Chebyshev Neural Network
Term Paper - Scientific Computing (MA-579H )
Suraj Kesharwani
Roll: 204161016
M.Tech - Data Science (2020 - 2022)
Indian Institute of Technology, GuwahatiINTRODUCTION
There is a wide range of Neural Network applications in classification and
regression problems and neural network shows that it is efficient in those areas.
The Multilayer Perceptron(MLP) with Backpropagation is very robust to perform
a complex task in a non-linear environment but the drawback is, that it can be
trapped in local minima. The problem can be resolved by the Chebyshev Neural
network. As it produces better accuracy and fast convergence rate with the
backpropagation algorithm compared with the MLP.
Structure
The structure of the Chebyshev Neural Network is based on the Chebyshev polynomials.
The architecture consists of two parts. ie
1. Numerical transformation Part
2, Learning Part
‘The Numerical transformation part is used for giving the input to the hidden layer by
expanding the input as a finite number of Chebyshev polynomials of the first order. The
result of the Chebyshev polynomial can be viewed as a new input vector. The Chebyshev
polynomial of the first kind T,(Z) can be defined by contour integral function.
xg 4 —Pyxent dt
TO tae
ma
The architecture consists of an input layer, a hidden layer, and an output layer. The
Weights between the input layer and hidden layer are set to 1 and the weights from
hidden to output are randomly generated in the range of -1 to 1. The hidden layer has 4
neurons and input to the hidden layer are the transfer functions which is a group of
Chebyshev orthogonal polynomial functions Tn) where X is the input
The Chebyshev polynomials can be generated by using the following recursive formula is
Ty
‘The first few Chebychev polynomials of the first kind are
2 * X x TX) - T, 00
TAM=1; TCQ=XTO= 2X1; TO = 4x? -3xXFigure 1, Architecture of CNN
Set the Chebyshev neural network as follows,
Input layer: $ = X
‘The input of hidden layer neuron is Net, = S
The output of hidden layer neuron is Out, = T, (Net)
Output layer: ¥ = W, T,
‘The activation function is sigmoid function ie fos) = =
Error can be calculated as E = f(X,)-Y,
In this proposed network there is no need to do the weight updation between the hidden
and input layer. To converge the network to produce the desired output, the weight
updation can be done only between the hidden and output layer by using the following
formula.
&(W) = (Output, - Target) * Output, (1 - Output)
Adjust the weights, Weights = Weights,+ 8(W))Experiment Results
‘We will take the non-linear function of
Y=0.5 *Tanh(25*V) - Tanh(V) + 0.2 * Tanh(V) + 0.03 * V
The result shows Chebyshev Neural networks converge and performs better than MLP.
‘The number of samples for training the network is 50. The learning rate set as 0.5. The
number of epochs taken for the network to realize the function is 200 and produces the
mean square error as E = 1.0 x 10°. The weight between the hidden and output are
updated in each epoch. The final approximation curve produced by the network along
with the actual output curve is plotted.
Train Data
(045 | — NN Output
Train Data
040
035:
» oo 02 i) 06 08
X- Axis — Input Y-axis - Output
Figure 2. CNN Plot for Training Dataset
Train Data
— WN output
04) ow Train Data
02
1)
09 02 04 06 08
X- Axis— Input; Y-axis - Output
Figure 3. MLP Plot for Training Dataset0.45
0.40
035
045
0.40
035
Test Data
— NWtest Output
Test Data
Vv
02 03 04 05 06 07
X- Axis — Input; Y-axis - Output
Figure 4. CNN Plot for Testing Dataset
Test Data
Nitest ouput
wo Test Data
02 03 04 05 06 07 1
X- Axis — Input; Y-axis - Output
Figure 5. MLP Plot for Testing Dataset
Results
Neural Network No of Samples No of Epoch Mean Square Error
Chebyshev (Testing | .. 200 6* 10?
Data)
Chebyshev (Test | , *102
Data) * : ereMLP (Training 50 1000 1*107
Data)
MLP (Testing Data) | 20 1 5* 10%
Summary
We have approximated a function with two different Neural Network which is
Multi-layer Neural Network and Chebyshev Neural Network. The result shows that
Chebyshev Neural Networks converges faster than the multilayer neural network and
also produces more accuracy. The advantage of Chebyshev based network is since the
network has a fixed structure, it can be implemented easily and the computation
complexity also drastically reduced. The result concludes that the proposed network is
five times faster.
REFERENCES
1, Yan, S.P, et al. "C2 concentration detection based on Chebyshev neural network
and best approximation theory." Instrument Technique and Sensor 6 (2011):
107-110
2, Mall, Susmita, and Snehashish Chakraverty. “Numerical solution of nonlinear
singular initial value problems of Emden-Fowler type using Chebyshev Neural
Network method.” Neurocomputing 149 (2015): 975-982.
3. MSornam1, V.Vanitha “Application of Chebyshev Neural Network for Function
Approximation” International Journal of Computer Science and Engineering
(2018), EISSN: 2347-2693