Feed forward neural network for sine

International Journal on Computational Sciences & Applications (IJCSA) Vol.4, No.2, April 2014
DOI:10.5121/ijcsa.2014.4201 01
Feed Forward Neural Network For Sine
Function With Symmetric Table Addition
Method Using Labview And Matlab Code
Fadhil A. Ali
Department of Electrical and Computer Engineering- Oklahoma State University
202 Engineering South Stillwater, OK 74078 USA- Tel:405-714-1084
Abstract
This work is proposed the feed forward neural network with symmetric table addition method to design the
neuron synapses algorithm of the sine function approximations, and according to the Taylor series
expansion. Matlab code and LabVIEW are used to build and create the neural network, which has been
designed and trained database set to improve its performance, and gets the best a global convergence with
small value of MSE errors and 97.22% accuracy.
Keywords
Neural networks, Symmetric Table Addition Method, LabVIEW , Matlab scripts
1.Introduction
1.1 Artificial Neural Network
An Artificial Neuron (ANN) is a model of biological neuron. Many of ANN receives signals from
the environment or other ANNs, and its gathering these signals by using activation function to
the signals sum. Input signals are excited through positive or negative numerical weights that
associated with each connection to the ANNs, thus firing of the ANN and the strength of the
exciting signal are controlled via a function referred to as activation function. The ANNs are
collecting all incoming signals and computing a net input signals as a function of the respective
weights. The net input renders to the activation function by calculated the output signal of the
ANNs, which is called the layered network. ANN may consist of input, hidden and output layers
[1], as shown in Figure (1).
Figure (1) Neural Network Architecture
The ANN is designed to work similar to neural tissue in the brain, where many independent
neurons (processing units) are interconnected into one large data-processing network. In the ANN

2
the “neurons” are represented as mathematical functions, which pass the data between themselves
in some defined and organized manner. The way of organization of the connections between the
“neurons” defines the features of the ANN, and its capability to perform certain tasks [2].
One of the simplest ANN’s is a feed-forward network, which it has worked in this investigation,
therefore other types of ANN will not be considered here. Such a network consists of: one input
layer, hidden layers (i.e. nodes)-(possibly more than one) and one output layer. The data are
transferred in a form of “signals”, it means values which are passed between the nodes. Each
connection among the nodes has a special variable associated with it – this is a weight. Each
neuron collects all the input values multiplied by the associated weights and processes them with
activation function:
S , = ∑ S , . W , ………….. (1)
where: f – Activation function, Sx,i – output signal of ith neuron in xth layer, Sx-1,i – output of ith
neuron in x-1th layer, Wx,i – weights connected to neuron Sx,i. The ‘response’ is an activation
function output of a given neuron, that is simply named the “neuron function”.
The most activation functions are:
1.Sigmoid function :
( ) =
(
σ
)
………………….. (2)
2.Gaussian function :
( ) = exp (−
σ
)…………………. (3)
Definitely, any neural network is training a number of times that will give the results in a stronger
weight of neurons but could not exceed the specific limits. This can be causes the network to
memorize instead of learning [3].
1.2 Symmetric Table Additional Method
Symmetric Table Addition Methods (STAMs) use two or more parallel table lookups followed by
multi operand addition to approximate the elementary functions. These methods stake advantage
of symmetry and leading sign bits in the table entries to drastically reduce the size of the table. In
[5] the method produces tables for the symmetric table addition method for approximating a
function f(x). The inputs n0, n1, n2, and n3, which corresponds to the number of bits in x0, x1,
and x2 are taken by STAM , where x = x0 + x1 + x2 + x3. It also takes as input the number of
guard digits," ng ", and a variable " f" , which indicates the function to be implement. This
procedure computes the coefficients produced by this approximation method, and reports on the
maximum and average error of the approximations.
According to Nihal Koc-Sahan et al[6], the multi-layer perceptron networks, the inputs to the
sigmoid function and its derivate correspond to sums of weighted values. And x typically has the
form;
x=bm-1bm-2….b0b-1b-2…b-p ………….(4)
Where x is n-bit fixed point number.

3
Then STAMs divide the input operand, x, into m+1 bit partitions, x0,x1,…xm. This approximation
is based on two term Taylor series expansion.
1.3 Decision Tree
The decision tree (DT) is an algorithm tool that uses a tree-like graph or model for taken
decisions and their possible consequences, including chance event outcomes. The DTs are
commonly used in operations research and decision analysis, to help identification strategy of the
most likely to reach a goal. This goal is to create a model that predicts the value of a target
variable based on several input variables.
The tree also can be "learned" by splitting the source set into sub-sets based on an attribute value
tests. This process is repeated on each derived sub-set in a recursive manner , which is recursive
partitioning. The recursion process is completed when the sub-set at a node for all that has the
same value of the target variable, or when splitting no longer adds value to the predictions.
In data mining, many trees algorithms can be described as the combination of mathematical and
computational techniques to aid the description, categorization and generalization of a given set
of data. Data refers in the specific records of the form:
(x, Y)=(x 1, x 2, x 3… x k, Y) ……................…. (5)
Y is the dependent variable, which is the target variable that must be understand and classify.
Also, a vector x is composed of the input variables ( x1, x2, x3...xk ), those are used for such task of
calculations [7, and 8].
2.Proposed method
In traditional neural network with sigmoid activation function has the advantage, it can
approximate any continuous function but gets the fail in ending up as large network as possible.
The goal in this paper is to design a neuron synapse of multi-layer network. The proposed method
uses sine function as building blocks, and combine together by adapting STAMs parameters to
build a final function. This has achieve a fitting performance for the training data of the present
network, and it has taking a tree formulation (DT) which the function was modeled. Since that
many literatures have proposed and used tree form of neural networks [9, 10, and 11].
In computer algebra and symbolic computations a function can be represented as tree
representation [12]. Its tree form of sine function and operation gets our neural network of trained
data to build the node. The node represents a sine function applied to the sum of STAM’s
parameters, and each terminal node represents input variables and gets its child nodes.
By using Taylor series expansion of sine function E ( );
E( )= − !
+ !
− !
+ ⋯,
= 0 …….....…...........………. (6)
Where is the input argument and are parameters (weights) in our network.
Let us construct the tree
Y=E( 2+ E ( 1)) + E ( 2) .......………. (7)

4
The training data set has taken an incremental way to perform new data set. The algorithm is used
the following steps for getting a forward network [13];
1- Start from a blank network. Initialize all weights to zero.
2- Testing the effect of adding a layer to the network performance, as follows:
a- Set up the current network configuration.
b- Add a layer to the current network by selecting the layer and terminating node types.
c- The nodes weights for the existing network are initialized by STAM’s values,
d- The optimization algorithm is applied to the new network scheme and find out the
resulting network performance accordingly.
e- Keeping the new added layer.
f- Repeat step 2 to get all possible sine function Taylor series terms.
The database is selected as randomized values during the test interval, in order to make sure that
the network learns instead of memorize. Insufficient training may lead to incorrect results. Thus,
training must be done until 100% training accuracy is achieved. Figure (2) shows the flowchart
for training and testing the database.
.
Figure (2) Flowchart of the training and testing data
On Taylor series, there are expansions for sin function, which are given in terms of odd powers of
x, in this work 6 input nodes have been taken with x,x3
,x5
,x7
,x9
, and x11
. Therefore, all the powers
of x with alternating zero weights; (Sin: 0, 1, 0, -1/6...,). Therefore, the built up neural network
has got each input to one corresponding node, and multiplied it by its weight to fed entire possible
outputs to the output nodes, it would be capable of the 6-terms Taylor's expansion of a sine
function.

5
Inputs Hidden layer Outputs
x1
--------------- H0 --------------------///
x3
--------------- H1 --------------------///
x5
--------------- H2 --------------------///
x7
--------------- H3 --------------------///
x9
--------------- H4 --------------------///
x11
--------------- H5 --------------------///
Figure (3) explicit representation of the presented neural network
3 .Symmetric table additional method (STAM) - Neuron
The conventional neural network shown in Figure (3) has shown the single variable weight i is
replaced by STAM block diagram shown in Figure (4) [6], which a multi operand adder has
cancelled to get a STAM neuron been established and it is shown in Figure(5).
Figure (3) a conventional neural network

6
Figure (4) STAM block diagram
Figure (5) STAM‘s Neuron
4.LabVIEW Design
In this section, the system architecture is described in LabVIEW block diagram. It has four
processing stages as seen in Figure (6).The system is initialized when a data collected from
datasheets, followed with neural network configuration, draw the network , resize the inputs and
finally initiate nodes. LABVIEW is used as a design platform for the system that is illustrated in
flowchart steps.
Figure (6) NN LabVIEW System’s flowchart
The processing of collected data includes inputs and operands of STAM output. This work
focuses on application of STAM algorithm gets the design of neuron synapses of the Neural
Network for a sine function which is the third stage of the designed system.

7
A database was created to accomplish Neural Network System (NN system) comprising data used
for experimental purpose. The database is pre-trained and tested using Matlab script function[14],
which is integrated with LabVIEW for real- time testing.
5.Results & Discussion
This section detaches the parameter setting for the neural network architecture system. In addition
to that, the (training, testing and real- time) data resulted are also discussed in some details.
5.1 Parameters Setting
After many attempts, it was found that the following parameters yield around 100% of training
accuracy. The error measure used to assess the network performance is Mean Square Error
(MSE) as follows [13];
MSE =
( ) ( )
……...............................................................(8)
where y (j), d (j) are the network output and the desired output j respectively, and J is the
size of the data set. Most of error measures and the number of weights in the resulted
network, which is given to compare the network complexities. Table (1) shows the most
parameters used with the training results.
Table (1) Details of Training and Testing for ANN
The data training has been conducted, and tan-sig is chosen as a transfer function for the layers as
it converges. Also, the number of hidden neurons has been set to 6 neurons. The Learning rate
and performance goal is set to a lower value , so that the network learns completely. A
momentum factor of 0.9 is added in the network which can be recover back for training even if it
falls into local minimum.
5.2 Training Results
In figure (7), 100% training accuracy met with 50 iterations. It shows the neural network toolbox
illustrating the transfer function, iterations and performance goal that is achieved during training.
The amount of 100 terms out of 112 inputs are 89% of the data were successfully trained as seen
Feed forward neural network algorithm
Defined Parameters
Transferred function Tan sig
Number of Hidden Neurons 6
Set the Performance Goal 1e-10
The Momentum Factor 0.9
The Learning Rate 0.01
The Total number of terms in the Taylor series
(sine function): 200
Training and Testing Parameters
The Training Epoch 50
The Training Time 0:00:01
The Training Accuracy 100%
The Testing Accuracy 97.22%

8
in the performance and regression plots in figure (8) and figure (9), with 100% training accuracy
respectively. Regression plot has shown the data are fitted the line when the training is complete.
Also, training state is shown in figure (10) indicates the closer in the accuracy percentage.
Figure (7) NN system toolbox
Figure (8) NN system performance

9
Figure (9) NN system regression plots
Figure (10) NN system training state plots
5.3 Testing Results
All the terms in the database which is 112 were used for testing purpose. The data was
randomized before testing so that the system has an accurate performance for sine function
(Target). The testing accuracy that has been achieved is 97.22% for 112 inputs of the system.
According to this result, that achieved a higher rates, figure (11) is shown the sine function in
both exact and predicted by the present NN system, while figure (12) illustrates the fitting curves
of the resulted sine function.

10
Figure (11) sine function plots
Figure (12) Fitting of sine function plot
5.4 Real-Time Testing
Real- time testing has been done for the complete system. The trained network was integrated in
LABVIEW. A GUI panel was designed in LABVIEW indicating the neural network graph in
database as set in Table (1). Random numbers is generated if targeted function data was found not
to be in database. This can be seen from figures (13, and 14) respectively, the frontal system was
designed in LABVIEW for a real- time application with a high accuracy.

11
Figure (13) configuration frontal of NN system
Figure (14) NN system training frontal
6 .Conclusion
This paper gives a proposed method to create symmetric table additional method (STAM) neuron.
Which it represented the general aspects of a sine function behavior and indicates the neuron
synapses in neural network design for a real-time application used and designed in LABVIEW
integrating the Neural Network from MATLAB codes embedded in the system. A minimized
MSE rate to give the presented value 97.22% has been achieved of the neural network accuracy.
7.Acknowledgement
Many thanks to Oklahoma State University and – School of Electrical and Computer Engineering
Department for hosting me as visitor scholar. My deep appreciation to Dr. James E. Stine for his
valuable notes. Also, many thanks to SRF/IIE for giving me all possible supports.
References
[1] Andries P. Engelbrecht, Computational Intelligence: An Introduction - second edition, John Wiley &
Sons Ltd, 2007
[2] D.T. Pham, X. Liu, Neural networks for identification, prediction and control, Springer-Verlag
London Ltd. UK, 1995
[3] N. F. Naim, A. I. M. Zakaria, N. A. Wahab, Classification of Thumbprint using Artificial Neural
Network (ANN), IEEE International Conference on System Engineering and Technology (ICSET),
Shah Alam, 2011, pp. 231- 234
[4] A. Reda, B. Aoued, Artificial Neural Network Based Face Recognition, First International
Symposium on Control, Communications and Signal Processing, Algeria, 2004, pp. 439- 442
[5] J.E.Stine and M.J.Schulte, The Symmetric Table Addition Method for Accurate Function
Approximation, Journal of VLSI Signal Processing Systems for Signal, Image, and Video
Technology, vol.21,1999,pp.167-177
[6] Nihal Koc-sahan , Jason A. Schlessman , Michael J. Schulte, Symmetric table addition methods for
neural network approximations,Proc. SPIE 4474, Advanced Signal Processing Algorithms,
Architectures, and Implementations XI, 2001,pp.126-133

12
[7] G. P. J. Schmitz, C. Aldrich, and F. S. Gouws, ANN-DT An Algorithm for Extraction of Decision
Trees from artificial Neural Networks, IEEE Transactions on Neural
Networks,vol.10,No.6,1999,pp.1392-1401
[8] I. K. Sethi, Layered Neural Net Design Through Decision Trees, IEEE International Symposium on
Circuits and Systems, vol.2,1990,pp.1082-1085
[9] S. Behnke, and N. B. Karayiannis, Competitive neural trees for pattern classifications IEEE
Transactions on Neural Networks,vol.9,No.6,1998,pp.1352–1369
[10] A. Sankar, and R. J. Mammone, Growing and pruning neural tree networks, IEEE Transactions on
Computers,vol.42,No.3,1993, pp.291–299
[11] B.T. Zhang, A Bayesian evolutionary approach to the design and learning of heterogeneous neural
trees, Integrated Computer- Aided Engineering,vol.9,2002,pp.73–86
[12] J. S. Cohen, Computer algebra and symbolic computation: Elementary algorithms. Natick: A. K.
Peters, Ltd,2002
[13] George S. Eskander and Amir F. Atiya, neural networks, Neural Networks ,vol.22 ,2009,pp. 395–404
[14] Forrester W. Isen , DSP for MATLAB and LabVIEW: Fundamentals of Discrete Signal Processing,
Morgan & Claypool publishers ,2009

Feed forward neural network for sine

Recommended

More Related Content

What's hot (19)

Viewers also liked (8)

Similar to Feed forward neural network for sine (20)

Recently uploaded (20)

Feed forward neural network for sine