Cellular Automata Machine For Pattern Recognition Pradipta Maji 1 Niloy Ganguly 2 Sourav Saha 1 Anup K Roy 1 P Pal Chaudhuri 1 1 Department of Computer Science & Technology , Bengal Engineering College ( D . U ) , Howrah , West Bengal , India 711103 2 Department of Business Administration , Indian Institute of Social Welfare and Business Management , Calcutta , West Bengal , India 700073
37
Embed
Cellular Automata Machine For Pattern Recognition Pradipta Maji 1 Niloy Ganguly 2 Sourav Saha 1 Anup K Roy 1 P Pal Chaudhuri 1 1 Department of Computer.
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Cellular Automata Machine For Pattern Recognition
Pradipta Maji1 Niloy Ganguly 2
Sourav Saha1 Anup K Roy1 P Pal Chaudhuri 1
1 Department of Computer Science & Technology , Bengal Engineering College ( D . U ) , Howrah ,
West Bengal , India 711103
2Department of Business Administration , Indian Institute of Social Welfare and Business Management , Calcutta ,
West Bengal , India 700073
CA Research Group (BECDU)
The Problem• Pattern Recognition - Study how
machines can learn to distinguish patterns of interest
• Conventional Approach - Compares input patterns with each of the stored patterns learn
A
B
C …
Z
Bookman Old StyleA
Comic Sans MS
CA Research Group (BECDU)
CA Research Group (BECDU)
The Problem
A
Comic Sans MS
A A BA
B
C …
Z
Bookman
old Style
Grid by Grid Comparison
CA Research Group (BECDU)
CA Research Group (BECDU)
The Problem
A A BGrid by Grid Comparison
0 0 1 00 0 1 00 1 1 11 0 0 11 0 0 1
0 1 1 00 1 1 00 1 1 01 0 0 11 0 0 1
No of Mismatch = 3
CA Research Group (BECDU)
CA Research Group (BECDU)
The Problem
A A BGrid by Grid Comparison
0 0 1 00 0 1 00 1 1 11 0 0 11 0 0 1
1 1 1 00 1 0 10 1 1 10 1 0 11 1 1 0
No of Mismatch = 9
CA Research Group (BECDU)
CA Research Group (BECDU)
The Problem• Time to recognize a pattern -
Proportional to the number of stored patterns ( Too costly with the increase of number of patterns stored )
Solution - Associative Memory Modeling
CA Research Group (BECDU)
CA Research Group (BECDU)
The Problem• Time to recognize a pattern -
Proportional to the number of stored patterns ( Too costly with the increase of number of patterns stored )
Solution - Associative Memory Modeling
A
B
C
CA Research Group (BECDU)
Transient
A
AA
AA
A
Transient
Transient
CA Research Group (BECDU)
Associative Memory• Entire state space - Divided into some
pivotal points.• State close to pivot - Associated with
that pivot.• Time to recognize pattern-Independent
of number of stored patterns.
A
B
CTransient
A
AA
AA
A
Transient
Transient
CA Research Group (BECDU)
CA Research Group (BECDU)
Associative MemoryTwo Phase : Learning and DetectionTime to learn is higherDriving a car Difficult to learn but once learnt it
becomes natural
A
B
CTransient
A
AA
AA
A
Transient
Transient
CA Research Group (BECDU)
CA Research Group (BECDU)
Associative Memory (Hopfield Net)
Densely connected Network - Problems to implement in Hardware
VLSI Domain • India under Prof. P Pal Chaudhuri• Late 80’s - Work at Indian
Institute of Technology Kharagpur• Late 90’s - Work at Bengal
Engineering College Deemed University, Calcutta
Book - Additive Cellular Automata Vol I, IEEE Press
CA Research Group (BECDU)
CA Research Group (BECDU)
Cellular Automata A computational Model with discrete
cells updated synchronously
………..
outputInput
Combinational Logic
Clock
From Left Neighbor
From Right Neighbor
0/1
2 - State 3-Neighborhood CA Cell
CA Research Group (BECDU)
CA Research Group (BECDU)
Cellular AutomataCombinational Logic can be of 256 typeseach type is called a rule
Each cell can have 256 different rules
………..
CA Research Group (BECDU)
98 236 226 107
4 cell CA with different rules at each cell
CA Research Group (BECDU)
State Transition Diagram
1
21163 135
12 15 14 4 810 7
9
0
9 15
613
7 12
3 14
11
5
2 8
1 4
10
0
CA Research Group (BECDU)
CA Research Group (BECDU)
Generalized Multiple Attractor CA
A
B
CTransient
A
AA
AA
A
Transient
Transient
The State Space of GMACA – Models an Associative Memory0100 1000
1010 0001
0101
0011
0010
0000
11010111
1100 1001
10110110
1110
1111
P1 attractor-1
P2 atractor-2
Rule vector:<202,168,218,42>
CA Research Group (BECDU)
CA Research Group (BECDU)
Generalized Multiple Attractor CA
0100 1000
1010 0001
0101
0011
0010
0000
11010111
1100 1001
10110110
1110
1111
P1 attractor-1
P2 atractor-2
Rule vector:<202,168,218,42>
CA Research Group (BECDU)
Pivot PointsPivot Points
Dist =3
Dist =1
The state transition diagram breaks into disjoint attractor basin Each attractor basin of CA should contain one and only one pattern to be learnt in its attractor cycle The hamming distance of each state with its attractor is less than that of other attractors.
CA Research Group (BECDU)
Synthesis of GMACA Reverse Engineering Technique Phase I: Random Generation of a set of directed Graph
Basin 1
0100 1000
0001
0010 0000
1110 11011011
0111
1111
Basin 2
Patterns to be learnt P1 = 0000 P2 = 1111
Number of bits of noise = 1
CA Research Group (BECDU)
1
0
CA Research Group (BECDU)
Synthesis of GMACA Reverse Engineering Technique
Phase II: State transition table from Graph
Basin 1
0100 1000
0001
0010 0000
CA Research Group (BECDU)
PresentState
NextState
01001000001000010000
00010001000100000010
CA Research Group (BECDU)
Synthesis of GMACA Reverse Engineering Technique
Phase II: State transition table from Graph
CA Research Group (BECDU)
PresentState
NextState
11101011110101111111
01110111011111111111
1110 11011011
0111
1111
Basin 2
CA Research Group (BECDU)
Synthesis of GMACA Reverse Engineering TechniquePhase III: GMACA rule vector from State transition table
CA Research Group (BECDU)
PresentState
NextState
11101011110101111111
01110111011111111111
PresentState
NextState
01001000001000010000
00010001000100000010
Basin 1 Basin 2
CA Research Group (BECDU)
Synthesis of GMACA Reverse Engineering TechniquePhase III: GMACA rule vector from State transition table
CA Research Group (BECDU)
Basin 1 Basin 2PrestState
NextState
010100001000000
00000
PrestState
NextState
111101110011111
11111
NEIGHBORHOODSTATE
111 110 101 100 011 010 001 000
Next State
CA Research Group (BECDU)
Synthesis of GMACA Reverse Engineering TechniquePhase III: GMACA rule vector from State transition table
CA Research Group (BECDU)
Basin 1 Basin 2PrestState
NextState
010100001000000
00000
PrestState
NextState
111101110011111
11111
NEIGHBORHOODSTATE
111 110 101 100 011 010 001 000
Next State
111 1
111 1
1
000 0
000 0
01
101 1
0
010 0
1
110 1
0
001 0
1
011 1
0
100 0
Rule 232
CA Research Group (BECDU)
Synthesis of GMACA Reverse Engineering TechniquePhase III: GMACA rule vector from State transition table
CA Research Group (BECDU)
PresentState
NextState
11101011110101111111
01110111011111111111
PresentState
NextState
01001000001000010000
00010001000100000010
Basin 1 Basin 2
NEIGHBORHOODSTATE
111 110 101 100 011 010 001 000
Next State
000 0
000 1
0/1?Collision
CA Research Group (BECDU)
Synthesis of GMACA Reverse Engineering TechniquePhase III: GMACA rule vector from State transition table
CA Research Group (BECDU)
NEIGHBORHOODSTATE
111 110 101 100 011 010 001 000
Next State0/1?Collision
Less the number of collision better the design.
Design Objective :
Design GMACA so that there is minimum number of collision during rule formation
Simulated Annealing to attain the design
•
CA Research Group (BECDU)
Objective Reduce Collision Increment of Cycle Length
Simulated Annealing Program Mutation Technique - 1
1110 1101 0111
1011
1111
Cycle Length = 2
1110 1101 0111
1111
Cycle Length = 1
1011
•
CA Research Group (BECDU)
Simulated Annealing Program Increment of Cycle Length
Present State
Next State
1110 1101 0111 1011 1111
1011 1011 1011 1111 1111
111 1
111 0
NEIGHBORHOOD STATE
111 110 101 100 011 010 001 000
Next State
*1 *0 *0*0/1?
1110 1101 0111
1111
Cycle Length = 1
1011
•
CA Research Group (BECDU)
Simulated Annealing Program Increment of Cycle Length
Present State
Next State
1110 1101 0111 1011 1111
1011 1011 1011 1111 1011
NEIGHBORHOOD STATE
111 110 101 100 011 010 001 000
Next State
Next State
*1 *0 *0*0/1?
111 0
111 0
1110 1101 0111
1011
1111
Cycle Length = 2
0 *1 *0 *0*
•
CA Research Group (BECDU)
Reduction of Cycle Length
Simulated Annealing Program Mutation Technique - 2
Cycle Length = 4
1110 1101 1011
01111111
Cycle Length = 3
1110 1101 1011
01111111
•
CA Research Group (BECDU)
Simulated Annealing Program Decrement of Cycle Length
PresentState
NextState
11101101101101111111
11011101011111111101111 0
111 1
NEIGHBORHOOD STATE
111 110 101 100 011 010 001 000
Next State
0/1? *0 *0 *1*
Cycle Length = 4
1110 1101 1011
01111111
•
CA Research Group (BECDU)
Simulated Annealing Program Decrement of Cycle LengthNEIGHBORHOOD STATE
111 110 101 100 011 010 001 000
Next State
Next State
*1 *0 *0*0/1?
PresentState
NextState
11101101101101111111
11011101011111111011111 1
111 1
1 *1 *0 *0*
Cycle Length = 3
1110 1101 1011
01111111
•
CA Research Group (BECDU)
• Memorizing Capacity
• Evolution Time
• Identification / Recognition Complexity
Performance of GMACA Based Pattern Recognizer
•
CA Research Group (BECDU)
Memorizing Capacity
Pattern Length
GMACA Hopfield Net
20 5 3 40 10 6 60 13 9 80 18 12 100 23 15
Conclusion : GMACA have much higher capacity than Hopfield Net
•
CA Research Group (BECDU)
Evolution Time
Pattern N o ofPatterns
In itialTem p
EvolutionTim e(m in)
20 5 15 1.06
40 10 20 3.01
60 13 25 4.52
80 18 35 7.45
100 23 40 15.08
•
CA Research Group (BECDU)
Identification / Recognition Complexity
• Cost of Computation for Recognition / Identification - Constant
CA Research Group (BECDU)
Achievements
1.Cellular Automata - A powerful machine in designing the pattern recognition tool
2.Storage Capacity of CA - Higher than Hopfield Net
3.A clever reverse engineering technique is employed to design Cellular Automata based Associative Memory
CA Research Group (BECDU)
PublicationsStudy of Non-Linear Cellular Automata For Pattern
Recognition To be published in IEEE Transaction, Man, Machine and Cybernetics, Part - B
Generalized Multiple Attractor Cellular Automata(GMACA) Model for Associative Memory Niloy Ganguly, Pradipta Maji, Biplab k Sikdar and P Pal Chaudhuri To be published in International Journal for Pattern Recognition and Artificial Intelligence
Error Correcting Capability of Cellular Automata Based Associative Memory, IEEE Transaction, Man, Machine and Cybernetics, Part - A