This presentation introduces the Self-Organizing Incremental Neural Network (SOINN), which can represent the topological structure of input data through online incremental learning without predefined network structure or size. It then discusses Self-Organizing Maps (SOM) and issues with other related algorithms. The detailed SOINN algorithm is presented, including adaptive threshold updating, weight updates, and handling new data. An Adjusted SOINN Classifier (ASC) is also introduced for supervised learning tasks. Experimental results demonstrate SOINN's ability to learn prototypical representations of artificial and real-world data.
2. Contents of this presentation
What is SOINN
Why SOINN
SOM
Detail algorithm of SOINN
ASOINN
F.Noorbehbahani - Isfahan University of Technology - Agust 2013
2
3. What is SOINN
SOINN: Self-organizing incremental neural network
Represent the topological structure of the input data
Realize online incremental learning
F.Noorbehbahani - Isfahan University of Technology - Agust 2013
3
4. Background: Networks for topology
representation
SOM(Self-Organizing Map): predefine structure and size of the
network
NG(Neural Gas): predefine the network size
GNG(Growing Neural Gas): predefine the network size; constant
learning rate leads to non-stationary result.
F.Noorbehbahani - Isfahan University of Technology - Agust 2013
4
5. Self-Organizing Maps (SOMs)
Self-Organizing Map (SOM) is an unsupervised learning algorithm.
SOM is a visualization method to represent higher dimensional data in
an usually 1-D, 2-D or 3-D manner.
SOMs have two phases:
Learning phase: map is built, network organizes using a competitive process
using training set.
Prediction phase: new vectors are quickly given a location on the
converged map, easily classifying or categorizing the new data.
F.Noorbehbahani - Isfahan University of Technology - Agust 2013
5
13. Characteristics of SOINN
Neurons are self-organized with no predefined network structure
and size
Adaptively find suitable number of neurons for the network
Realize online incremental learning without any priori condition
Find typical prototypes for large-scale data set
Robust to noise
F.Noorbehbahani - Isfahan University of Technology - Agust 2013
13
14. Detail algorithm of
SOINN
• Two-layer competitive network
• First layer: Competitive for input
data
• Second layer: Competitive for
output of first-layer
• Output topology structure and
weight vector of second layer
F.Noorbehbahani - Isfahan University of Technology - Agust 2013
14
15. Training flowchart of
SOINN
• Adaptively updated threshold
• Between-class insertion
• Update weight of nodes
• Within-class insertion
• Remove noise nodes
F.Noorbehbahani - Isfahan University of Technology - Agust 2013
15
16. First layer: adaptively updating
threshold Ti
F.Noorbehbahani - Isfahan University of Technology - Agust 2013
16
18. Artificial data set: topology
representation
F.Noorbehbahani - Isfahan University of Technology - Agust 2013
18
19. SOINN for supervised learning (ASC)
Automatically learn the number of prototypes needed to represent
every class
Only the prototypes used to determine the decision boundary will
be remained
Realize both types of incremental learning
Robust to noise
F.Noorbehbahani - Isfahan University of Technology - Agust 2013
19