Action Potential bandwidth
iain at 10xinc.com
Sun Aug 20 18:16:18 EST 2000
Richard N> The real question (as the alternative Richard points out)
Richard N> would have to take into account the information content of
Richard N> all that energy consumption.
I'm getting there... hang on.
Richard N> I would guess that your 1 GHz Pentium-III undergoes the
Richard N> same number of gate switches whether it is calculating an
Richard N> incredibly complex algorithm, downloading dirty pictures
Richard N> from the internet, or just sitting in a loop executing
Richard N> NOPs.
This is what got me going on this subject in the first place!
Microprocessors like the older Pentiums and modern workstation RISC
chips (Alpha 21264, UltraSPARCs, MIPS R12K) are pretty casual about
power dissipation. Their designers made few attempts to conserve
power when not computing things. That works great when you can
afford the heat sink. Modern Pentium-IIIs would burn 140 watts if
they were designed this way, and power-saving design techniques cost
less now than the heat sinks that would be necessary otherwise.
In a modern Pentium-III, the floating-point adder is prevented from
dissipating power unless it's actually computing a floating-point
add. The multimedia extensions and so on all work like this too, and
all told there are probably more than a dozen hefty chunks of the
chip that are enabled only when all the data necessary for execution
If the brain is to do more computation than a Pentium-III, those ion
channels switching are going to have to do more than a transistor
does when it switches. I can add two 64-bit numbers with just a few
thousand transistors -- less if I don't care about latency. If a
neuron has a thousand or more gated ion channels that participate in
the propagation of an action potential, that AP had better carry more
than a few bits of data. My guess is that an AP carries thousands or
perhaps tens of thousands of bits. It doesn't make sense otherwise:
why would a neuron need 100,000 inputs and outputs if it's only
moving a single bit?
So if an AP carries lots of bits, it doesn't make much sense to think
of the AP itself as data. My guess is that action potentials are
something like the enable pulses inside a Pentium-III: they turn on
the computation elements, briefly, when sufficient data has arrived
to justify a new computation. And I think those computation elements
are the ion channels.
Here's why: The ion channels have really huge energy density. The
neuron has ion pumps chugging away all the time, burning ATP to
establish an electrochemical energy reservoir that the ion channels
drain in a very short period of time.
Back to your question: what is the information content of all that
energy consumption? Put another way: what are the ion channels doing
with those ions? Do they just allow the ions to freely accelerate
through the water, eventually blowing off their energy as heat once
they cross the membrane? This seems fairly wasteful.
Perhaps, instead, the ions are staged down the ion channel proteins
much the same way electrons are staged along mitochondrial proteins
during the Kreb's cycle.
So if the ion channels are the computation elements, how are they
wired together? Certainly the rapidly changing electric field of an
AP is one way, but that's fairly coarse, and as I've said, I think
that's just the enable pulse. If they were communicating by
exchanging small molecules, these molecules would have to move from
ion channel to ion channel pretty fast, in order to keep up with the
AP. I haven't heard of this happening.
So I propose the ion channels are communicating by exchanging photons,
using the cell membrane as a waveguide. The ion channels get the
energy to transmit the photons from staging the ions down the
electric gradient. And the big deal is that the ion channels in this
proposal aren't just repeaters -- one can imagine that an ion
channel might have several sites for receiving several different
wavelength photons. Only when specific combinations of photons have
been received does the ion channel turn on, pumping out about one
million photons of its particular wavelength in reponse to about
one million ions flowing through.
So in this scheme, the nodes of Ranvier (that's where APs start,
right?) are just the enable logic. Synapses aren't just information
transport, they're latches to hold the data until the receiving
neuron triggers. And long-term memories might be state stored in
the configuration of ion channel proteins between action potentials.
And all this gives the brain a lot more computational complexity
than a $200 CPU.
-Iain McClatchie 650-364-0520 voice
http://www.10xinc.com 650-364-0530 FAX
iain at 10xinc.com 650-906-8832 cell
More information about the Neur-sci