| Usenet Newsgroups |
Re: What are neural nets good at, what are they bad at? (What trips them up?)
From: chi.fu@air.com.au (Tomasso)
On Jul 20, 5:19 am, "Tomasso" <chi...@air.com.au> wrote:
JordanF98765 wrote:
I feel like they're just this super artificial intelligence that can
learn anything. I mean, they're not self aware (we hope) but, from
what I've had of the idea explained to me, they seem like they should
be able to 'figure out' any data you give them.
Is this true?
They are non-linear function approximators. But because they have high
degrees of freedom they need lots of data to fit good approximations.
This can be interpreted as NNs are good modelling solutions. This is a
reasonable position for regions of the input space that are sampled well
(lots of data).
If missing values is an issue, NNs are not so good.
NNs rock at interpolation.
NNs suck at extrapolation.
I wish I'd said that... ...spot on, Greg!
Greg
