[Martin Taylor 960816 15:30]
Bill Powers (960731.0730 MDT)
I know of no evidence that neurons are capable of performing
matrix inversions, except by the kinds of methods we learn in school --
by manipulating symbols according to rules, a very slow process
involving large parts of the brain.
A neuron can do approximate addition and multiplication, and neurons can
and apparently do do nice linear correlations. Matrix inversion is a
process of combining such processes, and with its multiplicity of inputs
I wouldn't find it at all surprising if one neuron couldn't provide an
output that is equivalent to one line in the written-out form of an
inversion--or at least of a large matrix multiplication, which amounts to
much the same thing.
There's no reason to think of "symbols" in the performance of such operations.
Martin