Georgopoulos (1995) surveys research on population
coding in motor cortex of the direction of arm motion.
The population codes are naturally treated as fields, and the
transformations of directions are simple field computations.
We consider a region
in motor cortex in which activity
is observed in anticipation of reaching motions.
Each cell
has a preferred direction
in three-dimensional space.
Cell activity
falls off with the cosine of the angle
between the reaching direction
and the
preferred direction
.
Since (for normalized vectors) the cosine is equal to the
inner product of the vectors,
, we can express the activity:
for some constants a and b.
Thus the motor cortex represents a vector field
of the preferred directions, and the population coding of an
intended motion
is a scalar activity field
given by the inner product of the motion
with the preferred-direction field.
There is another way of looking at the population coding
of a motion
, which is sometimes more
illuminating.
Since all the neurons have the same receptive field profile,
we may rewrite
Eq. 3
in terms of a radial function
of the difference
between the preferred and intended direction vectors:
[ _u = (_u - r) , ]
where
[
(v)
= a + b - b \| v \|^2/2 .
]
This is because the Euclidean distance is related to the
inner product in a simple way:
(provided
).
Now let
be the direction field, defined over
three-dimensional space, that corresponds to
.
That is, the value of
at neural location u equals
the value of
at spatial location
,
or
, which we may abbreviate
.
For simplicity we suppose
is one-to-one, so we
can define
by
.
Notice that
effects a change of coordinates from
neural coordinates to three-dimensional space.
The direction field
can also be expressed as the result of convolving
the receptive field
with an idealized direction field
, a Dirac delta,
which has an infinite spike at
but is zero elsewhere:
[ = _r . ]
This is because convolving
with
effectively translates the center of
to
;
equivalently, the convolution blurs the idealized direction
field
by the receptive field profile
.