SV: Capacity of the brain

Mark James mrj at cs.usyd.edu.au
Thu Sep 9 16:12:01 EST 1999


Bill Skaggs wrote:

> 10^15 bytes is high but possibly not by more than a factor of 10.  There
> are around 10^14 connections in the human brain, and it is possible
> that each one holds a byte of memory.  I personally think it is more
> likely that each one holds a small fraction of a byte of usable
> memory, leading to a storage capacity on the order of 10^12 bytes, but
> all of this is quite speculative.  In any case the memory capacity of
> the brain is quite substantial, but it may not be orders of magnitude
> beyond the capacity of current computer hardware.

These estimates seem reasonable.  It is likely that many of the
connections in the cortex can be treated as having a fixed weight,
and so need not be stored individually, but can be described by
formulae (that perhaps have a pseudo-random component reproducible
from a fixed seed).

 
> Judging processing speed is another matter, because the numbers you
> get depend on what you call a "calculation", and this is essentially
> arbitrary.  Is a calculation a single decision by a neuron whether to
> fire?  This yields around 1000 calculations per neuron per second, and
> 10^13 calculations per second for the brain.  Or is it the input that
> a single connection provides to a neuron?  This yields around 10
> million calculations per neuron per second, and 10^17 calculations per
> second for the brain.

I think this second estimate of processing load is far too high.

First, as you suggest, the calculation of synaptic currents
dominates the updating of neuron outputs. However, only synapses
receiving a spike on each iteration (may) need to be calcuated --
and, at any one time, most neurons (and hence synapes) are inactive.

Second, there is no need to update each neuron's output at a rate
determined by the temporal resolution of spikes.  Each neuron can
be stepped asynchronously, whith asynchronous "interrupts" when
a spike arrives, as implemented in the Yale Neuron simulator.

Mark



More information about the Neur-sci mailing list