Wesley R. Elsberry
Last revised: 941030
Abbreviated title: ANS models of hearing.
Name of Institution: Wildlife and Fisheries Science
Texas A&M University
College Station TX 77843
Statement of Objectives
The objective of this project is to:
Develop a set of artificial neural system models of hearing based
on alternative hypotheses of neural function.
Evaluate the ANS models of hearing developed using echo returns gathered using echolocation pulses with the characteristics of a variety of dolphin species.
Statement of Approach
Alternative hypotheses of auditory neural function will be explored using artificial neural system modeling techniques.
Analysis of the anatomical and neurophysiological literature available on mammalian auditory systems will provide the basis for constructing neural population models of function at the level of the basilar membrane up through the auditory nerve. Categorization and recognition functions of a more central nature will be approximated through the use of extant ANS models.
Dolphin echolocation has been closely scrutinized in the captive setting. This research has typically examined the limits of resolution of this sensory system on various man-made tasks. The approach taken treats dolphin echolocation capabilities as a kind of "black box", where few, if any, of the internal mechanisms are apprehensible by the researcher. A clearer understanding of dolphin echolocation will be more likely to emerge if a "white box" method of approach is followed, where models of dolphin echolocation are formulated and tested as candidate mechanisms to fill the "white box". In this approach, these models would be expected to lead to further predictions, tests, and refinements of the models. One promising avenue of research in this domain is that of artificial neural system (ANS) models of dolphin echolocation, especially in the context of signal classification. ANSs are part of a class of approaches in artificial intelligence which comprise the "bottom-up" school. Known also as "sub-symbolic computation" and "connectionism", ANSs rely upon many low-level processing elements with a high degree of modifiable connectivity. Models by Roitblat, Nachtigall, Au, and Moore demonstrate the feasibility of this approach to approximating the function of classification of echolocation returns in a manner similar to that of dolphins. A possible critique of that work is that the models have little inherent biological plausibility; that is, the correspondence of the internal functioning of the model to known neurobiological principles is not high.
Different dolphin species have different characteristic echolocation pulses. Several species with differing characteristics will be selected and have an archetype echolocation pulse recorded for later playback as well as taking many natural echo returns from targets.
Echo returns from targets will be divided into two sets, one for use in training, and the other for use in testing the ANS models.
The ANS models will be trained using the training set of echo returns. After training is completed, the ANS models will be tested on the test set of echo returns.
Statement of Significance
The performance of the ANS models may indicate a relative difference in utility of the alternative hypotheses of hearing so tested.
Certain hypothesized mechanisms of hearing may be shown to have greater import in the analysis of signals of different characteristics.
Using parsimony with available neurobiological evidence as a key characteristic should enable the development of refinements of hypotheses in dolphin echolocation and other auditory systems.
Key Words
Dolphin, artificial neural network, hearing, echolocation
Introduction
The auditory system in mammals is well characterized anatomically, but the underlying neurophysiological function is not so well understood. Various hypotheses of hearing attempt to explain the layout and response of sensory cells along the basilar membrane and the interaction of neural response from those sensory cells leading to the complex representation of activity in the auditory nerve.
Artificial neural system (ANS) models have encapsulated some of the features common to biological neural systems in a computable form. ANS models exist which are geared to providing analogues of memory (Instar, Outstar, BAM, SDM, BSB), Pavlovian conditioning (Drive Reinforcement, Barto-Sutton), low-level neural interaction (gated dipole, on-center-off-surround, BCS), categorization (counter-propagation, SOM, ART), and general functions (Back-propagation).
Previous work on modeling dolphin echolocation with ANSs has dealt with the use of counter-propagation networks and variants of back-propagation networks (Roitblat et al. 1989, Moore et al. 1991). These studies show good correspondence between dolphin performance and model performance under the conditions specified for the models.
The representation of inputs for the counter-propagation ANS used by Roitblat et al. (1989) consisted of amplitude data for each of 20 constant interval frequency ranges.
The representation of inputs for the gateway integrator network used by Moore et al. (1991) consisted of amplitude data for each of 30 constant interval frequency ranges.
Technical Approach
The development of ANS models of hearing is dependent upon analysis of the literature on auditory neurophysiology. The goal is to make models which capture the relevant functional characteristics of that neurophysiology.
For creating a model encapsulating the "place" theory of auditory function, the representation of input data will probably follow that of Roitblat et al. (1989) and Moore et al. (1991) in that amplitude data values for particular frequency ranges will be the most important section of inputs. Two forms can be developed for input signals, the previously mentioned constant interval, and a constant Q representation. The constant Q representation was developed as more appropriately following studies of auditory perception. The issues involved in determining the size of Q for ANS applications has been taken up by Au (1994).
Data will consist of echo returns taken from targets ensonified by captive or wild dolphins. Eighty percent of the recorded data will be used in training the model ANSs, and the remaining twenty percent set aside to be used for testing the model ANSs.
Equipment
Necessary equipment will be audio recording gear capable of recording the high frequencies found in dolphin echolocation signals and returns; data storage media; digitization and signal processing analysis equipment to produce constant interval and constant Q representations of the data; and computer equipment suitable for hosting the suite of ANS models and the training and test data sets.
Other necessary arrangements will be to have acccess to dolphins trained in matching-to-sample, or go/no-go matching to standard paradigms.
References