This document provides an introduction to Adaptive Resonance Theory (ART), an unsupervised neural network architecture. It discusses the stability-plasticity dilemma that ART aims to address, by allowing new patterns to be learned without overwriting previously learned patterns. The basic ART architecture uses input, cluster, and reset layers. An algorithm is described that initializes parameters, processes input vectors to select a winning cluster unit, and updates weights based on input-cluster similarity and a vigilance parameter. Supplemental units are also introduced to help control the reset mechanism. Applications mentioned include face recognition, image compression, and medical diagnosis.
3. Introduction to ART
It is a self organising neural network architecture.
It is an unsupervised learning.
This type of network remains open to new learnings without washing
away previously learned code.
ART n/w and algo. Maintain the plasticity required to learn new
patterns, while preventing the modification of pattern that have been
learned previously.
ART is a vector classifier, it accepts input vector and categorise
depending upon which of stored pattern it match most or resembles the
most.
If input is not match to any pattern than a new category is made by
storing this new pattern.
No stored pattern is modified if i/p is not matched with any pattern.
4. Stability-Plasticity-Dilemma
• Real world facing a situation where data is continuously changing.
• It produce stability-plasticity dilemma
• Stability – system behaviour does not changes after irrelevant events.
• Plasticity – system adapts its behaviour according to significant
events.
• Dilemma – Preservation of learned Knowledge.
5. Basic Architecture
1. Input processing field – F1 LAYER
2. Cluster unit – F2 LAYER
3. Reset mechanism – It controls degree of similarity of patterns
placed on some cluster.
The input processing layer is divided into two parts
F1(a) – It present input vector given.
F1(b) – It is a interface portion between F1(a) and F2
The signal b/w F1 and F2 is used to comparing similarity of input signal
to weight of vector unit which is selected for storing pattern or for
recognising pattern.
6. Continue…
• F2 is connected to F1(b) through top down weights t(ji).
• F(b) is connected to F2 through bottom up weights b(ij).
• Cluster unit with largest i/p weight is selected for learning the new
pattern and all other cluster’s activation is set to zero.
• Each F2 layer neuron/cluster have associated weight with it, that
cluster will be allowe to learn a new pattern whose dot product of
own weight and input vector is maximal. Only that neuron will
respond back.
• Reset: Information from the input unit is combined with interface unit
, depending upon similarity b/w top down weight and i/p vector
cluster may or may not be allowed to learn a new pattern.
8. Continue…
The above shown diagram is computational unit(F1(a),F2(b),F2, but it
have drawback. i.e :-
1. These units have to behave differently at different stages.
2. Also problem with reset mechanism is that F2 unit has to be
inhibited in certain conditions but have to be returned to availability on
subsequent learning trails.
So, Supplemental unit (G1 and G2) is introduced in addition to reset.
These units send their signal from and to all others units in structure.
• Inhibitory signals shown by “-”
• Excitory signal shown by “+”
10. Algorithm
Parameters used in algorithm are :
n – Number of component in i/p vector.
m – Maximum number of clusters that can be formed.
b(ij) – Bottom up weight.
t (ji) – Top down weight.
p – Vigilance parameter ( set by user tester for testing the acceptance
level for cluster to train if net i/p is less than p then it will not trained and
vice-versa.)
s – Binary input vector.
x – Activation vector for F1(b).
||x|| - norm of vector x ( sum of all x(i)).
11. Description of Algorithm.
1. Binary i/p vector is applied to F1(a) then it is passed to F1(b) and it
is send to F2 over weighted interconnections.
2. Each neuron at F2 will calculate the net i/p i.e dot product of inpuct
vector and its own weight.
3. The neuron with highest net i/p will win and learn new pattern and
activation d=1 for that neuron.
12. Steps to Perform.
Step 1 : Initialize parameter
L>1 and 0<p<=1
Initialize weights
0<b(ij)<L/(L-1+N), t(ji)(0)=1
Step 2 : while stoping condition is false repeat step 3 to14
Step 3 : for each testing i/p repeat step 3 to 13
Step 4 : Set activation of all F2 units to zero
set F1(a) activation to input vector s.
13. Continue…
Step 5 : compute norm of s.
||s||= 𝑖 S(i)
Step 6 : set i/p signal from F1(a) to F1(b)
x(i)=s(i)
Step 7 : for each node of F2 that is not inhibited,
if y(j)≠-1, y(j)= 𝑏(𝑖𝑗)𝑥(𝑖)
14. Continue…
Step 8 : while reset is true do step 9 to 12
Step 9 : find J such that y(J) ≥ y(j) for all J.
if y(j) = -1, then all oods are inhibited and the selected
cluster can not be trained.
Step 10 : recompute activation of x F1(b).
x(i)= s(i)t(ji)
Step 11 : compute norm of x.
||x||= 𝑖 𝑥(𝑖)
15. Conitnue…
Step 12 : Test for reset
||x||/||s|| < p then,
y(j)=-1 cluster can not be trained , do step 8 again.
else, proceed to step 13.
Step 13 : Update weights of node j
b(ij) new = Lx(i)/(L-1+||x||)
t(ji) new = x(i)
Step 14 : Test for stopping condition.
16. Applications of ART
• Face recognition
• Image compression
• Signature Verification
• Medical diagnoses
• Target recognition
• Land cover classification