.11 shows that only one of the subnets records the subpattern as
a new pattern. Other subnets recall the index value of 1, which is the index
for the stored character pattern “A”. The decision of whether the pattern is
a recall or store is made based on the cumulative decisions of the distributed
HGN subnets using the recall value. Equation 4.13 shows the formula for
the recall, R
c
of a distributed HGN scheme with s subnets. Note that n
(r,i)
represents the neurons that produce an index that is similar to the index of
the targeted pattern class, r and n
(t,i)
represents each neuron in subnet i.
s
i=1
n
(r,i)
s
i=1
n (t, i)
R
c
=
(4.13)
Using the example from Figure 4.11, the recall value for a 1-bit distortion
pattern of character “A” is (4 + 121 + 16) ÷ (16 + 121 + 16) = 141 ÷ 153 =
0.9216. Therefore, its recall percentage is 92.16%.
The distribution of patterns into multiple HGN subnets might improve the
recall accuracy of the scheme. According to [64], the recall percentages of 1-bit
distorted patterns are significantly higher in the distributed HGN approach
than the HGN. This behavior is attributed to the encapsulation effect of the
distributed HGN, i.e., the effects of a distortion in a particular subnet do
not affect the other subnets. Figure 4.12 shows the encapsulation effect. It
also shows the internal state of the subnets from the 1-bit distorted pattern
of character “A.” The effects of the distortion are limited to the subnet that
