Machine Recognition in Complex Domain - High Dimensional Neurocomputing: Growth, Appraisal and Applications - page 134

7.1.2.3 ACONN: All-Classes-in-One-Neural Network
An extensive survey into the litrature of the neural network tool for pattern classi-
fication point out the fact that the first architecture in a way as to ape the network
arrangement in the classifier is one single neural network having as many outputs
as the number of classes in the database. Thus, first structure viz. All-Classes-in-
One-Neural Network (ACONN) is designed to classify all the classes (subjects) in a
database through single huge network.
7.1.2.4 OCONN: One-Class-in-One-Neural Network
In the second architecture as a tool for pattern classification, an ensemble of neural
networks is used, where each network is dedicated to recognize one particular class
in the database. Thus, the second structure viz One-Class-in-One-Neural Network
(OCONN), there are asmany neural networks as the number of classes in the database.
It has been experimentally verified in the literature [ 50 , 52 , 53 ] that the second
classifier structure achieves better performance in comparison to the first structure.
It leads to flexible structure of classifier, as a new network may easily be added when
a new subject arrives in the database for recognition.
Neural networks in a complex domain are becoming very attractive for solving many engi-
neering problems. For a problem of same complexity, one needs a smaller network topology
and lesser training time to yield better accuracy in comparison to equivalent neural networks
in a real domain. Various experiments presented in this chapter on face recognition explore
the computational power of the neurons in complex domain presented in Chap. 4 .
7.1.2.5 OCON: One-Class-in-One-Neuron
The pattern recognizer (classifier) presented in this chapter is designedwith an ensem-
ble of higher-order neurons instead of ensemble of neural network. Therefore, each
such neuron in complex domain is sufficiently powerful so that can take the responsi-
bility for the recognition of individual subject of the database, hence lead a new kind
of classifier, named as One-Class-in-One-Neuron (OCON). In this third architecture
as a tool for pattern classification, each neuron is dedicated to recognize one partic-
ular class in the database. It lead to most compact but flexible structure of classifier
where a new neuron may easily be added as a new subject arrives in the database for
recognition.
So far, as implementation is concerned, a neuron in OCON or a network in
OCONN is created with all the training samples of that class as positive example,
termed as class-one; and the samples of other class as negative example, constitute
the class-two. Therefore, recognition problem is a two class partitioning problem.
The structure of each component (neuron or network) of classifier remains same for
all classes and only weights wary. The weights of trained classifier are kept in a file