Spherical Topology Self- Organizing Map (SOM) Neural Networks for Visualization …courses.cecs.anu.edu.au/courses/CS_PROJECTS/11S2/Final... · 2012-01-03 · Spherical Topology Self-
Post on 28-May-2020
8 Views
Preview:
Transcript
Spherical Topology Self-Organizing Map (SOM) Neural Networks for Visualization of
Complex Data
Huajie Wu u4932885 Supervisor: Professor Tom Gedeon
Proposed by Kohonen in 1982
Regular Arrangement: 1D and 2D
But 1D and 2D have a “border effect”.
2
Therefore, many spherical SOMs are proposed H-SOM(proposed by Hirokazu in 2006), S-SOM (proposed by Sangole and Leontitsis in 2006).
H-SOM arrange the neurons along a helix, but calculating neighbors is quite complicated. My project CSSOM, mainly based on S-SOM.
3
The arrangement & neighborhood structure: Predefined grid units The number of vertices can be calculated as:
NN=2+10*4n
The representations of distortions and colors: Two significantly different colors Two clumps between the
boundaries of the colors
4
S-SOM has limitations. Arrangement and representation of a sequence of events. Concentric spherical SOM solved those problems. CSSOM Structure: ‘a’ in sphere1 is the winning neuron
5
Plot the spheres in form of a chain (Plot ‘Chain’ Glyph) Analyze a specified sphere and adjacent. Good for analyzing the spheres as a whole.
6
Plot an arbitrary continuous equal-size spheres (Plot ‘Equal’ Glyph)
Solved the problem in displaying “Plot ‘Chain’ Glyph” schema.
Provide users optimal visual effects.
7
The difference between parallel training and sequence training
8
Spheres Times S1 S2 S3
1 P2 P2 P2 2 P3 P3 P3
3 P1 P1 P1 4 P5 P5 P5
5 P4 P4 P4
… … … …
Spheres Times S1 S2 S3
1 P1 P2 P3 2 P2 P3 P4
3 P3 P4 P5 4 P4 P5 P1
5 P5 P1 P2
… … … …
Quantization Error and Topological Error:
QE evaluate how well the neural network map fits the input patterns. Average distance between vectors and BMU.
TE is used to measure the topology preservation. Proportion of all data vectors for 1st and 2nd BMU are not adjacent.
9
The quality of SOMs
Focus on comparing between SSOM and CSSOMs. “ECSH” dataset is used. 3,641 patterns. Read paragraphs in Easy, Calm, Stressful, Hard order. It reflects 60Hz recordings and spans one minute. The representations of SOMs
10
SOMs structure The number of spheres Units/per layer Total units
S-SOM 1 2,562 2,562 CSSOM(4S) 4 642 2,568 CSSOM(16S) 15 162 2,592 CSSOM(61S) 61 42 2,562 CSSOM(214S) 214 12 2,568
Continued…
1498
376
684
1595
1267
0200400600800
10001200140016001800
Average number of neighborhoods per unit
SSOM
CSSOM(4S)
CSSOM(15S)
CSSOM(61S)
CSSOM(214S)
30
7 4 2 1 0
5101520253035
The number of units with different neighborhoods
the number of unitswith differentneighbors
11
Figure 1. Figure 2.
Table 1: QE and TE of SSOM and CSSOM using dataset “ECSH”
ERROR SSOM CSSOM(4S) CSSOM(16S) CSSOM(61S) CSSOM(214S)
QE 193.77 381.71 188.25 587.73 426.14 TE 0.101 0.328 0.179 0.092 0.105
Parallel training and sequence training
“ECSH” dataset is used. The results are not expected.
12
Table 1: QE and TE of Parallel Training and Sequence Training using dataset “ECSH”
ERRORS Parallel Training Sequence Training
QE 188.25 1511.52
TE 0.179 0.496
Comparisons with parallel training and sequence training in visualization
Figure 1:Visual glyph for “Parallel Training”
Figure 2:Visual glyph for “Sequence Training” 13
Fulfilled the objectives
Implemented working CSSOM Evaluation code to test Sequence training needs more work e.g. For each epoch, the first pattern selected randomly for training, but patterns’ order (time series) still kept sequential.
Other future work More diverse dataset R and epochs are adjustable automatically, depending on the dataset’s characteristics…
14
top related