David S.
Touretzky is a Senior Research Scientist in the Computer
Science Department and the Center for the Neural Basis
of Cognition at Carnegie Mellon University. He received
his B.A. from Rutgers University in 1978, and his Ph.D.
from Carnegie Mellon in 1984. Dr. Touretzky's research
focuses on the study of representations, both in computers
and in nature. His earlier work was on symbol processing
in connectionist networks using distributed representations.
In the last few years his interests have shifted toward
computational neuroscience. He is presently engaged
in studies of animal learning and navigation, and the
role of the hippocampus in spatial representation.
Summary:
Landmark-based
navigation is a rich domain for exploring issues of
representation and processing in neural systems. At
the behavioral level, there is a wealth of data on how
animals use landmarks to locate food or return to their
homes. At the neurophysiological level, the responses
of hippocampal pyramidal cells, the well known "place"
cells and of head direction cells in thalamus and the
subicular comple provide striking neural correlates
to behavioral variables.
Systems-level
theories fill the gap between these two modes of description.
To construct such a theory for navigation, we must first
determine a set of computations at some reasonably abstract
level that can reproduce the observed behavior, and
then show how these computations could be realized in
neural tissue. We are of course a long way from this
goal. However, in the present paper we describe a theory
of landmark-based navigation in rodents that is constrained
by both behavioral and neurophysiological data. The
theory is embodied in a computer model called CRAWL,
allowing us to replicate various experiments in the
animal navigation literature and make predictions about
place cell responses in novel situations. Portions of
the model have also been implemented on a mobile robot.
Collett,
Cartwright, and Smith (1986) trained gerbils to search
for a food reward at a fixed position relative to an
array of identical cylindrical landmarks. The array
was translated but not rotated from trial to trial,
and the animals were released from different starting
points to ensure that the landmarks would provide the
only reliable cues to the reward location. After training
to criterion, probe trials were introduced in which
the food was absent and the distribution of the animals'
search efforts was measured.
For a single
cylinder, the animals learned to search at the correct
bearing and distance from that landmark. Bearing information
was presumably measured with respect to the internal
compass, because the arena was designed to minimize
stimuli that could serve as directional cues. The room
walls were painted black, and there was only a single
overhead light, in the center of the ceiling. The light
illuminated a circular region of the floor but left
the walls in shadow.
Experiments
with another group of animals using pairs of cylinders
produced similar results: well-trained animals would
go directly to the goal location. Now Collett et al.
could test the animal's representation of the environment
by introducing occasional probe trials with modifications
to the landmark array.
When one
landmark was missing on a probe trial the gerbils searched
alternately in two locations, each at the correct bearing
and distance from one of the landmarks they had observed
during training. It was as if they were binding the
cylinder to one and then the other of the two remembered
landmarks. When the distance between the two cylinders
was doubled on a probe trial (the "split landmark array"
case, the animals also searched in two locations, each
associated with one of the two landmarks. They did not
search in the center of the expanded array.
Collett et
al. proposed a mechanism that could account for these
results, and several others involving more complex arrays
of landmarks. We shall refer to it as the "vector voting
hypothesis." According to this hypothesis, when the
gerbils reach the goal location and find food there,
they note the vector between each landmark and their
present position. Then, at the beginning of a new trial,
when they first emerge from the "start box" and see
the landmark array, they apply every learned vector
to every currently perceived landmark. The locations
receiving the most votes are the ones they choose to
search.
However,
the vector voting hypothesis alone does not account
for the split landmark result, where the gerbils searched
in two locations instead of four. The four sites should
receive one vote each when the landmark array is split.
Collett et al. concluded that the gerbils must be using
their perception of the entire array to distinguish
the east from the west landmark, and binding remembered
vectors to only the corresponding landmarks during the
probe trials.
This explanation
accounts for the data, but it introduces a binding mechanism
whose characteristics are left unspecified. Our model
reproduces all the effects in the cases above and a
variety of similar experiments without resorting to
explicit binding mechanisms to disambiguate landmarks.
There are
several functional subsystems to our model, which should
not be presumed to correspond to individual anatomical
sites. VISUAL INPUT provides range and egocentric bearing
information for a set of currently visible landmarks.
The animal's HEAD DIRECTION, i.e. its "internal compass,"
is updated using vestibular cues; in the block labeled
LOCAL VIEW, allocentric landmark bearings are calculated.
The animal
is also assumed to maintain an estimate of its position
in some internal coordinate system. This value is updated
by the PATH INTEGRATOR as the animal moves, based on
vestibular cues and an efferent copy of motor commands.
There is evidence for path integration abilities in
a wide variety of animals. Unfortunately, its precise
neural substrate is presently unknown.
Finally,
the role of the PLACE UNITS in our theory is to maintain
an association between perceived landmark positions
and path integrator coordinates, so that either can
be reconstructed from the other. This explains several
other questions raised by neural recording experiments
using rats:
How do rats
self-localize when reintroduced into a familiar environment
at a random spot? Our theory says they use visual landmarks
to activate a place code, which in turn evokes a set
of coordinates used to reinitialize the path integrator.
How are place
fields able to persist in the dark? The path integrator
is updated with each motion the animal makes. Our theory
says that the output of the path integrator may be used
to drive place cells. Errors will eventually accumulate,
but the system may be kept reasonably calibrated if
other sorts of cues are available, such as tactile information.
How is drift
in the path integration and head direction systems corrected?
The place units representing a location keep track of
the allocentric bearings and distances of landmarks
visible from that location. If landmarks appear at the
correct distances but their bearings are off by a consistent
amount, this indicates drift in the internal compass.
If the path integrator's output differs somewhat from
the coordinates derived from the place units, this indicates
drift in the path integrator.
For more
information, see the related papers.
Collett,
T. S. Cartwright, B. A. and Smith, B.A (1986) Landmark
learning and visuo-spatial memories in gerbils. Journal
of Comparative Psyiology A, 158: 835-851.
Touretzky,
D.S., Wan, H.S., and Redish, A.D. (1994) Neural representation
of space in rats and robots. in Zurada, J. M, and Marks,
R.J., eds Computational Intelligence: Imitating Life.
IEEE Press.
Wand,
H. S. Touretzky D. S. and Redish, A. D. (1994) Towards
a computational theory of rat navigation. in Proc. 16th
Annual Conference of the Cognitive Science Scoiety.
Lawrence Erlbaum Associates.
