Capacity of the brain

Mark James mrj at cs.usyd.edu.au
Mon Sep 6 04:30:40 EST 1999


Bill Skaggs wrote:

> Yes, about 200 megabytes!  It's a cool paper, but he wasn't actually
> trying to measure the capacity of the brain, he was trying to measure
> the amount of information that a person knows.  He also assumed
> optimal compression of the information.  It seems unlikely that the
> brain stores information in anything like an optimally compressed
> form.

The brain may not optimally compress the information,
but it doesn't mean a computer brain simulator can't.

A computer simulator can save memory space by storing
synaptic weight vectors in compressed form (at the cost
of processing time), and by "swapping out" inactive
synapses to disk [the retrieval latency should be low
enough for real-time simulation].

Memory and CPU time can be saved by aggregating
"redundant" synapses into a single stronger synapse.
(I won't go into why I think this *may* be feasible).

Overall, if a number of major assumptions about the
operation of the brain hold**, I agree with Kurzweil
that we will have the hardware for the real-time
simulation of a human brain by 2010. It'll be an
interesting race to see whether the price of the
memory or the price of the processors dominates.

Under the assumptions, processing load is proportional
to the average number of spikes per second in an active
brain. Current PCs can process billions of linear synaptic
activations per second.  The number of these in real brains
may be as low as 1000 billion (after aggregation).
100GB mem & 1TB disk would be needed for the weights & states.
That is, we could run a real-time simulation of a whole brain
now, given a few million $ (or a faster Internet), and of
course the right software.

Kurzweil suggests use of scanning technology to determine
the circuitry (including weight settings), though I am
sceptical of this, preferring a reverse-engineering
approach.

Interesting discussion -- Mark

** Linear summing models of synapses/dendritic trees sufficient,
   & certain constraints on the updating of synaptic weights.
   If these things have to be simulated in a more complex way,
   *much* more computation would be required.



More information about the Neur-sci mailing list