On Mar 8, 4:04 pm, casey <jgkjca... from yahoo.com.au> wrote:
> On Mar 9, 3:38 am, feedbackdroid <feedbackdr... from yahoo.com> wrote:
>> > Unfortunately, current NNs do little more than solve toy problems.
> > In the recent presentation by Hinton, he mentioned it took him 17
> > years to figure out how to properly make something [boltzmann
> > probabilistic networks] that works significantly faster and better
> > than backprop networks. And what do his marvelous new networks
> > solve? The recognition of the numbers 1 to 9 in various distorted
> > forms. 17 more years down the tubes.
>> So were you impressed by what Donahoe's people have shown as described
> in the cool paper that GS recommended?
In line with my prior comment, 10 simplistic simulated neurons does
not a brain make.
>> The brain does implement its action in the form of neural networks
> so somehow we know at least biological neural networks have the
> right stuff.
>> I don't know that much about current ANN's and they haven't interested
> me much for the reasons you mention above. I have wondered if for any
> computational process there is an ANN equivalent. My understanding is
> that ANN's do some kind of multivariate statistics?
>> My reading of the evolution of biological neural nets is that maybe
> random neural nets produced computational abilities and the ones that
> enhanced the organisms survival were selected. I imagine the first
> networks would have been simple reflexes until intermediate nets could
> produce useful things like a central pattern generator that could be
> modulated by its inputs from external and internal sources.
>> I have sometimes amused myself with the simple problem of character
> recognition using GOFAI and have thought about how any implementation
> might be translated into a neural network/s.