Information Technology Reference
In-Depth Information
Table 3.
Classification of sonar targets by neural network
Hidden units
% Right on Training Set
% Right on Test Set
0
79.3
73.1
2
96.2
85.7
3
98.1
87.6
6
99.4
89.3
12
99.8
90.4
24
100.0
89.2
According to [1], the authors of this work [9] further report that a nearest neighbor
classifier on the same data gave an 82.7% probability of correct classification.
The immunochip emulator for intrusion detection has also been trained and tested
by the "aspect-angle dependent" sets. Classification results of the immunochip
emulator using only the 1st stage training (see Section 3) are shown in Tab. 4.
Table 4.
Classification of sonar targets by immunochip emulator
Dimension of SFIN
(
q
)
Training time
(s)
% Right on
Training Set
% Right on
Test Set
Total
Errors
3
0.02
100.0
76.9
24
5
0.03
100.0
84.6
16
7
0.03
100.0
89.4
11
8
0.06
100.0
90.3
10
9
0.08
100.0
93.2
7
10
0.08
100.0
92.3
8
Brief comparison between Tab. 3 and Tab. 4 shows that the best classification of
the immunochip emulator (93.2%) is better than that of the neural network (90.4%).
Besides, the emulator does not make mistakes on the training set (this is guaranteed
by Proposition 2).
Note very low training time of the emulator in Tab. 4 (for AMD Athlon 1.53 GHz).
Unfortunately, the training time of neural network in Tab.3 is unavailable from [1] or
[9]. However, it can be estimated indirectly by the work [22], which uses the same
sonar benchmark.
The authors of this work [22] report 58 s or 72 s (for Pentium 350 MHz) for their
genetic algorithm applied to the neural networks with 3 or 4 hidden units respective
and note that "This method is efficient because the time cost for evolution is about 2
or 3 orders less than that spent in training the networks." Such estimation confirms
that the training time of the immunochip emulator is far lower than that of neural
networks.
7 Hardware Implementation
A perspective way of hardware implementation of the immunochip can be provided
by DSP of new TigerSHARC family. Such DSP is compatible with the standard PC,
where it can be connected via PCI bus. Therefore, a hardware emulator of the
