[comp.ai.neural-nets] number of hidden nodes needed ?

martink@whiteface.crd.ge.com (Kenneth Martin) (06/17/91)

In doing some development work on forming completely enclosed
boundaries, I stumbled across a lower limit on the number of 
hidden nodes required to form a completely enclosed boundary
in an N dimensional input space using a feed forward ANN with
a monotonically increasing activation function. No one around
here has heard of such a lower limit.  Do any of you know if 
such a limit has already been written up ? If so where ?

Thanks in advance

- Ken Martin 
- martink@whiteface.crd.ge.com
- (518) 387-7014

esrmm@warwick.ac.uk (Denis Anthony) (06/18/91)

In article <20652@crdgw1.crd.ge.com> martink@whiteface.crd.ge.com (Kenneth Martin) writes:
>In doing some development work on forming completely enclosed
>boundaries, I stumbled across a lower limit on the number of 
>hidden nodes required to form a completely enclosed boundary
>in an N dimensional input space using a feed forward ANN with
>a monotonically increasing activation function. No one around
>here has heard of such a lower limit.  Do any of you know if 
>such a limit has already been written up ? If so where ?

I think the following may be relevant :-

@article{
	title="Bounds on the number of hidden neurons in multilayer perceptrons",
	author="Huang S and Huang Y",
	year=1991,
	volume=2,
	number=1,
	pages="47--55",
	journal= "{IEE} transactions on neural networks",
	}

ps. What is your answer for lower limit ?