Re: Finding useful functions- part 1

From: Stargazer (fuckoff_at_spammers.com)
Date: 10/30/04 

Date: Sat, 30 Oct 2004 18:22:25 -0300

Wolf Kirchmeir wrote:
> Stargazer wrote:
>
> > Stephen Harris wrote:
> >
> > > Supervised learning means there are external constraints applied
> > > to the processing of the network while it is processing which
> > > influences the eventual output.
> > >
> > > Unsupervised learning means that the network does not receive
> > > external inputs while it is processing. The processing works only
> > > with the internal structure of the network, then produces output.
> >
> >
> > Just a small observation here. In unsupervised learning, the network
> > in fact receives external input (otherwise it would not do much). It
> > does not receive a user-supplied output (training signals or
> > training sets) with which to compare its own output in order to
> > derive error correction. Supervised systems, on the other hand,
> > receive this user-supplied sets, which is employed (usually) as an
> > error correction to be back-fed (at least in traditional multilayer
> > perceptrons).
>
> If that is the case, the supervised ans unsupervised NNs must have
> different architectures. Or so it seems to me. In that case,
> supervised NNs must contain unsupervised sub-NNs. Or?

Or... none of the above, they just have different learning paradigms.
In unsupervised networks, neurons adjust their connection weights
according to rules that have nothing to do with what the user
wants. Neurons act more like coincidence detectors. In that sense,
a typical unsupervised NN may have comparable architecture with
a typical MLP (multilayer perceptron). There are several
unsupervised learning algorithms (hebbian, self-organizing
maps, etc). Different architectures, on the other hand, imply
different styles of connection among neurons. An MLP, for instance,
is architecturally different from recurrent networks (such as SRNs,
Elman's Simple Recurrent Networks). And both of them are supervised
NNs. In other networks, the supervised/unsupervised distinction may
be a bit blurred (such as in Hopfield networks, which are totally
recurrent and don't have explicit inputs or outputs).

*SG*