Sound Localization of World and Head-Centered Space in Ferrets

Abstract

The location of sounds can be described in multiple coordinate systems that are defined relative to ourselves, or the world around us. Evidence from neural recordings in animals point toward the existence of both head-centered and world-centered representations of sound location in the brain; however, it is unclear whether such neural representations have perceptual correlates in the sound localization abilities of nonhuman listeners. Here, we establish novel behavioral tests to determine the coordinate systems in which ferrets can localize sounds. We found that ferrets could learn to discriminate between sound locations that were fixed in either world-centered or head-centered space, across wide variations in sound location in the alternative coordinate system. Using probe sounds to assess broader generalization of spatial hearing, we demonstrated that in both head and world-centered tasks, animals used continuous maps of auditory space to guide behavior. Single trial responses of individual animals were sufficiently informative that we could then model sound localization using speaker position in specific coordinate systems and accurately predict ferrets' actions in held-out data. Our results demonstrate that ferrets, an animal model in which neurons are known to be tuned to sound location in egocentric and allocentric reference frames, can also localize sounds in multiple head and world-centered spaces.SIGNIFICANCE STATEMENT Humans can describe the location of sounds either relative to themselves, or in the world, independent of their momentary position. These different spaces are also represented in the activity of neurons in animals, but it is not clear whether nonhuman listeners also perceive both head and world-centered sound location. Here, we designed behavioral tasks in which ferrets discriminated between sounds using their position in the world, or relative to the head. Subjects learnt to solve both problems and generalized sound location in each space when presented with infrequent probe sounds. These findings reveal a perceptual correlate of neural sensitivity previously observed in the ferret brain and establish that, like humans, ferrets can access an auditory map of their local environment.

Keywords: auditory; behavior; ferret; hearing; modeling; sound localization.

Figure 1.

Experimental setup. A, Task arena in which ferrets approached the center of a speaker ring to initiate presentation of a 250-ms broadband noise burst from one of twelve speakers. Values indicate speaker angle relative to the arena (world) coordinate system. B, Infrared images showing one ferret at the center spout as platform angle is varied. Response ports around the arena periphery contain IR emitters and are thus highlighted here, but were not illuminated within the visible spectrum during testing. Values show platform angles (and thus head directions) within the world. C, Dissociation of speaker position in head and world-centered space that occurs with platform rotation.

Figure 2.

Task design. A, World-centered task in which subjects approached the center of a speaker ring to initiate presentation of a 250-ms broadband noise burst from a speaker either at the North (N) or South (S) of the arena, or later from probe speakers around the remainder of the arena (gray) by responding at East (E) or West (W) response ports. Arrows show the position of correct responses, which remained constant as the central platform was rotated. F numbers (F1701, etc.) refer to ferrets trained in the North-South discrimination. B, Dissociation of speaker angle relative to the head and speaker angle in the world as the platform angle was rotated at 30° intervals. In addition to test sounds, probe sounds were also presented from untrained speaker locations on a random subset (10%) of trials. C, Variant of the task for an additional ferret (F1902) in which we altered the world-centered locations to be associated with each response. D, Head-centered task in which subjects discriminate 250-ms broadband noise bursts from a speaker either at the Front (F) or Back (B) of the head, by visiting response ports either to the Left (L) or Right (R) of the head. Arrows show the direction for correct responses. E, Dissociation of speaker angles in head and world-centered space as platform angle was rotated. F, Variant of the task in which we altered the head-centered locations associated with each response for one ferret.

Figure 3.

World-centered task performance. A, Performance discriminating sounds at trained locations for each ferret as a function of platform angle (n = 400 trials per platform angle). Data shown as mean percent correct across bootstrap resampling (n = 100 iterations). Dashed lines show chance performance (50%). Insets show the training configurations [either North vs South (F1701, F1703, F1811), or South-East vs North-West (F1902)]. B, Performance measured only from the first ten trials of sessions after platform rotation (n = 50 trials per platform angle). Data shown as in A. C, Performance on sessions immediately before and after swapping the speakers at test locations. Observed data (full lines) is compared with predictions made if animals were responding based on speaker identity (dashed lines) or sound location (dotted gray lines). Predictions based on speaker identity were made by subtracting performance before swap from 100%. Predictions based on sound location were simply the same performance before and after swap.

Figure 4.

World-centered response probability. A, B, Probability of responding at the West response port for test and probe sounds as a function of sound angle in the world (A), or relative to the head (B; n = 36 trials per angle) in ferrets trained in the world-centered task. C, Response probability as a joint function of head and world-centered sound angle (n = 432 trials over 144 locations).

Figure 5.

Models of world-centered task performance. A, Probability of responding at East or West response port under four example conditions in which platform angle and speaker location is varied. Values show probability of responding at East and West ports, expressed as percentage. Exclamation marks indicate trials for which the model would attempt to respond at an inactive (null) port. B, Performance of three ferrets trained with the same pair of world-centered sound locations (F1701, F1703, and F1811) in terms of overall accuracy (top row: % correct), probability of responding at the West port as a function of sound angle in the world (middle) and West response probability as a function of sound angle in head and world-centered space (bottom). C–G, Corresponding predictions from simulations of each model (for details of model parameters in each simulation, see Materials and Methods).

Figure 6.

Model fit to single trial behavior. A, B, Model validation performance showing accuracy in predicting single trial behavior from held-out data on test (A) and probe trials (B). Performance shown for data collected from all ferrets trained to discriminate sounds from North and South locations (F1701, F1703, F1811). Box plots show median and interquartile range; individual data points show validation performance for each fold (n = 20). C, D, Performance of each model as a function of sound angle in the world (C) and for each combination of sound angle in head and world-centered space (D). Performance shown as the percentage of individual trials that the model correctly predicted the animal's behavior (% correct). Data shown as median across 20-fold cross-validation. Dashed lines in C show chance performance.

Figure 7.

Head tracking and response time analysis. A, Screenshots showing tracking of head and body position using DeepLabCut. Response locations are labeled, for example, East (E) and West (W) ports. B, Trajectories of head position during trials as animals responded to test and probe sounds. Data shown from responses in the first 10 sessions in which probe sounds were presented. Markers show positions on each frame; lines show linear interpolation between frames. C, Path lengths for data shown in B. Scatter plots show path lengths for individual trials, with lines showing mean and SE for each ferret. D, Comparison of median reaction times (RTs) with probability of responding at the West response port. Chance performance = 0.5. Median RTs were calculated across trials for a given combination of speaker location in head and world-centered coordinates (n = 144 conditions per ferret). E, RTs as a function of adjusted response probability (p'; see Materials and Methods). Data shown as in E.

Figure 8.

Head-centered sound localization. A, Performance discriminating sounds at trained locations for each ferret as a function of platform angle (n = 400 trials per angle). Data shown as mean percent correct across bootstrap resampling (n = 100 iterations). Dashed lines show chance performance (50%). Cartoons show the training configurations [either Front vs Back (F1901, etc.), or Left vs Right (F1810)]. B, Probability of making Left responses to test and probe sounds as a function of sound angle in the world, or relative to the head (n = 36 trials per angle). C, Response probability as a joint function of head and world-centered sound angle, shown for individual animals (C, n = 432 trials over 144 locations). D, Comparison of ferret behavior with simulated behavior of models that linked head or world-centered sound location to the probability of responding at the Left response port. Data shown for ferrets (F1901 and F1905) trained in Front-Back discrimination. Model parameters for simulations are given in Table 1 (For model parameters, see Materials and Methods). E, Model validation performance predicting single trial behavior from held-out data (20-fold cross-validation). Box plots show median and interquartile range; individual data points show validation performance for each fold. F, Misalignment of head direction relative to platform shown in the distribution of offset values across trials (n = 87) for animals trained to discriminate front and back sounds in head-centered space [F1901 (n = 57) and F1905 (n = 30)]. In both cases, median offsets were significantly greater than zero (Wilcoxon rank sum, p < 0.001). Data shown as box plots indicating median and interquartile range, with individual trials shown as separate markers.