jonathan reagan umass dartmouth csums summer 11 august 3 rd 2011
TRANSCRIPT
Learning Algorithms of Neural Networks and
ApplicationsJonathan Reagan
Umass Dartmouth CSUMS Summer 11
August 3rd 2011
Outline What is a Neural Network? How does it work? Why do we care? Results Issues encountered Future work
Neural Network
InputLayers
HiddenLayers
OutputLayers
Idea of Learning Not realistic to study every possible
case Smaller sample can be used to model
the entire case Assume connections hold
Learning Methods (input)=[age, income, credit
score, etc] (output)=[dependability]
We want weights of α’s X*α(Hidden)=Y
iXiY
yxSign ii )(
Learning Methods-Cont. Use the learning method to find α
I Y-Xα I=0
tttttt XYXSign ))((2
11
Preliminary ResultsPerceptron Least Square
N AccuracysFailed Trials N Accuracys
Failed Trials
Min AVG Max N/Total Min AVG Max N/Total
2 0.4048 0.5296 0.6081 0/100 2
.4081
.5310
.6306
0/100
3 0.3968 0.5247 0.6161 0/100 3
.4000
.5248
.6177
0/100
4 0.3968 0.5292 0.6306 0/100 4
.3984
.5195
.6194
0/100
5 0.4081 0.5312 0.6274 0/100 5
.4048
.5154
.6306
0/100
6 0.3984 0.5446 0.6161 2/100 6
.4081
.5248
.6226
0/100
7 0.4145 0.5312 0.6194 9/100 7
.3645
.5308
.6306
0/100
8 0.4194 0.544 0.6177 20/100 8
.3790
.5374
.6306
0/100
9 0.3952 0.5444 0.621 34/100 9
.3742
.5410
.6323
0/100
10 0.4048 0.5465 0.629 49/100 10
.3726
.5412
.6371
0/100
Speed of Convergence
1 2 3 4 5 6 7 8 9 10 110
100
200
300
400
500
600
700
800
900f(x) = 0.00314824074223472 exp( 1.37608992318393 x )R² = 0.943699577315672
Time vs. N
100 TrialsExponential (100 Trials)
Speed of Convergence-cont.
5.5 6 6.5 7 7.5 8 8.5 9 9.5 10 10.50
100
200
300
400
500
600
700
800
900
f(x) = 197.05 x − 1176.92R² = 0.986531193868044
Time vs. N
100 TrialsLinear (100 Trials)
Why have limits on Iterations?2 million Convergence Failed Trials
4 million convergence Failed Trials
N T(Time)Seconds N/Total N T(Time)Seconds N/Total
2 0.03 0/50 2 0.03 0/50
3 0.04 0/50 3 0.05 0/50
4 0.23 0/50 4 0.24 0/50
5 0.93 0/50 5 0.88 0/50
6 10.49 0/50 6 10.4 0/50
7 50.24 .2/50 7 79.8 .2/50
8 143.07 .8/50 8 260.2 .7/50
9 259.8 15/50 9 484.7 14/50
10 363.6 22/50 10 727.5 22/50
11 483.9 28/50 11 928.2 28/50
12 560.1 34/50 12 1096.1 33/50
13 661.7 40/50 13 1287.2 38/50
14 732.7 43/50 14 1416.2 42/50
15 750.4 44/50 15 1463.7 44/50
16 778.1 46/50 16 1508.9 46/50
17 819.3 48/50 17 1607.7 48/50
18 832 50/50 18 1665.4 50/50
Data Testing Random Data can’t be learned Deterministic Data can be learned Adding Random variance decreases
Accuracy More values of N the Better But more values of N take Longer
Future Work Increase the speed of the Neural
Network Find more applicable data for testing
of the Neural Network Try multiple layer Neural Networks
and Compare
Questions?