> This also seems to me to be the right approach - creating
> communicative ‘layers’ for different tasks. However, I suspect it
> will be necessary to add *at least* a third layer. This two-layer
> approach sounds like traditional client-server architecture compared
> to three-tier approaches. In this regard, Pat Hayes (a recognized AI
> ‘authority’) had this to say:
>
> < Minksy and Papert's demolition of the Perceptron depended crucially
> on that machines having only two 'levels'. At the time it was generaly
> assumed that this wasnt really important, but we now know that it was.
> Three levels and you can do all kinds of wonderful things, quite
> escaping the M&P critiques. It took a good few years to realise this.>
> (http://www.zynet.co.uk/imprint/online.html in "Hayes Response to
> Sutherland")
This is a bit comparing apples and oranges; what Minsky and Papert
(and likely Hayes) talk about are layered feed-forward networks
where information is passed from A to B to C, while Thaler is
using a recurrent net where information moves between A and B.
Recurrent networks can do more than few-layer feedforward networks,
and Thalers scheme could actually be implemented in a single
layer if needed (essentially you divide the connection matrix
into semi-independent parts).
-- ----------------------------------------------------------------------- Anders Sandberg Towards Ascension! asa@nada.kth.se http://www.nada.kth.se/~nv91-asa/main.html GCS/M/S/O d++ -p+ c++++ !l u+ e++ m++ s+/+ n--- h+/* f+ g+ w++ t+ r+ !y