[Martin Taylor 2018.09.29.10.47]
[Rick Marken 2018-09-28_22:26:02]
[Martin Taylor 2018.09.28.16.12]
MT: Indeed, what we know of neurophysiology tells us
that there is no single place within the brain where
that [perceptual] value exists for some future
technology to measure.
RM: The work of Hubel & Wiesel suggests that
perceptions, in the form of neural firing rates, are
located in afferent neurons and other single neurons in
the brain.
It's interesting that you should mention H&W, since their work
was published when I was in graduate school, transitioning from
being an engineer to a perception psychologist. When it was
reported, the effect on the group I associated with was that finally
the physiologists acknowledged that neurons could do
quasi-mathematical operations that psychologists believed they had
to be able to do. But the implications of their work was not as you
describe. It was mainly that they proved it was possible for neurons
to recode very large numbers of inputs into an informationally
efficient form that kept the important contingencies among the
inputs (important meaning frequently occurring) while losing the
contingencies that were probably (on statistical grounds) due to
irregular fluctuations in the visual field.
At that time, a colleague and I had been asked to produce a report
for a weekly seminar group on the Bush-Mosteller theory of memory.
We used a Hebbian approach, but I can’t remember whether that was
because B&M had done so or whether we just thought it was a
neato idea. Regardless, that was our expectation for the synaptic
variation in the inputs to visual neurons (and their distribution of
outputs to the main part of the brain in the visual cortex). This
was a good part of the reason the H&W findings came as a relief
rather than as a surprise.
The result of this was always that it demonstrated that the
perceptions we form depended on recoding the mix of
on-centre-off-surround, oriented edges, corners, and so on signals
that were reported along the optic nerve to the brain, and were
never located in any single place thereafter, any more than any
particular component of a Fourier Transform locates a value of a
signal at any one moment.
In other words, what you say is simply not true. Whether at some
future time some perception or other is found to be represented in
one and only one place in the brain, it is not true that the
findings of H&W suggest that it is.
MT: The "perception" in Bill's model is a rather
vague average rate of firings of nerves in a “bundle”
with a fuzzy boundary.
RM: I don't see what's vague about it. In Bill's model
a perception – or, more properly, the state of a
perceptual variable – is the average firing rate in an
afferent nerve (a bundle of neurons) at any instant.
Which does directly contradict your earlier assertion, does it not?
Incidentally, would you like to tell us what the average firing rate
in a nerve bundle is at an instant when none of the neurons in the
bundle are firing?
Would you like also to discuss how it is determined which neurons
are deemed to belong in any particular “bundle” and which do not,
for any particular perceptual signal? Remember, when you do this,
that most, if not all, effective neurons have thousands of synaptic
inputs, and probably no two have all their inputs coming from the
same set of source neurons, which means that each has a different
pattern of inputs that produces its maximum firing rate. Put another
way, the firing rate for any input pattern will produce a wide
distribution of firing rates across all the neurons that respond at
all to it, rather than there being a sharp cut-off between neurons
that do and neurons that don’t respond to that pattern.
The average firing rate is a variable called the
perceptual signal, p, and it varies along with variations
in in the aspect of the environment that is defined by the
perceptual function that produces the perceptual signal as
an output.
Well, at least you wrote something on which we can agree.
This seems very precise to me.
But not to me.
RM: In PCT models, p is allowed to go negative, which,
of course, neural firing rates can’t do. So in order to
make our models true to the facts of physiology, every
control system should really be two control systems
controlling the same variable, one system acting when the
perceptual signal goes above the reference signal and the
other acting when the perceptual signal goes below the
reference signal. I’ve attached a simple hierarchical
model that implements the control systems in this way –
the physiologically correct way. If you run this model
note that the none of the signals in the model – all of
which are presumed to be neural firing rates – go
negative. But while this parallel loop model is truer to
the physiology than is the more familiar single loop
version where signals can go negative (as in the Live
Block Diagram: https://www.dropbox.com/s/sizvbwso44mastu/LiveBlock.exe?dl=0
) it’s a lot more tedious to create and gives the same
result as the single loop model.
Well, in many respects it does, but there's more to it, as we
discussed a year or two ago. The opposition between the two
+positive and -positive sides of the balance introduces a new degree
of freedom because of the slowing effect of the leaky integrator
output. I chose to call this effect “stiffness” by analogy with the
case of two opposing control systems that allow a parade flag-bearer
to keep the flagstick vertical in a cross-wind by tightening her
grasp with each hand to increase the oppositional conflict-induced
gain of the virtual controller for deviations in that direction, and
at the same time in the orthogonal (stiffness) direction.
RM: But the fact that perception is modeled as a neural
current in PCT makes it difficult to conceive of how
control of complex perceptions, like sequences and
programs – perceptions that are defined over a fairly
long period of time – can work.
Ah, yes. That's an issue I discussed with Bill at some length when I
first began to learn about PCT. I don’t remember ever becoming clear
about it at that time, though he may have been, but I think I am a
little clearer now, though not perfectly clear. Probably another
thread would b a better place to discuss it. Maybe we can get it
straight with the participation of the general CSGnet readership.
That was why I posted my post entitled "Control of
Higher Level Perceptions". I think it’s important that we
develop models that can control such perceptions to show
how control of these higher level perceptions could work.
Yes.
Martin