Citation: van der Linden M, Murre JMJ, van Turennout M (2008) Birds of a Feather Flock Together: Experience-Driven Formation of Visual Object Categories in Human Ventral Temporal Cortex. PLoS ONE 3(12): e3995. https://doi.org/10.1371/journal.pone.0003995
Editor: Martin Giurfa, Centre de Recherches su la Cognition Animale - Centre National de la Recherche Scientifique and Université Paul Sabatier, France
Received: June 24, 2008; Accepted: November 18, 2008; Published: December 24, 2008
Copyright: © 2008 van der Linden et al. This is an open-access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Funding: This research was supported by a grant from the Netherlands Organisation for Scientific Research (NWO 400-03-338). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.
Competing interests: The authors have declared that no competing interests exist.
A crucial property of the human object-recognition system is its capacity to group different-looking objects into the same category, and to assign similar-looking objects to different categories. Pineapples and berries look very different, but they are both members of the category ‘fruits’. In contrast, berries and beads can look similar, but belong to different categories. Someone more skilled in recognizing fruits might be able to discriminate between similar sub-exemplars of berries (e.g., salmonberries and raspberries), suggesting that the neural representation of object categories is plastic and changes as a result of experience. The present study investigates the neural mechanisms mediating experience-induced formation of visual object categories in the human brain.
There are strong indications both from neuropsychological and functional brain imaging experiments that the ventral temporal cortex is involved in the representation of category-specific information , , , . Differential neural responses within occipitotemporal cortex have been demonstrated for a wide range of object categories , , , . However, the neural mechanisms mediating the formation of category-specific representations in human occipitotemporal cortex are still largely unknown. Animal studies have revealed that learning and experience can shape neural response properties of cells in inferior temporal cortex, possibly resulting in category-specific representations. For example, after monkeys were trained to categorize visual stimuli, inferior temporal neurons responded selectively to stimuli belonging to the trained category . Furthermore, other electrophysiological recordings from monkey cortex revealed increased selectivity in responses from inferior temporal neurons for visual stimulus features diagnostic for trained object categories , as well as for combinations of features in learned objects . Functional imaging of the human brain has shown that visual as well as functional experience with novel object categories alters neural responses in occipitotemporal cortex , , , . Recently, fMRI data provided evidence for increased neural sensitivity in occipitotemporal cortex after categorization training . It remains unclear, however, whether, and how training-related neuronal changes are linked to the formation of behaviourally relevant object categories.
In the present study, we investigate neural mechanisms of object category formation in human occipitotemporal cortex. We directly compare neural changes mediating the formation of behaviourally relevant object categories with neural changes following visual exposure to objects in the absence of category formation. Our findings provide evidence for learning-related increases in selectivity of neural responses to object properties that are relevant for categorization.
We designed a stimulus set consisting of six highly similar bird shapes that are difficult to distinguish without training (Figure 1). To directly test for neural correlates of category formation, a discrete category-boundary between similar-looking birds was established by training (Figure 2a). In addition to this categorization training, subjects performed a control task in which they were visually exposed to two other bird types, but to hinder category learning, the feedback they received was random . Subjects were not informed that the feedback could be correct or incorrect. This manipulation allowed us to investigate neural changes specifically related to the formation of an object category compared with changes occurring as a result of repeated visual exposure. To investigate neural correlates of object-category formation, pre- and post-training fMRI time-series were obtained while the participants viewed exemplars of the different bird types (Figure 2b). We predicted that if category formation is mediated by increased neuronal responsiveness in occipitotemporal cortex, this increase should occur only for those birds for which a discrete category-boundary has been established, compared with visually similar birds for which no such boundary has been learned. Critically, this effect should be distinct from general training effects, such as increased familiarity and visual object-learning.
Figure 1. Construction of the stimulus set.
(A) Pictures of non-existing but plausible bird shapes were constructed in a 3D model manipulation program. From a base-bird we derived six colorless prototype birds (A, B, C, D, E, F) that differed in trunk, belly, tail, beak, head shape, cheeks, brow, and eye position. Each bird was rendered under the same lighting and camera settings to make sure that shading and scale was identical for all birds. (B) Exemplars were created by systematically morphing each of the six prototype birds with all other birds. Shown is an example of morphing bird type A and bird type B at morph ratios of 90∶10, 80∶20, 70∶30, 60∶40. The category boundary was set at 50∶50.
Figure 2. Training and fMRI paradigms.
(A) During the training sessions participants were presented with a series of bird exemplars. They performed a 1-back task in which they indicated whether two consecutive birds were the same type or not. In the category-training condition implicit category learning was established by providing corrective feedback after each trial. In the visual-exposure condition random feedback was given after each trial, hindering category learning while keeping visual exposure to the birds equal to the category-training condition. (B) In the pre and post-training fMRI scanning sessions the bird types were presented in blocks of five exemplars at morph ratios of 60∶40, 75∶25, and 90∶10. Each image was presented for 3 seconds with a mean inter-stimulus-interval of 1 s. Experimental blocks alternated with rest periods of 10 s. Subjects were instructed to view the birds attentively.
Twelve neurologically healthy right-handed participants, not bird experts (ten females, mean age 20.7 years, range 18–25) with no neurological history participated in the experiment. All subjects had normal or corrected-to-normal vision. Subjects were paid for their participation. All subjects gave written informed consent. The study was approved by the local ethics committee (CMO region Arnhem-Nijmegen, the Netherlands).
The stimuli consisted of pictures of birds that were constructed in a 3D model manipulation program (Poser 4 by Curious Labs, Santa Cruz, CA). First, six prototype birds were constructed from a base-bird (Songbird Remix by Daz3d, Draper, UT). Parts of the bird that were manipulated included its back, belly, tail, beak, head shape, cheeks, brow, and eye position. Next, each of the six birds was morphed with all other birds (at ratios of 95∶5, 90∶10, 80∶20, 75∶25, 70∶30, 65∶35, 60∶40, and 55∶45) analogous to the procedure used by Freedman and colleagues to investigate category formation in the monkey brain . The category boundary was set at 50%. As a result, stimuli that were near opposite sides of a category boundary, though visually similar, belonged to different categories. Morphing happened smoothly between corresponding points on the birds. Each bird was colourless, rendered under the same lighting and camera settings, and exported as an image. Images had identical colour, shading and scale. In addition, using the same software, a set of control images of six different faces was constructed. The images measured 300 by 300 pixels in the training sessions and were slightly reduced in size (250 by 250 pixels) in the scanning sessions.
Procedure and experimental paradigm
The six bird types were divided into pairs, and each pair was assigned to one of three conditions: 1) category training, where subjects received correct feedback to their responses, 2) visual exposure, where the amount of exposure to the birds was equal to the amount of exposure to the category trained birds, but category learning was hindered by random feedback, 3) no training. Assignment of bird types to the three conditions was counterbalanced over subjects in such a way that each bird type appeared equally often in each of the training conditions. The experiment was constructed using dedicated experimental software (Presentation by Neurobehavioral Systems, Albany, CA) and was run on a Pentium 4 with a 2.80 GHz processor and 2 GB of RAM.
Training included three sessions, each of which lasted approximately two hours, on three consecutive days. During a training session, subjects sat comfortably in a soundproof cabin in front of a 19” computer screen. They performed a 1-back task on a series of bird images, in which they indicated with the index and middle finger of their right hand whether two consecutive birds were the same bird type or not. Subjects received feedback to their responses consisting of a printed text centred on the screen in coloured Arial font in size 16 (green: “right”, red: “wrong”, and yellow: “too late”). Bird exemplars were morphed at 55, 65, 70, 80, and 95% with all other bird types (e.g. bird type A at 95% morphed with B, C, D, E, and F at 5%). In total there were 25 exemplars (each bird type was morphed at five morph levels with the other five bird types) for each of the four bird types presented during training. Each exemplar was presented 30 times per training session. The average morph distance between birds was 58,67%. The proportion of same and different responses was fifty-fifty. In each trial, stimuli were presented for 1000 ms after which a response could be given during 2250 ms. Feedback was presented for 250 ms. Stimuli onset asynchrony was 4000 ms. A training session consisted of 10 blocks of 150 trials. In each block, 30 trials of category training (correct feedback) were alternated with 30 trials of visual exposure (random feedback). Subjects were not informed on this alternation of correct and random feedback conditions. Each block of 150 trials was followed by a small self-paced pause after which a subject could continue the experiment by pressing a button.
Subjects participated in an fMRI scanning session one day prior to training, and in an identical fMRI scanning session one day after training. During scanning, bird exemplars from each of the three conditions (category-training, visual exposure, and no training) were presented and subjects were instructed to view the birds attentively.
Bird exemplars were different from the exemplars encountered during training and included morphs at 60, 75, and 90%. Birds were presented in blocks. Each block contained 5 images of one bird type at a certain morph level. Images within one block were morphed with different bird types so that they were not identical to each other. For example, a block could consist of five images of 60% of bird-type A morphed with 40% of bird type B, C, D, E, or F. Each image was presented for 3 seconds with a mean inter-stimulus-interval of 1 s (varying between 600 and 1400 ms in steps of 200 ms between). Experimental blocks alternated with rest periods of 10 s for sampling the baseline. Experimental blocks were repeated six times, resulting in 108 blocks (6 bird types * 3 morph levels * 6 repetitions). In addition, six blocks were included that contained five images of artificial faces. Blocks were presented in pseudorandom order. Total scan time was 54.7 minutes.
Participants read the instructions for the scan session from a piece of paper before going into the scanner. They were instructed that they were going to watch pictures of objects presented in series of five and that these were followed by a few seconds of blank screen. They should watch these pictures carefully. To keep the subjects alert, we included catch trials. After each block a catch trial could occur. The chance of such an occurrence was on average, one out of six blocks. Subjects were instructed that once in a while, after the five pictures in the block were shown, an additional picture could appear after a cue. This picture was either an exemplar of the same bird type, but at a different morph level or an exemplar of a different bird type at the same or a different morph level as the bird exemplars in the previous block. They were instructed to judge whether this picture was the same bird type as the birds presented before the cue. The subjects indicated with a button-press on an MR-compatible response box (Lumitouch by Photon Control, Burnaby, Canada) whether this image was the same as the previously seen images (right index finger) or not (right middle finger). Subjects' heads were fixated and they were shielded from the scanner noise with earplugs. A beamer projected mirror-reversed stimuli on a screen at the end of the bore, which the subject was able to see through a mirror attached to the head coil.
For each subject, 1575 whole-brain images (echo-planar imaging, 34 slices, 3 mm thick with 10% gap, repetition time = 2180 ms, voxel size = 3×3×3 mm, echo time = 30, flip angle = 70°, field of view = 19.2 cm, matrix size = 64×64) were acquired on a 3T whole body MR scanner (Magnetom TRIO by Siemens Medical Systems, Erlangen, Germany). In addition, a high resolution structural T1-weighted 3D magnetization prepared rapid acquisition gradient echo sequence image was obtained after the functional scan (192 slices, voxel size = 1×1×1 mm).
Training data analysis
Response times for the correct trials and the percentage of correct trials were computed for each subject. These dependent variables were submitted to a training condition×morph level×session multivariate analysis of variance (MANOVA) with repeated measures. Training condition consisted of two levels (visual exposure and category training), morph level consisted of five levels (55, 65, 70, 80 and 95%), and session consisted of three levels (first, second, and third training session). To investigate the differentiation between training conditions over time, additional 2 (training condition)×5 (morph level) MANOVA's were performed for each of the training days. All significant interactions were explored with appropriate F-tests.
The presence of a category boundary was investigated by comparing the proportion of ‘same’ responses for bird pairs with an equal morph distance for cases in which the birds were from the same or from a different category. This was done for responses in the final training session, separately for the category training and visual exposure condition. Analyses of these data comprised a 2 (within or between category)×4 (10, 25, 30, 40 % distance) MANOVA for both the category training and visual exposure condition.
fMR imaging data analysis
Imaging data analysis was done using BrainVoyager QX (by Brain Innovation, Maastricht, The Netherlands). The first two volumes were discarded to allow for T1 signal equilibrium. The following preprocessing steps were performed: slice scan time correction (using sinc interpolation), linear trend removal, temporal high-pass filtering to remove low-frequency non-linear drifts of 3 or fewer cycles per time course, and 3D motion correction to detect and correct for small head movements by spatial alignment of all volumes to the first volume by rigid body transformations. Estimated translation and rotation parameters were inspected and never exceeded 3 mm. Co-registration of functional and 3D structural measurements was computed by relating functional images to the structural scan, which yielded a 4D functional data set. Structural 3D and functional 4D data sets were transformed into Talairach space .
To dissociate between effects of training and effects of repetition we performed within-session analyses , . Since objects were repeated in the training as well as in the control conditions, within-session differences between these conditions can not be due to repetition effects but must result from specific effects of training. Therefore, to examine specific training effects we compared responses to bird types in the different training and control conditions, separately for the pre- and post-training session.
Regressors of interest were modelled using a gamma function (tau of 2.5 s and a delta of 1.5) convolved with the blocks of experimental conditions  and multiple regression was performed using the general linear model (GLM). In order to correct for multiple comparisons, the false discovery rate (FDR) controlling procedure was applied on the resulting p values for all voxels. The value of q specifying the maximum FDR tolerated on average was set to .05. With this value, a single-voxel threshold is chosen by the FDR procedure which ensures that from all voxels shown as active, only 5% or less are false-positives , . To further eliminate false-positives in the whole brain analysis, analyses were constrained to only those cortical areas that were responsive to viewing objects as compared with rest. To this end a conjunction analysis with a standard “minimal t-statistic” approach  was used, which is equivalent to a logical AND of the contrasts at the voxel level. For general training effects we used the contrasts: (Category training+Visual exposure<No training)∩(All objects>Rest) to detect training-related decreases in activity and (Category training+Visual exposure>No training)∩(All objects>Rest) to detect training-related increases in activity. For the specific effects of category training we used the contrast: (Category training>Visual exposure)∩(All objects>Rest) to detect increases in activity and (Category training<Visual exposure)∩(All objects>Rest) to detect decreased activity. To test for a main effect of session we contrasted (All objects pre-training)>(All objects post-training). All contrasts were calculated on data that were normalized using a z-transformation.
To further investigate responses within voxel populations (>50 mm3) that showed a significant effect of training, voxel-averaged beta-weights (i.e. regression coefficients) were extracted from these populations for each condition and morph level, separately for the pre- and post-training sessions and averaged over subjects. Random effects GLMs were computed using these regionally-averaged beta-weights. Specific effects of interest were tested with linear contrasts. All reported t-tests are two-tailed. The ROI time-courses were standardized, so that beta weights reflected the BOLD response amplitude of one condition relative to the variability of the signal.
To test for modulation of morph level we extracted the event-related responses to all bird conditions (category training, no training, and visual exposure) at all morph levels (10, 25, 40, 60, 75, and 90 %) from the region in the right middle fusiform gyrus that showed a category training effect. As an example, for the 10 % morph levels of category trained birds (if a subject had bird types A and B assigned to category training) we used responses to the following birds in the calculation: 90A:10B, 90C:10B, 90D:10B, 90E:10B, 90F:10B, 90B:10A, 90C:10A, 90D:10A, 90E:10A, 90F:10A. Each of these bird exemplars occurred six times in the experiment. In total there were 60 trials per morph level per condition. We then used ANOVA's to compute the linear relation between the morph levels and the brain response (beta weights).
Behavioural training results showed that participants became proficient in categorizing the bird exemplars, but only after receiving correct feedback (Figure 3, a and b). In the first session, percentage of correct responses was equally low in both conditions [F(1,11) = 3.76, p = n.s.]. The percentage of correct responses increased as training progressed over time, but only in the category-training condition [F(2,10) = 29.27, p<.001, and not in the visual exposure condition [F(2,10) = 0.03, p = n.s.]. A similar pattern of results was found for response times. In the first session, no differences in response times were observed. Training-related decreases in response times were observed in the category-training condition [F(2,10) = 9.04, p<.01], whereas in the visual training condition response times remained stable over time [F(2,10) = 0.52, p = n.s.]. Significant differences in reaction times and accuracy between category-training and visual exposure conditions were obtained in session 2 (accuracy: [F(1,11) = 26.40, p<.001] reaction times: [F(1,11) = 8.60, p<.05]) and session 3 (accuracy: [F(1,11) = 40.45, p<.001]; reaction times: [F(1,11) = 5.80, p<.05]). By the end of training subjects had developed categorical perception for bird types trained with correct feedback. In the visual-exposure condition performance hovered between 55% and 65%. In the category-training condition, performance improved to around 90% correct for morphs close to the prototype. Even for morph ratios near the category boundary (55:45 morphs), performance exceeded 80% at the end of training. Thus, even though a 55:45 exemplar of, say, bird type A had only 55% of A properties (and 45% of either B, C, D, E, or F properties) it was nonetheless categorized as type A 80% of the time.
Figure 3. Training results.
(A) Mean percentage of correct responses and (B) mean response latencies to the 1-back task, as a function of morph level, plotted for each of the three training days. (C, D) Proportion of “same” responses (see methods) as a function of physical distance between birds in a pair, separately for bird pairs that belonged to the same category (within) and bird pairs that belonged to different categories (between). The left histogram (C) presents the results for the category-training condition, the histogram on the right (D) the visual-exposure condition.
In the third training session, a significant effect of morph level [F(4,8) = 21.40, p<.001] was obtained. Responses were more accurate for bird exemplars with higher morph levels (close to the prototype) than for bird exemplars with lower morph levels (close to the category boundary). This effect of morph level was larger in the category training condition than in the visual exposure condition, as revealed by a condition×morph level interaction [F(4,8) = 6.02, p<.05]. In addition, responses were faster to bird exemplars closer to the prototype than to bird exemplars closer to the category boundary, but only in the category-training condition [F(4,8) = 6.87, p<.05].
The presence of a category boundary was investigated by comparing the proportion of ‘same’ responses for bird pairs with an equal morph distance for cases in which the birds were from the same or from a different category. This was done for responses in the final training session, separately for the category training and visual exposure condition. As expected, for category training we obtained a significant effect of the category boundary (Figure 3, c and d): Subjects were much more likely to rate bird pairs to be the same when they belonged to the same side of the category boundary than equal distance bird pairs belonging to different sides of the category boundary [F(1,11) = 115.86, p<.0001]. For visual exposure the effect was also present [F(1,11) = 4.97, p<.05] but smaller [F(1,11) = 5.22, p<.05]. Importantly, for category training there was no effect of physical distance [F(3,9) = 2.45, p<.05], and no interaction between distance and category boundary [F(3,9) = 0.88, p = n.s.]. The sharp difference in responses for within and between category pairs was maintained over decreasing physical distance between bird pairs (see Figure 3c), clearly indicating category formation. Furthermore, this result shows that the slightly greater performance for the more extreme morphs does not simply reflect a greater average distance between these morphs and their comparison stimuli. For the visual exposure condition a significant effect of distance [F(3,9) = 4.56, p<.05] was obtained. A higher proportion of ‘same’ responses was observed for bird pairs with a small distance than for bird pairs with a large distance (see Figure 3d). See Text S1 and Figure S1 for additional d prime analyses.
Analyses of the pre-training fMRI data showed no significant differences in activity between the bird types. All birds elicited similar patterns of activity, indicating that initially, no differentiation between the birds was made on the basis of their physical features. To test for neural correlates of training-induced category formation, we analyzed post-training responses for the different bird types within object-responsive regions, that is, regions that were active for viewing objects as compared with rest (see methods). See also Text S2 and Figure S2 for the fMRI analysis and discussion of a main effect of session.
General effects of training.
To test for general effects of training, we compared post-training fMRI responses to all trained bird types (category-training and visual-exposure conditions), with post-training fMRI responses to not-trained bird types.
In the post-training session larger responses for trained compared with not trained bird types were obtained in the left posterior fusiform gyrus at a threshold of p<.05 (False Discovery Rate corrected) see Figure 4a and Figure S3. Additional random-effects multivariate analyses of the beta weights extracted from this region for each of the training conditions in both scanning sessions revealed a significant interaction between scanning session and training condition [F(2,10) = 10.64, p<.005]. The response to category-trained bird types was reduced in the post-training session compared to the pre-training session (t(11) = 2.90, p<.05, for the visual-exposure condition the response was also reduced but did not reach significance (t(11) = 2.00, p = .07). Whereas before training, conditions did not differ significantly, after training responses were significantly larger for training as compared with no-training conditions. Direct contrasts of post-training conditions showed that compared with no training, responses were enhanced in the category-training condition [t(11) = 2.58, p<.05] as well as in the visual-exposure condition [t(11) = 3.62, p<.005], see Figure 4b. In these voxel populations, no significant difference was found for category-training and visual exposure conditions [t(11) = 1.05, p = n.s.].
Figure 4. General effects of training.
(A) Group-averaged activation maps from post-training scanning overlaid on lateral (top) and ventral (bottom) views of Talairach-normalized inflated hemispheres. In red, regions showing an effect of training as compared with no training at p<0.05 (False Discovery Rate corrected). In blue, brain regions showing decreased activity following training as compared with no training. (B) Group-averaged time-course of the BOLD response (percent signal changed) averaged over all voxels in the left fusiform gyrus (Talairach coordinates of the centre of mass: x = −33, y = −69, z = −18) that showed a general training effect. Shown are the group-averaged responses for each of three conditions in the pre and post-training scanning session (red: category training, green: no training, blue: visual exposure). Error bars represent standard error of the mean.
In addition to this general training-related enhancement of responses we observed general training-related decreases in activity in frontal, parietal, and occipitotemporal regions at a threshold of p<.05 (False Discovery Rate corrected), see Figure S4 and Table S1. Additional random-effects analyses showed a significant interaction between scanning session and training condition in the right inferior temporal, bilateral fusiform, inferior occipital gyri, the right inferior and middle frontal gyrus, and the bilateral intraparietal sulcus. Whereas before training, conditions did not differ significantly, after training responses were significantly decreased for both for the category-training and the visual-exposure condition, as compared with the no-training condition (Table S1). In addition, these analyses revealed that these decreases in brain activity were independent of training condition. No differences were observed between responses in category-training and visual-exposure conditions.
Specific effects of category training.
To directly test for specific effects of category-training, we contrasted post-training responses to category-trained birds with post-training responses to visual-exposure birds. This contrast revealed significantly larger neural responses for category-trained birds in right middle fusiform gyrus and in the right lateral occipital gyrus (Figure 5a). A random effects analysis revealed significant greater activity for category-trained birds as compared with visual-exposure birds in the right fusiform gyrus [t(11) = 3.26, p<.01], but not in the lateral occipital gyrus [t(11) = 2.07, p = n.s.]. In addition to this increase in activity, decreases in activity for category-trained bird types as compared with visual exposure bird types were observed in occipitotemporal, inferior frontal, and parietal brain regions, see Table S2 andFigure S5. See also Table S3 for areas that were more active for category training compared with no training and visual exposure compared with no training (and vice versa).
Figure 5. Specific effects of category training.
(A) Group-averaged activation maps from post-training scanning overlaid on lateral (top) and ventral (bottom) views of Talairach-normalized inflated hemispheres. In red, regions showing a specific effect of category training as compared with visual exposure at p<0.05 (False Discovery Rate corrected). In blue, brain regions showing decreased activity following category training as compared with visual exposure. (B) Group-averaged time-course and mean beta-weights of the BOLD response in the right middle fusiform gyrus (Talairach coordinates of the centre of mass: x = 36, y = −35, z = −16) in percent signal change. Shown are the group-averaged responses for each of three conditions in the pre and post-training scanning session (red: category training, green: no training, blue: visual exposure). Error bars represent standard error of the mean.
To further analyze the category-specific increase in activity, regions in the right middle fusiform gyrus showing a category-training related increase in activity were defined per subject (Figure S6). Mean beta-weights were extracted from these regions for each condition and morph level, separately for the pre- and post-training session (Figure 5b). A random-effects multivariate analysis of the regionally-averaged beta-weights showed a significant main effect of training condition [F(2,8) = 9.70, p<.01], as well as a significant interaction between session (pre- and post-training) and training condition [F(2,8) = 35.62, p<.0001]. Before training the right fusiform gyrus did not differentiate between the bird types. After training responses were significantly larger for the category-trained bird types than for visual-exposure and not-trained birds. Direct comparisons of the responses in the different training conditions revealed that responses for category-trained birds were significantly larger than responses for visual-exposure bird types [t(9) = 11.32, p<.0001], as well as not-trained bird types [t(9) = 3.06, p<.05]. In addition, significantly smaller responses were found for the visual-exposure condition as compared with the no-training condition (t(9) = 3.00, p<.05).
If the category-training related increase in the right middle fusiform gyrus is specifically related to sensitivity of neuronal populations to the diagnostic features of the category, we should see a positive linear relation between morph level and brain response. This relation should be present for the category trained birds, post-training but not pre-training, and also not for birds from the visual exposure condition for which category-learning was hindered. In addition, if the effect of morph level is specific to category learning it should not be present in the left fusiform gyrus, as this region showed a general training effect. To test this prediction, we investigated whether responses in the right middle and left posterior fusiform showed a linear increase as a function of morph level. As can be seen in Figure 6, a clear linear relationship of morph level and brain response was obtained in the post-training scan session for the category trained birds in the right fusiform only. Before training there was no linear relation between morph level and right middle fusiform response in the category training condition [F(1,4) = 0.09, p = n.s.; R = 0.15], birds from the no training condition [F(1,4) = 0.00, p = n.s.; R = 0.29], or for birds from the visual exposure condition [F(1,4) = 0.11, p = n.s.; R = 0.16]. After training there is still no linear relation between brain response and morph level for birds that were not trained [F(1,4) = 0.17, p = n.s.; R = 0.20]. However, for birds that were category trained there was a significant linear relation between morph level and beta-weight [F(1,4) = 15.87, p<0.05; R = 0.89] and interestingly for birds in the visual exposure condition there existed a negative linear relation between morph level and brain response [F(1,4) = 7.96, p<0.05; R = −0.82]. The responses in the left fusiform gyrus for category trained and visual exposure bird types showed no linear relation with morph level before [category training: F(1,4) = 0.11, p = n.s.; R = 0.05; visual exposure: F(1,4) = 0.30, p = n.s.; R = 0.27] or after training [category training: F(1,4) = 4.95, p = n.s.; R = 0.74; visual exposure: F(1,4) = 1.99, p = n.s.; R = 0.58]. This finding confirms that the effect of morph level in the right fusiform is specific for category learning and not a general consequence of training.
Figure 6. Fusiform responses as a function of morph level.
The effect of morph level is plotted for voxels in the right middle fusiform gyrus showing a specific training effect and voxels in the left posterior fusiform gyrus that showed a general training effect in the post-training scan. For each training condition (red: category training, green: no training, blue: visual exposure) the regionally-averaged brain responses (mean beta-weight) are plotted as a function of morph level (%) in pre-and post-training scan sessions. Lines represent the optimal linear fit between morph level and brain response. Error bars represent standard error of the mean.
Our data provide evidence for experience-induced shaping of neural responses in ventral temporal cortex. Before training, all birds elicited similar patterns of activity, indicating that initially, no differentiation between the birds was made on the basis of their physical features. After training activity in occipitotemporal cortex was modulated as a function of experience. Post-training, activity in the left fusiform gyrus was significantly larger for trained as compared with not-trained bird types (Figure 4). This differentiation in responses occurred after category training as well as after visual exposure. Importantly, category training led to a relative increase in right fusiform responses. Post-training, bird types for which a sharp category-boundary was established during training elicited larger right fusiform responses than not-trained birds. In contrast, visual exposure alone resulted in reduced responses in the right fusiform gyrus (Figure 5b). This clearly shows that the increase in activity for category-trained bird types in the right fusiform gyrus was not caused by mere visual exposure, but mediates the formation of category-specific representations.
These results fit well with functional brain imaging data demonstrating increased activity in occipitotemporal cortex as a function of improved object recognition and visual expertise. Training-related increases in activity in occipital cortex have been reported to follow perceptual discrimination training with nonnatural nonsense objects , . In addition, increased activity in the fusiform gyrus has been found after subjects became proficient in individuating a homogeneous set of nonsense objects . Moreover, increased fusiform activity has been reported after subjects had learned to perform functional tasks with a set of novel stimuli . In addition, larger fusiform responses were observed in individuals that were highly skilled in recognizing a particular class of objects such as birds, cars, or Lepidoptera (butterflies and moths) , , . Although these results clearly show the involvement of occipitotemporal cortex in visual object learning they do not necessarily imply category formation. By dissociating between general effects of visual exposure and specific effects of category training we show that increased activity in the right fusiform gyrus is related to category formation.
Functional imaging data of humans  as well as electrophysiological recordings from monkey cortex ,  have shown increased neural responses in ventral temporal cortex as a function of increased object familiarity. Recently, event-related potential data have shown distinct neural effects for object learning at basic and subordinate levels . While training at a basic object level resulted in improved encoding of coarse visual features, training at a subordinate level resulted in additional encoding of more fine-grained visual object features. The present results show that on the first day of training, performance in the 1-back task was slightly above chance in both training conditions suggesting improved object coding as a function of visual experience. During the second and the third training session performance dramatically improved but only when subjects received correct feedback on their responses. This is in line with the idea that successful categorization of highly similar objects is mediated by learning fine-grained object features indicative of category membership. Indeed, whereas sensitivity in category discrimination was high for the category-trained bird types, for the visual-exposure bird types category-discrimination ability was very poor. In the visual exposure condition, the proportion of same responses was slightly higher for within- as compared with between-category bird-pairs. However, this small effect differed significantly from the sharp boundary effect obtained after category training. Consistent with the behavioural results, we found a clear neural dissociation between general effects of visual training and the formation of an object category. Whereas post-training training-related increases in activity in the left posterior fusiform gyrus occurred independently of category formation, increased responses in the right middle fusiform gyrus were only observed for bird-types for which a sharp category-boundary was established. This dissociation suggests that the left fusiform gyrus is probably involved in the encoding of general shape information, and the right fusiform is encoding fine-grained visual information required for category formation.
Our results are consistent with electrophysiological recordings from the inferior temporal cortex in monkeys suggesting that object category formation is mediated by a learning induced sharpening of neuronal stimulus selectivity , , . Our behavioural data showed that responses were more accurate and faster for birds at higher morph levels, reflecting that birds close to the prototype are more distinctive than birds close to the category boundary. This implies that the closer to the prototype, the more apparent the features that determine to which category a bird belongs. Recently, it has been shown that neuronal selectivity in monkey inferior temporal cortex is shaped by those object features that were most relevant during categorization training . In addition, single-cell recordings from monkey cortex have demonstrated that discrimination training enhances the selectivity of neurons in inferior temporal cortex not only for features in isolation but also for whole objects . In line with these findings from monkey cortex, our findings suggest that after category training, neuronal populations in the right fusiform gyrus differentiated between object features that were informative of a category and features that were uninformative. Right fusiform activity was modulated by morph level (Figure 6). Responses were positively related with the morph-level of category trained birds and negatively related with the morph-level of birds for which category-learning was hindered by random feedback. This means that the higher the percentage of features trained to be relevant for categorization, the larger the responses in the right fusiform gyrus. In contrast, the higher the percentage of features trained to be irrelevant for categorization training, the smaller the right fusiform responses. Moreover, the left fusiform gyrus that showed a general training effect did not show a positive linear relation between morph level and responses, indicating that the effect of morph level is specific for category learning and does not occur as general consequence of visual exposure. One of the neural mechanisms that could explain this pattern of enhanced responsiveness to relevant category features and suppressed responses to irrelevant features involves increased tuning of neuronal populations to informative combinations of visual features. Op de Beeck et al.  have shown that the largest effects of training occur in regions that already process stimulus properties that are relevant during training. This suggests that increased tuning of neuronal populations concerns those features that were most relevant during training. However, since the present fMRI data reflect overall magnitude of response of relatively large neuronal clusters, no direct conclusions can be drawn on whether the results indeed reflect increased neural tuning. One way to investigate neuronal sensitivity with fMRI is by using an adaptation paradigm. Recent studies using this paradigm showed narrow shape tuning of neural populations in occipitotemporal cortex to sub-exemplar faces ,  and trained car stimuli . This suggests that neural populations in this brain region are highly specialized to dissociate between fine-grained visual features, which fits nicely with our interpretation of the results.
The location of our post-training training-related increase in activity in the right fusiform gyrus seems to be close to the location of the fusiform face area (FFA), a region that has been claimed to be specifically involved in face recognition , . This claim has been challenged by findings relating FFA activity to increased expertise in object recognition , . However, since we did not localize the FFA in our subjects we should be cautious about whether the current results directly address the debate regarding the function of the FFA. It is unclear whether the exact same region is involved here. The FFA is neighboured by regions that prefer other stimuli, such as bodies , . Also, birds have faces and previous studies have shown that the FFA responds to animal faces to a considerable extent , . Our subjects might have found the features in the bird's head extra useful for categorization. Therefore, the training-effects may have occurred in regions that process facial features. Note however, that not all features informative of a bird's category were located in its head and we cannot be certain that during training the facial features received indeed the most attention. Nevertheless, should the increase we observe for category-trained birds be attributed to the presence of a face in the stimuli, this does not deter from our novel finding of an increase that is specific to only those bird types for which category boundaries were formed during training.
In addition to training-related increases in activity, in some areas neural responses were significantly reduced for bird types from both category training and visual exposure conditions. These opposite patterns of responses in different brain regions might reflect two different learning mechanisms. While the underlying mechanism for the relative increase in the right middle temporal gyrus might be increased neuronal tuning for those object features relevant for category learning, a different mechanism could explain lower responses for trained compared with not-trained birds. Reduced occipitotemporal responses have consistently been reported to follow repeated exposure to visual objects , even over a delay of several days , . This so-called repetition-suppression effect has been argued to reflect a learning process in which stimulus representations are optimized. Repeated exposure to the same stimulus causes neurons coding non-specific stimulus features to drop out of the responsive pool, while neurons tuned optimally to the stimulus continue their activity , , . As a consequence, the total number of responsive neurons decreases, leading to a reduced overall response. In line with this idea, the reduced neural response for trained birds could reflect the formation of sharper object representations. Since reduced responses occurred in both the visual exposure and the category-training condition, this sharpening process is not related to object-category formation but probably reflects object-specific visual learning. In addition to general training-related decreases in activity, some occipitotemporal regions showed reduced responses for category-training as compared with visual-exposure conditions. This shows that applying random feedback not only hindered category learning , but also affected sharpening of object-specific representations. Although repetition suppression occurs as a result of repeated visual exposure, differences in encoding as a result of receiving correct or random feedback, might have led to differential changes in stimulus-specific representations , .
Our data provide evidence for learning-related formation of visual object category representation in occipitotemporal cortex. However, occipitotemporal cortex is not the only brain region that has been implicated in object-category learning. Monkey data have shown that neurons in prefrontal cortex respond selectively to members of a learned category, irrespective of within category variations . These data were obtained while monkeys were actively involved in a categorization task. Although in our paradigm subjects may have been implicitly categorizing the birds throughout the scan session in order to successfully perform the task, this did not elicit training-specific increases in prefrontal cortex. Recently, it has been shown that prefrontal cortex shows a category-dependent response only when human subjects were performing a categorization task and not when performing a displacement detection task . The exact relationship between the nature of a categorization task and category-selective responses in human cortex remains to be determined. Data from network models on object category learning suggest that during learning, the top-down influence of prefrontal cortex enhances the selectivity of the neurons in inferior temporal cortex encoding the behaviourally relevant features of the stimuli , . Presumably, category-learning requires collaboration between these different brain structures, with the occipitotemporal cortex storing characteristic features of objects belonging to a learned category, and the prefrontal cortex being involved in explicit retrieval of category information.
Main effect of session. Group-averaged activation maps of the between-session effect overlaid on lateral (top) and ventral (bottom) views of Talairach-normalized inflated hemispheres. In grey with a black outline, regions showing less activity for all objects after training as compared with activity to the same objects before training at p<0.01 (False Discovery Rate corrected).
(8.44 MB TIF)
Single-subject data showing a general effect of training. In red the areas that showed a higher response to trained as compared with not trained birds (p<0.05) overlaid on the axial slices from the corresponding normalized structural images. Structural images are in neurological convention.
(1.64 MB TIF)
General effects of training.(A) Group-averaged activation maps from post-training scanning overlaid on lateral (top) and ventral (bottom) views of Talairach-normalized inflated hemispheres. In red, regions showing an effect of training as compared with no training at p<0.05 (False Discovery Rate corrected). In blue, brain regions showing decreased activity following training as compared with no training. (B) Mean beta-weights (i.e., estimates of signal amplitude) for voxel populations in left and right occipitotemporal cortex showing a general decrease in activity for trained birds as compared with not-trained bird types. Shown are the group-averaged responses for each of three conditions in the pre- and post-training scanning sessions. Error bars represent standard error of the mean.
(4.49 MB TIF)
Specific effects of category training.(A) Group-averaged activation maps from post-training scanning overlaid on lateral (top) and ventral (bottom) views of Talairach-normalized inflated hemispheres. In red, regions showing a specific effect of category training as compared with visual exposure at p<0.05 (False Discovery Rate corrected). In blue, brain regions showing decreased activity following category training as compared with visual exposure. (B) Mean beta-weights for voxel populations in left and right occipitotemporal cortex showing a specific decrease for category-trained birds as compared with birds from the visual-exposure condition. Shown are the group-averaged responses for category-training, no training, and visual-exposure conditions in the pre- and post-training scanning sessions. Error bars represent standard error of the mean.
(3.65 MB TIF)
Single-subject data showing a specific effect of category training. In red the areas that showed a higher response to birds from the category training condition as compared with visual exposure birds (p<.05) in the right middle fusiform gyrus overlaid on the axial slices from the corresponding normalized structural images. Structural images are in neurological convention.
(1.30 MB TIF)
Brain regions showing a significant decrease in activity after category training and visual exposure as compared with no training, as well as a significant interaction between training condition and scanning session in a random effects analysis. For each region, mean Talairach coordinates, corresponding Brodmann's areas (BA), averaged t-values (df = 11) for the contrast between (category training+visual exposure) and (no training) are reported, separately for the pre- and post-training sessions. In addition, averaged t-values (df = 11) are reported for the interaction between training condition and scanning session.
(0.05 MB DOC)
Brain regions showing significantly less activity for category-trained birds as compared with birds from the visual exposure condition, as well as a significant interaction between training condition and scanning session in a random effects analysis. For each region, mean Talairach coordinates, corresponding Brodmann's areas (BA), averaged t-values (df = 11) for the contrast between category training and visual exposure are reported, separately for the pre- and post-training sessions. In addition, averaged t-values (df = 11) are reported for the interaction between training condition and scanning session.
(0.05 MB DOC)
Pre-roost murmuration displays by European starlings Sturnus vulgaris are a spectacular example of collective animal behaviour. To date, empirical research has focussed largely on flock movement and biomechanics whereas research on possible causal mechanisms that affect flock size and murmuration duration has been limited and restricted to a small number of sites. Possible explanations for this behaviour include reducing predation through the dilution, detection or predator confusion effects (the “safer together” hypotheses) or recruiting more birds to create larger (warmer) roosts (the “warmer together” hypothesis). We collected data on size, duration, habitat, temperature and predators from >3,000 murmurations using citizen science. Sightings were submitted from 23 countries but UK records predominated. Murmurations occurred across a range of habitats but there was no association between habitat and size/duration. Size increased significantly from October to early February, followed by a decrease until the end of the season in March (overall mean 30,082 birds; maximum 750,000 birds). Mean duration was 26 minutes (± 44 seconds SEM). Displays were longest at the start/end of the season, probably due to a significant positive relationship with day length. Birds of prey were recorded at 29.6% of murmurations. The presence of predators including harrier Circus, peregrine Falco peregrinus, and sparrowhawk Accipiter nisus was positively correlated with murmuration size (R2 = 0.401) and duration (R2 = 0.258), especially when these species were flying near to, or actively engaging with, starlings. Temperature was negatively correlated with duration but the effect was much weaker than that of day length. When predators were present, murmurations were statistically more likely to end with all birds going down en masse to roost rather than dispersing from the site. Our findings suggest that starling murmurations are primarily an anti-predator adaptation rather than being undertaken to attract larger numbers of individuals to increase roost warmth.
Citation: Goodenough AE, Little N, Carpenter WS, Hart AG (2017) Birds of a feather flock together: Insights into starling murmuration behaviour revealed using citizen science. PLoS ONE 12(6): e0179277. https://doi.org/10.1371/journal.pone.0179277
Editor: Charlotte K. Hemelrijk, Rijksuniversiteit Groningen, NETHERLANDS
Received: March 9, 2017; Accepted: May 27, 2017; Published: June 19, 2017
Copyright: © 2017 Goodenough et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All files are available from the University of Gloucestershire Research Repository eprints.glos.ac.uk (accession number 4617).
Funding: The authors received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
Collective animal behaviour, where multiple individuals of a species act in a highly-coordinated manner, is both taxonomically and ecologically widespread . Examples include fish schooling to distract predators (e.g. banded killifish Fundulus diaphanous) , ants collectively foraging using pheromone trails , and herds of migrating blue wildebeest Connochaetes taurinus . Many collective behaviours exhibit characteristics of self-organisation, whereby relatively simple repeated interactions produce complex emergent patterns [5–7]. In many cases, environmental factors modulate individual interactions, such that different collective behaviours emerge under different conditions . Collective behaviours are assumed to be adaptive  and understanding the mechanisms by which they emerge is a key area of research.
One spectacular example of collective behaviour is the pre-roost aerial displays undertaken by European starlings Sturnus vulgaris [9–10]. These murmurations (so-called because of the sound produced by multiple wingbeats) can involve thousands of individual birds forming a coherent three-dimensional murmuration “cloud” within which the movement of each individual bird is highly cohesive and synchronized. This synchronised movement means that the group forms a range of different shapes [11–12] including spheres, planes and waves [13–14] while remaining more-or-less static with respect to a focal point on the ground (generally the roosting site). Typically occurring at sunset, murmurations generally end with the birds descending en masse to roost.
As highlighted by King and Sumpter , murmuration behaviour is of considerable interest not only to biologists, but also to physicists, engineers and mathematicians. To date, attention has tended to focus on the mechanisms of murmurations to determine the rules governing individual “agent” behaviour and understanding how these rules lead to collective behaviour [15–16]. Much of this work has used computer analysis of rapid-burst photographs or video footage of murmuration events [11–12; 17–18] as well as agent-based simulation modelling [19–22]. For example, photography of murmurations in Rome has allowed insights into the morphology and orientation of starling flocks . Analysis of videos from the same sites has revealed that the distance (correlation length) over which velocity is correlated among neighbours within a murmuration is dependent on flock size, such that the behavioural state of one bird affects, and is affected by, that of all other birds involved in the same event . Tracking individual birds using 3D reconstruction has shown that direction-switching is spatially localised (i.e. the birds that turn first are all co-located within the group) and then propagates through the group through bird-to-bird information transfer . The 3D reconstruction approach has also been used to show that each bird interacts with, and moves according to, six or seven nearest neighbours and it is the proximity of those neighbours, as opposed to all birds within a fixed distance, which dictates individual movement .
Somewhat surprisingly given the amount of work on the mechanisms of murmurations, we know little about their adaptive value or what factors affect size and duration . One potential explanation is the “warmer together” hypothesis. As murmurations occur immediately before roosting, and during the late autumn and winter months, it is possible that they act to “advertise” a roost site so the roost becomes warmer as more birds gather (producing an inverse relationship between temperature and murmuration duration/size). This would also allow the possibility of individual birds following more successful or experienced individuals to good feeding areas as the roost disperses. This has been seen both for socially-roosting red-billed quelea and cliff swallows Hirundo pyrrhonota [23–24] and could be particularly important in cold weather . The “warmer together” hypothesis has not been tested empirically for starlings, but Davis and Lussenhop  showed that small flocks 'funnelled' into progressively larger ones along flight-lines to the roost and concluded that social stimulation from aerial displays was important for creating larger roosts. A similar situation has been found in common bushtits Psaltriparus minimum where birds form denser roosts in colder weather  and in wrens Troglodytes troglodytes where birds vocally advertise communal roosts in cold weather .
The second hypothesis (or, more correctly, group of hypotheses) is that murmurations are an anti-predator strategy. Starlings are potential prey for a range of avian predators, especially hawks and falcons. Murmurations could reduce predation risk in one of several ways, which are not mutually exclusive:
- Dilution effect: as group size (N) increases, the chance of any one individual suffering predation (1/N) decreases, favouring larger groups [29–30]. Individual birds could further decrease their individual risk through a separate (but complementary) strategy of seeking more sheltered locations within the flock, with birds on the edge of the flock continually moving towards the centre within a Hamiltonian “selfish herd” [31–32]. Flocking behaviour in other species has been shown to be consistent with selfish herd dynamics (e.g. sand fiddler crabs Uca pugilator  and European minnows Phoxinus phoxinus in structurally-simple habitats ).
- Detection effect: group anti-predator vigilance increases as the number of individuals in the group increases [35–36]. This remains the case even when each individual in a group contributes less vigilance than a typical solitary individual, as shown for yellow-eyed juncos Junco phaeonotus and starlings in small feeding flocks [37–38, respectively]. The detection effect often interacts with the dilution effect, as shown by as shown for captive socially-flocking red-billed quelea Quelea quelea  and red deer (elk) Cervus elaphus .
- Predator confusion effect: most aerial predators hunt by targeting a specific bird within a group. The constant movement within murmurations might reduce a predator’s ability to “lock onto” an individual; an effect known as target degeneracy . This has been seen in silvery minnows Hybognathus nuchalis, whose movement within groups of almost eliminated depredation by largemouth bass Micropterus salmoides . Group size and density might also be important, with predators finding it harder to isolate individuals in larger, denser groups.
There has been little work to test the “safer together” hypotheses for starling murmurations. The dilution and confusion hypotheses have been supported by agent-based modelling of within-flock movements [32, 42], as well as in a computer game experiment on the ability of a (human) predator to “capture” a target starling in a biologically-realistic model . Empirical studies have shown that predation pressure can result in waves of agitation (dark bends moving away from the predator ) as a result of changes in bird orientation  and in the shape of the murmuration itself , but these studies did not analyse the effect of predators on murmuration size or duration. The studies also only considered one predator species (peregrine falcon Falco peregrinus) and were based on data from just two sites.
Obtaining detailed data on multiple murmurations at broad spatiotemporal scales presents a considerable challenge . Here, we use a citizen science approach  to harness the efforts of thousands of volunteers to record murmurations. Conducted over two years, and amassing >3,000 records from 23 different counties, the project gathered information on murmuration size and duration in relation to location, season, time of day, and habitat (year one) as well as temperature and predator presence and behaviour (year two). This approach overcame the typical limitations inherent in murmuration research to allow, for the first time, a robust analysis of potential triggers of murmuration behaviour.
Citizen science surveys
The first online survey, promoted through the websites of the Royal Society of Biology and the University of Gloucestershire, was created using Survey Monkey and prefaced with information on starling murmurations and photographs of murmuration displays. The survey opened on 17th October 2014 with participants able to document records from the beginning of that month (1 October = day 1). The survey remained open until 31st March 2015, with the final record being submitted on 23rd March 2015 (day 174). This spanned the autumn/winter period when murmurations occur. Participants were asked to log observations of murmurations and provide the following baseline information: date; time; location (as postcode, Ordnance Survey grid reference, latitude/longitude or address); whether the location was urban, rural or suburban; and the habitat over which the birds were murmurating. Participants were asked to estimate murmuration size (number of birds), duration (minutes from the start of observation until the display ended), and what happened at the end of the murmuration—if this was observed—to determine whether birds went down to roost or dispersed from the location. Participants could report multiple murmurations over the survey period, including murmurations occurring at the same location on different nights. Participants were advised before the submission step that by continuing with submission they were providing informed consent for participating in the study and their data being used in subsequent published research.
A second online survey was launched on 3rd November 2015 with participants being submit records from the beginning of the preceding month (1 October = day 1). The survey remained open until 31st March 2016 with the final record being submitted on 27th March 2016 (day 179 due to leap year). Participants were again asked to record murmuration size, duration, and what happened at the end of the event but this time were asked for details of temperature (°C) and the presence of other birds. The species that could be recorded were all potential predators (or species that might, in silhouette, look similar to potential predators): kestrel Falco tinnuculus, peregrine Falco peregrinus, sparrowhawk Accipiter nisus, red kite Milvus milvus, buzzard Buteo buteo, harrier Circus spp., owl (order: Strigiformes); corvids (order: Corvidae), or gulls Larus spp. The order of species was randomised for each survey to reduce the risk of under- or over-representation by virtue of order of appearance. There was also opportunity for other species to be listed. The activity of potential predators was recorded from the options of: (1) perching and silent; (2) perching and calling; (3) flying but neither close to, nor engaging with, the starlings; and (4) flying very close to, or actively engaging, with the starlings (active engagement being defined as a bird of prey flying through the starling flock or making a predator strike, whether or not this was successful). Finally, people were asked for very brief details of location (nearest town) and when the murmuration had occurred from the options of “today”, “within the last week” or “within the last month” from a drop-down list. This very basic way of collecting location and date information made the survey quicker and avoided duplication with year one but allowed international records to be identified and excluded from certain analyses (e.g. analysis of predators in cases where the potential predator community differed) and exclusion of records of murmurations from more than one month previously when memory of temperature and predator activity could be unreliable.
In both years, there was extensive media promotion and coverage of the study including mentions in most UK national newspapers as well as articles in over 200 UK local and regional newspapers and national magazines. The survey was featured on UK regional and national TV news and countryside-related programmes as well the local BBC radio network. A variety on online news sites, blog posts and Facebook sites were used to promote the survey and via a dedicated Twitter account @Starling_Survey. Twitter was also searched daily for people mentioning murmuration displays and those users were sent direct tweets asking them to complete the survey by following an embedded link. The majority (ca. 95%) of Twitter invitations were sent in the first three months of the survey each year; in February and March there were fewer mentions of murmurations on Twitter in general and most of these were from people who had received the invitation in response to previous murmuration tweets. Anecdotal observations did not suggest murmurations that were the subject of tweets–and thus invitations to complete the survey–were exceptional in any way (e.g. particularly large or long in duration). Rather, tweeters frequently mentioned that their observation was the first murmuration they had seen, that they happened to get a good photograph or it, or simply that they had enjoyed a display and wanted to alert others to its existence.
Data from both years were carefully standardised. Historical records (i.e. people recording a murmuration remembered from >1 month previously) were deleted (n = 40), as were incomplete records (n = 167). Records that were obviously inaccurate or unusable were also discarded. These included records where there were no birds present (n = 2), occasions when murmuration duration was given as zero minutes (n = 8), and reports of individuals feeding in gardens/parkland (n = 67). Where location had been provided as a postcode, address or grid reference, this was translated into latitude and longitude. Records outside the UK were highlighted so that they could be included in spatial analyses but excluded from analyses of predators to avoid location being a confounding factor. In each year, a subset of the data was created, which only contained observations of murmurations of ≥500 birds that descended to roost at the end of the display. These were likely to be true murmurations based on both the size of the flock and the ultimate roosting behaviour, rather than simple flocks of birds moving to a new location. The use of this threshold was based on work in Rome where 448 birds and 428 birds were the minima to required produce a definite murmuration pattern [11, 15].
For the UK subset data, average day length was calculated for the geographical centre of the UK as defined by the Ordnance Survey (Dunsop Bridge, Lancashire, UK, 53.9419°N 2.5369°W) using the Sunrise-Sunset mobile application (version 1.03; Peter Smith, petesmith.co.nz/sunrise-sunset-modern). These data were added for each record based on the day that the murmuration was sighted. Although day length was specific only to the date of the murmuration record and not its location, analysis showed that day length in the UK only varies across the survey period by an average of ca. 10 minutes from the centre of the UK to the north of Scotland or the far south of England.
All analyses were undertaken in R Version 3.3.1 , SPSS Version 22 for Windows (IBM), or QGIS Version 2.16  using underlying base maps from Natural Earth that are in the public domain. To analyse spatial patterns, QGIS was used to display location data at both global and UK scales. Pearson correlation was then used to quantify any relationship between: (1) the number of records received from a location and average murmuration size; and (2) the size and duration of murmuration events. Separate one-way ANOVA were used to determine any association between habitat and: (1) murmuration size; and (2) murmuration duration. To compare murmuration size and duration in the UK with non-UK sites, we used independent t-tests. To identify any patterns in murmuration size or duration with location across the UK, we correlated these variables against both latitude and longitude.
Temporal patterns in the size and duration of murmuration events in the 2014/15 data were analysed using curvilinear regression, with day as the independent variable (1 = 1 October 2014; 174 = 23rd March 2015). The optimal model was selected based on R2. Linear regression was used to establish any pattern between day length and murmuration duration. As day length was averaged spatially over the UK (see above) this analysis was undertaken using data aggregated into weeks rather than using the raw daily data. This reduced the risk of days with atypically clustered data points (e.g. all records on an individual day coming, by chance, from the north of Scotland) biasing analysis.
To analyse the effect of temperature and predator presence/activity on murmuration size and murmuration duration, Regression with Empirical Variable Selection (REVS) was used . This is superior to stepwise algorithms, which although highly intuitive, can be inconsistent, only test a small number of possible models, and can miss the optimal model because of the one-at-a-time nature of adding variables [49–52]. The alternative of all-subsets regression, whereby numerous models are created and compared using Akaike’s Information Criterion (AIC) , is more robust but the number of models generated increases exponentially with the numbers of predictors. The results can also be difficult to interpret when multiple (often very different) models are generated that have similar support, as shown by work by on bird-habitat associations . REVS involves a series of n models being created (n = the number of predictors); the first containing the variable with most empirical support, the second containing that and the next most-supported, and so on. The resultant models are compared post-hoc using (AIC). This means: (1) the number of models needing comparison is lower (as n = the number of predictors rather than many times that number, as are typically generated with other approaches) and (2) all competing models have many variables in common (i.e. the “core” is the same; just minor differences in presence/absence of additional variables), which makes interpretation easier. REVS has been used in a range of ecological studies [54–55], including those using citizen science data .
Here, the REVS process was run four times, twice with murmuration size as the dependent variable and twice with murmuration duration as the dependent variable. The first analysis in each case included temperature, the presence of each predator species recorded (kestrel, peregrine, sparrowhawk, red kite, buzzard, harrier, owl) and the presence of avian species that could be mistaken for potential predators (corvids and gulls): n = 10 predictors. The second analysis in each case included temperature and the activity of each predator species (perching silent, perching calling, flying, or engaging with starling flock): n = 29 predictors (7 species * 4 behaviours = 28, plus temperature). It was necessary to run these analyses separately because presence and activity of each predator species were highly correlated (indeed, on a per-species basis, presence was the sum of all four activity data columns), which confounded orthogonality constraints . Models were compared post-hoc using AIC based on ΔAIC < 2, while R2 was used to assess the biological significance of models. Although p values are arguably not important in AIC-driven analyses, they were given at model-level so overall statistical significance could also be assessed. For each dependent variable under consideration, three models were reported: (1) Minimum adequate model—the most parsimonious model (fewest predictors whilst still attaining ΔAIC < 2); (2) Optimal—the model that best balanced the number of variables and explanatory power (ΔAIC = 0); and (3) Maximum—the model that increased R2 to the maximum possible within the ΔAIC < 2 limit. Spatiotemporal autocorrelation was checked for each model using the Durbin-Watson test with the underlying data formatted so records were ordered by time within location to provide an appropriate data structure; in all cases the values were between 1.5 and 2.5 (a value of 2 signifies no autocorrelation and ± 0.5 is well within the accepted range given by Field ). There were no duplicate records for the same site on the same night and thus pseudoreplication was avoided.
It was not necessary to seek or obtain ethical clearance for this study since all work was solely based on non-invasive observation. No animals were handled, approached, or otherwise affected by the work undertaken.
In total, 3,211 records were submitted with the records split between years as detailed in Table 1. The end of the murmuration event was seen in 67.1% of cases. Based on data from 2014/15, when the exact date that the murmuration had been observed was provided, there was no relationship between the lag time from sighting and reporting a murmuration event (number of days: min = 0; max = 44; mean = 3) and its size or duration (Pearson correlation: r = 0.040, n = 553, p = 0.346 and r = 0.044, n = 553, p = 0.305, respectively).
The survey was envisioned as UK-based but 191 records (~6% of total) were submitted from outside the UK from 22 different countries. In total, 29 international records were not considered to be actual murmurations based on our data cleaning protocol (see Methods). The remaining records were from: USA (70), Canada (30), Netherlands (13), Eire (10), Italy (10), Spain (10), France (5), Belgium (3), Mexico (3), Bulgaria (2), Greece (2), Hungary (2), India (2), Australia (1), Germany (1), Jordan (1), Pakistan (1), Switzerland (1), and Ukraine (1); Fig 1A.
Reported starling murmurations: (a) = international distribution (; (b) UK basic distribution; (c) = UK distribution showing number of records; (d) = UK distribution showing mean size of murmuration. All base maps from Natural Earth (freely available in the Public Domain); all starling data from Starling Survey run by authors and freely available–see Data Availability Statement. Maps created using QGIS under CC BY.
In terms of the UK, records came from as far south as Penzance in Cornwall 50.1190° N, 5.5370° W; as far north as Thurso in the Scottish Highlands, 58.5960° N, 3.5210° W; as far east as Lowestoft in Suffolk, 52.4800° N, 1.7500° E (the most easterly point of the UK); and as far west as Ballymoney, County Antrim, Northern Ireland 55.0710° N, 6.5080° W (Fig 1B). There were also records from offshore islands: Orkney, Skye and Arran (Scotland); Anglesey (Wales); and the Isle of Wight (England). There were some interesting patterns in the number of murmuration records from different sites (Fig 1C), with popular hotspots including Brighton and Aberystwyth piers (Southeast England and West Wales, respectively) and Gretna Green on the west coast near the English/Scottish border. There were also spatial patterns in the average size of murmuration events (Fig 1D), with large murmurations reported on the Somerset Levels in Southwest England, near Berwick-upon-Tweed on the east coast of the English/Scottish border, and Anglesey off the Welsh coast. There was no correlation between the average size of murmurations at a given location and the number of records received from that location (r = 0.044, n = 308, p = 0.443).
There was no significant difference in average murmuration size between the UK and either all non-UK records or the subset of non-UK murmurations that occurred outside the species’ native range in the USA/Canada (independent t-test: t = 0.604, d.f. = 545, p = 0.546 and t = 0.995, d.f. = 530, p = 0.320, respectively). However, UK murmurations lasted longer (mean = 26 minutes ± 44 seconds SEM) than all non-UK murmurations (mean = 18 minutes ± 2.5 minutes SEM) or the USA/Canada subset (mean = 16 minutes ± 3 minutes SEM). These differences were significant (UK vs. non-UK: t = -2.849, d.f. = 548, p = 0.005; UK vs. USA/Canada: t = -3.077, d.f. = 533, p = 0.002). These analyses were all based on data from 2015/16 when most international records were submitted and used only the reduced datasets (following criteria outlined above: ≥500 birds; end of murmuration seen). Within-UK patterns were analysed using 2014/15 data as exact location was only requested in the first survey. There was no significant relationship in murmuration size with either latitude (F1,549 = 2.510, p = 0.114) or longitude (F1,549 = 0.326, p = 0.568). There was a significant relationship between murmuration duration and latitude (F1,549 = 11.693, p = 0.001) with murmurations lasting longer further north but the amount of variance explained was low (r2 = 0.021). The was no relationship between duration and longitude (F1,549 = 3.114, p = 0.078).
In 2014/15, participants recorded habitat for each murmuration event. Of the UK murmurations, 61.2% occurred in rural areas, 19.2% in suburban and 19.6% in urban. The distribution of murmurations across habitat types is shown in Fig 2. Analysis on the subset of the data with ≥500 birds and the end of the murmuration was observed revealed no association between the habitat and either the size of the murmuration or its duration (ANOVA: F11,645 = 0.941, p = 0.500 and F11,641 = 1.217, p = 0.272, respectively). Roosting sites included elevated vegetation (especially coniferous trees and hedgerows), tall vegetation (especially when surrounded by water, as in the case of reedbeds or emergent marsh vegetation) and elevated manmade structures such as piers and building ledges.
Fig 2. Number of murmurations associated with different landscapes (terrestrial, anthropogenic, aquatic) and different habitats with those landscapes based on data from survey year one (2014/15); n = 1,293.
Based on temporal data collected in 2014/15, there was a clear pattern in the number of records over time (Fig 3A), but this largely coincided with the main dates of publicity rather than any potential underlying seasonal pattern. There was a clear temporal trend in the mean size of murmurations, with the average number of birds increasing throughout the season until a peak in early February, after which size decreased again (Fig 3B). This was best explained by a quadratic curvilinear regression (F2, 573 = 4.671, p = 0.010; R2 = 0.127), which essentially described a negatively-skewed bell-shaped curve (y = 3.539x12 + 1264x1−53688; Fig 3A).
Temporal patterns in: (a) number of murmurations (n = 1,644); (b) mean number of birds per confirmed murmuration (n = 1,293); and (c) mean duration of murmuration only including records where the end of the murmuration event was recorded (n = 553). All data based on data from survey year one (2014/15); more details on sample sizes are given in Table 1. Dotted lines show annual means. Error bars show standard error.
There was also a temporal trend in the duration of murmurations, which were longer at the beginning and end of the season and shorter in the middle (Fig 3C). This was best explained by a quadratic curvilinear regression (F2,569 = 21.877, P < 0.001, R2 = 0.267), which essentially described a shallow u-shaped distribution (y = 0.001x12–0.065x1 + 19.915; Fig 3C). This pattern reflected seasonally variation in day length with murmurations being at their shortest around the winter solstice (i.e. in the middle of the study period). Reflecting this, there was a strong positive correlation between murmuration duration and day length (F1,28 = 17.488, P < 0.001, R2 = 0.384; Fig 4). It should, however, be noted that longer durations at the start/end of the season (>30 minutes) were based on fewer records and so the precision of the mean is reduced (see larger error bars in Fig 3C). As murmurations in early October were recorded only after the survey launched in mid-October (year 1) or early-November (year 2), recording bias might be partly responsible for early-season murmurations being above-average in duration by virtue of being more memorable. However, given that only 2–4 weeks elapsed between survey launch and the earliest accepted record of 1st October, and that there was no relationship between sighting-to-reporting lag time and murmuration duration (see above), this is unlikely to be an important bias.
Fig 4. Relationship between murmuration duration (weekly mean duration in minutes) and day length in 2014/15 based on records where the end of the murmuration event was recorded (N = 553).
Birds of prey were recorded at 29.6% of murmurations. The most common species was sparrowhawk followed by buzzard, marsh harrier, hen harrier, and peregrine falcon (Fig 5). In addition, 15.8% of observers mentioned corvids and 17.6% mentioned gulls. These are not birds of prey, but were included because they might appear similar in silhouette and elicit a predator-avoidance response, especially given the low-light conditions at dusk.
Explanatory models: Murmuration size and duration
The presence of potential predators was positively correlated with murmuration size. Harrier presence was added into the explanatory model first, followed by presence of buzzard, peregrine and owl. Adding sparrowhawk presence improved the final model, which explained 35% of variance in murmuration size and was highly significant (Table 2). All relationships were positive, such that murmurations were larger when potential predators were present (or, potentially, predators were more likely to be found at larger murmurations: see Discussion). Temperature and presence of non-predator species (corvids/gulls) were not significant. When presence of predators was replaced with their activity, similar models were created. Interestingly, although four activities were recorded for all predator species (perching silent, perching calling, flying near to the starling flock, and activity engaging with starlings by flying through the flock or making a predatory strike), it was only “flying” and “engaging” that ever met the model entry criteria. Harrier (engaging; flying) and peregrine (flying; engaging) were the four most important variables, followed by buzzard (engaging). All relationships were positive. The final model explained 40% of variance in murmuration size, which was higher than models using predator presence alone (Table 2). Again, temperature was not significant.
Table 2. Hierarchical regression models to explain murmuration size (number of birds) and murmuration duration (minutes) based on either temperature and predator presence or temperature and predator activity (perching silent, perching calling, flying, interacting with flock).
In all cases, three models were created: (1) Minimum Adequate Model (MAM)–the most parsimonious model (i.e. the model that had fewest predictors whilst still attaining ΔAIC < 2); (2) Optimal—the model that best balanced the number of variables and explanatory power (i.e. ΔAIC = 0); and (3) Maximum—the model that increased adjusted R2 to the maximum possible within the ΔAIC < 2 limit. See methods for more details.
Both predators and temperature were significantly related to murmuration duration. The first model (using predator presence) showed that murmuration duration was significantly related to sparrowhawk presence (positive) and temperature (negative). Adding presence of harrier, buzzard and peregrine (all positive) improved the model but the variance explained was fairly low at 12%; presence of non-predatory corvids and gulls was not significant (Table 2). When the presence-only model was replaced by one that accounted for the behaviour of potential predators, the variance explained increased substantially. As with models of murmuration size, the only predator activity variables entered into the duration model were those that involved potential predators either flying near to the starling flock or actively engaging with it. The initial minimum model included sparrowhawk (engaging and flying; both positive) and temperature (negative). This model was improved by the addition of harrier (engaging and flying), peregrine (engaging), buzzard (flying), and kite (flying); all relationships were positive. The final model explained 26% of variance (Table 2). Analysis of temperature and duration alone (regression: F1,439 = 9.611, P = 0.002, R2 = 0.144) was weaker than the relationship between duration and day length (P < 0.001, R2 = 0.384; see above for full details).
The end of the murmuration event
The frequency of birds descending en masse after murmurating was substantially and significantly higher, and the frequency of the murmuration dispersing and the birds flying away was lower, when a bird of prey was present compared to when a bird of prey was not present (chi square test for association: x2 = 33.600, d.f. = 2, P < 0.001). There was no difference in murmuration ending relative to temperature (one-way ANOVA F2,487 = 1.755, P = 0.174).
Using a citizen science approach, we gathered more than 3,000 records of starling murmuration behaviour from across the UK (and internationally) and covering two years. This is a far larger murmuration dataset both in terms of number of events and spatiotemporal scale than has previously been possible [13, 58]. Even after applying stringent filtering criteria, enforcing a minimum estimated size of 500 birds and requiring observers to see the end of the murmuration, there were >1,000 analysable murmuration observations.
Our analyses show that murmurations are not only widespread in terms of geographical scale, but also occur across a variety of habitats. This suggests that suitable roosting sites are widely distributed and are not confined to a few land-use types (unlike winter foraging, where starlings show a strong preference for permanent fields with abundant high-energy prey such as leatherjackets Tipulidae ). This implies that birds might travel some distance between feeding and roosting sites, an idea supported by individual birds regularly travelling ≥8 km between roost and feeding grounds  and exceptionally up to 50 km in favourable weather . Individuals can switch roost (and thus murmuration) site to take advantage of overnight opportunities near their chosen feeding ground , which leads to turnover in the individuals involved in murmurations at specific sites on consecutive nights. Using citizen science could have led to geographical spread being under-recorded (since people are most likely to observe murmurations that are most accessible), but we found no relationship between the number of records from a location and murmuration size such that we can discount “honeypot murmurations” (i.e. specific murmurations becoming natural attractions) becoming over-represented.
Our data allowed empirical testing of competing hypotheses for the purpose of murmuration behaviour in starlings. The “warmer together”, or thermal, hypothesis predicts that murmuration size and duration will be inversely correlated with temperature since the need for roost advertising should be greater in colder conditions. However, although low temperatures have previously been found to be important in promoting flocking behaviour [9, 35], we found only weak support for this hypothesis in the specific case of starling murmurations. Temperature was not a significant predictor of murmuration size although it was significantly negatively related to duration (colder evenings had longer murmurations). Duration was, however, more strongly related to day length than temperature (R2 = 0.384 versus R2 = 0.144), with longer murmurations tending to occur at the beginning and end of the season and shorter murmurations occurring in the darkest (and usually coldest) winter weeks.
Predator presence has been shown to affect shapes and waves within murmurations [13–14, 44]. Our data have allowed, for the first time, robust analysis of multiple predator species on murmuration size and duration. Our analyses support predators being important in murmuration behaviour (the “safer together” hypotheses) but determining cause and effect is problematic. Predators were present at ~30% of murmurations and their presence was significantly positively associated with both size and duration. However, larger and longer murmurations are undoubtedly far more visible to birds of prey than smaller, shorter, murmurations. Thus, while the dilution effect means individual predation risk is lower in a larger group (as found for other taxa, including marine insects, European minnows, and sand fiddler crabs [30, 33–34]), being part of a larger group may actually increase the per-group strike rate, as found in whirligig beetles Gyrinidae . The murmuration-predator relationship could, therefore, be as much (potentially more) driven by murmurations attracting predators–and larger murmurations being more attractive–than by predators causing starlings to murmurate. Interestingly, though, it was only birds of prey, rather than crows or gulls, that were linked to murmuration size and duration, suggesting that murmurating starlings might be able to distinguish predators from visually similar species (although note that the sample sizes were much smaller for crows and gulls, which might have had an effect). Additional evidence for the causal role of predator defence in murmuration behaviour is provided by the observation that only when predators were flying near, or actively engaging with, the murmuration (i.e. when they could pose a direct and immediate predation risk) did they affect murmuration size/duration. Our analyses also show that predator presence was related to starling behaviour at the end of murmuration: it was more likely that birds would descend en masse if a bird of prey was present. In contrast, starlings were more likely to disperse in the absence of predators. This terminal behaviour is, in many ways, more revealing than size/duration correlations when considering the role of predators: if murmurations were not in some way related to predator defence then terminal behaviour is unlikely to be correlated with predator presence.
A citizen science approach enabled a substantial dataset to be collected over a wider spatiotemporal scale than would have otherwise been possible. However, this approach is not without limitation. In addition to the location of records potentially reflecting the location of populations rather than, or in addition to, the focal phenomenon (see above), there is a risk of recording error among citizen scientists. Here, untrained observers were asked to give accurate estimates of bird numbers and, in common with much citizen science research, there was no way to validate the data. Inter-observer variation is likely; moreover because larger displays are likely to be harder to estimate accurately, recording error might not be independent of display size leading to heteroscedasticity. Any such patterns in the data would not be identifiable. One way to address this in future studies would be to ask respondents to submit a photograph or–better in the case of behaviour–video footage so data could be extracted by a single trained expert rather than a large number of untrained members of the public. This would increase researcher time but, where this is possible, would help improve data quality and consistency. This approach has been used previously to confirm species identification in citizen science surveys . Where this is not possible, use of an online quiz to assess observers’ skill so that this could be used to weight their data post-submission (as per ) might be helpful.
Despite the limitations outlined above, our findings suggest that the collective behaviour observed in starling murmurations is primarily an anti-predator adaptation rather than a way of attracting larger numbers of individuals to a roost for warmth. Suitable roosting sites attract large numbers of birds who would be vulnerable flying to the roost individually. Murmurating above the roosting site provides multiple advantages in terms of the dilution effect , increased vigilance leading to the detection effect [37–39] and predator confusion . This model of murmuration relies on having a critical mass of birds arriving at more-or-less the same time to initiate the murmuration and further study of the behaviour of starlings at the start of the murmuration (and indeed, just before the start of the murmuration) would be valuable in unravelling how this behaviour develops from a relatively few number of individuals into a spectacular collective behaviour comprising potentially tens of thousands of individuals.
The authors thank Adam Blackmore and Richard Rolfe from the University of Gloucestershire and Jon Kuldlick from the Royal Society of Biology for assistance with study promotion, real-time mapping, and publicity, and the thousands of members of the public who reported starling murmurations.
- Conceptualization: AEG.
- Data curation: AEG.
- Formal analysis: AEG WSC.
- Methodology: AEG AGH.
- Project administration: NL.
- Visualization: WSC.
- Writing – original draft: AEG AGH.
- Writing – review & editing: AEG AGH WSC.
- 1. Sumpter DJ. The principles of collective animal behaviour. Philosophical Transactions of the Royal Society of London B: Biological Sciences. 2006; 361(1465):5–22. pmid:16553306
- 2. Krause J, Godin JGJ. Shoal choice in the banded killifish (Fundulus diaphanus, Teleostei, Cyprinodontidae): effects of predation risk, fish size, species composition and size of shoals. Ethology. 1994; 98: 128–136.
- 3. Hart A, Jackson DE. U-turns on ant pheromone trails. Current Biol. 2006; 16: 42–43.
- 4. Thirgood S, Mosser A, Tham S, Hopcraft G, Mwangomo E, Mlengeya T, et al. Can parks protect migratory ungulates? The case of the Serengeti wildebeest. Anim Conserv. 2004; 7: 113–120.
- 5. Nicolis G, Prigogine I. Self-organization in nonequilibrium systems. New York, New York: Wiley; 1977.
- 6. Camazine S. Self-organization in biological systems. Princeton, New Jersey: Princeton University Press; 2003.
- 7. Sumpter DJ. Collective animal behavior. Princeton, New Jersey: Princeton University Press; 2010.
- 8. Deneubourg JL, Goss S. Collective patterns and decision-making. Ethol Ecol Evol. 1989; 1: 295–311.
- 9. Emlen JT. Flocking behaviour in birds. Auk. 1952; 69: 160–170.
- 10. King AJ, Sumpter DJ. Murmurations. Current Biol. 2012; 22: 112–114.
- 11. Ballerini M, Cabibbo N, Candelier R, Cavagna A, Cisbani E, Giardina I, et al. Empirical investigation of starling flocks: a benchmark study in collective animal behaviour. Anim Behav. 2008; 76: 201–215.
- 12. Cavagna A, Cimarelli A, Giardina I, Parisi G, Santagati R, Stefanini F, et al. Scale-free correlations in starling flocks. Proc Nat Acad Sci. 2010; 107: 11865–11870. pmid:20547832
- 13. Carere C, Montanino S, Moreschini F, Zoratto F, Chiarotti F, Santucci D, et al. Aerial flocking patterns of wintering starlings, Sturnus vulgaris, under different predation risk. Anim Behav. 2009; 77: 101–107.
- 14. Procaccini A, Orlandi A, Cavagna A, Giardina I, Zoratto F, Santucci D, et al. Propagating waves in starling, Sturnus vulgaris, flocks under predation. Anim Behav. 2011; 82: 759–765.