Information Technology Reference
In-Depth Information
h is SOM training algorithm produces a network where nodes in SOM represent
the structure of normal and abnormal samples for training, and nodes are labeled
on the basis of this representation. Most importantly, the fi rst phase uses NS algo-
rithm for generating normal space and this normal space is used to train an SOM
to produce a map that only refl ects the self-space leaving the nonself space.
During the second phase, primary response, if unlabeled nodes are present, the
network can classify them as normal or abnormal on the basis of fi nding the node,
which is present near the input (the winner code) where the inputs are classifi ed on
the basis of the label of this node. Node labels are assigned by fi nding the closest
labeled sample for each node and assigning the sample's labels for the correspond-
ing node. h is strategy would generalize k -nearest neighbor classifi cation.
In the third phase, secondary response, new unlabeled nodes are classifi ed on
the basis of winner node. In this phase, the network is supposed to work more pre-
cisely by producing a specifi c response and by identifying more accurately about the
normal or specifi c kind of anomaly.
h e fi rst stage is executed only once, but the second and third phases are repeated
until there are new samples available. A visual representation of the feature space
can be generated by a two-dimensional (2-D) grid corresponding to the network
and by assigning diff erent colors to each node depending on the category it repre-
sents. González et al. (2005) experimentally showed that this model can capture
the structure of the normal samples used for training and by the third phase, it was
seen that this model incredibly improved the discrimination by the a nity matura-
tion process that is carried on the second phase of the model.
6.6 Summary
h is chapter fi rst discusses the DT, then DT- and DC-based methods are described.
Other recent works presented include a multilevel immune algorithm, MHC-based
approaches, and cytokine networks. Recent works in immunology show that as
the antigen evolve toward imitating self-molecules, antigens became invisible to
the antibodies' defense mechanisms pointing to the necessity of other means of
protection probably constituted by T cells (Dasgupta, 2007). Moreover, the phy-
logenies of immune system (Warr and Cohen, 1991), the evolutionary development
of immune functions play an important role in keeping diversity. h e last section
discusses a hybrid approach, which combines NS and neural network methods to
design a classifi cation algorithm.
6.7 Review Questions
1. Explain the main ideas behind DT.
2. How can a danger zone be defi ned?
3. Illustrate with diagrams self-nonself (SNS), infection/noninfectious (INS),
and danger signal models.
Search WWH ::




Custom Search