Scribd
Upload
80 views

Neural Network Lectures RBF 1

This document provides an overview of radial basis function (RBF) networks. It begins by discussing linear models and their limitations, as well as methods like kernel methods and large margin techniques that make linear models more powerful. It then introduces various types of basis functions that can be used in RBF networks, including sigmoids, Gaussians, and polynomials. The document explains that RBF networks learn the parameters of the basis functions, making them more powerful than linear models but also more complex. It concludes by defining RBFs mathematically and listing examples of globally and compactly supported RBFs.

Uploaded by

Yekanth Ram
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
80 views

Neural Network Lectures RBF 1

This document provides an overview of radial basis function (RBF) networks. It begins by discussing linear models and their limitations, as well as methods like kernel methods and large margin techniques that make linear models more powerful. It then introduces various types of basis functions that can be used in RBF networks, including sigmoids, Gaussians, and polynomials. The document explains that RBF networks learn the parameters of the basis functions, making them more powerful than linear models but also more complex. It concludes by defining RBFs mathematically and listing examples of globally and compactly supported RBFs.

Uploaded by

Yekanth Ram
Copyright
© © All Rights Reserved
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 44

Neural Network Lectures

ME 60033 IMS 2014-15


Radial Basis Functions
Dr C S Kumar, Robotics and Intelligent
Systems Lab

Is it a Neural Network?

Unsupervised Learning

References
1. Simon Haykins Neural Networks and Learning
Machines
2. Christopher Bishop Neural Networks in Pattern
Recognition
3. Kevin Gurney Introduction to Neural Networks
4. Hertz, Krogh, Palmer Introduction to theory of
Neural computation (Santa Fe Institute Series)
5. Handbook of Brain Sciences & Neural Networks

Linear models
It is mathematically easy to fit linear models to data.
We can learn a lot about model-fitting in this relatively simple case.
There are many ways to make linear models more powerful while
retaining their nice mathematical properties:
By using non-linear, non-adaptive basis functions, we can get
generalised linear models that learn non-linear mappings from input
to output but are linear in their parameters only the linear part of
the model learns.
By using kernel methods we can handle expansions of the raw data
that use a huge number of non-linear, non-adaptive basis functions.
By using large margin kernel methods we can avoid overfitting even
when we use huge numbers of basis functions.
But linear methods will not solve most AI problems.
They have fundamental limitations.

Some types of basis function in 1-D

Sigmoids

Gaussians

Polynomials

Sigmoid and Gaussian basis functions can also be used in multilayer


neural networks, but neural networks learn the parameters of the basis
functions. This is much more powerful but also much harder and much
messier.

Two types of linear model that are equivalent


with respect to learning
bias

y (x, w ) = w0 + w1 x1 + w2 x2 + ... = w T x
y (x, w ) = w0 + w11 (x) + w22 (x) + ... = w (x)
T

The first model has the same number of adaptive coefficients as


the dimensionality of the data +1.
The second model has the same number of adaptive coefficients
as the number of basis functions +1.
Once we have replaced the data by the outputs of the basis
functions, fitting the second model is exactly the same problem
as fitting the first model (unless we use the kernel trick)
So its silly to clutter up the math with basis functions

The loss function


Fitting a model to data is typically done by finding the
parameter values that minimize some loss function.
There are many possible loss functions. What criterion should
we use for choosing one?
Choose one that makes the math easy (squared error)
Choose one that makes the fitting correspond to maximizing
the likelihood of the training data given some noise model
for the observed outputs.
Choose one that makes it easy to interpret the learned
coefficients (easy if mostly zeros)
Choose one that corresponds to the real loss on a practical
application (losses are often asymmetric)

Minimizing squared error


y = w x
T

error =

2
T
(

)
t
w
x
n
n
n
T

w = ( X X)
*

optimal
weights

inverse of the
covariance
matrix of the
input vectors

X t

vector of
target values

the transposed
design matrix has
one input vector per
column

A geometrical view of the solution


The space has one axis for each
training case.
So the vector of target values is a
point in the space.
Each vector of the values of one
component of the input is also a
point in this space.
The input component vectors
span a subspace, S.
A weighted sum of the input
component vectors must lie
in S.
The optimal solution is the
orthogonal projection of the
vector of target values onto S.

3.1 4.2
1.5 2.7
0.6 1.8
component
vector

input vector

When is minimizing the squared error equivalent to


Maximum Likelihood Learning?
Minimizing the squared residuals
is equivalent to maximizing the
log probability of the correct
answer under a Gaussian
centered at the models guess.
t = the

yn = y ( x n , w )

correct
answer

p (t n | yn ) = p ( yn + noise = t n | x n , w ) =
log p (t n | yn ) = log 2 + log +
can be ignored if
sigma is fixed

y = models

estimate of most
probable value

1
2

( t n yn ) 2
2 2

(t n yn ) 2
2 2

can be ignored if
sigma is same for
every case

Multiple outputs
If there are multiple outputs we can often treat the learning
problem as a set of independent problems, one per output.
Not true if the output noise is correlated and changes from
case to case.
Even though they are independent problems we can save
work by only multiplying the input vectors by the inverse
covariance of the input components once. For output k we
have:

w *k = ( XT X) 1 XT t k
does not depend
on a

Least mean squares: An alternative approach for really


big datasets

+1

weights after
seeing training
case tau+1

= w En ( )
learning
rate

vector of derivatives of the


squared error w.r.t. the weights
on the training case presented
at time tau.

This is called online learning. It can be more efficient if the dataset is very
redundant and it is simple to implement in hardware.
It is also called stochastic gradient descent if the training cases are picked
at random.
Care must be taken with the learning rate to prevent divergent
oscillations, and the rate must decrease at the end to get a good fit.

Regularized least squares


~
1 N
E (w ) = { y (x n , w ) t n }2 +
2
n =1

|| w ||2

The penalty on the squared weights is mathematically compatible with the


squared error function, so we get a nice closed form for the optimal weights
with this regularizer:

w * = ( I + XT X) 1 XT t
identity
matrix

Radial Basis Functions


Let : R + R be a continous function with (0) 0. If xi , let

i (xi ) = x xi ,
where is the Euclidean norm. Then i is called the RBF.
Linear:

Cubic:

r3

Multiquadrics:

r 2 + c 2 where c is a shape parameter.

r 2 n log r , n 1, in 2D,
2 n 1
in 3D.
r , n 1,

Polyharmonic Spines:
Gaussian:
2014/8/25

cr 2
29

Globally Supported RBFs


=1+r

2014/8/25

= r 2 + c2

30

Compactly Supported RBFs


= (1 r ) 2+

2014/8/25

= (1 r ) 4+ (4r + 1)

31

You might also like