Echolocating whales use ultra-fast echo-kinetic responses to track evasive prey

Visual predators rely on fast-acting optokinetic reflexes to track and capture agile prey. Most toothed whales, however, rely on echolocation for hunting and have converged on biosonar clicking rates reaching 500/s during prey pursuits. If echoes are processed on a click-by-click basis, as assumed, neural responses 100x faster than those in vision are required to keep pace with this information flow. Using high-resolution bio-logging of wild predator-prey interactions we show that toothed whales adjust clicking rates to track prey movement within 50-200ms of prey escape responses. Hypothesising that these stereotyped biosonar adjustments are elicited by sudden prey accelerations, we measured echo-kinetic responses from trained harbour porpoises to a moving target and found similar latencies. High biosonar sampling rates are, therefore, not supported by extreme speeds of neural processing and muscular responses. Instead, the neuro-kinetic response times in echolocation are similar to those of tracking reflexes in vision, suggesting a common neural underpinning.


Introduction 25
Response speed critically determines the outcome of interactions between mobile prey and pursuit 26 predators. Prey must react rapidly to survive while predators must counter evasive prey movements 27 quickly to gain sustenance. The fitness implications of these interactions have led to an evolutionary 28 escalation of response times with the fastest-responding individuals being the most likely to survive 29 and reproduce (Dawkins and Krebs, 1979). However, sensory and motor requirements are 30 asymmetric for predators and prey. Responding to close-approaching predators, prey may trade 31 accuracy for speed, relying on imprecise ballistic motor actions triggered by strong sensory cues 32 that require little neural processing (Domenici and Batty, 1997; Turesson et al., 2009). This has led 33 to extremely fast responses employing short efferent pathways linking sensors directly to muscles, 34 such as the Mauthner-cell-mediated C-start response of teleost fish to fluid motion from oncoming 35 predators (Eaton and Hackett, 1984). C-start responses are characterised by sudden accelerations 36 (Domenici and Blake, 1997) and unpredictable trajectories (Moore and Biewener, 2015), with 37 response latencies as low as 5-10 ms (Eaton et al., 1977). In contrast, predators must sacrifice speed 38 for accuracy, typically requiring greater sensory resolution and motor-planning capabilities to track 39 and successfully pursue evasive prey. The increased processing needed to locate prey in complex 40 natural scenes, along with the typically larger body size of predators, inevitably result in slower 41 movement responses compared to prey, and these are often partly offset by ingenious capture tactics 42 function to the optokinetic reflexes in vision, leading us to hypothesise that echo information must 76 guide two inter-related control loops during hunting (Fig. 1): a kinematic response loop controls the 77 heading, posture and locomotor rate so as to intercept prey, while a sensor-motor response loop 78 maintains attention fixed on the moving target by continual adjustment of the clicking rate. On that 79 basis, we predict that sudden prey movements during close approaches should provoke tightly-80 reported clicking rate dynamics of wild harbour porpoises during buzzes that appeared to be 85 associated with prey movements but offered no analysis. The only reported study of predator 86 responses during echolocation buzzes is for wild bats approaching suspended prey that were moved 87 by hand. Rather than adjust clicking rates, these bats aborted buzzes some 100 ms after strong target 88 movements (Geberl et al., 2015) perhaps indicative of a neural processing delay, but the decision to 89 end a buzz may well result from different neural processing than that involved in prey tracking.  Crucially, these tags sample the sensory scene at exactly the same rate as it is acquired by the 101 animal, i.e., the rate at which clicks are produced. 102 Here we used high-resolution DTAGs on two species of echolocating toothed whales living in very 108 different habitats to study biosonar responses to prey movements in the wild. Specifically, we tested 109 the hypotheses that prey movements trigger neuro-motor feedback during buzzes, and that this 110 feedback operates at the extreme speeds needed to keep pace with the high clicking rate in buzzes. 111 We show that both species make stereotyped biosonar adjustments when prey attempt to escape 112 during close approaches, but the apparent latency of these echo-kinetic responses is much longer 113 than the inter-click-intervals (ICI) in buzzes. Given the evolutionary importance of such a feedback 114 system, we further hypothesized that it would be stimulated by any rapidly moving target during a 115 close approach, facilitating controlled studies of biosonar responses. To test this, we trained harbour 116 porpoises to approach a moveable target while wearing a biologging tag, enabling direct 117 measurement of biosonar-mediated sensory and kinematic response latencies as a function of target 118 movement.

Biosonar responses to prey movement in free-ranging animals 121
Echograms visualising the acoustic scene during prey capture buzzes in wild harbour porpoises (Pp) 122 (Phocoena phocoena) and Blainville's beaked whales (Mesoplodon densirostris) (Md) frequently 123 show evidence of evasive prey that launch sudden escape attempts as the predator approaches ( Fig.  124 2a-b). Prey that accelerate away from the predator can quickly move beyond the acoustic depth of 125 field (i.e., the buzz inter-click interval, ICI, times one half of the sound speed) requiring a rapid 126 increase in ICI by the predator to avoid ambiguous target ranging. The ICIs used in buzzes when 127 targeting evasive prey show dynamics that seem to correspond to changes in prey-range ( Fig. 2A-128 B), suggesting a tight sensor-motor feedback loop. To verify that these ICI changes are linked to 129 prey escapes, we plotted the proportion of outward depth-of-field adjustments (i.e., increasing ICIs)  Energetic prey targeted by harbour porpoises can make multiple escape attempts within a buzz 146 providing an opportunity to examine ICI responses to repeated cues (Fig. 3A). Plotting prey range 147 against the unambiguous depth of field (equivalent to an input-output plot of a control system) 148 revealed distinctive counter-clockwise loops due to the ICI response latency (Fig. 3B). To estimate 149 this latency, we advanced the ICI time series in 5 ms steps until the areas within the loops were 150 minimised. An advance of 90 ms collapsed most of the loops in the delay compensation plot (Fig.  151 3C) suggesting that successive prey movements elicit responses with self-similar latency. In with prey escape movements occurred in some cases (Fig. 3A), over-compensation was more 154 typical ( Fig. 2A) leading to depths of field that extended well beyond prey range when prey 155 attempted to escape. However, echoes from schools of prey can have dynamics which greatly 156 exceed the speed of single prey requiring extensive outward ICI adjustments to avoid range 157 ambiguity (Fig. 4). 158 close prey approaches (buzzes) show prey escape attempts (white arrows) and dynamics in inter-160 click-interval (ICI, yellow arrows) that appear to be causally related. These plots show echo 161 strength (colour-coded by Signal-to-Noise Ratio, SNR, in decibels (dB) as a function of distance 162 from the predator (vertical axis) and time with respect to the start of the buzz (horizontal axis). See 163 Supplementary Material for a guide to interpreting these plots. Boxplots below show the proportion 164 of positive changes in ICI (δICI, i.e., an expansion of the acoustic depth of field) as a function of 165 time before/after the first prey response in 74 buzzes by 6 harbour porpoises (C) and 36 buzzes by 7 166 Blainville's beaked whales (D). Due to differences in click repetition rate, 50 ms bins are used for 167 harbour porpoise and 100 ms bins for Blainville's beaked whales. Light grey boxes show the 168 changes in ICI for randomised control data. * indicate bins in which >95% of observed proportions 169 exceeded the randomised proportion of positive δICI for control data. 170 prey. The prey makes a series of escape attempts, marked with white arrows, and the ICI (the first 172 yellow-red line above 1 m distance, here converted by the echogram scaling to the acoustic depth of 173 field in metres, i.e., ICI x sound speed / 2) varies cyclically with each prey movement. (B-C): Plots 174 of prey range versus acoustic depth of field for the same echogram with the times of prey escapes 175 indicated by black dots. Anti-clock-wise loops in (B) indicate that ICI lags behind prey movements. 176 Advancing the ICI by 90 ms flattens the loops (C) indicating that the ICI is consistently delayed by 177 about this amount with respect to prey movements throughout the buzz. 178 to rapid changes in overall school echo cross-section rather than the movements of the closest prey 181 (for example at 9.5-10s when the school moves away but the target fish stays at approximately the 182 same range). 183

Echolocation responses of trained animals to target movements 184
In wild predator-prey interactions, predators frequently strike at prey at about the same time that the 185 prey responds, raising the possibility that ICI adjustments are timed based on predator strikes (i.e., 186 implying an anticipatory or feed-forward control scheme) rather than based on prey movement. To 187 exclude this potential confound, we designed an experiment in which captive harbour porpoise 188 approached a target that could be moved suddenly. Echograms during target approaches (Fig. 5) 189 show that this experimental design successfully replicated the sharp speed changes of prey in wild 190 porpoise chases and thus provide reliable cues for timing biosonar and kinematic responses. showed similar latency as for wild porpoises with response times of 50-100 ms for Freja and 100-202 150 ms for Sif (Fig. 6 a-b). Whole body kinematic responses, inferred from the differential of on-203 animal acceleration measurements (jerk) were detectable with latencies of 0-50 ms for the two 204 porpoises although strong jerk responses were more clearly evident at 50-100 ms (Fig. 6 C-D). exceeded the randomised proportion of positive δICI for control data. 211 As the target was pulled at varying speeds by hand, the trials could be ranked according to the 212 magnitude of the initial target movement cue. Using root-mean-square (RMS) target acceleration as 213 a proxy for this movement cue, we plotted ICI and RMS jerk as a function of cue magnitude (Fig.  214   7). Both signals showed a latency that tended to decrease with increasing target acceleration. Using 215 a 5-ms threshold (corresponding to a 3.75 m depth of field) to detect strong outward ICI 216 adjustments, the latency of these large-scale responses was inversely correlated with RMS target 217 acceleration (r 2 =0.3, p<0.0001). Kinematic response latencies, using a 300 m/s 3 threshold on RMS 218 jerk to detect the onset of strong responses, were similarly inversely correlated with RMS target 219 acceleration (r 2 =0.2, p<0.001), albeit with more variability. This leads us to propose that acute sensor-motor feedback during buzzes is a fundamental feature of 257 toothed whale echolocation that has enabled hunting of nutritionally-valuable muscular but reactive 258

prey. 259
The measured response latencies (Fig 2, 6) show that tight sensor-motor feedback in echolocation 260 buzzes can be achieved without extreme neural processing speeds. Echo-kinetic response latencies 261 of 50-200 ms in our study are comparable to optokinetic reflexes in primate vision (Land, 1999;262 Kirchner and Thorpe, 2006) and to vocal response latencies to passive acoustic cues in dolphins than 20x longer than the 2.5-3.5-ms ICI during buzzes, demonstrating that echolocating whales 265 process and respond to echo information during prey approaches much more slowly than they 266 acquire it. Put another way, the maximum information bandwidth (i.e., 1/(2xICI) Hz, by Nyquist 267 theorem) is some 40x greater than the maximum control bandwidth that can be achieved given the 268 response delay, i.e., approximately 1/(4x response latency) Hz (Astrom, 1997). This implies 269 strongly that echo processing and control decisions during buzzes are decoupled from click rate 270 rather than occurring on a click-synchronous basis as widely assumed (Au, 1993). This conjecture is 271 consistent with the proposed processing mode of click packets produced during long-range The high clicking rate in buzzes enables rapid detection of prey responses but may provide other 285 benefits, when combined with appropriate feedback mechanisms, such as: i) signal-to-noise ratio 286 improvement of weak echoes by integration over multiple clicks (i.e., using integral control); ii) 287 speed-based processing of echoic scenes to predict target motion (via differential control), and (iii) 288 detection of modulations in echo level (e.g., due to prey tail-beats, Wisniewska et al, 2016) that may 289 be the earliest cues of prey responses, while avoiding aliasing in this discrete-time sensor. The high 290 click rate in buzzes effectively forms a temporal fovea, akin to the spatial fovea in many visual 291 predators, matched to the burst movement rates of the relatively small prey targeted by most toothed 292 whales. This ensures the observability of prey behaviour and enables control tactics that counter 293 unpredictable prey movement. 294 that may be employed. When the change in target range is dominated by predator movements, a 296 gradual upward adjustment of clicking rate is sufficient to track the changing spatial relationship of 297 predator and prey. In comparison, unexpected rapid prey movements often provoke large 298 adjustments in the biosonar rate in which the acoustic depth of field is rapidly expanded at the 299 expense of temporal resolution. This suggests a layered control with slow tracking during stalking, 300 but saccade-like ballistic increases in clicking rate during chases. Such rapid control actions may 301 also accommodate the dynamics of schooling prey, which can quickly switch between cohesion 302 when being pursued and dispersion when escaping (Couzin and Krause, 2003). 303 While information bandwidths in echolocation are likely linked to prey dynamics, the control 304 bandwidths (i.e. the speed with which the system can respond to changes) may be more matched to 305 the size and manoeuvrability of the predator given that size influences both the rotational inertia of 306 the body and the length, and therefore contraction rate, of muscles (Domenici, 2001). We 307 hypothesised the existence of two control loops in echolocation-guided hunting, controlling, 308 respectively, the acoustic depth of field and the swimming kinematics (Fig. 1). We have been able to 309 demonstrate the biosonar feedback loop in both wild and controlled settings, but full-body 310 kinematic responses to prey movement are confounded in wild predator-prey interactions by the 311 predator's own striking actions. However, our controlled studies demonstrate that porpoises make an 312 accelerative response to target movement with latency roughly comparable to the biosonar response 313 and it seems very likely that wild animals would have similar kinematic responses. Therefore, our 314 finding of longer biosonar response latencies in Blainville's beaked whales, which are three times 315 the size of harbour porpoises, suggests that control bandwidths scale inversely with predator size. In 316 effect, selection pressure on higher control bandwidths may be opposed by the increasing energetic 317 cost of fast movements in large animals. 318 Thus, despite the overt differences between echolocation and vision, the response bandwidths and suggests that extreme sensory sampling rates, guiding fast echo-kinetic responses, may have been a 324 critical development, parallel to optokinetic reflexes in visual predators, enabling echolocation to be 325 used to hunt agile prey, as opposed to just navigation and prey search. Our results, therefore, reveal 326 strongly convergent neural sensor-motor feedback loops between vision and echolocation that are 327 key for sensing dynamic spatial relationships with small prey. The non-invasive experimental real-world problems opening the way for a deeper understanding of ecological drivers on sensor 330 performance. 331

Materials and methods 332
Echolocation in free-ranging animals 333 unclear prey echo traces or with substantial interference (e.g., echoes from the sea-surface, sea-floor 360 or other organisms) were rejected. The remaining echograms were examined for indications of prey 361 escape attempts. These appear as sudden changes in the slope of prey echo traces (Fig. 2a)  362 reflecting a step change in the closing speed between predator and prey as the prey accelerates away 363 (Wisniewska et al., 2016). As prey reactions are typically fast, the onset time of the slope change in 364 the echo trace is usually well-defined. The first such reaction time in each buzz echogram was 365 selected manually with ~10 ms accuracy, and traces with unclear or gradual slope changes were 366 rejected. Potential biosonar adjustments to these prey movements were quantified by the proportion 367 of positive changes in ICI (suggesting an outwards adjustment of the depth of field) in 50-ms (Pp) 368 or 100-ms (Md) bins, spanning from 500 ms before to 500 ms after each prey response time. As ICI 369 varies continuously throughout buzzes, these bin sizes were chosen as a compromise between 370 temporal resolution and rejection of noise from routine ICI variations. The wider bin size for Md 371 reflects the longer ICIs produced during buzzes by these larger animals. 372 To determine the probability of chance associations between target movement and ICI changes a 373 bootstrap method was applied for each species. The same biosonar response metric was computed 374 1000 times for randomly-selected pairs of buzzes in which the prey movement time from one buzz 375 was applied to another buzz. Specifically, the time elapsed between the start of the buzz and the 376 onset of the prey response in the first buzz of each pair was added to the start of buzz time in the 377 second buzz to give a mock prey move time from which to reference the analysis time bins. A 378 significant deviation from chance was concluded for each time bin in which >95% of the observed 379 proportions exceeded the randomised proportions. The sphere contained an embedded hydrophone (flat (±2 dB) frequency response from 1 to 160 385 kHz) and two-axis accelerometer (flat (+0/-3 dB) frequency response from 0 to 2 kHz, axes oriented 386 horizontally), and was suspended in the water via a 0.8-mm nylon string to a depth of 387 approximately 1.5 m. A 1.2-mm-diameter screened cable carrying the accelerometer and 388 hydrophone signals from the target, was loosely attached to the nylon string and connected out of 389 the water to a three-channel synchronous 16-bit analog-to-digital convertor sampling at 500 kHz 390 (National Instruments, Austin, TX). A second nylon line running horizontally from the sphere to the 391 sound and movement recording tag attached 5 cm behind the blow-hole with silicone suction cups. 393 This tag contains a single hydrophone sampled at 576 kHz (flat (±2 dB) frequency response from 1 394 to 150 kHz, clipping level of 175 dB re 1μPa) together with a tri-axial accelerometer sampled 395 synchronously at 200 Hz and a tri-axial magnetometer and pressure sensor sampled at 50 Hz. 396 For each session, one of the two porpoises was introduced into the pool and stationed approximately 397 8 m from the target until given a cue to perform the target interception task. If the animal 398 intercepted the target by touching it, it was bridged with a whistle and received a fish reward upon 399 returning to station. In randomly selected trials, the target was moved manually approximately 30 400 cm by pulling vigorously on the horizontal line when the porpoise approached within one body 401 length. The line was held slack prior to this to limit any early anticipatory target movement. Target 402 movement was selected pseudo-randomly for each trial between fast, slow, and no movement, with 403 a maximum run of two equal conditions. Up to 20 trials were performed with each animal per day 404 for a total of 150 trials over 4 days for the two animals. 405 Echolocation clicks were detected in the animal-attached DTAG recordings using a supervised 406 detector. The time offset between the tag and the National Instruments recordings was then 407 determined for each trial by matching click sequences between the tag and the target hydrophones 408 (max. timing error due to acoustic propagation of ~1 ms). Echograms were then assembled from the 409 tag data as described above. Sudden changes in the closing speed between the target and the 410 porpoise due to rapid movement of the target generated the same distinctive slope changes in echo 411 traces as observed in wild predator-prey interactions. To maximise timing accuracy, trials were only 412 selected for analysis if the target attained a speed greater than the porpoise's approach speed 413 (approx., 1 m/s) resulting in a V-shaped echo trace. The apex of the V was then taken as the 414 reference time for calculating response latencies. The target acceleration (as measured by the 415 embedded accelerometer) began ~100 ms before this due to tightening of the line and rotation of the 416 spherical target to align the tie point with the pull direction. Given the thin line and spherical target, 417 neither of these movements generated significant echo signatures and the porpoise was therefore 418 unlikely to detect the target motion until it is underway. We accordingly view the apex of the V in 419 the echogram as a close indicator of the time at which the porpoise was first able to detect the target 420 movement. As the target movement varied in each trial, the root-mean-squared target acceleration 421 was calculated as a proxy for target motion. This was computed from the accelerometer embedded 422 in the target by removing the fixed gravity component from each axis and then summing the 423 squared signals from both axes over the 500 ms following the first acceleration transient.
Biosonar responses to target movement were quantified as for the wild toothed whales using the 425 proportion of positive ICI changes in 50-ms intervals. Locomotor responses to target movement 426 were assessed from the norm-jerk (Ydesen et al., 2014) calculated from the DTAG accelerometer 427 data sampled at 200 Hz. To assess the probability of chance associations between target movement 428 and ICI changes, the same bootstrap method was applied as for wild harbour porpoise and beaked 429 whales (see above) i.e. randomly-selecting 1000 pairs of trials and applying the elapsed time 430 between buzz start and target movement from one trial to the buzz of another trial. 431