i.kurtzer (2002.04.20.2130)
Here's an article that is relevant to identifying controlled perceptual
vairables in ant navigation. I could only print the text.
Nature 392, 710 - 714 (1998)
Multiple stored views and landmark guidance in ants
S. P. D. JUDD1 AND T. S. COLLETT1
Sussex Centre for Neuroscience, School of Biological Sciences, Brighton BN1
9QG, UK
Correspondence and requests for materials should be addressed to S.P.D.J. or
T.S.C.
ABSTRACT
Under some circumstances, Diptera and Hymenoptera learn visual shapes
retinotopically, so that they only recognize the shape when it is viewed by
the same region of retina that was exposed to it during learning1,2. One use
of such retinotopically stored views is in guiding an insect's path to a
familiar site3-5. Because the retinal image of an object changes with
viewing distance and (sometimes) direction, a single stored view may be
insufficient to guide an insect from start to goal. Little, however, is
known about the number of views that insects store. Here we show that wood
ants take several 'snapshots' of a familiar beacon from different vantage
points. An ant leaving a newly discovered food source at the base of a
landmark performs a tortuous walk back to its nest during which it
periodically turns back and faces the landmark. The ant, on revisiting the
familiar landmark, holds the edges of the landmark's image steady at several
discrete positions on its retina. These preferred retinal positions tend to
match the positions of landmark edges that the ant captured during its
preceding 'learning walks'.
To investigate the possibility that wood ants (Formica rufa) store multiple
views, we analysed their behaviour as they approached a cone. Our first step
was to demonstrate that ants learn to distinguish between upright and
inverted cones. An ant that had been trained to find sucrose close to either
an upright or an inverted cone was presented with a choice between the two
cones placed 12 cm apart and 40 cm from the ant (Fig. 1a). The ant
approached the pair veering from one to the other. At a mean distance of
15.8 cm (s.d. = 7.2 cm, n = 76) from the cones, it tended to select the
rewarded cone type and to walk to its base (Fig. 1b, c). Mostly, the ant
aligned its body axis so that it faced the cone, suggesting that any views
of the cone that the ant may have recorded were 'fixed' to the front of the
eye.
Figure 1 Recognizing cones. Full legend
High resolution image and legend (29k)
Evidence for the storage of multiple views of the rewarded cone comes from a
fine-grain analysis of the approaches of individual ants towards cones and
towards a single black-and-white edge that was much longer than the border
of the cone but oriented at the same 15� angle from the vertical. Let us
start with edges appropriate to upright cones (Fig. 2a, d). If ants match
the image of the edge to a single stored template, they should adjust their
orientation during their approach so that the edge is kept appropriately
positioned on the template in a single preferred retinal position (Fig. 1d).
But if there are multiple templates, the edge should have several preferred
retinal positions, one for each template (Fig. 1e, f).
Figure 2 A wood ant's approaches to extended black-and-white edges and a
cone, and its return path from the cone. Full legend
High resolution image and legend (90k)
Over some segments of the approach, the image of the edge is held relatively
steady on the retina, but during other segments the edge moves rapidly as
the ant scans its environment (Fig. 2b, e). We plotted how long the edge
dwelt in different retinal positions during the plateau periods (a plateau
is defined as a region of the plot where the edge remained stationary on the
retina within a window of 3� for a distance of >6 mm of approach) (Fig. 2c,
f). The resulting histograms are multipeaked in a way that suggests that the
ant takes several snapshots of the cone and that it does so when it fixates
the cone frontally. First, the two edges (Fig. 2a, d), corresponding to the
right and left sides of the cone from the ant's perspective, are viewed
predominantly by the right and left retina, respectively. Second, the peaks
in the two plots are distributed symmetrically about the midline.
Translated into viewing distance, the positions of these peaks correspond to
the cone viewed from 4, 7 and 17 cm. Similar data from seven more ants also
showed multiple peaks. Peaks were identified automatically and the
distribution of interpeak intervals shown in Fig. 3a was assembled from the
approaches of all the ants to edges. The average angular distance between
peaks is 13.32� (s.d. 4.86�), which is about three times the 4�
interommatidial separation, as measured in Cataglyphis6.
Figure 3 Distribution of peaks. Full legend
High resolution image and legend (24k)
If these peaks are indeed evidence of multiple stored views, approaches to
cones should generate peaks in similar retinal positions to approaches to
black-and-white edges. With edges, each putative template can be matched
throughout the approach by placing the edge in the appropriate position
(Fig. 1d). However, with upright cones there is a unique matching position
only at places where views are recorded (Fig. 1e, f). Away from those
places, the cone is free to bounce between the two edges of the template.
Nonetheless, the retinal plots of approaches to edges and cones (Fig. 2g, h)
are similar. The two sets of peaks tend to coincide (compare Fig. 2c, f, i)
and have similar distributions of interpeak intervals. To assess the
statistical significance of this similarity, we measured the intervals
between the closest peaks in the edge and cone distributions from the same
ant (Fig. 3c). A Monte Carlo method indicates that the two sets of peaks
obtained from each ant are aligned more closely than would be expected by
chance (P < 0.002 for data from eight ants).
A good check of our interpretation of multiple peaks is to examine where
ants that are accustomed to approaching inverted cones position edges on
their retina (Fig. 4). As the ant's horizon is only just above the ground
plane, the apex of the inverted cone will be imaged at the same vertical
position at all distances so that the bottom part of all the stored views
will overlie each other (Fig. 1h). Consequently, there should be only one
preferred position for an edge. The pattern of results matches this
prediction: the distribution of dwell-times for approaches to inverted cones
and to related edges shows a single peak, which is centred on the midline
(Fig. 4c, f).
Figure 4 Approaches to inverted cones and black bars. Full legend
High resolution image and legend (37k)
Further proof that ants trained to approach upright and inverted cones put
the edges of these cones in different learnt retinal positions came from
recording approaches of ants to a 1-cm-wide black bar oriented at 15� to the
vertical. Ants trained to approach inverted cones held the bottom of the bar
on the midline as though they were centering the apex of the cone (Fig. 4i),
whereas ants trained to approach upright cones kept the bottom of the bar
positioned away from the midline (Fig. 4l), as would be appropriate when
fixating an upright cone.
How does an ant approaching a familiar cone choose a template from its
collection? If it uses the template that best matches its current view, it
should engage small templates when far from the cone and larger ones close
to. However, this mechanism of template selection would not necessarily lead
to the same choice when the ant approaches a single extended edge: extended
edges will, for example, match large templates equally well over the whole
approach. In Fig. 5, we plot the mean retinal positions of the edge of the
cone and of the extended edge as a function of the ant's distance from these
edges. In both cases, the edges are, on average, relatively close to the
midline when the ant is at the start of its journey and they move towards
the periphery of the retina as the ant nears the stimulus. The ant tends to
select a template suited to a distant cone when it views either the edge or
the cone at a distance, and to select a template suited to a close cone when
it is near. The similarity of the two plots indicates that cues in addition
to static-image matching may be involved in template selection.
Figure 5 Retinal position of the edge of an upright cone or of a
black-and-white edge is plotted against the ant's distance from: a, an
upright cone (66 approaches); and b, a black-and-white edge (20 approaches).
Full legend
High resolution image and legend (24k)
Bees and wasps perform elaborate learning flights on first leaving their
nest or a newly discovered feeding place. During these flights they turn
back and look at the goal7-10. Wood ants do something similar. On their
first few departures from the cone, their path is very tortuous, with much
turning back. The darkened line in the departure of Fig. 2j indicates path
segments in which the ant has turned back and is walking towards the
landmark. When the ant is close to the landmark, the retinal image
transforms rapidly and these inspections of the cone are frequent and occur
over a large range of viewing directions, perhaps helping the ant separate
the landmark from the background. Further away, the ant's view of the
landmark changes more gradually; inspections become correspondingly rarer
and the range of viewing directions is narrower (Fig. 2j,k).
One indication that learning may occur during inspections of the cone on
homeward journeys is the distribution of retinal positions of the edge of
the cone when the ant looks back at it. Peaks occur in similar locations to
the comparable peaks derived from the same ant's approaches to edges. A
Monte Carlo method shows that the intervals between the closest matching
peaks on approaches and departures (Fig. 3d) are smaller than would be
expected by chance (P < 0.004 for data from seven ants). Some of the peaks
found on approaches may be missing from the distribution of peaks generated
during departures (compare Fig. 2a, l). The absence of some peaks indicates
that, as in bees9, the acquisition of landmark information cannot be
restricted to departures.
Our results indicate that object recognition by insects and humans may have
interesting parallels. In both cases, there is evidence11-15 that the visual
system encodes familiar objects in terms of stored two-dimensional views and
that the problem of recognizing the same object from several viewpoints is
solved, in part, by acquiring several two-dimensional views. The plots in
Fig. 2c, f and k indicate that ants may have adopted a strategy of sampling
more often close to the cone, where image size changes rapidly, and less
frequently further away, where image size is less dependent on range.
Sampling the cone at equi-angular separations gives the ant a series of
views that evenly covers the transformation of the shape from its first
sighting to the ant's arrival at its base (Fig. 1e, f, g).
Methods
Housing, training and testing ants. Ant nests were kept in large plastic
tanks. Individually marked foragers collected sucrose on a flat shelf (120
cm by 80 cm) above the nest which they reached by means of a paper
'drawbridge' (Fig. 1a). Ants from one colony foraged for sucrose at the base
of a black upright cone (7-cm base, 12-cm height). The cone was inverted for
ants from another colony. Both cone types were present during the training
for the choice experiment (Fig. 1) and the positions of the cones were
swapped frequently. One cone was associated with the sucrose and the other
was an unrewarded distractor. The shelf was wiped with alcohol after each
foraging trip to remove possible trail pheromones.
Videorecording and analysis. The 30-cm trips to and from the feeder were
videorecorded from above and analysed automatically to obtain the ant's
position and the orientation of its body axis. The retinal position of the
edge ofthe cone was calculated on the assumption that the head is aligned
with the body. High-magnification videorecordings of ants as they walked
straight and turned corners showed that the head is kept mostly aligned with
the body. Onstraight paths, the mean angle between head and body was 0.013�
(s.d. = 1.99�, n = 90 frames), and while the ant was performing a 83�
turnthe mean angle between head and body was 2.57� (s.d. = 3.78�, n = 23
frames).
Plateauex were selected using a computer algorithm that fitted straight-line
segments to the plots of the retinal position of edges. The plateauex
collected from all the approaches made by one ant to the same stimulus were
sorted into a histogram, which displays how long the edge was held in
different retinal positions. Peaks in the histogram were then found using
another computer algorithm. The data were first smoothed with a binomial
filter of width 1, and the peak positions were given by the positions of the
remaining local maxima.
Assessing the similarity of multimodal distributions. Statistical tests were
performed on the data from each ant using the positions of the peaks. We
compared the similarity of peak positions from approaches to cones and
approaches to extended edges, or from departures from cones and approaches
to extended edges. The computer picked out the closest matching pairs of
peaks in the two distributions as follows: for any given peak in the
edge-derived data, the closest peak in the cone-derived data was found.
Provided that the reverse also held, that is, that the original peak in the
edge-derived data was the closest peak to the matched peak in the
cone-derived data, the interpeak difference was recorded and the pair was
removed from the data set. This was continued until there were no further
matches. The process was repeated with data from the other ants, and the
standard deviation of all of the peak differences, assuming a mean
difference of zero, was taken as a measure of the similarity of the
distributions. The better the match between the peak distributions, the
smaller this similarity score was.
5,000 sets of artificially generated cone-derived data were then matched to
the real edge-derived data using the same procedure. For each ant, the
artificial data sets each had the same number of peaks as the real data, but
the peaks were distributed randomly between 0� and 80� from the midline,
with the constraint of a minimum separation of 6� and a maximum separation
of 24� between peaks. We computed a similarity score across all ants for
each data set. The number of sets out of the 5,000 that scored less than the
real data was taken as the probability of getting the real match by
chance16.
Received 3 December 1997;accepted 19 February 1998
References 1. Dill, M., Wolf, R. & Heisenberg, M. Visual pattern recognition
in Drosophila involves retinotopic matching. Nature 365, 751-753 (1993). |
PubMed | ISI |
2. Wehner, R. in Handbook of Sensory Physiology Vol. VII/6C (ed. Autrum, H.)
287-616 (Springer, Berlin, Heidelberg, New York, 1981).
3. Collett, T. S. & Land, M. F. Visual spatial memory in a hoverfly. J.
Comp. Physiol. 100, 59-84 (1975). | ISI |
4. Wehner, R. & R�ber, F. Visual spatial memory in desert ants, Cataglyphis
fortis (Hymenoptera, Formicidae). Experientia 35, 1569-1571 (1979). | ISI |
5. Cartwright, B. A. & Collett, T. S. Landmark learning in bees: experiments
and models. J. Comp. Physiol. 151, 521-543 (1983). | ISI |
6. Zollikofer, C. P. E., Wehner, R. & Fukushi, T. Optical scaling in
conspecific Cataglyphis ants. J. Exp. Biol. 198, 1637-1646 (1995). | PubMed
ISI |
7. Vollbehr, J. Zur Orientierung junger Honigbienen bei ihrem ersten
Orientierungsflug. Zool. Jb. 79, 33-69 (1975).
8. Zeil, J. Orientation flights of solitary wasps (Cerceris; Sphecidae;
Hymenoptera): I. Description of flight. J. Comp. Physiol. A172, 189-205
(1993). | ISI |
9. Lehrer, M. Why do bees turn back and look? J. Comp. Physiol. A172,
549-563 (1993). | ISI |
10. Collett, T. S. & Lehrer, M. Looking and learning: a spatial pattern in
the orientation flight of the wasp Vespula vulgaris. Proc. R. Soc. Lond. B
252, 129-134 (1993). | ISI |
11. Rock, I., DiVita, J. & Barbeito, R. The effect on form perception of
change of orientation in the third dimension. J. Exp. Psychol. Hum. Percept.
Perform. 7, 719-732 (1981). | Article | PubMed | ISI |
12. Rock, I. & DiVita, J. Acase of viewer centred object perception. Cogn.
Psychol. 19, 280-293 (1987). | ISI |
13. Tarr, M. J. & Pinker, S. Mental rotation and orientation-dependence in
shape recognition. Cogn. Psychol. 21, 233-282 (1989). | ISI |
14. B�lthoff, H. H. & Edelman, S. Psychophysical support for a
two-dimensional view interpolation theory of object recognition. Proc. Natl
Acad. Sci. USA 89, 60-64 (1992). | PubMed | ISI |
15. Ullman, S. High Level Vision: Object Recognition and Visual Cognition
(MIT Press, Cambridge, MA, 1996).
16. Manly, B. F. J. Randomization and Monte Carlo Methods in Biology
(Chapman and Hall, London, 1991).
Acknowledgements. We thank M. F. Land and D. Osorio for valuable comments.
Financial support came from the BBSRC and Human Frontier Science Program.
S.P.D.J. received a BBSRC research studentship