1

34
Simple Associative Network
a hardlim wp b + ( ) hardlim wp 0.5 ( ) = =
p
1 stimulus ,
0 no stimulus ,

= a
1 response ,
0 no response ,

=
Associative Learning Associative Learning
1
35
Binary Association Binary Association
1
36
Real Value Association Real Value Association
1
37
Bipolar Association Bipolar Association
1
38
Hamming Distance Hamming Distance
1
39
Enhanced Hamming Distance Enhanced Hamming Distance
1
40
Enhanced Binary Association Enhanced Binary Association
1
41
Interference Interference
1
42
XX--OR Interference OR Interference
1
43
Banana Associator Banana Associator
p
0
1 shape detected ,
0 shape not detected ,

= p
1 smell detected ,
0 smell not detected ,

=
Unconditioned Stimulus Conditioned Stimulus
1
44
Unsupervised Hebb Rule Unsupervised Hebb Rule
w
ij
q ( ) w
ij
q 1 ( ) oa
i
q ( ) p
j
q ( ) + =
W q ( ) W q 1 ( ) oa q ( ) p
T
q ( ) + =
Vector Form:
p 1 ( ) p 2 ( ) . p Q ( ) , , ,
Training Sequence:
from (7.5)
Define the rate of
association
1
45
Banana Recognition Example Banana Recognition Example
w
0
1 w 0 ( ) , 0 = =
Initial Weights:
p
0
1 ( ) 0 = p 1 ( ) , 1 = { } p
0
2 ( ) 1 = p 2 ( ) , 1 = { } . , ,
Training Sequence:
w q ( ) w q 1 ( ) a q ( ) p q ( ) + =
a 1 ( ) h a r d l i m w
0
p
0
1 ( ) w 0 ( ) p 1 ( ) 0.5 + ( )
h a r d l i m 1 0 0 1 0.5 + ( ) 0 (no response)
=
= =
First Iteration (sight fails):
w 1 ( ) w 0 ( ) a 1 ( ) p 1 ( ) + 0 0 1 + 0 = = =
o = 1
1
46
Example Example
a 2 ( ) h a r d l i m w
0
p
0
2 ( ) w 1 ( ) p 2 ( ) 0.5 + ( )
h a r d l i m 1 1 0 1 0.5 + ( ) 1 (banana)
=
= =
Second Iteration (sight works):
w 2 ( ) w 1 ( ) a 2 ( ) p 2 ( ) + 0 1 1 + 1 = = =
Third Iteration (sight fails):
a 3 ( ) hardlim w
0
p
0
3 ( ) w 2 ( ) p 3 ( ) 0.5 + ( )
hardlim 1 0 1 1 0.5 + ( ) 1 (banana)
=
= =
w 3 ( ) w 2 ( ) a 3 ( ) p 3 ( ) + 1 1 1 + 2 = = =
Banana will now be detected if either sensor works.
1
47
Problems with Hebb Rule Problems with Hebb Rule
Weights can become arbitrarily large (continue
updating on presenting input)
There is no mechanism for weights to decrease (a
problem occurs in case of facing any noise)
1
48
Hebb Rule with Decay Hebb Rule with Decay
W q ( ) W q 1 ( ) oa q ( ) p
T
q ( ) W q 1 ( ) + =
W q ( ) 1 ( )W q 1 ( ) oa q ( ) p
T
q ( ) + =
This keeps the weight matrix from growing without bound, which can be
demonstrated by setting both a
i
and p
j
to 1:
w
i j
max
1 ( ) w
i j
max
oa
i
p
j
+ =
w
i j
max
1 ( ) w
i j
max
o + =
w
i j
max o

--- =
from (7.45)
1 0 ; rate decay s s =
1
49
Example Example: : Banana Associator Banana Associator
a 1 ( ) h a r d l i m w
0
p
0
1 ( ) w 0 ( ) p 1 ( ) 0.5 + ( )
h a r d l i m 1 0 0 1 0.5 + ( ) 0 (no response)
=
= =
First Iteration (sight fails):
w 1 ( ) w 0 ( ) a 1 ( ) p 1 ( ) 0.1w 0 ( ) + 0 0 1 0.1 0 ( ) + 0 = = =
a 2 ( ) h a r d l i m w
0
p
0
2 ( ) w 1 ( ) p 2 ( ) 0.5 + ( )
h a r d l i m 1 1 0 1 0.5 + ( ) 1 (banana)
=
= =
Second Iteration (sight works):
w 2 ( ) w 1 ( ) a 2 ( ) p 2 ( ) 0.1w 1 ( ) + 0 1 1 0.1 0 ( ) + 1 = = =
= 0.1 o = 1
1
50
Example Example
Third Iteration (sight fails):
a 3 ( ) hardlim w
0
p
0
3 ( ) w 2 ( ) p 3 ( ) 0.5 + ( )
hardlim 1 0 1 1 0.5 + ( ) 1 (banana)
=
= =
w 3 ( ) w 2 ( ) a 3 ( ) p 3 ( ) 0.1w 3 ( ) + 1 1 1 0.1 1 ( ) + 1.9 = = =
0 10 20 30
0
10
20
30
0 10 20 30
0
2
4
6
8
10
Hebb Rule Hebb with Decay
w
i j
max
o

---
1
0.1
------- 10 = = =
1
51
Problem of Hebb with Decay Problem of Hebb with Decay
Associations will decay away if stimuli are not occasionally presented.
w
ij
q ( ) 1 ( )w
ij
q 1 ( ) =
If a
i
= 0, then
If = 0, this becomes
w
i j
q ( ) 0.9 ( )w
i j
q 1 ( ) =
Therefore the weight decays by 10% at each iteration
where there is no stimulus.
0 10 20 30
0
1
2
3
1
52
Instar Instar ((Recognition Network Recognition Network))
(A solution to decay rate problem)
A vector input, a scalar output performing recognition tasks.
1
53
Instar Operation Instar Operation
a hardlim Wp b + ( ) hardlim w
T
1
p b + ( ) = =
The instar will be active when
w
T
1
p b >
or
w
T
1
p w
1
p u cos b > =
For normalized vectors, the largest inner product occurs when the angle between the
weight vector and the input vector is zero -- the input vector is equal to the weight
vector.
The rows of a weight matrix represent patterns
to be recognized.
1
54
Vector Recognition Vector Recognition
b w
1
p = If we set
the instar will only be active when u = 0.
b w
1
p > If we set
the instar will be active for a range of angles.
As b is increased, the more patterns there will be (over a wider range of u) which
will activate the instar.
w
1
(recognize only pattern
1
w)
or
b > p w
1
1
55
Instar Rule Instar Rule
w
ij
q ( ) w
ij
q 1 ( ) oa
i
q ( ) p
j
q ( ) + =
Hebb with Decay
Modify so that learning and forgetting will only occur
when the neuron is active - Instar Rule:
w
i j
q ( ) w
i j
q 1 ( ) oa
i
q ( ) p
j
q ( ) a
i
q ( ) w q 1 ( ) + =
i j
w
ij
q ( ) w
ij
q 1 ( ) oa
i
q ( ) p
j
q ( ) w
i j
q 1 ( ) ( ) + =
w q ( )
i
w q 1 ( )
i
oa
i
q ( ) p q ( ) w q 1 ( )
i
( ) + =
or
Vector Form:
; =
1
56
Graphical Representation Graphical Representation
w q ( )
i
w q 1 ( )
i
o p q ( ) w q 1 ( )
i
( ) + =
For the case where the instar is active (a
i
=

1):
or
w q ( )
i
1 o ( ) w q 1 ( )
i
op q ( ) + =
For the case where the instar is inactive (a
i
=

0):
w q ( )
i
w q 1 ( )
i
=
1
57
Example Example
p
0
1 orange detected visually ,
0 orange not detected ,

=
p
shape
texture
weight
=
1
58
Training Training
W 0 ( ) w
T
1
0 ( )
0 0 0
= =
p
0
1 ( ) 0 = p 1 ( ) ,
1
1
1
=
)

`


p
0
2 ( ) 1 = p 2 ( ) ,
1
1
1
=
)

`


. , ,
First Iteration (o=1):
a 1 ( ) hardlim w
0
p
0
1 ( ) Wp 1 ( ) 2 + ( ) =
a 1 ( ) h ardlim 3 0
0 0 0
1
1
1
2 +
\ .
|
|
|
| |
0 (no response) = =
w 1 ( )
1

w 0 ( )
1
a 1 ( ) p 1 ( ) w 0 ( )
1
( ) +
0
0
0
0
1
1
1
0
0
0

\ .
|
|
|
| |
+
0
0
0
= = =
1
59
Further Training Further Training
(orange)
h a 2 ( ) h a r d l i m w
0
p
0
2 ( ) W p 2 ( ) 2 + ( ) = a r d l i m 3 1
0 0 0
1
1
1
2 +

)
1 = =
w 2 ( )
1

w 1 ( )
1
a 2 ( ) p 2 ( ) w 1 ( )
1
( ) +
0
0
0
1
1
1
1
0
0
0

\ .
|
|
|
| |
+
1
1
1
= = =
a 3 ( ) h a r d l i m w
0
p
0
3 ( ) W p 3 ( ) 2 + ( ) =
(orange)
h a r d l i m 3 0
1 1 1
1
1
1
2 +

)
1 = =
w 3 ( )
1

w 2 ( )
1
a 3 ( ) p 3 ( ) w 2 ( )
1
( ) +
1
1
1
1
1
1
1
1
1
1

\ .
|
|
|
| |
+
1
1
1
= = =
Orange will now be detected if either set of sensors works.

.
1
60
Kohonen Rule Kohonen Rule
w q ( )
1
w q 1 ( )
1
o p q ( ) w q 1 ( )
1
( ) + = for i X q ( ) e ,
Learning occurs when the neurons index i is a member of the set X(q).
We will see in the next chapter that this can be used to train all
neurons in a given neighborhood.
1
61
Outstar Outstar ((Recall Network Recall Network))
A scalar input, a vector output performing recall tasks.
1
62
Outstar Operation Outstar Operation
W a
-
=
Suppose we want the outstar to recall a certain pattern a*
whenever the input p = 1 is presented to the network. Let
Then, when p = 1
a satlins Wp ( ) s atli ns a
-
1 ( ) a
-
= = =
and the pattern is correctly recalled.
The columns of a weight matrix represent patterns
to be recalled.
1
63
Outstar Rule Outstar Rule
w
ij
q ( ) w
i j
q 1 ( ) oa
i
q ( ) p
j
q ( ) p
j
q ( )w
ij
q 1 ( ) + =
For the instar rule we made the weight decay term of the Hebb
rule proportional to the output of the network. For the outstar
rule we make the weight decay term proportional to the input of the network.
If we make the decay rate equal to the learning rate o,
w
i j
q ( ) w
i j
q 1 ( ) o a
i
q ( ) w
ij
q 1 ( ) ( ) p
j
q ( ) + =
Vector Form:
w
j

q ( ) w
j

q 1 ( ) o a q ( ) w
j

q 1 ( ) ( ) p
j
q ( ) + =
1
64
Example Example -- Pineapple Recall Pineapple Recall
1
65
Definitions Definitions
a s atl ins W
0
p
0
Wp + ( ) =
W
0
1 0 0
0 1 0
0 0 1
=
p
0
shape
texture
weight
=
p
1 if a pineapple can be seen ,
0 otherwise ,

=
p
pi neap ple
1
1
1
=
1
66
Iteration Iteration 11
p
0
1 ( )
0
0
0
= p 1 ( ) , 1 =
)

`


p
0
2 ( )
1
1
1
= p 2 ( ) , 1 =
)

`


. , ,
a 1 ( ) satli ns
0
0
0
0
0
0
1 +
\ .
|
|
|
| |
0
0
0
(no response) = =
w
1
1 ( ) w
1
0 ( ) a 1 ( ) w
1
0 ( ) ( ) p 1 ( ) +
0
0
0
0
0
0
0
0
0

\ .
|
|
|
| |
1 +
0
0
0
= = =
o = 1
1
67
Convergence Convergence
a 2 ( ) satlins
1
1
1
0
0
0
1 +
\ .
|
|
|
| |
1
1
1
(measurements given) = =
w
1
2 ( ) w
1
1 ( ) a 2 ( ) w
1
1 ( ) ( ) p 2 ( ) +
0
0
0
1
1
1
0
0
0

\ .
|
|
|
| |
1 +
1
1
1
= = =
w
1
3 ( ) w
1
2 ( ) a 2 ( ) w
1
2 ( ) ( ) p 2 ( ) +
1
1
1
1
1
1
1
1
1

\ .
|
|
|
| |
1 +
1
1
1
= = =
a 3 ( ) satli ns
0
0
0
1
1
1
1 +
\ .
|
|
|
| |
1
1
1
(measurements recalled) = =
1
68
Hamming Network
From figure 3.5
Instar layer
Competitive layer
Competitive Networks Competitive Networks
1
69
Layer Layer 11 ((Correlation Correlation))
p
1
p
2
. p
Q
, , , { }
We want the network to recognize the following prototype vectors:
W
1
w
T
1
w
T
2
.
.
.
w
T
S
p
1
T
p
2
T
.
.
.
p
Q
T
= = b
1
R
R
.
.
.
R
=
The first layer weight matrix and bias vector are given by:
The response of the first layer is:
The prototype
closest to the
input vector produces
the largest response.
a
1
W
1
p b
1
+
p
1
T
p R +
p
2
T
p R +
.
.
.
p
Q
T
p R +
= =
: (Recognize multiple patterns)
1
70
Layer Layer 22 ((Competition Competition))
a
2
0 ( ) a
1
=
a
2
t 1 + ( ) pos li n W
2
a
2
t ( ) ( ) =
w
ij
2
1 if i j = ,
c otherwise ,

= 0 c
1
S 1
------------ < <
a
i
2
t 1 + ( ) poslin a
i
2
t ( ) c a
j
2
t ( )
j i =

\ .
| |
=
The neuron with the largest
initial condition will win the
competiton.
The second layer is
initialized with the output
of the first layer.
Lateral inhibition
For each neuron :
1
71
Competitive Layer Competitive Layer
a
i
1 i i
-
= ,
0 i i
-
= ,

=
n
i
-
n
i
> i , i
-
i s n
i
n
i
-
= ,
a compet n ( ) =
n W p
w
1
T
w
2
T
.
.
.
w
S
T
p
w
1
T
p
w
2
T
p
.
.
.
w
S
T
p
L
2
u
1
cos
L
2
u
2
cos
.
.
.
L
2
u
S
cos
= = = =
1
72
Competitive Learning Competitive Learning
w
i
q ( ) w
i
q 1 ( ) oa
i
q ( ) p q ( ) w
i
q 1 ( ) ( ) + =
w
i
q ( ) w
i
q 1 ( ) = i i
-
=
w
i
-
q ( ) w
i
-
q 1 ( ) o p q ( ) w
i
-
q 1 ( ) ( ) + =
w
i
-
q ( ) 1 o ( ) w
i
-
q 1 ( ) op q ( ) + =
For the competitive network, the winning neuron has an
ouput of 1, and the other neurons have an output of 0.
Instar Rule
Kohonen Rule
1
73
Graphical Representation Graphical Representation
w
i
-
q ( ) w
i
-
q 1 ( ) o p q ( ) w
i
-
q 1 ( ) ( ) + =
w
i
-
q ( ) 1 o ( ) w
i
-
q 1 ( ) op q ( ) + =
1
74
Example Example
1
75
Four Iterations Four Iterations
1
76
Typical Convergence Typical Convergence ((Clustering Clustering))
Before Training After Training
Weights
Input Vectors
1
77
Dead Units Dead Units (p14-9)
One problem with competitive learning is that neurons with initial weights far
from any input vector may never win.
Solution: Add a negative bias to each neuron, and increase the magnitude of the
bias as the neuron wins. This will make it harder to win if a neuron has won often.
This is called a conscience.
Dead Unit
1
78
Stability Stability (p14-10)
1
w(0)
2
w(0)
p
1
p
3
p
2
p
5
p
6
p
7
p
4
p
8
1
w(8)
2
w(8)
p
1
p
3
p
2
p
5
p
6
p
7
p
4
p
8
If the input vectors dont fall into nice clusters, then for large learning rates the
presentation of each input vector may modify the
configuration so that the system will undergo continual evolution.
1
79
Competitive Layers in Biology Competitive Layers in Biology
w
i j ,
1 if i j = ,
c if i j = ,

=
w
i j ,
1 if d
i j ,
0 = ,
c if d
i j ,
0 > ,

=
Weights in the competitive layer of the Hamming network:
Weights assigned based on distance:
On-Center/Off-Surround Connections for Competition
1
80
Mexican Mexican--Hat Function Hat Function
1
81
Feature Maps Feature Maps
w
i
q ( ) w
i
q 1 ( ) o p q ( ) w
i
q 1 ( ) ( ) + =
w
i
q ( ) 1 o ( ) w
i
q 1 ( ) op q ( ) + =
i N
i
-
d ( ) e
Update weight vectors in a neighborhood of the winning neuron.
N
i
d ( ) j d
i j ,
d s , { } =
N
13
1 ( ) 8 12 13 14 18 , , , , { } =
N
13
2 ( ) 3 7 8 9 11 12 13 14 15 17 18 19 23 , , , , , , , , , , , , { } =
1
82
Self Organizing Maps: Clustering Self Organizing Maps: Clustering
With unsupervised learning there is no instruction and the network is left to
cluster patterns. All of the patterns within a cluster will be judged as
being similar.
Cluster algorithms form groups referred to as clusters and the arrangement
of clusters should reflect two properties:
Patterns within a cluster should be similar in some way.
Clusters that are similar in some way should be close together.
1
83
The Self The Self- -organizing Feature Map organizing Feature Map
The input units correspond in number to the dimension of the training
vectors.
The output units act as prototypes.
This network has three inputs and five cluster units. Each unit in the input layer is connected to every
unit in the cluster layer.
1
84
How does a cluster unit act as a prototype? How does a cluster unit act as a prototype?
w
2,1
w
1,1
x
1
x
2
p
1
x
2
coordinate x
1
coordinate
w
1,1
w
2,1
The input layer takes coordinates from the input space. The weights adapt during training, and when learning
is complete each cluster unit will have a position in the input space determined by weights.
1
85
Radius 1
Radius 2
Winning unit, only one to update if radius=0
Units to be updated when radius=1
Units to be updated when radius=2
The topology determines which units fall within a
radius of the winning unit.
The winning unit will adapt its weights in a way that
moves that cluster unit even closer to the training
vector.
Weight updates within a radius Weight updates within a radius
1
86
Algorithm Algorithm
initialize weights to random values.
while not HALT
for each input vector
for each cluster unit calculate the distance from the training vector
find unit j with the minimum distance
update all weight vectors for units within the radius according to
check to see if the learning rate or radius needs updating
check HALT.

=
i
i ij j
x w d
2
) (
)] ( )[ ( ) ( ) 1 ( n w x n n w n w
ij i ij ij
+ = + q
1
87
Self Self- -Organizing Maps Organizing Maps
On each training input, all output nodes that are within a topological distance, d
T
,
of D from the winner node will have their incoming weights modified.
d
T
(y
i
,y
j
) = # nodes that must be traversed in the output layer in moving between
output nodes y
i
and y
j
.
D is typically decreased as training proceeds.
Fully Interconnected
Input
Output
Partially
Intraconnected
1
88
Example Example
An SOFM has a 10x10 two-dimensional grid layout and the radius is initially set to
6. Find how many units will be updated after 1000 epochs if the winning unit
is located in the extreme bottom right-hand corner of the grid and the radius
is updated according to
r = r-1 if current_epoch mod 200 = 0
Assume that epoch numbering starts at 1.
1
89
Example Example
1
90
Example Example 22
1
91
Example Example 22
1
92
Example Example 22
1
93
Example Example 33
1
94
Example Example 33
Figure 3.16 The weight vector positions converted to (x,y) coordinates after 5000 cycles through
all the training data.
1
95
Example Example
33
Figure 3.18 The weight vector positions converted to (x,y) coordinates after 20000 cycles
through all the training data.
Figure 3.17 The weight vector positions converted to (x,y) coordinates after 10000 cycles
through all the training data.
1
96
Convergence Convergence
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
-1 -0.8 -0.6 -0.4 -0.2 0 0.2 0.4 0.6 0.8 1
-1
-0.8
-0.6
-0.4
-0.2
0
0.2
0.4
0.6
0.8
1
1
97
There Goes The Neighborhood There Goes The Neighborhood
D = 1
D = 2
D = 3
As the training period progresses, gradually decrease D.
Over time, islands form in which the center represents the centroid C of a set
of input vectors, S, while nearby neighbors represent slight variations on C and
more distant neighbors are major variations.
These neighbors may only win on a few (or no) input vectors, while the island
center will win on many of the elements of S.
1
98
Self Organization Self Organization
In the beginning, the Euclidian distance d
E
(y
l
,y
k
) and Topological distance d
T
(y
l
,y
k
) between output nodes y
l
and y
k
will
not be related.
But during the course of training, they will become positively correlated: Neighbor nodes in the topology will have
similar weight vectors, and topologically distant nodes will have very different weight vectors.
d
E
(y
l
, y
k
) = w
li
w
ki
( )
2
i=1
n

Euclidean Neighbor
Emergent Structure of Output Layer
Before After
Topological Neighbor
1
99
Example Example 44
An SOFM network with three inputs and two cluster units is to be trained using the four training vectors:
[0.8 0.7 0.4], [0.6 0.9 0.9], [0.3 0.4 0.1], [0.1 0.1 02] and initial weights
The initial radius is 0 and the learning rate is 0.5 . Calculate the weight changes during the first cycle
through the data, taking the training vectors in the given order.
(
(
(

5 . 0 8 . 0
2 . 0 6 . 0
4 . 0 5 . 0
q
weights to the first cluster unit
1
100
Solution Solution
The euclidian distance of the input vector 1 to cluster unit 1 is:
The euclidian distance of the input vector 1 to cluster unit 2 is:
Input vector 1 is closest to cluster unit 1 so update weights to cluster unit 1:
( ) ( ) ( ) 26 . 0 4 . 0 8 . 0 7 . 0 6 . 0 8 . 0 5 . 0
2 2 2
1
= + + = d
( ) ( ) ( ) 42 . 0 4 . 0 5 . 0 7 . 0 2 . 0 8 . 0 4 . 0
2 2 2
2
= + + = d
) 8 . 0 4 . 0 ( 5 . 0 8 . 0 6 . 0
) 6 . 0 7 . 0 ( 5 . 0 6 . 0 65 . 0
) 5 . 0 8 . 0 ( 5 . 0 5 . 0 65 . 0
)] ( [ 5 . 0 ) ( ) 1 (
+ =
+ =
+ =
+ = + n w x n w n w
ij i ij ij
(
(
(

5 . 0 60 . 0
2 . 0 65 . 0
4 . 0 65 . 0
1
101
Solution Solution
The euclidian distance of the input vector 2 to cluster unit 1 is:
The euclidian distance of the input vector 2 to cluster unit 2 is:
Input vector 2 is closest to cluster unit 1 so update weights to cluster unit 1 again:
( ) ( ) ( ) 155 . 0 9 . 0 6 . 0 9 . 0 65 . 0 6 . 0 65 . 0
2 2 2
1
= + + = d
( ) ( ) ( ) 69 . 0 9 . 0 5 . 0 9 . 0 2 . 0 6 . 0 4 . 0
2 2 2
2
= + + = d
) 60 . 0 9 . 0 ( 5 . 0 60 . 0 750 . 0
) 65 . 0 9 . 0 ( 5 . 0 65 . 0 775 . 0
) 65 . 0 6 . 0 ( 5 . 0 65 . 0 625 . 0
)] ( [ 5 . 0 ) ( ) 1 (
+ =
+ =
+ =
+ = + n w x n w n w
ij i ij ij
(
(
(

5 . 0 750 . 0
2 . 0 775 . 0
4 . 0 625 . 0
Repeat the same update procedure for input vector 3 and 4 also.
1
102
The angle between vectors as the measure of The angle between vectors as the measure of
similarity similarity
| |
n
v v v v ......
2 1
=
| |
n
w w w w ......
2 1
=
| |
n n
w v w v w v w v ...... .
2 2 1 1
=
The dot product which is sometimes called the inner product or scalar
product, of two vectors v and w
is given as follows:
p
1
p
2
a
The angle between non-zero vectors v and w is:
|
|
.
|

\
|

w v
w v.
cos
1
2 2
2
2
1
...
n
v v v v vector of magnitude or norm v + + + = =
1
103
Weight update for dot product as measure of Weight update for dot product as measure of
similarity similarity
The winning prototype index for an input vector x:
index(x)=max p
j
.x for all j
The weight updates are given by:
x n w
x n w
n w
j
j
j
q
q
+
+
+ +
) (
) (
) 1 (
1
104
Self Self- -Organized Maps for Robot Navigation Organized Maps for Robot Navigation
Owen & Nehmzow (1998)
Task: Autonomous robot navigation in a laboratory
Goals:
1. Find a useful internal representation (i.e. map) that supports an intelligent choice of
actions for the given sensory inputs
2. Let the robot build/learn the map itself
- Saves the user from specifying it.
- Allows the robot to handle new environments.
- By learning the map in a noisy, real-world situation, the robot
will be more apt to handle other noisy environments.
Approach:
Use an SOM to organize situation-action vectors.
The emerging structure of the SOM then constitutes the robots functional internal
representation of both the outside world and the appropriate actions to take in different
regions of that world.
1
105
The Training Phase The Training Phase
R
1. Record Sensory Info
Turn Right
& Slow Down
2. Get correct actions
3. Input Vector = Sensory Inputs & Actions
Input
Output
4. Run SOM on Input Vector
5. Update Winner & Neighbors
1
106
The Testing Phase The Testing Phase
R
1. Record Sensory Info
2. Input Vector = Sensory Inputs & No Actions
Input
Output
3. Run SOM on Input Vector 4. Read Recommended Actions from the
Winners Weight Vector
A
1
107
Clustering of Perceptual Signatures Clustering of Perceptual Signatures
The general closeness of successive winners shows a correlation between points & distances in the
objective world and the robots functional view of that world.
Note: A trace of the robots path on a map of the real world (i.e. lab floor) would have ONLY short
moves.
The sequence of winner nodes during the testing phase of
a typical navigation task.
1
108
SOM for Navigation Summary SOM for Navigation Summary
SOM Regions = Perceptual Landmarks = Sets of similar perceptual patterns
Navigation = Association of actions with perceptual landmarks
Behavior is controlled by the robots subjective functional interpretation of
the world, which may abstract the world into a few key perceptual categories.
No extensive objective map of the entire environment is required.
Useful maps are user & task centered.
Robustness (Fault Tolerance): The robot also navigates successfully when a
few of its sensors are damaged => The SOM has generalized from the specific
training instances.
Similar neuronal organizations, with correlations between points in the visual
field and neurons in a brain region, are found in many animals.
1
109
Learning Vector Quantization Learning Vector Quantization
The net input is not computed by taking an inner product of the prototype vectors
with the input. Instead, the net input is the negative of the distance between the
prototype vectors and the input.
1
110
Subclass Subclass
For the LVQ network, the winning neuron in the first layer
indicates the subclass which the input vector belongs to. There
may be several different neurons (subclasses) which make up
each class.
The second layer of the LVQ network combines subclasses into
a single class. The columns of W
2
represent subclasses, and the
rows represent classes. W
2
has a single 1 in each column, with
the other elements set to zero. The row in which the 1 occurs
indicates which class the appropriate subclass belongs to.
w
k i ,
2
1 = ( ) subclass i is a part of class k
1
111
Example Example
W
2
1 0 1 1 0 0
0 1 0 0 0 0
0 0 0 0 1 1
=
Subclasses 1, 3 and 4 belong to class 1.
Subclass 2 belongs to class 2.
Subclasses 5 and 6 belong to class 3.
A single-layer competitive network can create convex classification regions. The second
layer of the LVQ network can combine the convex regions to create more complex
categories.
1
112
LVQ Learning LVQ Learning
w
1
i
-
q ( ) w
1
i
-
q 1 ( ) o p q ( ) w
1
i
-
q 1 ( ) ( ) + =
a
k
-
2
t
k
-
1 = =
w
1
i
-
q ( ) w
1
i
-
q 1 ( ) o p q ( ) w
1
i
-
q 1 ( ) ( ) = a
k
-
2
1 t
k
-
= 0 = =
If the input pattern is classified correctly, then move the winning weight toward the
input vector according to the Kohonen rule.
If the input pattern is classified incorrectly, then move the winning weight away
from the input vector.
LVQ learning combines competive learning with supervision.
It requires a training set of examples of proper network behavior.
p
1
t
1
, { } p
2
t
2
, { } . p
Q
t
Q
, { } , , ,
1
113
Example Example
p
1
0
1
= t
1
1
0
= ,
)
`

p
4
0
0
= t
4
0
1
= ,
)
`

W
2
1 1 0 0
0 0 1 1
=
W
1
0 ( )
w
1
1
( )
T
w
1
2
( )
T
w
1
3
( )
T
w
1
4
( )
T
0.25 0.75
0.75 0.75
1 0.25
0.5 0.25
= =
p
2
1
0
= t
2
0
1
= ,




`
)
p
3
1
1
= t
3
1
0
= ,




`
)
(Random setting)

1
114
First Iteration First Iteration
a
1
c o m p et
0.25 0.75
T
0 1
T

0.75 0.75
T
0 1
T

1.00 0.25
T
0 1
T

0.50 0.25
T
0 1
T

)
c o m p et
0.354
0.791
1.25
0.901

)
1
0
0
0
= = =
a
1
c o m p et n
1
( ) c o m p et
w
1
1
p
1

w
1
2
p
1

w
1
3
p
1

w
1
4
p
1

)
= =
1
115
Second Layer Second Layer
a
2
W
2
a
1
1 1 0 0
0 0 1 1
1
0
0
0
1
0
= = =
w
1
1
1 ( ) w
1
1
0 ( ) o p
1
w
1
1
0 ( ) ( ) + =
w
1
1
1 ( )
0.25
0.75
0.5
0
1
0.25
0.75

\ .
|
| |
+
0.125
0.875
= =
This is the correct class, therefore the weight vector is moved
toward the input vector.
1
116
Figure Figure
1
117
Final Decision Regions Final Decision Regions
1
118
LVQ LVQ22
If the winning neuron in the hidden layer incorrectly classifies the current input, we
move its weight vector away from the input vector, as before. However, we also adjust
the weights of the closest neuron to the input vector that does classify it properly. The
weights for this second neuron should be moved toward the input vector.
When the network correctly classifies an input vector, the weights of only one neuron
are moved toward the input vector. However, if the input vector is incorrectly
classified, the weights of two neurons are updated, one weight vector is moved away
from the input vector, and the other one is moved toward the input vector. The
resulting algorithm is called LVQ2.
1
119
LVQ LVQ2 2 Example Example
1
120
Radial Radial--basis function networks basis function networks
RBF = radial-basis function: a function
which depends only on the radial distance
from a point
So RBFs are functions taking the form
Where is a nonlinear activation scalar
function of distance between vector x and
x
i
(|| ||)
i
x x |
p = -3:.1:3;
a = radbas(p);
a2 = radbas(p-1.5);
a3 = radbas(p+2);
a4 = a + a2*1 + a3*0.5;
plot(p,a,'b-',p,a2,'b--',p,a3,'b--',p,a4,'m-')
title('Weighted Sum of Radial Basis Transfer Functions');
xlabel('Input p');
ylabel('Output a');
1
121
Radial Radial--basis function networks basis function networks
x
M
Input
x
1
x
2
Three-layer networks
Hidden layer
w
i
d
output = S w
i
adjustable parameters are weights w
i
number of hidden units = number of data points
Form of the basis functions decided in advance
(|| ||)
i
x x |
(|| ||)
i
x x |