1 of 20

Mahsa Famil Barraghi

Department of Computer Science

University of Tabriz, Iran

ICAISV-2023

Habib Izadkhah

Department of Computer Science

University of Tabriz, Iran

Deep Parallel Self-Organizing Maps

for Visual Classification

2 of 20

SELF

ORGANIZING

MAP

Related

work

Deep self-

organizing

maps

Experimental

results

& discussion

1/18

Concolusion

Content

Deep Self-Organizing Maps for Visual Classification Mahsa Famil Barraghi & Habib Izadkhah

3 of 20

2/18

Definition

Self-Organizing Map

Better capability of revealing the overlapping structure in clusters

Kohonen 1982

Visualize high-dimensional data sets

Deep Self-Organizing Maps for Visual Classification Mahsa Famil Barraghi & Habib Izadkhah

4 of 20

3/18

Algorithm

Self-Organizing Map

SOM Algorithm

Initialize all weights of the SOM

Select a vector x randomly in the

training set.

Find the best matching unit BMU

Update the weight vector of BMU

and its neighboring

Deep Self-Organizing Maps for Visual Classification Mahsa Famil Barraghi & Habib Izadkhah

5 of 20

4/18

Liu, Wang and Yong 2015, Wickramasinghe,

Amarasinghe and Manic 2019

Related Works

Deep Self-Organizing Maps for Visual Classification Mahsa Famil Barraghi & Habib Izadkhah

6 of 20

5/18

Architecture

DeepSOM

Deep Self-Organizing Maps for Visual Classification Mahsa Famil Barraghi & Habib Izadkhah

7 of 20

6/18

Detail

DeepSOM

Deep Self-Organizing Maps for Visual Classification Mahsa Famil Barraghi & Habib Izadkhah

8 of 20

7/18

Training

DeepSOM

Combines information of maps.

The array is reshaped to a 2D grid

Step 1: Weight initialization

Step 2: SOM Layer

Step 3: Sampling Layer:

This process is repeated until the last layer

Step 4: Repeat Steps 2-3

Steps are repeated until the maximum number of iterations (epochs) has been reached

Step 5: Repeat 2-4:

Weights of the network are randomly initialized.

 

Deep Self-Organizing Maps for Visual Classification Mahsa Famil Barraghi & Habib Izadkhah

9 of 20

8/18

Parameters

Results

Layers

Kernel size

Stride

SOM map Size

Hidden Layer 1

10×10

2

Changeable

Hidden Layer 2

6×6

1

Changeable

Output Layer

5×5

1

8×8

model

Map sizes of the first layer

Map sizes of the second layer

1

20×20

14× 14

15× 15

14× 14

2

22× 22

16× 16

15 ×15

14 ×14

3

22 ×22

14 ×14

16 ×16

14 ×14

4

24 ×24

16 ×16

16 ×16

14 ×14

MNSIT

  • Significantly smaller training set of 3000 images
  • Test set of 1000 images

Deep Self-Organizing Maps for Visual Classification Mahsa Famil Barraghi & Habib Izadkhah

10 of 20

9/18

Accuracy

Results

Model

Map sizes of the first layer

Map sizes of the second layer

Train accuracy

Test accuracy

1

20×20

14× 14

15× 15

14× 14

84.47

81.56

2

22× 22

16× 16

15 ×15

14 ×14

85.75

82.26

3

22 ×22

14 ×14

16 ×16

14 ×14

87.81

84.19

4

24 ×24

16 ×16

16 ×16

14 ×14

86.63

83.53

Model

Map sizes

 

Train accuracy

Test accuracy

1

8×8

62.23

61.70

2

12×12

63.20

62.56

3

16×16

65.43

65.03

4

20×20

66.97

66.35

5

24×24

69.45

68.89

Deep Self-Organizing Maps for Visual Classification Mahsa Famil Barraghi & Habib Izadkhah

11 of 20

10/18

Tools

Visual Data Mining

Hit maps

U-Matrix

Class label distribution

Deep Self-Organizing Maps for Visual Classification Mahsa Famil Barraghi & Habib Izadkhah

12 of 20

11/18

U-Matrix

Visual Data Mining

Yellow areas

mean that they

are far apart

It shows the distance between weight vectors of neighboring neurons

Dark blue area, means that the nodes in this area are very

similar

Deep Self-Organizing Maps for Visual Classification Mahsa Famil Barraghi & Habib Izadkhah

13 of 20

12/18

U-Matrix

Visual Data Mining

Deep Self-Organizing Maps for Visual Classification Mahsa Famil Barraghi & Habib Izadkhah

14 of 20

13/18

Hit maps

Visual Data Mining

Based on the number of times a particular neuron becomes a winning neuron for a particular class

How often a neuron is selected as the winning neuron

Deep Self-Organizing Maps for Visual Classification Mahsa Famil Barraghi & Habib Izadkhah

15 of 20

14/18

Hit maps

Visual Data Mining

Deep Self-Organizing Maps for Visual Classification Mahsa Famil Barraghi & Habib Izadkhah

16 of 20

15/18

Class label distribution

Visual Data Mining

The similar the two numbers are to each other, the closer they get to each other on the map

Numbers that are similar to each other are placed close together

Deep Self-Organizing Maps for Visual Classification Mahsa Famil Barraghi & Habib Izadkhah

17 of 20

16/18

Class label distribution

Visual Data Mining

Deep Self-Organizing Maps for Visual Classification Mahsa Famil Barraghi & Habib Izadkhah

18 of 20

17/18

Concolusion

It is easy to use and understand compared to other deep models.

It improves the accuracy of simple SOM by 23.65%.

It improves accuracy by 27.5% compared to the three-layer convolutional method.

Two different map sizes creates a balance between over-fitting and under-fitting.

It improves the accuracy of the model of Wickramasinghe by 1.09%

Deep Self-Organizing Maps for Visual Classification Mahsa Famil Barraghi & Habib Izadkhah

19 of 20

18/18

References

[1] T. Kohonen.Self-Organizing Maps.Book published by Springer-Verlag Berlin Heidelberg New York. (2001)

[2] D. Alahakoon, S. Halgamuge, B. Srinivasan. Dynamic self-organizing maps with controlled growth for knowledge discovery. IEEE Trans Neural Netw

11(3):601-14. doi: 10.1109/72.846732. (2000)

[3] A. Rauber, D. Merkl,. M. Dittenbach. The Growing Hierarchical Self-Organizing Map: Exploratory Analysis of High-Dimensional Data. IEEE TRANSACTIONS

ON NEURAL NETWORKS. (2002)

[4] A. Fonseka, D. Alahakoon. Exploratory data analysis with Multi-Layer Growing Self-Organizing Maps. Fifth International Conference on Information and

Automation for Sustainability. (2010)

[5] H. Dozono, G. Niina, S. atoru Araki. Convolutional Self Organizing Map. International Conference on Computational Science and Computational

Intelligence (CSCI). (2016)

[6] M. Sakkari, M. Zaied1. A Convolutional Deep Self-Organizing Map Feature extraction for machine learning. Springer Science Business Media, LLC, part of

Springer Nature (2020)

[7] S. Aly, S. Almotairi. Deep Convolutional Self-Organizing Map Network for Robust Handwritten Digit Recognition. IEEE 10.1109/ACCESS.2020

.3000829, Page(s): 107035 - 107045. (2020)

[8] L. Elend, O. Kramer. Self-Organizing Maps with Convolutional Layers. Springer Nature Switzerland AG WSOM 2019,AISC 976,pp. 23–32.(2020)

[9] N. Liu, J. Wang, Y. Yong. Deep Self-Organizing Map for Visual Classification. 978-1-4799-1959-8/15/ IEEE. (2015)

[10] C. Wickramasinghe, K, Amarasinghe, M. Manic. Parallalizable Deep Self-Organizing Maps for Image Classification. DOI

10.1109/TII.2019.2906083, IEEE (2019)

Deep Self-Organizing Maps for Visual Classification Mahsa Famil Barraghi & Habib Izadkhah

20 of 20

THANK YOU!

ICAISV-2023