[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: comparing cochlear models ?
Thank you for all the informative comments. I agree with you on the
fact that using the ability of a particular model in reproducing
physiological and psycho-acoustical characteristics of the biological
cochlea can be a sensible metric (as it's the most common metric used
apparently), but it is not exactly what i was asking for. as a
computational modeler, since it is not clear how all of these
physiological and psycho-acoustical properties associated to cochlea
play role "computationally" and how they reflects cochlea's information
processing capabilities or information representation in general.
So what I'm interested in is a handy quantitative metric that I can
use to choose between different cochlear models available in order to
maximize the performance of the whole partcular auditory processing
task that i might be interested in (e.g. source separation, Speech
recognition,...). of course I'm not asking for the obvious answer, that
"the best cochlear model is the one that maximises the overall
performance of the system". Instead I'm interested in being able to
have an evaluation of different blocks of the system individually,
specifically in this case "the cochlea".
So, do you think there can be such an evaluation method ?
Sorry for being so long.
On Fri, May 21, 2010 at 1:33 PM, Bastian Epp <Bastian.Epp@xxxxxxxxxxxxxxxx>
As the comparison and evaluation of cochlear models is exactly what bothered my during the last year (or so), I think I want to share my points:
In general, it is very interesting to compare and to evaluate models of the cochlea...even though that is sometimes quite hard due to different approaches.
I agree with the view, that one way to evaluate cochlear models is to look at their performance to account for many different datasets (physical (OAE), physiological(I/O) and even psychoacoustical) simultaneously with a "decent" amount of fixed parameters....but especially for the psychoacoustic paradigms, the potential for arguing starts at the metric which is evaluated to account for the data. So maybe one should start with physiological data and/or use very simple (and still feasable) approaches for psychoacoustics. Some ideas as a starter:
Evaluation of the shape of the excitation pattern (slopes basal/apical, Q-factor, ...) in a panoramic view (whole cochlea) and at a single place for different frequencies (place-fixed view).
Evaluation of the response of the model at a fixed place for different levels of the same frequency
Propagation of the cochlear activity through a simple model of the middle ear (which should be comparable as long as the used middle ear is linear).
Phase-issues with evoked (e.g. CEOAE) emissions?
Here it's getting neccessary to define the metric which is evaluated, but there should be the phenomenon of a nonlinear interaction of two tones...independently of the analysis which is applied.
All these measures should be relatively robust against absolute scaling and hence be comparable and (if the model is supposed to be "realistic" in one sense or the other) in line with data.
I see a big chance in that, since different models account (and are developped for) different aspects of cochlear functioning, and all reflect a special case of THE cochlea which was implemented by nature.
We did the things mentioned above (and some more) using the Duifhuis-Transmission line in a modifed form. I would be happy to share ideas with anybody who is also interested. We are also preparing a manuscript and I am also happy to share that as soon as it is in a readable form. (BTW: Thx Jon for sharing the manuscript, nice collection of data and a good start for discussions....)
If anybody read the message until here: Sorry for the length.
Best from Oldenburg