154th ASA Meeting, New OrleansLA

[ Lay Language Paper Index | Press Room ]


Sound localization by echolocating bats: Are auditory signals enough?


Murat Aytekin1,aytekin@umd.edu
Jonathan Z. Simon2 - jzsimon@umd.edu
Cynthia F. Moss1  - cmoss@psyc.umd.edu

1. Department of Psychology,
University of Maryland,
Biology/Psychology Bldg. Rm 1148
College Park, MD 20742

2. Department of Electrical & Computer Engineering
A.V. Williams Building
University of Maryland, College Park
College Park,  MD  20742

Popular version of paper 2pAB4
Presented Wednesday afternoon, Nov 28, 2007
154rd ASA Meeting, New Orleans, LA

Animation
Figure 1: Bats monitor their environment by emitting brief ultrasound pulses and listening to the echoes reflected from objects around them.


Echolocation and spatial hearing

Echolocating bats emit brief ultrasound pulses and listen to the echoes reflected from objects surrounding them for navigation and prey detection and capture. Bats can determine relative positions of the echo sources (objects) in terms of distance and direction. Distance is measured as a function of the time elapsed between the outgoing sound and the incoming echo. To determine direction, bats use the acoustic cues in echo signals received at each ear. The focus of this paper is on how bats extract directional properties of echoes.


Echolocation Pulse Spectrum
Figure 2: Time waveform (a) and spectrogram (b) of a typical echolocation pulse of an big brown bat (Eptesicus fuscus). Echolocation pulses are brief in time (1-5ms) and between 20 kHz and 100 kHz in frequency.

How do animals localize sound?

Sounds impinging on the ears are transformed by the shape of the head and ears. External ears, acting like antennas, amplify or weaken sounds based on their direction of arrival. These transformations create acoustical cues that the animal’s brain can use to determine sound source location in space. For example: sounds coming from the right will be received earlier and with larger intensity at the right ear. These interaural differences in signal intensity and arrival time (IID and ITD) decrease as the sound source moves towards the front and increase in the opposite direction as it reaches the left side. Therefore, the auditory system in the brain can use IID and ITD as cues for horizontal position of the sound source. For echolocating bats IID cues can also be used for the localization of the vertical direction – provided that acoustic signals are wideband [6, 5, 1]. Other direction-dependent features due to directional modifications can be found in the spectrum of the acoustic signals at each ears.

Localization_Cues
Figure 3: Sound sources are received with different intensity and at different times (tr and tl) depending on the relative direction of the sound source with reference to the animal’s head. Intensity and the time difference of the acoustic signals at the two ears are direction-dependent. Echolocating bats can determine both the horizontal and the verticaldirections of the sound source using IID cues, when the sound is wideband.

Echolocating bats and active sound localization

Bats share similar computational scheme for passive sound localization with other mammals. But the echoes they detect partly depend on the spectral and spatial properties of the echolocation pulses. Bats ensonify the space in front of them to probe their environment. Acoustic pulses generated by bats propagate towards different directions with different intensities. Most bats use echolocation pulses comprised of multiple frequencies. Sound beams generated at high frequency are directionally more focused, whereas sound beams generated at low frequencies are directionally more diffused. Low frequency sounds are not reflected easily from small objects (like insects) and travel further. In contrast, high frequency sounds decay faster yet are easily reflected by the small objects. Information extracted from echoes containing multiple frequencies provides bats high resolution representation of their environment. Measurements of sonar emission patterns from unrestrained and freely vocalizing bats show variation in the echolocation beam shape across individual vocalizations [3]. Consequently, the bat auditory system needs to dynamically adapt its computations to accurately localize objects using changing acoustic inputs.

Echolocation_Beams
Figure 4: Echolocation pulses propagate towards different directions with different intensities. Profiles of these intensities also called acoustic beams (shown by shaded areas) depict the sound intensity traveling towards a given direction (direction of the arrows) at a fixed distance. The beam profiles, given for two frequency regimes, are hypothetical and are not obtained from real measurements. Bats might be capable of controlling the shape of the echolocation beam. Length of the arrows represent the intensity of the outgoing sound towards the selected directions. 

Statement of the Problem

It is largely accepted that the computation of the sound localization by the auditory system is based solely on the acoustic signals received at the ears. Schemes of how the brain might compute sound location are thus mostly limited to identifying acoustic cues that map the signal arriving at the ears to a location in space. Like many other mammals, bats can move their ears, and therefore acoustic cues for sound localization even change with ear movements. To accurately compute the direction of the echo-return, the auditory system also needs to take ear position into account. Non-auditory signals, carrying information about ear position and beam shape, are therefore required for auditory localization of sound sources. The auditory system must learn associations between these extra-auditory signals (e.g. head and ear position) and the acoustic spatial cues. These demands on the nervous system suggest that sound localization is achieved through the interaction of behavioral control and acoustic inputs.

Standard approaches that rely on acoustic information alone cannot explain the effects of this interaction. Moreover, unlike the approach we propose here, by taking an outside observer’s point of view, they do not address how the auditory system ever acquires the knowledge of the spatial coordinates to utilize these acoustic cues. We propose a sensorimotor approach to the general problem of sound localization, with an emphasis on learning that allows spatial information to be acquired and refined by a naive organism. Thus, the sensorimotor approach does not require prior knowledge of the space by the system.

A sensorimotor approach for sound localization

Our goal is to show that merely from  an organism’s observations of the sensory consequences of its self-generated motions, there is sufficient information to capture the spatial properties of sounds. The goal is not to find a way to match acoustic inputs to corresponding spatial parameters but rather show how the animal could learn the spatial properties of the acoustic inputs. Normal sensory development requires relevant sensory inputs  and self-generated actions are also essential for the development of spatial perception [4] and to adapt to changes that might occur after its acquisition [7]. Perceptual spatiality of sound sources is achieved under unnatural acoustic conditions only if a relation between auditory signals and self-generated motion existed [8]

From an animal’s point of view, the environment surrounding the animal is stable as it moves within the environment. This requires the ability to distinguish sensory input changes caused by self-generated movements from those that are the result of changes in the environment. This can be achieved with the proprioceptive (sense of the relative positions of the body parts) information, plus the ability to predict sensory consequences of the organism’s actions, that is, sensorimotor expertise. We assert that sensorimotor early experience, during development, is necessary for accurate sound localization.

The sensorimotor approach [2] is based on premises. We first assume that the organisms are initially unaware of the spatial properties of the sound sources. Second, we limit the external sensory information to auditory signals only (e.g. not to visual input). Third, we postulate an interaction between the auditory system and the organism’s motor state, that is, proprioception and motor actions. The first two premises may be viewed as the worst-case-scenario for sound localization, ignoring any sound localization mechanisms that might be hardwired in the brain or aided by vision. But they don’t significantly constrain the approach. The third premise, in contrast, is crucial to the proposed computational scheme.

Demonstration: Obtaining spatial coordinates by echolocating bats

Our demonstration is a proof of concept. The problem is simplified for the purpose of clarity but does not effect generality. We make three assumptions:

  1. We assume that an organism can distinguish the differences between the auditory and proprioceptive inputs. These inputs can be separated and classified since auditory inputs can change when the organism is motionless and similarly proprioceptive inputs can be generated through motion of the body when there is no auditory signal.
  2. Based on the classification and separability of sensory inputs in these two groups, an organism can identify the dimension of space through its interaction with the environment [10]. Thus, we also assume that the dimension of auditory space is known. Auditory space is 2-dimensional since a sound source direction can be determined by two parameters: horizontal and vertical angles (we ignore the third dimension of distance to simplify the example).
  3. We assume that the organism can distinguish the motor actions that can induce a displacement in the direction of a sound source. An algorithm that demonstrates how a naive system can identify these special movements is given in [11].

We measured the directional transformations of sounds due to head and the external ears for big brown bats. These measurements are used to simulate acoustic signals received from different directions. Limiting our simulated bat’s movements to head motions. Simulated sounds with random spectra are presented from different directions to the artificial bat. When a sound is present the simulated bat makes three 1o-head movements (up-down, left-to-right, tilt to the right) and records the acoustic inputs changes in relation to its head movements. These measurements are fed to an unsupervised-learning algorithm. The algorithm generates a 2-dimensional representation of the data set, a map (Fig.5). Each point on this map  represents a direction in space. Note that our computational method does not use any a-priori information about the directions of the sounds – only the measurements taken by the simulated bat. Learning of this map implies that our simulated bat can learn the geometric relations between the points in space. What identifies a point in space is the unique relationship between the acoustic input changes and the known set of head movements that generate them. Auditory input changes caused by a group of head movements will be identical even if the sound sources themselves are not.


Map
Figure 5: Spatial locations of the presented sounds are depicted as color coded points in the stereographic projection of the frontal hemisphere (Inset). The latitude and longitude lines indicating horizontal and vertical coordinates and the color code allow comparisons with the map obtained from the unsupervised learning method and the spatial coordinates of the sound sources. Note that the points on the map obtained with our computational scheme maintain the same neighborhood relationships of spatial points in the frontal space of the bat.

Conclusion

We have proposed a computational method for learning the properties of auditory space using sensorimotor theory [9], a previously unexplored issue of the problem of sound localization. We have argued that a computational theory of sound localization should be able to explain the experience-dependent nature of the computation as well as its dependence on other sensory inputs. The computational method described here provides a framework under which integration of experience-dependent plasticity and multisensory information processing aspects of sound localization can be achieved. In conclusion, we have shown that a naive organism can learn to localize sound based solely on dynamic acoustic inputs and their relation to proprioceptive (motor) states.

References

[1]   M. Aytekin, E. Grassi, M. Sahota, and CF. Moss. The bat head-related transfer function reveals binaural cues for sound localization in azimuth and elevation. J Acoust Soc Am, 116(6):3594–3605, Dec 2004.

[2]   M. Aytekin, J. Z. Simon, and C. F. Moss. A sensorimotor approach to sound localization. Neural Computation. (In press).

[3]   M. Aytekin. Sound Localization by Echolocating Bats. PhD thesis, University of Maryland, College Park, Maryland, May 2007.

[4]   J. Campos, D. Anderson, M. Barbu-Roth, E. Hubbard, M. Hertenstein, and D. Witherington. Travel broadens the mind. Infancy, 1:149–219, 2000.

[5]   Z. M. Fuzessery and G. D. Pollak. Neural mechanisms of sound localization in an echolocating bat. Science, 225(4663):725–728, Aug 1984.

[6]   A. D. Grinnell and V. S. Grinnell. Neural correlates of vertical localization by echo-locating bats. J Physiol, 181(4):830–851, Dec 1965.

[7]   R. Held. Shifts in binaural localization after prolong exposures to atypical combinations of stimuli. Am J Psychol, 68:526–548, 1955.

[8]   J. M. Loomis, C. Hebert, and J. G. Cicinelli. Active localization of virtual sounds. J Acoust Soc Am, 88(4):1757–64, Oct 1990.

[9]   J. K. O’Regan and A NoŽ. A sensorimotor account of vision and visual consciousness. Behav Brain Sci, 24(5):939–73; discussion 973–1031, Oct 2001.

[10]   D. Philipona, J. K. O’Regan, and J-P. Nadal. Is there something out there? Inferring space from sensorimotor dependencies. Neural Comput, 15(9):2029–2049, Sep 2003.

[11]   D. Philipona, J. K. O’Regan, J-P Nadal, and O. J.-M.D. Coenen. Perception of the structure of the physical world using multimodal unknown sensors and effectors. Advances in Neural Information Processing Systems, 15, 2004.


[ Lay Language Paper Index | Press Room ]