the responses on this digest to this item are bizarre.
On Thu, Jun 18, 2015 at 7:15 AM, João Ferreira <[log in to unmask]>
wrote:
> I stumbled upon some research by google. I think this might be interesting
> to anyone interested in artificial intelligence, visual communication and
> art.
>
>
> http://googleresearch.blogspot.co.uk/2015/06/inceptionism-going-deeper-into-neural.html
>
I found the blog post to be an exciting demonstration of the reversal of
a deep-layer of hidden layers in a standard connectionist (neural)
network, showing that although the networks were designed for
pattern recognition, they could be reversed to do pattern production.
Pattern production? Fantastic!
The terms deep learning and hidden layers are deeply technical and cannot
be readily inferred by what the English interpretation might appear to mean.
I was associated with the early beginnings of this technology and one of
the difficulties scientists faced was understanding the
weight structure of the hidden layers (again, a technical question).
In English, we could configure neural nets to do wonderful things after a
period of learning, but we had almost no idea of how they configured
themselves. All we knew was that after practice and learning, they had
determined a set of weights (numerical representations) on each of their
thousands (today millions) of connections that worked.
This blog was showing pictorial representations of the weight structure.
This is interesting but not so novel. The most impressive part was
the demonstration that each higher level layer of connections had a
higher-level aspect of the visual picture being represented. How did this
happen? It was not pre-programmed that way -- it is a natural result of the
architecture, which implies that related architectures in the human brain
might have similar levels of abstraction.
And as I said, the fact that the network could be reversed and recreate
abstractions of the images that had been learned is very exciting;.
---
this blog has nothing whatsoever to do with art. Or design. But it
is brilliant science and engineering. But to understand the brilliance
and the importance does require a basic understanding of
connections/neural networks and their modern incarnation as multiple layers
of connections -- the so-called "deep learning" pioneered by Geoff Hinton
(A PostDoc of mine who has far surpassed anything i could ever even
imagine, let alone do.)
Don
Don
Don Norman
Prof. and Director, DesignLab, UC San Diego
[log in to unmask] designlab.ucsd.edu/ www.jnd.org <http://www.jnd.org/>
-----------------------------------------------------------------
PhD-Design mailing list <[log in to unmask]>
Discussion of PhD studies and related research in Design
Subscribe or Unsubscribe at https://www.jiscmail.ac.uk/phd-design
-----------------------------------------------------------------
|