Radial Basis Function Networks - PowerPoint PPT Presentation


PPT – Radial Basis Function Networks PowerPoint presentation | free to download - id: 4b7870-OGM2M


The Adobe Flash plugin is needed to view this content

Get the plugin now

View by Category
About This Presentation

Radial Basis Function Networks


In the name of God Institute for advanced studies in basic sciences Radial Basis Function Networks Yousef Akhlaghi || xi xj || = [(xi1 - x1j)2 + (xi2 x2j)2 ... – PowerPoint PPT presentation

Number of Views:397
Avg rating:3.0/5.0
Slides: 29
Provided by: pey3
Learn more at: http://www.iasbs.ac.ir


Write a Comment
User Comments (0)
Transcript and Presenter's Notes

Title: Radial Basis Function Networks

In the name of God
Institute for advanced studies in basic sciences
Radial Basis Function Networks
Yousef Akhlaghi
This seminar is an introduction to radial basis
function networks as linear neural networks.
Definition Radial basis function (RBF)
networks are a special class of single
hidden-layer feed forward neural networks for
application to problems of supervised learning.
Linear models have been studied in statistics for
about 200 years and the theory is applicable to
RBF networks which are just one particular type
of linear model. However the fashion for neural
networks which started in the mid-80s has given
rise to new names for concepts already familiar
to statisticians.
Supervised Learning
Problem to be solved Estimate a function from
some example input-output pairs with little or no
knowledge of the form of the function.
Nonparametric regression
Function approximation
System identification
Inductive learning
Supervised learning
Two main regression problems in statistics
- parametric regression
- nonparametric regression
Parametric Regression
In parametric regression the form of the
functional relationship between the dependent and
independent variables is known but may contain
parameters whose values are unknown and capable
of being estimated from the training set.
For example fitting a straight line
to a bunch of points
Important point The free parameters as well as
the dependent and independent variables have
meaningful interpretations like initial
concentration or rate.
Nonparametric regression
There is no or very little a priori knowledge
about the form of the true function which is
being estimated.
The function is still modeled using an equation
containing free parameters but Typically this
involves using many free parameters which have no
physical meaning in relation to the problem. In
parametric regression there is typically a small
number of parameters and often they have physical
In neural networks including radial basis
function networks - Models are nonparametric
and their weights and other parameters have no
particular meaning in relation to the problems to
which they are applied. -The primary goal is to
estimate the underlying function or at least to
estimate its output at certain desired values of
the input
Linear Models
The model f is expressed as a linear
combination of a set of m fixed functions often
called basis functions by analogy with the
concept of a vector being composed of a linear
combination of basis vectors.
Nonlinear Models
if the basis functions can change during the
learning process then the model is nonlinear.
An example
Almost the simplest polynomial is the straight
which is a linear model whose two basis functions
and whose parameters (weights) are
The two main advantages of the linear character
of RBF networks
1- Keeping the mathematics simple it is just
linear algebra (the linearly weighted
structure of RBF networks)
2- There is no optimization by general purpose
gradient descent algorithms (without involving
nonlinear optimization).
Radial Functions
Their characteristic feature is that their
response decreases (or increases) monotonically
with distance from a central point. The centre,
the distance scale, and the precise shape of the
radial function, are parameters of the model, all
fixed if it is linear.
A typical radial function is the Gaussian
Its parameters are c
centre r width
Local modelling with radial basis function
networks B. Walczak , D.L. Massart
Chemometrics and Intelligent Laboratory Systems
50 (2000) 179198
From exact fitting to RBFNs
(No Transcript)
In matrix notation the above equation is
For a large class of functions, the matrix F is
non-singular, and eq. 3 can be solved
The basis functions can have different forms. The
most popular among them is the Gaussian function
controlling the smoothness properties of the
interpolating function.
(No Transcript)
xi xj
  • (xi1 - x1j)2 (xi2x2j)2 (xin-xnj)20.5

Distance (x)


Hidd. nodes objects
Exact Fitting
In practice, we do not want exact modeling of the
training data, as the constructed model would
have a very poor predictive ability, due to fact
that all details noise, outliers are modeled.
To have a smooth interpolating function in which
the number of basis functions is determined by
the fundamental complexity of the data structure,
some modifications to the exact interpolation
method are required.
1) The number K of basis functions need not
equal the number m of data points, and is
typically much less than m.
2) Bias parameters are included in the linear
3) The determination of suitable centers becomes
part of the training process.
4) Instead of having a common width parameters,
each basis function is given its own width sj
whose value is also determined during training.
Architecture of RBFN
RBFN can be presented as a three-layer
feedforward structure.
  • The input layer serves only as input distributor
    to the hidden layer.
  • Each node in the hidden layer is a radial
    function, its dimensionality being the same as
    the dimensionality of the input data.
  • The output is calculated by a linear combination
    . i.e. a weighted sum of the radial basis
    functions plus the bias, according to

In matrix notation
Training algorithms
RBF network parameters
  • The centers of the RBF activation functions
  • The spreads of the Gaussian RBF activation
  • The weights from the hidden to the output layer

subset selection
Different subsets of basis functions can be drawn
from the same fixed set of candidates. This is
called subset selection in statistics.
  • forward selection
  • starts with an empty subset
  • added one basis function at a time (the one that
    most reduces the sum-squared-error)
  • until some chosen criterion stops
  • backward elimination
  • starts with the full subset
  • removed one basis function at a time ( the one
    that least increases the sum-squared-error)
  • until the chosen criterion stops decreasing

Orthonormalization of basis functions
Initially, the basis functions are centered on
data objects.
Nonlinear training algorithm
  • Apply the gradient descent method for finding
    centers, spread and weights, by minimizing the
    cost function (in most cases squared error).
    Back-propagation adapts iteratively the network
    parameters considering the derivatives of the
    cost function with respect to those parameters.

Drawback Back-propagation algorithm may require
several iterations and can get stuck into a
local minima of the cost function
Radial Basis Function Neural Net in MATLAB
function net, tr newrb (p, t, goal, spread,
mn, df)
NEWRB adds neurons to the hidden layer of a
radial basis network until it meets the specified
mean squared error goal.
P - RxQ matrix of Q input
vectors. T - SxQ matrix of Q
target class vectors. GOAL - Mean
squared error goal, default 0.0. SPREAD
- Spread of radial basis functions, default
1.0. MN - Maximum number of
neurons, default is Q. DF - Number
of neurons to add between displays, default 25.
The following steps are repeated until the
network's mean squared error falls below GOAL or
the maximum number of neurons are reached
1) The network is simulated with random weight
2) The input vector with the greatest error
is found 3) A neuron (basis function) is
added with weights equal to that vector.
4) The output layer weights are redesigned to
minimize error.
About PowerShow.com