Original SOINN

22
An incremental network for on-line unsupervised classification and topology learning Shen Furao Osamu Hasegawa Neural Networks, Vol.19, No.1, pp.90-106, (2006)

Transcript of Original SOINN

Page 1: Original SOINN

An incremental network for on-line

unsupervised classification and

topology learning

Shen Furao Osamu Hasegawa

Neural Networks, Vol.19, No.1, pp.90-106, (2006)

Page 2: Original SOINN

Background: Objective of unsupervised learning (1)

Clustering: Construct decision boundaries

based on unlabeled data.

– Single-link, complete-link, CURE

• Computation overload

• Much memory space

• Unsuitable for large data sets or online data

– K-means: • Dependence on initial starting conditions

• Tendency to result in local minima

• Determine the number of clusters k in advance

• data sets consisting only of isotropic clusters

Page 3: Original SOINN

Background: Objective of unsupervised learning (2) Topology learning: Given some high-dimensional data

distribution, find a topological structure that closely reflects the topology of the data distribution

– SOM: self-organizing map

• predetermined structure and size

• posterior choice of class labels for the prototypes

– CHL+NG: competitive Hebbian learning + neural gas

• a priori decision about the network size

• ranking of all nodes in each adaptation step

• use of adaptation parameter

– GNG: growing neural gas

• permanent increase in the number of nodes

• permanent drift of centers to capture input probability density

Page 4: Original SOINN

Background: Online or life-long learning

Fundamental issue (Stability-Plasticity Dilemma): How can

a learning system adapt to new information without

corrupting or forgetting previously learned information

– GNG-U: deletes nodes which are located in regions of

a low input probability density

• learned old prototype patterns will be destroyed

– Hybrid network: Fuzzy ARTMAP + PNN

– Life-long learning with improved GNG: learn number

of nodes needed for current task

• only for supervised life-long learning

Page 5: Original SOINN

Objectives of proposed algorithm • To process the on-line non-stationary data.

• To do the unsupervised learning without any priori

condition such as:

• suitable number of nodes

• a good initial codebook

• how many classes there are

• Report a suitable number of classes

• Represent the topological structure of the input

probability density.

• Separate the classes with some low-density overlaps

• Detect the main structure of clusters polluted by noises.

Page 6: Original SOINN

Proposed algorithm

Input

pattern

First Layer

Growing

Network

First

Output

Second Layer

Growing

Network

Second

Output

Insert

Node

Delete

Node Classify

Page 7: Original SOINN

Algorithms

• Insert new nodes

– Criterion: nodes with high errors serve as a criterion to insert a new node

– error-radius is used to judge if the insert is successful

• Delete nodes

– Criterion: remove nodes in low probability density regions

– Realize: delete nodes with no or only one direct topology neighbor

• Classify

– Criterion: all nodes linked with edges will be one cluster

Page 8: Original SOINN

Experiment • Stationary environment: patterns are randomly chosen

from all area A, B, C, D and E

• NON-Stationary environment:

Environment

I II III IV V VI VII

A 1 0 1 0 0 0 0

B 0 1 0 1 0 0 0

C 0 0 1 0 0 1 0

D 0 0 0 1 1 0 0

E1 0 0 0 0 1 0 0

E2 0 0 0 0 0 1 0

E3 0 0 0 0 0 0 1 Original Data Set

Page 9: Original SOINN

Experiment: Stationary environment

Original Data Set Traditional method: GNG

Page 10: Original SOINN

Experiment: Stationary environment

Proposed method: first layer Proposed method: final results

Page 11: Original SOINN

Experiment: Non-stationary environment

GNG-U result GNG result

Page 12: Original SOINN

Experiment: Non-stationary environment

Proposed method: first layer

Page 13: Original SOINN

Proposed method: first layer

Experiment: Non-stationary environment

Page 14: Original SOINN

Experiment: Non-stationary environment

Proposed method: first layer Proposed method: Final output

Page 15: Original SOINN

Experiment: Non-stationary environment

Number of growing nodes during online learning

(Environment 1 ~ Environment 7)

Page 16: Original SOINN

Experiment: Real World Data

Facial Im

age

(AT

T_FA

CE

)

(a) 10 classes

(b) 10 samples of class 1

Page 17: Original SOINN

Experiment:Vector

Vector of (a)

Vector of (b)

Page 18: Original SOINN

Experiment: Face Recognition results

10 clusters

Stationary

Correct

Recognition

Ratio: 90%

Non-Stationary

Correct

Recognition

Ratio: 86%

Page 19: Original SOINN

Experiment: Vector Quantization

Original Lena (512*512*8) Stationary Environment: Decoding

image, 130 nodes, 0.45bpp,

PSNR = 30.79dB

Page 20: Original SOINN

Experiment: Compare with GNG

Number

of Nodes bpp PSNR

First-layer 130 0.45 30.79

GNG 130 0.45 29.98

Second-layer 52 0.34 29.29

GNG 52 0.34 28.61

Stationary Environment

Page 21: Original SOINN

Experiment: Non-stationary Environment

First-layer: 499 nodes, 0.56bpp,

PSNR = 32.91dB

Second-layer: 64 nodes, 0.375bpp,

PSNR = 29.66dB

Page 22: Original SOINN

Conclusion

• An autonomous learning system for

unsupervised classification and topology

representation task

• Grow incrementally and learn the number of

nodes needed to solve current task

• Accommodate input patterns of on-line non-

stationary data distribution

• Eliminate noise in the input data