Four adult male big brown bats (E. fuscus), naïve to the laboratory flight room in Experiment 1, were trained to fly through an obstacle array (Figure 3). The bats were wild-caught (collecting permits #2012-34 and #2013-32, Rhode Island Department of Environmental Management). Bats were housed in a colony room (22–24°C, and 60–75% relative humidity) set to a 12-h/12-h reversed dark/light cycle in order to perform diurnal experiments on nocturnal bats. Bats ate enough larval Tenebrio molitor to maintain a mass between 15 and 18 g and had free access to vitamin-supplemented water (Poly-Vi-Sol, Enfamil) refreshed daily. The same four bats were flown in Experiments 1 and 2. The study was carried out in accordance with the recommendations of the Brown University Institutional Animal Care and Use Committee (IACUC). The protocol was approved by the Brown University IACUC.

All experiments were conducted in a customized flight room (Figure 3A) that was insulated acoustically and electrically from outside noise. The flight room was covered in acoustic foam (Sonex “One” panels, West General LLC, San Jose, CA, United States) to absorb 20–25 dB of the sound energy from echolocation calls. An infrared camera (Photon 320, FLIR, Billerica, MA, United States) was positioned to view the bats in flight so that experimenters in the adjacent electronic control room could monitor in real time while controlling the sound-recording system. The flight room measured about 8 m long × 4 m wide × 3 m high, and the active flight area was 7 m long. Rows of multiple black plastic chains (Petrites et al., 2009; Barchi et al., 2013; Knowles et al., 2015) were hung from crossbars near the ceiling of the flight volume to act as quantifiable surrogates for vegetation-like acoustic reflectors (Petrites et al., 2009; Yovel et al., 2009). These chains spanned nearly the entire floor-to-ceiling distance, and acted as a dense obstacle matrix through which the bats could not fly except for an open flight corridor.

In Experiment 1, the chains were configured as shown in Figure 3. Rows of chains were spaced 20 cm apart, and chains within a row were spaced 30 cm apart. Rows of chains were added to either side of the 100 cm central corridor to create a 70- and a 40-cm-wide corridor. In this experiment, flight paths down the central corridors were straight. Detailed analysis of the results from this experiment (Wheeler et al., 2016) shows that the proximity of the nearest obstacles had a significant effect on call timing as measured by IPI and the ratio of pulse intervals following and preceding each call (post-IPI/pre-IPI ratio).

In Experiment 2, chains throughout most of the flight room were spaced identically to the 40 cm configuration in Experiment 1. The chains forming the central corridor were hung from flexible, 1/4″ diameter PVC pipes, which were secured to the ceiling to maintain a consistent corridor width of 40 cm. Two movable joints in these pipes were spaced 1.8 m apart and 90 cm from the midpoint of the chain corridor. The shape of this corridor was manipulated by positioning each joint in one of three positions (central, left, and right), allowing for a total of nine curved corridor configurations (Figure 3). Stoppers were placed 40 cm from the ends of each corridor edge so that the central joints would slide only 40 cm to either side, and the central position was marked such that the experimenter could move both joints to any of the three positions quickly and accurately. This allowed for a change in corridor configuration between each trial/flight while avoiding intrusion from chains located in neighboring rows.

Bats were released at the front of the room (Figure 3A) through an opening in a net and flew across the room to land on the back wall. As a control, bats also flew in the open room with no chains. Bats were rewarded with a mealworm for each successful flight. A successful flight was defined as when the bat navigated the obstacle matrix in its entirety without deviating from the corridor or crashing into or landing on one of the chains. Each bat was allowed a maximum of 10 flights per day. All flights were performed in the dark to exclude visual cues from the bats’ perception. A single long-wavelength (> 650 nm) red light allowed experimenters to find bats after each trial. In Experiment 1, bats flew down a 100-, 70-, and 40-cm-wide straight corridor in the flight room. Figure 3B shows a typical echolocation call pattern from the 70-cm-wide corridor. Notice that the time between echolocation calls (IPIs) decreases as the bat traverses the obstacle array and approaches the wall. Since the wall is the bat’s landing target, and the distance to the wall decreases as the bat approaches it, pulse intervals long enough to allow echoes to return from the far wall become shorter. The amplitude of the calls shown in Figure 3B increases because the bat is approaching the microphone used to record echolocation sounds. Echolocation calls in this example are primarily grouped into doublets, or sonar sound groups having two calls within a group.

Experiment 2 used the same 40-cm-wide corridor from Experiment 1, but the path shape varied (Figure 3C). Experiment 2 was conducted in two broad phases: the fixed phase and the variable phase. Figure 3D shows example trials from both phases of Experiment 2. This experiment was designed this way in order to determine whether expectation could influence the timing of biosonar calls emitted by the bats. In the fixed phase, bats could expect the same corridor configuration from trial to trial; but in the variable phase, bats could not predict the next corridor configuration after each trial. The hypothesis was that SSIMS could differentiate predictable, expected (fixed) flights from unpredictable, unexpected (variable) flights based on biosonar sound timing patterns.

In the fixed phase, the configuration of the corridor did not change from trial to trial. Bats flew down the same configuration for up to 10 successful trials per day. First, the corridor was fixed at 40 cm wide, and bats flew down a straight corridor (shape 1) for 5 days. Then, bats were trained to fly down a partially curved corridor (shape 2) for 1 day before flying down a maximally curved corridor (shape 3) for 3 days. Next, bats were trained to fly in shape 5 for 1 day, then a full S-curve (shape 6) for several days. In each of these conditions, the corridor did not change from trial to trial within a day. A total of 470 trials or flights were successfully completed in the fixed phase, while 2.9% (n = 14) of all attempted flights (N = 484) were unsuccessful.

In the variable phase, the corridor configurations changed from trial to trial. In this phase, the bat was transferred from the colony room into the flight room in an acoustically insulated box so that the bat could not scan the room upon entering and before being released for flight. In preparation for releasing the bat into the chain array, the experimenter held the bat in one hand, and with the other hand, held a large piece of acoustic foam in front of the bat until the moment of release. This precaution was taken to prevent the bat from gathering pre-flight information about the configuration of the chains. The variable phase was conducted in two parts. In the first part, the corridor conditions were pseudo-randomly varied between configurations 1, 3, and 6. In the second part, the corridor configuration varied pseudo-randomly between all nine possible shapes (Figure 3C). The randomization did not allow more than one repeat of the same configuration from flight to flight. Both parts of the variable phase were combined for analysis. Fewer flights were conducted in configurations 2, 4, 5, 7, 8, and 9 than the other configurations, so no comparison was made for these configurations in the fixed versus variable phases. Shapes 4, 5, 7, 8, and 9 were only presented in the variable phase. All corridor configurations were new to the bats in Experiment 2 with the exception of shape 1. The flight path shapes were pseudo-randomly presented to decrease the probability that bats could predict which configuration they would be required to navigate. A total of 379 successful trials were recorded, and 2.8% (n = 11) of all attempted flights (N = 390) were unsuccessful.

Two electret ultrasonic microphones (MEMS SPM0404UD5, Knowles Electronics, Itasca, IL, United States) were mounted on custom preamplifier and high-pass filter boards that were located on the far wall directly in front of the corridor exit. To minimize backscatter from the wall, behind each microphone-preamplifier unit was a square, 20-cm foam baffle. These microphones detected sounds emitted by bats during flight (gray diamonds, Figure 2). Outputs of the 2 channels were digitized (one PCIe-424 with two accompanying HD192, all from MOTU, Cambridge, MA, United States) at a sampling rate of 192 kHz and stored as 16-bit wav files. All electronic wiring from microphones and the infrared camera were fed through a small opening to an adjacent control room. Experimenters in the control room began the sound recording when instructed by the experimenters in the flight room, and stopped recording when the bat landed on the far wall as visualized in real time by the infrared camera. Using this procedure resulted in sound recordings that included some echolocation calls from before the bat entered the chain array, and after the bat exited the chain array.

Sounds recorded by the microphone that had the best consistency of echolocation signal amplitude throughout each successful flight were analyzed for IPIs. The single-channel recordings used in these experiments were always from one of two microphones located in the center of the far wall. The higher amplitude recording was chosen from the two possible microphones. All wav files were truncated to include only in-flight sounds (Figure 3B), which removed the terminal buzz and any pre-flight calls. A custom program (MATLAB, MathWorks, Cambridge, MA, United States) identified and low-pass filtered frequencies below 35 kHz to exclude communication sounds, and high-pass filtered frequencies above 50 kHz to include only ultrasonic echolocation calls. The envelope of each sound was computed using the Hilbert transform. The program extracted IPIs from the time between peak amplitudes for successive sounds. To exclude any echoes or sounds emitted outside of the flight, only IPIs between 12 and 100 milliseconds were included in subsequent analysis.

Details of the original SSIMS algorithm can be found in a previous publication (Vargas-Irwin et al., 2015a), but are summarized in Figure 2. The SSIMS algorithm embeds point-process data into a high-dimensional pair-wise similarity space, and then projects the data into a more compact (low-dimensional) representation, which facilitates statistical analysis as well as data visualization while still capturing the relationship between individual data points. Originally designed for neural data analysis, SSIMS uses the spike train metric proposed by Victor and Purpura (1996) to quantify similarity (Victor, 2005). The method evaluates differences in spike trains based on how spikes in one train have to be shifted, deleted, or inserted to make it precisely match another train (Figures 2A,B). This cost function is similar to the “edit distance” or “Levenshtein distance” used to compare letter strings or genetic sequences. The algorithm can be directly applied to bat biosonar pulses by treating each pulse as an action potential. This is equivalent to treating each bat as a single neuron, and each flight down the corridor as a spike train recorded from a neuron in response to a stimulus.

The use of spike train metrics makes it possible to analyze long time periods (on the order of seconds) while preserving the fine temporal structure of call patterns. Note that only the timing of each pulse is considered and not the spectral characteristics. The cost of shifting a pulse by 20 ms was equivalent to inserting and deleting a pulse (1/q = 20 ms). This value for 1/q was arrived empirically – SSIMS achieved optimal classification performance when 1/q = 20 ms. Different values of q resulted in qualitatively similar results, but with slightly lower classification values (data not shown). The high-dimensional space defined by pair-wise distances is projected into a low-dimensional representation t-distributed stochastic neighbor embedding (t-SNE) (Figures 2B,C) (van der Maaten and Hinton, 2008). This method is well suited to this type of analysis because it is based on pair-wise similarity estimates and explicitly seeks to preserve the structure within local neighborhoods – in this case, clusters of flights with similar call patterns. A single point in the resulting SSIMS projection represents the timing of echolocation pulses recorded during a specific flight (Figure 2C). The distance between points represents the degree of similarity between the call patterns they represent. Two identical call patterns correspond to the same point in this space; the more different they are, the farther apart they lie in the SSIMS projection. Projecting the data into a low-dimensional space makes it easy to visualize the similarity between different experimental conditions in a way that also displays information about each individual flight. Dimensionality reduction also increases the density of the data (relative to the dimensionality of the space), making it easier to characterize data distributions and perform pattern recognition (i.e., classification).

In Experiment 1, SSIMS analyzed sequences of pulses from individual flights, and the resulting SSIMS plots (Figures 2C, 4B) allowed clusters of points (call sequences or flights) to be classified as belonging to the open room, and the 100-, 70-, and 40-cm-wide corridors in the obstacle array. Calls from all bats were projected onto a single common similarity space (dotted line in Figure 4B) such that the representation of each call was calculated using similarity estimates across the entire set of recordings. Raster plots in Figure 4A display data from one example bat to illustrate that bats emitted more calls with shorter IPIs (histograms on the right) as obstacle array clutter increased and flight corridor width decreased. Histograms in Figure 4A show that the distribution of IPIs shifted toward shorter time intervals with decreasing corridor width, indicating that more calls were made in groups.

We quantified the differences between call patterns using a nearest-neighbor (NN) classifier implemented using leave-one-out cross validation. The NN classifier was chosen because it is designed to capture the local structure of complex data patterns, making it well suited to interpret data processed using SSIMS (which is designed to preserve exactly this kind of relationship). Classification was performed using 10D SSIMS projections. Asymptotic performance was achieved with between 2 and 6 dimensions for different bats (see Supplementary Figure 1). NN classification correctly assigned flights (N = 839) with 70–89% accuracy depending on the individual bat (Figure 4B).

In Experiment 1, all bats decreased the mean IPI as more obstacles were added to the flight room (see Table 1). Comparing the open room flights to the 40-cm-wide corridor flights, bats decreased the length of the mean IPI by 29–62% (a decrease of 13–44 ms), depending on the bat.

The recorded call sequences revealed that two pairs of bats – Bats 1 and 3, and Bats 2 and 4 – were more alike than to bats in the other pair (Figure 5). The differences in the call patterns employed by different bats were evaluated by comparing distributions of pair-wise SSIMS distances over all flights between and within bats. Statistical significance was assessed using a Kruskal–Wallis (KW) test, while the effect size was summarized using a ratio of SSIMS distance medians (distance between biosonar sound sequences – points – in similarity space) (B/W ratio). For example, a B/W ratio equal to 2 indicates that the normalized median distance between call patterns across different bats was twice of that observed for call patterns for the same bat. Figure 5 shows that across all bats, the B/W ratio was 1.27 for Experiment 1 and 2.70 for Experiment 2. Although both values are statistically significant (KW p < 0.0001), the more than twofold difference in B/W ratio suggests an increase in bat-specific call pattern differences with increasing task difficulty.

To test whether within bat variability and between bat variability were significantly different between the two experiments, we performed a KW test, and found that normalized within bat distances for Experiment 2 were significantly smaller than in Experiment 1 (KW p < 0.0001). This result suggested that bats used more stereotyped call patterns in Experiment 2, which is likely because of the consistent 40 cm corridor width in this experiment. Additionally, the normalized between bat differences were significantly greater for Experiment 2 (KW p < 0.0001). This provided statistical support of our assertion that there were increased bat-specific differences in Experiment 2.

Pairs of bats were also analyzed for individual similarities, and only one pair was found not to be significantly different (KW p > 0.05): bat 2 and bat 4 in Experiment 1, suggesting that this pair of bats emitted very similar sound patterns for this experiment. All other bat pairs showed statistically significant differences for both Experiment 1 and Experiment 2 (KW p < 0.0001) (see Supplementary Figure 2).

Tables 2–4 address the differences in mean IPI, mean number of calls per second, and the average flight velocities in the 40-cm-wide straight corridor during Experiment 1 versus the fixed and variable phases of Experiment 2. The columns in Tables 2–4 are organized from left to right in the chronological order in which the experiments were conducted. The trends show that bats emitted calls having progressively longer mean IPIs (Table 2), fewer calls per second (Table 3), and slower flight speeds (Table 4) in the 40-cm-wide corridor as the experiments progressed. Between the straight 40 cm corridor flights in Experiment 1 and the variable phase of Experiment 2, bats decreased their average flight velocities by 11–35% depending on the individual bat (see Table 4). This could explain how bats navigated through more complex, unpredictable path shapes using fewer calls per second and longer mean IPIs. However, gross differences in mean IPI, calls/sec, and flight velocity between the fixed and variable phases of Experiment 2 were not readily apparent (see Tables 2–4). It could therefore not be determined from these parameters alone whether expectation had any effect on bat biosonar call patterning.

To evaluate whether expectation can impact the timing of active sonar probing by bats, SSIMS was used to assess call patterning in two phases of Experiment 2. In order to compensate for the greater variability in call patterns across bats for Experiment 2, (see Supplementary Figure 2) individual similarity spaces were generated for each subject such that a call pattern was represented in terms of the relationship to calls emitted by the same bat. In this experiment, the same four bats from Experiment 1 flew down 40-cm-wide corridors of varying shapes (Figure 3). In the fixed phase of Experiment 2, the bats flew down corridors having the same configuration from trial to trial. In the variable phase of Experiment 2, the corridor configuration was pseudo-randomly changed from trial to trial (see Figure 3D and see section “Materials and Methods”).

To determine whether the bats’ expectation of the corridor shape had any effect on echolocation behavior, the flights down the straight and S-curved corridors (shapes 1, 3, and 6, see Figure 3C) in the fixed phase and variable phase were compared. We evaluated differences between the fixed and variable conditions using a NN classifier. Fixed versus variable classification results were inconsistent for the S-curve shapes (exceeding chance in only 2 out of 8 comparisons), but were consistently above expected chance levels in the straight corridor configuration for all bats. Figure 6 shows that (1) during fixed straight (shape 1) flights bats produced stereotyped call patterns different from the curved corridor flights; and (2) that flights down the straight corridor during the variable phase were more like curved corridor flights than fixed phase straight corridor flights. In other words, unexpected straight flights were more like curved flights than expected straight flights.

The absence of clustering for all variable flights from all fixed flights (stars and dots, respectively, in Figure 6) suggests that expectation, or the ability of the bat to predict the flight path shape, did not determine overall temporal patterns emitted in the chain array. Rather, SSIMS plotted clusters of straight flights in the fixed phase apart from all other flights. Our results showed that the bats altered their call patterns in a similar way to address (1) more complex flight paths and (2) unpredictable flight paths. This finding suggests that bats adopted a more conservative (careful) strategy whenever a more challenging or unknown situation was presented.

To identify which features the SSIMS analysis may have detected to discriminate fixed versus variable straight flights but not fixed versus variable curved flights, the post-IPI to pre-IPI ratios and IPI distributions were plotted (Figure 7). In the straight corridor configuration (red), bats emitted a greater proportion of calls having a post-IPI/pre-IPI ratio between 1.5 and 2. Arrows in Figure 7B show where this is visible on the histograms. This ratio value indicates that the interval of time following these calls was roughly 1.5–2 times longer than the interval preceding the calls, which means that they likely represent the last call in a sonar sound group (Kothari et al., 2014; Wheeler et al., 2016). Figure 7 indicates that the bats in this experiment used more sonar sound groups in unpredictable learned surroundings (variable, straight path) than predictable learned surroundings (fixed, straight path). The inability of SSIMS to reliably categorize fixed versus variable curved flights (Figure 6), and the absence of visible differences in the IPI ratios for these flights (Figures 7C,D) provides evidence that memory plays a more important role than expectation in biosonar sound patterns emitted by big brown bats.

The SSIMS method ignores sonar sound groups entirely, and provides a clear visualization of the magnitude of the difference between call patterning in the fixed and variable flights as compared to the IPI ratio histograms (compare Figures 6, 7).

The advantage of using SSIMS to analyze temporal patterning of bat echolocation sounds across a flight is that the effect of an environment on the timing of vocalizations can be visualized easily. The major limitations of this method for such analyses are as follows: (1) other aspects of the echolocation signal such as duration, frequency spectra, and beamform are not incorporated into the similarity estimates; and (2) finer temporal details about sonar sound group production, while included in the analysis, are not explicitly reported on the SSIMS space projection.

SSIMS representations are useful in determining the degree to which echolocation call sequences are similar or different. However, SSIMS alone cannot pinpoint individual timing features such as sonar sound groups, flight velocity, IPIs, and IPI ratios. SSIMS analysis also ignores all other adaptable features of biosonar pulses, including frequency and amplitude.

SSIMS can provide a valuable initial screening such that further examination of individual features that may contribute to SSIMS distances is more directed. The outline of the previous analysis of Experiment 1 (see section “Results”) exemplifies the arduous process of finding meaningful pulse timing differences using other methods. Using SSIMS, biosonar sound sequences from an entire flight event can be represented in a single point, and easily compared to other flights on the same plot. This ease of visualization is advantageous because it reveals which comparisons of individual timing features might be most interesting.

The advantage of SSIMS over other methods such as linear mixed effects models (Falk et al., 2015; Wheeler et al., 2016), Markov Chain analysis (Petrites et al., 2009; Kloepper et al., 2014), and others, that it requires no boundaries or binning and very few assumptions. It preserves fine temporal structures (millisecond scale) even over relatively long time scales (second or minute scale).

Perhaps the greatest advantage of SSIMS is that the resulting scatter plots in SSIMS space are easy to interpret visually. For example, the SSIMS plot in Figure 6 clearly shows that expected straight corridor flights are different from unexpected flights in the same path. The Figure 7B histogram shows that the post- to pre-IPI ratio between 1.5 and 2 increases when the straight corridor flights are unexpected. Figure 7B does not illustrate differences in IPI ratio distributions for the curved flight paths. While Figure 7 shows an example of what types of differences SSIMS is detecting, the degree to which pulse patterning changes is more easily visualized in the similarity plot.

Past research has diligently characterized individual echolocation sound parameters (duration, mean IPI, repetition rate, starting and ending FM sweep frequencies, etc.,) that adapt to obstacle avoidance or target-detection conditions. Take, for example, sonar sound groups. Sonar sound groups, or strobe groups, are loosely described as a group of sounds having short time intervals between calls within the group, and longer time intervals delineating the group (Moss et al., 2011; Kothari et al., 2014). More specifically, these groups have been mathematically defined as having at least a 20% longer interval between groups (“island criterion”), and no more than 5% within-group variability (“stability criterion”) (Kothari et al., 2014). While useful for some classification purposes, these limitations are arbitrary, and have made it difficult to compare pulse interval patterning across experiments. This is partly because IPI values alternate between long and short to define sound groups, and these absolute time measurements appear to be scaled proportionally to the environment (Moss et al., 2006, 2011; Petrites et al., 2009; Aytekin et al., 2010; Kothari et al., 2014; Wheeler et al., 2016). This proportionality extends over a wide enough range in big brown bats that the interval categories of short versus long overlap. Thus, using set time interval sizes or too narrow a range of size ratios to define sound groups does not fully capture the bat’s behavior. In contrast, SSIMS can determine whether a particular environment or condition resulted in an observed change in echolocation call timing, without a priori restrictions.