Structural analysis and the principle of adaptive resonance in artificial neural networks

1.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
Structural analysis and the principle of
adaptive resonance in artificial neural
networks
Vadim Lutsiv*
*Vavilov State Optical Institute & Saint Petersburg University of Aerospace Instrumentation
1

2.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
The first artificial neural networks and other image
classification algorithms handled the images as
indivisible wholes
It was their weak point
2

3.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
Image structural description is very universal and robust
Roof
Wall
Window
Door
Robust structural description of buildings:
•Walls are somewhere below a roof
•Windows are somewhere inside a wall
•Door is somewhere inside a wall
•Door is somewhere aside the windows
•Door is somewhere below the windows 3

4.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
Structural assembling of patterns in the
“Neocognitron” neural network
A
Complex
cells
Simple
cells
Complex
cells
4

5.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
Fragment of the “Neocognitron” neural network
implementing a kind of structural matching and the
principle of adaptive resonance
5

6.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
Examples of noise suppression, signal correction,
and pattern recognition in the “Neocognitron” network
based on implicit structural matching and adaptive
resonance
6

7.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
Hopfield-Tank type artificial neural network for
structural matching the sets of structural elements
Additional features of element
Type of element
Orientation of
element
Ordinate of
element
Decision layer
(dynamic
correspondence
matrix)
Abscissa of
element
Feature
vector describing
separate structural
element
a) Initial image
6
7
8
5
9
11
12
4
3
13
2
1
b) Affine transformed
and distorted image
3
5
2
6
12
Second input layer
Single level neural structural classifier
1
7
4
First input layer
10
10
9
8
11
13
Example of correct
structural matching of
artificially built contours
7

8.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
Hierarchical Hopfild-Tank network for image
structural matching
1st local object in the second image
Higher level decision layer
1st local object
in the first image
mth local
object
First lower
level input layer
(for the first
input image)
divided to sublayers
nth local object in the second image
Top-down inhibition of
mistaken sublayers
Structural description of the 1st
local object
Structural elements of the
nth local object
Feature vector describing
the 3rd structural element
Second lower level input
layer divided to sub-layers
Lower level decision layer
divided to sub-layers
Inhibition links between different sub-layers (local objects)
8
in each row and column of the lower level decision layer

9.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
Correcting the lower level structural descriptions
based on the matching results in the higher
hierarchical level (using the principle of adaptive
resonance)
6
7
0
1
2
3
4
5
6
7
0
1
2
3
6
7
I-1
0
1
2
3
4
5
6
7
I
0
1
2
3
I+1
5 6 7a 0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
567
01234567
01234567
J-1
J-1
J
J+1
J
J+1
Moving the structural elements to
another group, giving a most
powerful correspondence sub-matrix
I-1
I
I+1
States of correspondence submatrices corresponding to a
correct structural description
9

10.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
The algorithm of teaching the multilayer
perceptrons was proposed by Hinton at al in 1986
It generated a new wave of neural networks
popularity
10

11.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
In 1998, LeCun at al proposed o prototype of
modern deep learning convolutional neural
networks
A principle of structural decomposition was implicitly implemented in
their network, but it was almost not noticed until 2012
11

12.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
In 2012, a deep convolutional neural network
of Hinton, Krizhevsky, and Sutskever won the
ImageNet competition
A principle of structural decomposition was implemented in their
network still insufficiently explicitly
12

13.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
The information about mutual spatial position
of detected image details was substantially
suppressed by max-pooling operation
The both image versions could be recognized as a face
13

14.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
At last, in 2017 Hinton proposed the capsule
networks implementing explicitly the principles
of structural analysis and adaptive resonance
14

15.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
New features of capsule network
U
15

16.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
By weighting the vector U corresponding to
detected nose by matrix W an expected position U
of the whole face is coded
U
16

17.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
Realization of adaptive resonance principle
• Capsule suppresses input from respective
previous capsule by scalar weighting with
coefficient C if the position of detected
structural element does not correspond to
expected position.
• The expected position of element and the
type of recognized object can be changed in
several iterations
V
17

18.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
Modified loss function applied in teaching of
network
18

19.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
Visualization of image recognition results
Decoding of object class
number to object image
Visualization of
recognition results
19

20.

Vavilov State Optical Institute
Saint Petersburg University of Aerospace Instrumentation
Thank you for kind attention
20