Visual Cues for the Retrieval of Landmark Memories by Navigating Wood Ants

Summary Background Even on short routes, ants can be guided by multiple visual memories. We investigate here the cues controlling memory retrieval as wood ants approach a one- or two-edged landmark to collect sucrose at a point along its base. In such tasks, ants store the desired retinal position of landmark edges at several points along their route. They guide subsequent trips by retrieving the appropriate memory and moving to bring the edges in the scene toward the stored positions. Results The apparent width of the landmark turns out to be a powerful cue for retrieving the desired retinal position of a landmark edge. Two other potential cues, the landmark's apparent height and the distance that the ant walks, have little effect on memory retrieval. A simple model encapsulates these conclusions and reproduces the ants' routes in several conditions. According to this model, the ant stores a look-up table. Each entry contains the apparent width of the landmark and the desired retinal position of vertical edges. The currently perceived width provides an index for retrieving the associated stored edge positions. The model accounts for the population behavior of ants and the idiosyncratic training routes of individual ants. Discussion Our results imply binding between the edge of a shape and its width and, further, imply that assessing the width of a shape does not depend on the presence of any particular local feature, such as a landmark edge. This property makes the ant's retrieval and guidance system relatively robust to edge occlusions.


Idiosyncratic Paths
We show here more paths of individual ants. Ants that are trained to find food at the edge of a shallow gradient exhibit striking individual differences in their paths (Figure S1). The paths of some ants resembled the mean path and went directly to the edge. Other ants went consistently to the left of the edge, and others went to the right.
Interesting differences in the routes acquired by individual ants were also found when ants were trained to approach a two-edged landmark that varied continuously in width. From trial to trial, landmark width was varied randomly between 20 and 80 cm, with the food inset 10 cm from the left edge on all trials. We recorded routes for 20, 40, and 80 cm landmarks ( Figure S2). Ants trained under this regime could perhaps have used the ratio of height to width to categorize the training widths and acquire a direct route to the feeder for different width categories. This outcome was not observed. Instead, the mean routes ( Figure S2A) differed between the three widths (L = 107, 8 ants, 3 conditions, p < 0.01), as though the ants had acquired a single look-up table, which they applied to landmarks of all widths. On average the routes were most direct when ants approached the 40-cm-wide landmark. However, there were marked individual differences. One ant approached the feeder most directly when the landmark was 40 cm wide ( Figure S2B), with paths resembling the population mean, and another when it was 20 cm wide ( Figure S2C). The two ants may have acquired sets of snapshots for different training landmarks. We simulated the behavior of the two ants with look-up tables based on their paths to the 40 cm landmark and then generated routes to the 20-and 80-cm-wide landmarks. The simulated routes correspond reasonably well with those of the individual ants.

How Well Do the Model and Data Match?
We have already shown that the test data match the model qualitatively in that the model predicts the relative bearings and the rough magnitudes of the ants' differing approaches to narrow and wide walls and gradients.
Here we attempt to see how well the simulated routes match the observed routes. The paths of different individuals differ in a consistent manner from individual to individual, but the paths of each individual also differ between runs (Figures S1 and S2). Thus, it seems appropriate to look for fits between the model and data at the level of the mean routes of individual ants.
Our procedure was to restrict analysis to individuals from which we had a reasonably large data sample (five or more training runs and four or more test runs) and then for each individual determine a look-up table Figure S1. Idiosyncratic Routes of Individual Ants Routes from three ants trained to find food at the edge of the shallow gradient, as in Figure 2A in the main text. In (A), the ant walked straight to the food, n = 18. In (B), (n = 15) and (C) (n = 16), the ants walked consistently to a point to the left or right of the food. from its mean training path. The individual's paths to the test landmarks were then simulated from this look-up table; these paths provided the spine of a corridor that varied in width between 2.5 and 20 cm either side of the spine. We scored whether 75% or more of the individual's mean test path lay within the model corridor. Table S1 shows the proportion of ants that met this criterion for the different corridor widths. In interpreting the table, bear in mind that when the training runs are variable and the sample small, the mean may not be a good representation of the ants' behavior. (B and C) Individual, mean, and simulated paths for two ants (20 cm, n = 8 and 6; 40 cm, n = 10 and 10; 80 cm, n = 6 and 7).