Developing Neural Network Applications Using Labview

12/5/4
Developing Neural Network Applications Using Labview
Scribd Upload a Document

Search books, presentations, business, academics... Search Documents
Explore
Documents
Books - Fiction Books - Non-fiction Health & Medicine Brochures/Catalogs Government Docs How-To Guides/Manuals Magazines/Newspapers Recipes/Menus School Work + all categories Featured Recent
People
Authors Students Researchers Publishers Government & Nonprofits Businesses Musicians Artists & Designers Teachers + all categories Most Followed Popular
We're using Facebook to personalize your experience. Learn MoreDisable View Public Profile My Documents My Collections My Shelf Messages Notifications Account Help Log Out
Welcome to Scribd - Where the world comes to read, discover, and share...
Were using Facebook to give you reading recommendations based on what your friends are sharing and the things you like. We've also made it easy to connect with your friends: you are now following your Facebook friends who are on Scribd, and they are following you! In the future you can access your account using your Facebook login and password. Learn moreNo thanks Some of your friends are already on Scribd:
www.scribd.com/doc/38040990/Developing-Neural-Network-Applications-Using-Labview
1/105
12/5/4
Graphene Coatings
Single Layer Graphene on SiO2 Wafer glass, PET, or your substrate
graphene-supermarket.com/CVD-grown-
2/105
12/5/4
3/105
12/5/4
4/105
12/5/4
5/105
12/5/4
6/105
12/5/4
7/105
12/5/4
8/105
12/5/4
9/105
12/5/4
10/105
12/5/4
11/105
12/5/4
12/105
12/5/4
13/105
12/5/4
14/105
12/5/4
15/105
12/5/4
16/105
12/5/4
17/105
12/5/4
18/105
12/5/4
19/105
12/5/4
20/105
12/5/4
21/105
12/5/4
22/105
12/5/4
23/105
12/5/4
24/105
12/5/4
25/105
12/5/4
26/105
12/5/4
27/105
12/5/4
28/105
12/5/4
29/105
12/5/4
30/105
12/5/4
31/105
12/5/4
32/105
12/5/4
33/105
12/5/4
34/105
12/5/4
35/105
12/5/4
36/105
12/5/4
37/105
12/5/4
38/105
12/5/4
39/105
12/5/4
40/105
12/5/4
41/105
12/5/4
42/105
12/5/4
43/105
12/5/4
44/105
12/5/4
45/105
12/5/4
46/105
12/5/4
47/105
12/5/4
48/105
12/5/4
49/105
12/5/4
50/105
12/5/4
51/105
12/5/4
52/105
12/5/4
53/105
12/5/4
54/105
12/5/4
55/105
12/5/4
56/105
12/5/4
CHAPTER 4 AREAS OF APPLICATION
4.1 Pattern Classification
Pattern Classification is part of the much broader field called Pattern Recognition. Recognition basically has 2 parts

Feature Extraction Pattern Classification
In this thesis, the LabVIEW VIs have been developed for pattern classification applications, assuming that the data is already defined over feature space (feature extracted form).
Classification is very important in numerous fields. A simple example is in a post
57/105
12/5/4
Classification is very important in numerous fields. A simple example is in a post office where mail to be delivered needs to be sorted. If one has an automatic system in which a computer can recognize the zip code and sort the mail based on zip codes then it would make the work much more efficient. What is needed is a scanner which reads the data into the system. Then, the system can use an effective algorithm to differentiate and sort the mail. There are many other applications where one could make use of such an automated data classifier. There are many algorithms and techniques to classify data. In this thesis, LabVIEW VIs have been constructed for pattern classification applications that have a useful GUI and are very user-friendly. This chapter will discuss some statistical classification techniques followed by a practical neural net approach and other clustering algorithms to classify data.
65
of 115
54
Basically, a feature extracted data set could be of any dimension. In this thesis, a popular IRIS data set [11] has been used. This feature extracted data has been extensively used in the past and has been thoroughly analyzed. It is known that this data set consists of 4-dimensional vectors belonging to three classes. Out of the total of 150 vectors, each of the three classes consists of 50 vectors, each class defined by a cluster. The idea of classification is to form clusters of data from a pool of data points. The statistical distribution of this IRIS data set has been studied in the past. Essentially, one could split the data analysis into two types: Parametric and non-parametric. In the
parametric case, good prior knowledge is available about the data, such as the probability density function, a priori probabilities of occurrence for each class, and perhaps one or more moments for each class. For the non-parametric case, such knowledge is not assumed, but rather classification must be based on clustering. Next, Bayes theorem will be addressed, and then the discussion will continue with building classifiers.
4.1.1 Statistics of Classification
Thomas Bayes was a mathematician who developed a method to classify data into www.scribd.com/doc/38040990/Developing-Neural-Network-Applications-Using-Labview
58/105
12/5/4
Thomas Bayes was a mathematician who developed a method to classify data into clusters based on the probability theory known as maximum likelihood. A Bayes classifier is a mathematical model or classification system. Since there are 3 classes and 50 points per class, then the a priori probability of occurrence for a sample value (point) to be from any one of the 3 classes can be taken to be 1/3, or P(Ck) = 0.33, k = 1, 2, 3. These probabilities are assumed a priori. The probability of a feature vector x given that class Ck has occurred is denoted by P( x / Ck ) which is called the class conditional probability. These conditional probabilities must be assumed or have to be
55
calculated based on a set of rules. Then, let the probability of a particular pattern, x, occurring, independent of any class be denoted by P ( x ) . Then, according to Bayes theorem, the probability that pattern class Ck is the true class, given feature vector x is denoted by P(Ck / x ) , the posterior probability. This probability is related to P(x) and
P( x / Ck ) as follows, P(Ck / x ) = P( x / Ck ) P(Ck ) P( x )
The denominator is called the normalization factor that assures that, P(C1/x ) + P(C2 /x ) + P(C3/x ) = 1 Essentially the theorem assigns a membership value between 0 and 1 to a particular feature corresponding to its membership in each of the 3 classes. The class Cp, corresponding to the largest value of P(Ck / x ) , k = 1, 2, 3 for a given measured x is the winner, which means that x belongs to class Cp. For classification [3], the basic idea is to estimate the density value of a particular feature vector in a particular region of feature space. Using parametric techniques as applied to the IRIS flower data set; one can determine the conditional probability densities and assume a priori probabilities. Using these and Bayes theorem, one can determine the posterior probabilities for all of the three clusters. The highest value of the
posterior probability P(C / x ) would be the winner, and x , then belongs to the cluster
59/105
12/5/4
posterior probability P(Ck / x ) would be the winner, and x , then belongs to the cluster
with the highest value. This procedure is normally applied in statistical classification, and it is a very effective procedure. Suppose that the data distribution is unknown. Then, one will need to employ non-parametric algorithms to classify the input data points. However, with the data set
56
available, one can employ unsupervised algorithms that can automatically capture features from the data and cluster the data. In classifying patterns in feature space based on geometric properties, the general procedure is to construct boundaries to separate clusters; the more well-placed that the boundaries are, the better the classification in terms of reducing the associated error of misclassification. If these boundaries are poorly chosen, the misclassification error will increase. These boundaries can be defined mathematically by discriminant functions. In general, classification techniques are aimed at constructing these discriminant functions based on the input feature vectors.
4.1.2 Statistics and Neural Nets an Analogy
Now, consider how a neural net functions in the process of pattern classification. Normally the standard feed forward neural net uses a non-linear transfer function to map the input feature space to the output decision space. For example, in this thesis the sigmoid function has been used for different nets. Suppose the input feature vector is x and the weight vectors are denoted by w as before, then the dot product and summation is given by
net = x.w
Applying the sigmoid function to net gives the result

f (net ) =
1 1 + exp(net )
as the output of the neuron. This f (net ) is nothing but the posterior probability function
www.scribd.com/doc/38040990/Developing-Neural-Network-Applications-Using-Labview 60/105
12/5/4
that was earlier obtained using the Bayes theorem.
57
Consider a two-class problem with equal a priori probabilities of occurrence. Then, the net value that was calculated using the inputs and the weights is just the likelihood ratio of the class conditional probabilities expressed as,
p ( x / c1 ) net = p ( x / c ) 2
If this likelihood ratio is greater than 1 then the x belongs to class C1. If the likelihood ratios is lesser than 1 then x belongs to class C2. Thus the net input to a neuron is the class conditional probability p ( x / ck ) where k= 1, 2 and the output is the posterior probability P (Ck / x ) . Thus, the neural net can be considered to be an implementation of Bayes theorem by analogy. The sigmoid function is called the S-shaped function which compresses any input value onto a small range, (0, 1). The sigmoid activation function acts a discriminant function to classify data in feature space. Although the sigmoid function is non-linear, the decision boundary it creates is still linear. Geometrically speaking, there are many kinds of data to be classified. Some data sets can be linearly separable. But some data distributions cannot be linearly separated into clusters or classes. Well-known examples of data sets that are not linearly separable are the binary XOR data and the IRIS flower data. In the case addressed here in regard to the IRIS flower data, it has been confirmed that for the three clusters or defined pattern classes, two of the classes or clusters overlap.
58
www.scribd.com/doc/38040990/Developing-Neural-Network-Applications-Using-Labview 61/105
12/5/4
62/105
12/5/4
63/105
12/5/4
64/105
12/5/4
65/105
12/5/4
66/105
12/5/4
67/105
12/5/4
68/105
12/5/4
69/105
12/5/4
70/105
12/5/4
71/105
12/5/4
72/105
12/5/4
73/105
12/5/4
74/105
12/5/4
75/105
12/5/4
76/105
12/5/4
77/105
12/5/4
78/105
12/5/4
79/105
12/5/4
80/105
12/5/4
81/105
12/5/4
82/105
12/5/4
83/105
12/5/4
84/105
12/5/4
85/105
12/5/4
86/105
12/5/4
87/105
12/5/4
88/105
12/5/4
89/105
12/5/4
90/105
12/5/4
91/105
12/5/4
92/105
12/5/4
93/105
12/5/4
94/105
12/5/4
95/105
12/5/4
96/105
12/5/4
97/105
12/5/4
98/105
12/5/4
99/105
12/5/4
100/105
12/5/4
101/105
12/5/4
102/105
12/5/4
103/105
12/5/4
Leave a Comment
104/105
12/5/4
Submit Characters: 400
Submit Characters: ... Upload a Document

Search books, presentations, business, academics... Search Documents
Follow Us! scribd.com/scribd twitter.com/scribd facebook.com/scribd About Press Blog Partners Scribd 101 Web Stuff Support FAQ Developers / API Jobs Terms Copyright Privacy Copyright 2012 Scribd Inc. Language: English
105/105

Developing Neural Network Applications Using Labview

Hochgeladen von

Dokumentinformationen

Originalbeschreibung:

Copyright

Verfügbare Formate

Dieses Dokument teilen

Dokument teilen oder einbetten

Freigabeoptionen

Stufen Sie dieses Dokument als nützlich ein?

Sind diese Inhalte unangemessen?

Copyright:

Verfügbare Formate

Developing Neural Network Applications Using Labview

Hochgeladen von

Copyright:

Verfügbare Formate

12/5/4