You are on page 1of 1

Adaptive Multilayer Networks

Network Building Algorithms: Constructive Algorithms


Neural Tree Algorithm

 Developed by Sirat and Nadal in 1990.


 The learning algorithm is inspired from a growth algorithm, the tiling algorithm.
 As in the former case, this is a constructive algorithm, for which convergence is guaranteed.
 The Neural Tree Network Organization can be viewed in 2 ways:
o The Structural Organization
o The Functional Organization
 The Structural Organization of the Tree as Decision Tree:
o The tree is given inputs that are patterns containing feature values, like we have been
considering in all algorithms.
o These inputs are fed to a ‘Decision Tree’. A decision tree is a Binary Tree. I.e., it has one root
node and can have two child nodes, one left and one right. These child nodes can further
have two child nodes, a left child and a right child. Therefore, the root node has a left sub-
tree and a right sub-tree.
o Each of the nodes in this decision tree is a Decision Node. Each decision node asks a
question. The output of each decision node is either a yes, or a no (0 or 1) as an answer to
this question.
o Based on this yes or no decision, either the left or the right sub-tree is selected.
o The inputs are then fed to the root node of the selected subtree. This node again asks a
yes/no question and the answer selects the next sub-tree. This continues till the answer is
the final class to which the input belongs.
 Consider the following diagram to understand the Decision Tree approach.

 The functional organization of the Tree as Neural Tree:


o The Neural Tree ‘Network’ consists of a set of perceptrons. These perceptrons are not
structurally connected; instead, they are functionally organized as a binary neural tree.
o In the neural tree, each neuron of a neural tree receives inputs from, and only from, the
input layer; its output does not feed into any other neuron, but is used to propagate down
the decision tree.
o Each uses unipolar binary signum activation function in the perceptrons.

 The main advantage of this approach is due to the local processing in restricted portions of input
space, during both learning and classification. Moreover, only a small subset of neurons have to be
activated during the classification stage.

You might also like