Re: What evidence is there that a neuron is digital computer ?

casey <jgkjcasey@xxxxxxxxxxxx> wrote:
On Jul 26, 11:38=A0pm, c...@xxxxxxxx (Curt Welch) wrote:
I said the pulses have infinite temporal resolution.
...
I've said nothing about pulse encoding sensors in
this thread and what they can or can not do.

So what was the use of observing that there is an
infinite temporal resolution if you couldn't use it?

Infinite temporal resolution is used the sensors.

A point in space also has "infinite resolution" but
in practice there is a limit to how you can use it.

When you build real mechanical gears, or a rope and pulley, it makes total
use of that infinite resolutions. There is no limit to "how you can use
it".

There is a limit to whether you _need_ to use it - not whether you _can_
and _do_ use it.

How do you set it and read it out?

I'm not sure why you are asking that.

There is no need to set it or read it out. That's a _human_ need and not a
physical fact of how continuous systems work. There is no value in
pretending continuous systems are not continuous just because a human can't
"set it" or "read it out".

A neuron works in the continuous domain of time just a a gear works in a
continuous domain of rotational space. The fact that you, as a human, are
unable to measure or set the position of the shaft to near infinite
precision doesn't mean the shaft doesn't have such state.

You can't pretend that neurons are digital just because you want them to be
John.

Really, what are you getting at here?

What I had in the back of my mind was all your talk
in the past about how your pulse sorting units could
take two pulses in at "once" and would be able to
sort them according to arrival time as in time there
would be a difference in their arrival time no
matter how small.

You are suggesting that I claimed that two pulses that happened at
different points in time were really happening "at once"? No, I've never
said anything so odd John. IF you thought I did, it was simply something
else you failed to understand.

What is the hardware for this?

The hardware for making two pulses that don't happen at the same time act
as if they are happening at the same time? I have no clue, you tell me.

Can you see it might have a problem coding small
values in succession as it has to take time off
to recharge?


Sure. If it can't fire again for 10 ms then it
can't code for anything less than 10 ms. Isn't
that obvious?

I thought so but you were saying the neuron makes
use of higher resolutions in time.

Yes. I did say that because it does. Honestly, after multiple messages
where I took the time to explain this to you in simple examples with real
numbers, and a few from Don as well, you still don't understand the
difference between max firing rate and max resolution?

It's not that it's just an infinite number of
values between 0 and 500 John, there's an infinite
number of values between 300 and 301.

But can you make use of the infinite values between
two points in space or time? I say there is a space
and time frame below which a particular piece of
hardware cannot discriminate between the values in
that frame.

Well, that sort of effect does show up, but it shows up down at the level
of the Heisenberg uncertainty principle and other sorts of quantum effects,
not at the level of the 10 ms max firing rate of a neuron. That is why I
keep saying "nearly" infinite because the true resolution is so far below
what we care about that there no point in thinking of it as anything but
infinite - until you get down to the scale where quantum effects become
important.

All this is about claims you made to me with your
pulse sorting unit that it could always determine
which pulse arrived first because the chance of
them having the same value in the sea of infinite
values is as good as zero. What I couldn't figure
out is what hardware could do this if the pulses
were very close in time.

All this is about my high resolution asynchronous pulse sorting ideas?
Really? You think that's what we have been discussing in this thread?

What I was trying to point out to you in the past (not in this thread -
except again last night in the long message I write), was that when you
switch to using continuous real time asynchronous pulse signals, instead of
using digital synchronous values, the notion of input values happening "at
the same time" goes out the window. In the analog domain, one pulse will
always be "before" the other. If you simply make an analog gate to detect
which pulse comes first, it will tell you which pulse came first, down to
infinite resolution. But that does not help you because the concept of
"came first" is simply defined by how that gate works. Pulses don't just
exist in time, they exist in space. The move though space as they travel
down a wire, or a nerve fiber. When you want to try and define "which came
first", you also have to define where in space the "finish line" is located
at. When you build hardware to detect which pulse came first, it defines,
by the nature of how it works, what it means by "came first". It simply
will always select one pulse or the other, and never say "tie" because it's
designed not to do that.

If you digitize the analog async pulses by assigning them a digital time
stamp, then you will be able to get duplicates. But with a very high
resolution time stamp, you we see so few duplicates, that there is not
need, or point, to thinking in terms of "pulses happening at the same
time".

My point in discussing these things with you in the past was not to try and
prove something about whether things would "never happen at the same time".
It was to try and open you mind to a way of thinking about this problem
different than your standard "everything is a synchronous digital finite
state machine" with "inputs happening at the same time".

If you are to build a unit that has a "temporal
function" it has to be physical and I assume, at
this stage, it will be built out of your standard
electronic components.


Ok. I guess. But now you are totally changing
the subject so I don't get what's going on in
your mind.

That you have to implement things in a physical form
and that rules out using your infinite resolution.

Well no. When I build analog hardware I AM using the infinite resolution
so it's not ruled out in the least. Real neurons have INFINITE RESOLUTION
and THEY USE IT. The question is not whether they use it - they DO. The
question is whether this particular problem needs infinite resolution to be
solved - which I assume it doesn't - but since no one has solved it, we
don't really know. The only machine that has solved it - the brain - does
use infinite resolution.

Again, you seem to keep tripping up over the difference of what YOU as a
HUMAN can UNDERSTAND and talk about and measure and set, vs what the neuron
itself is actually doing. The neuron (as does all analog hardware) makes
use of infinite resolution continuous effects in time and space. That's
just a fact about the entire universe. Humans however, at the high level,
need to digitize things in order to talk about them. That's what the whole
process of language is about - Making up unique symbols (discrete physical
digital effects) so that we can communicate them to each other, and write
them time, and talk to ourselves about them.

But the langauge needs of a human, are not the physical facts of a neuron,
or of a mechanical gear. The universe is full of continuous effects that
have effectively infinite precision.

It is unclear to me what you mean by "spatial function
computing neural networks". As far as I know the units
I play with are functionally just as "temporal" as
any real neuron.


Well then, lets start from scratch again.
...

After I posted I realised what you meant by spatial
function. My confusion was you implying I didn't talk
about dealing with temporal patterns but that is not
true. I suggested they were dealt with by being
converted to a spatial code which is what memory is
even if it is scattered through the whole net as
a volage charge on a capacitor or a digital value
in a counter.

What I've been talking about as temporal functions,
basically map to the idea of sequential logic. But
yet, there are subtle differences in how I normally
think about it because I think abstractly in terms
of processing real analog temporal pulse signals
instead of in terms of digital signals.

Yes and it this subtle difference in how you think about
it that has been the core problem when I try and imagine
how it could be implemented in real hardware.

Well, the difference is that analog machines don't every need to convert
everything to digital before it is "computed". It takes the input in analog
form, does some functional computation in analog, and produces an analog
output.

Though the pulses are digital in the spatial domain, they are analog all
though the brain in the temporal domain. The temporal problem of
intelligent behavior is solved by the brain in the analog domain. It's
never converted (fully) to digital, and then back to analog when it makes
our body move in a very analog way. The timing of when the pulses show up
to activate a muscle is as important as the fact that it did show up.

sequential logical is a concept from the pure digital domain. I tend to
think abstractly in terms of the continuous time domain, even though I end
up writing code to simulate it, in the pure digital domain.

The only neural networks you have talked about, have
no ability at all to solve that problem - and it's why
I have not yet been interested in the networks you have
talked about.

Not true. You just didn't like the idea of the temporal
pattern being converted to a spatial form. I assume it
is because you see limited resolution in this form then
the infinite possible number of positions in time that
a pulse may turn up at the unit.

I don't mind it being converted to a spatial form. In fact it _must_ be
converted to some spatial form. That's exactly what my pulse sorting
networks do. You have simply not yet suggested a technique for doing that
conversion that looks to me like it would be of any use in this
reinforcement learning problem. When you come up with such a conversion,
I'll stand up and pay attention to it. For a matter of fact, I don't
remember you every coming up with any conversion actually. All you have
ever said that I remember is, "you nee to add memory circuits".

< snip >

Does any of the above help you understand what I've been
talking about?

I have always believed I understood what you imagine about
infinite high resolution analog temporal coding with pulses
but what wasn't clear was the cold hard need to convert
this abstraction to an implementation in actual hardware
that could make use of it.

My pulse sorting networks are working hardware of the abstraction. I've
been showing you the "cold hard" (working!) hardware for years now. What
on earth is so hard for you to understand about all this?

--
Curt Welch http://CurtWelch.Com/
curt@xxxxxxxx http://NewsReader.Com/
.

Relevant Pages