Neural Representations of Trajectories
harry.erwin at sunderland.ac.uk
Mon Jul 23 06:27:24 EST 2001
Robert Gurk <bob.g at gmx.de> wrote:
> On 22 Jul 2001, Harry Erwin wrote:
> > Take a look at Avis Cohen's work.
> > <http://www.life.umd.edu/biology/faculty/cohen/index.html>
> > One nice thing about bat work is that the motor models are
> > fairly simple.
> > My question is basically: now that I'm fairly confident that
> > bats retain past position measurements in their trajectory
> > representation, and that those position measurements are in
> > general (3D+time) rather than head-centered or body-centered
> > coordinates, and that the trajectory representation can be
> > used in assessing nearness, how does a wet network that is
> > approximately 2D do all this?????
> Wait a minute. If the position measurements are neither head nor
> body-centered, how are they centered? There must be some point
> (0,0,0), right? And object-centered coordinates would not be
> very useful if the object is, say, a fly.
The measurements themselves seem to be range, azimuth, elevation, and
the time the echo was heard in a head-centered coordinate system. Note,
however, the bat often looks towards the target rather than directly
ahead and it may be banking/climbing/diving, so this coordinate system
is messy and varies extremely rapidly. The time is that of the echo
being heard, not the time the state update is incorporated, which is
about 100 msec later, or about a foot further along the curving
trajectory. Measurements are collected about every 50 msec along the
bat's trajectory, so integrating multiple raw measurements is
computationally painful and subject to severe non-linearities unless
converted at some stage to a common global coordinate system. Bats are
also known to dead-reckon accurately for extended periods (the
Wiederorientierung phenomenon), which suggests they have a
representation of their environment that is independent of the
coordinate system of recent measurements. Finally, and probably most
importantly, the various filters I've investigated produce more
behaviorally realistic localization performance if fed a converted
3D+time measurement than if fed a raw range/az/el/time measurement.
Note that the filters using a global 3D coordinate system are
indifferent to the origin or the specific basis vectors used.
> One more question: The wet network you are referring to, is
> related to hearing, I guess?
Yes, but the planning process seems more general than that. It can
incorporate vision as well.
> I can offer two more links:
Thanks greatly. We're not that advanced in bat acoustic studies.
Dr. Harry Erwin, harry at dherwin.org, http://world.std.com/~herwin
More information about the Neur-sci