Auditory model comparison and optimization using dynamic time warping

auditory modeling;optimization;model comparison;dynamic time warping;Gradient Descent

Sturgeon, Brantly ; Jones ; Douglas L.

关键词: auditory modeling;    optimization;    model comparison;    dynamic time warping;    Gradient Descent;   

Others  :  https://www.ideals.illinois.edu/bitstream/handle/2142/49471/Brantly_Sturgeon.pdf?sequence=1&isAllowed=y

美国|英语

来源: The Illinois Digital Environment for Access to Learning and Scholarship

【 摘 要 】

Auditory transduction modeling efforts have relied on detailed metrics such as phase-locking, adaptation time, and spike timings.A technique allowing summary comparison of auditory models is missing from the current body of research.We introduce a new technique based on the dynamic time warping algorithm as a distance metric.This technique is also applied in conjunction with a simple finite-difference gradient descent technique to generate better model parameters.These improved parameters reduce error due to poor parameter estimation and allow for a clearer evaluation of the underlying mechanics of a model.We evaluate this technique beginning with a simple model and ending with a cochlear model that exhibits three major transduction phenomena:frequency selectivity, compression, and a limited set of inner hair cell dynamics.We apply these techniques to related work and seek to identify the model that best describes the transduction of both naturally produced and spectrally reduced synthetic stop consonants.We produce and compare optimized models that harness the aforementioned major phenomena.Additionally, we find that the comparison technique predicts that incremental modeling of auditory phenomena will simulate more accurate neural ensembles.Results from this work show that the tested phenomena are crucial to cochlear modeling, but that a significant performance gap exists between the examined models and the natural auditory transduction process.