Scribd
Upload
675 views3 pages

Time Delay Neural Networks

The document discusses time delay neural networks. It describes how these networks take time-series inputs to recognize patterns over time. The network can selectively replicate previous input patterns by sending a stronger signal to activate the appropriate hidden neurons. A new structure is presented that allows the network to reset its weights, taking inputs sequentially to learn patterns and then selectively reproduce any learned pattern on demand. Mathematical equations are provided for the input, hidden and output layers to describe the network's operation.

Uploaded by

navinsehgal
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOC, PDF, TXT or read online on Scribd
675 views3 pages

Time Delay Neural Networks

The document discusses time delay neural networks. It describes how these networks take time-series inputs to recognize patterns over time. The network can selectively replicate previous input patterns by sending a stronger signal to activate the appropriate hidden neurons. A new structure is presented that allows the network to reset its weights, taking inputs sequentially to learn patterns and then selectively reproduce any learned pattern on demand. Mathematical equations are provided for the input, hidden and output layers to describe the network's operation.

Uploaded by

navinsehgal
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as DOC, PDF, TXT or read online on Scribd
You are on page 1/ 3

Selection of a pattern in time delay neural network based on input

techniques

Basics: A time delay neural network is largely based on time series which is largely
dependent on the input as a sequence of time inputs.

X(t)

X(t-1)

X(t+1)
X(t-2)

X(t-3)

A time delay neural network has an input equation based on z-transform.The input in
such a case is based on a time series.These time based input creates a distinct pattern in
the network.

Z-transform- The function notation for sequences is used in the study and application of
z-transform.Consider a function X(t) defined for t >=0 that is sampled at times t =
T,2T,3T,…… where T is the sampling period .We can write the sample as a sequence
using the notation { Xn = x(nT) n=0 to ∞ }

Inverse Z transform : Let x(z) be the transform of sequence { Xn}.Then Xn is given by


the formula
Xn = Ζ-1[ X(z)] where Z1,Z2,Z3…….Zk are the poles of f(z) = X(z) Z n-1
The functions corresponding to Z transform are given by
S = { b,bn,e2n,n,n2,nm,nbn,ne2n}
As an example find the Z-transform for the input equation X = bn
∞
X(z) = ∑bnZ
n =0
− n = z/ z-b

We can also define various properties of

(i) Linearity. .

(ii) Delay Shift. .

(iii) Advance
Shift. , or
Now we define a new type of neural network which not only takes inputs through a time
series trying to recognize the final input through the neural network .It also tries to
replicate selective patterns fed to it at previous times.

Model Structure of a time delay neural network

2 X(5)

The above time delay neural network takes input on each neuron of it’s input layer in the
format 1,2,3,4.
The output at any neuron in the hidden layer as well as the output layer is defined by the
Z-transform equation.
These inputs modify the weights and cause the learning process.The inputs in time series
1,2,3,4 cause the neurons in the hidden layer to be modified and 4 inputs the result is
obtained in the 5th time slot from the output neuron.

New Structure : Under the new structure the time delay network has four inputs
T0,T1,T2,T3

T0
H0
T1
H1 T4
T2 T0

T3

C
The above network has five neurons inputs T0,T1,T2,T3,C .Each of these inputs is
focused onto two neurons H0 and H1 of the hidden layer.

It is different from the previous neural networks because this network not only takes
input in time series but also this network can selectively produce any previous input
pattern however complex it is.

There is also a C input neuron which causes all the hidden neuron weights to be reset.

Selection of previous patterns: This can be achieved by first inputting patterns


T0,T1,T2,T3 one by one.
We use a threshold function due to which we can replicate any of the input patterns
selectively by sending an input signal with a higher value.

Mathematical Structure:

Inputs -> T0,T1,T2,T3


Weights -> w00,w01,w10,w11,w20,w21,w30,w31
Hidden-> H0,H1
Hidden weights-> W00,W01,W10,W11
Output-> O1

4
H 0(n) = ∑TnWno
n =1
4
H 1(n) = ∑TnWn 1
n =1
2
O ( k ) = ∑H ( k )Wk
k =1

You might also like