You are on page 1of 7

Multilayer Perceptron

1
Nonlinear Input-Output Mapping

Approximation of functions

input nodes output nodes

m0-dimensional inner product space M-dimensional IP space


Continuously differentiable (like the activation functions)

Number of hidden layers

2
Universal Approximation Theorem

Weights of a single Weights of a single Input to NN!


output neuron hidden layer
without nonlinearity
3
• Generalization of Fourier Series? (finite, hence
approximation)

• A single hidden layer is sufficient for a multilayer


perceptron to compute an appropriate approximation
to a given training set represented by the set of input?!

• Does not say that a single hidden layer is optimum in


the sense of learning time, ease of implementation, or
generalization.

Bounds…

Function approximation on training data, Estimator on the


test data!

4
Multi-dimensional Fourier Transform

Assuming integrability, first absolute moment of


Fourier magnitude distribution is defined

Euclidean

… bound on the approximation

5
Number of neurons in the
single hidden layer

This comes after all the conditions of the universal


approximation theorem is met!

Another!
Input nodes

Total hidden neurons


6
Curse of dimensionality
Minimization with respect to N

Curse of Rate of convergence


dimensionality to the minimization

Lesser availability of dense samples for the


estimation, while higher could be the function
complexity with increase in dimension!
7

You might also like