[From Erling Jorgensen (2013.11.17 2150EST)]
Rupert Young (2013.11.17 17.00 UT)
Hi Rupert,
I had a very similar reaction when I was first reading Jeff Hawkins' book,
_On Intelligence_. "It was going so well..."
I don't remember quite where, & I didn't quite formulate the words. But I
recall being bothered & a little dismayed that he took a "prediction"
route for how the neocortex does its business.
I did try to persist with the book, tough slogging at times, both
highlighting & writing notes in the margins. I remember thinking that it
was *almost* compatible with a PCT approach. I also flagged the whole
concept of the "name cells," because it occurred to me that this might be
an implementation of Bill Powers' concept of sending a reference signal
as essentially an "address" for the desired perception.
Even with PCT, I have long been fascinated by the "wiring"; in other words,
how the brain might implement even in simplified form some of the
required functions for stable control loops. One of the large loose ends
in PCT & its simulations, to my mind, has been the paucity of realistic
input functions.
Intensities we can do -- a number standing in for frequency of firing.
Weighted sum vectors we can do, to simulate how Sensations might be
created. Although even there, it has taken researchers some sophisticated
work over many decades to suggest how the brain itself might be
constructing its many forms of Sensations. But for the rest...
We simply do not know how Configurations could be constructed by neural
hardware. We don't know what combination of patterns constitute a
Transition. We don't know how that flow of experience gets partitioned
into Events. If we simply presume those lower components, we can
simulate some kinds of Relationships, but we don't really know how the
brain does it. Nor do we have much of an idea of how the brain classifies
things into Categories, despite some interesting work on semantic
networks. Perceptions like Sequences & Programs seem more amenable to
modeling efforts, but it is hard to define either a Principle or a System
Concept in language, let alone to understand how the brain might do it.
I realize this little rant tends to reify the levels of the proposed
PCT Hierarchy, & I'm sure the brain in actuality does something much more
complex & surprising than these simplifications. But the value of Bill's
outline of possible layers of perception is to give us some idea of what
to look _for_. And in that pursuit, along comes quite a brilliant thinker
like Jeff Hawkins.
My working sense is that a book like _On Intelligence_, & the subsequent
work by Numenta on Hierarchical Temporal Memory, may offer some plausible
steps in creating what we in CSG would call the "perceptual input
functions." I would love to see them inserted into actual PCT control
loops, hierarchically arranged, to see if negative feedback arrangements
of such variables could accomplish all that we expect they will. But I
don't yet see that kind of cross-fertilization of these two domains
happening. That is why I value these discussions we are having on the
CSGNet.
You do raise some pertinent questions about learning & memory.
Assuming that catching a ball is not something
that we can do straight off, what is it that changes, or is acquired,
through the process of learning? Is it the parameters of the (optical
variable) control systems that change? Or are the reference values of
the systems acquired through learning, and regarded as memory? Or does
the output function change?
I will need to think some more about the possibilities you raise. I do
not think we should regard the answers as self-evident.
Thanks for your contributions.
All the best,
Erling