Subject: Re: Autocorrelation From: Peter Cariani <peter(at)epl.meei.harvard.edu> Date: Tue, 18 Jul 2000 14:19:47 -0400Christian Kaernbach wrote: > Please forgive me that I will not enter into this discussion in this > detail. Whenever my daily e-mails takes me more than 10 minutes I am > convinced something is going wrong. My only and major point was and is: > K&D 1998 is important to modelers because it sets constraints (asymmetry > of 1st- and 2nd-order regularities). Up to the modelers to show that > there is no problem. And I am not interested in comments why this and > that model fails to reproduce this asymmetry (neglecting tails of tunig > curves or whysoever).... all this is pure speculation. I am interested > in models that succeed to produce this effect. I would not want to waste any more (than 10 minutes) of your time regarding these broader issues, which involve the neural substrates of auditory percepts that encompass phenomena other than the masking of these click trains that are produced by psychophysically-unresolved harmonics. Jont Allen wrote: > How do the tails become involved? The levels would need to be very large, wouldnt they? Also the neural filter would emphasis the tip region, so the spikes envelope would depend mainly on the cf region. At 60 dB SPL a harmonic complex consisting of harmonics below 1 kHz will impress the periodicity of its fundamental on fibers whose CF's are 2-10 kHz. These fibers may not fire at rates much higher than spontaneous, but their discharges are locked to the stimulus fundamental, so they contribute pitch-related intervals to the global interval distribution. Whether responses will be dominated by components near or far from CF will depend on their relative intensities, and also potentially on temporal precedence patterns. I brought up tails of tuning curves initially, because the population-interval representation is the distribution of all-order interspike intervals across the whole auditory nerve array, and that lower-frequency components at moderate levels effectively drive higher-CF fibers, which then contribute pitch-related intervals to the distribution as a whole. This is one (temporal) way of thinking about the upward spread of masking -- when one sums all of the intervals from lower frequency and higher frequency components, because of the asymmetry of tuning curves and better phase-locking to lower frequencies, more intervals related to the lower component are produced. The broadening of tuning curves means that a global interval representation can make use of the full dynamic range of the whole auditory nerve array -- the extent of a given temporal pattern spreads with higher level. The tails of tuning curves come into the present discussion partly because K & D used low-pass noise, and this low-pass noise also has the effect of masking the pitch produced by higher partials, thereby lowering its salience and making the task more like a pitch detection near threshold rather than the masking of a more salient pitch well above threshold (but as I said before, the more general masking result is quite robust and striking in the absence of the LP noise and well above threshold). If you don't factor in the tails and the low-pass noise nor the stronger degree of phase-locking to lower frequency components, as K & D decided not to do in their simple autocorrelation, then the relative height of the pitch-related autocorrelation peak is artificially higher than it would otherwise be. I don't fault them for using the simple autocorrelation model -- for many purposes, especially for resolved harmonics as one finds in musical contexts, it is not a bad rough approximation, and more involved neural models introduce all sorts of additional factors and distract from K & D's primary interest, which is in the psychophysics. However, if one knocks down a simple model (that nobody believes literally) for a limited range of high-frequency/harmonic number masking phenomena (that, in themselves, are relatively unimportant as far as pitch perception goes), one should not claim that one has falsified an entire theory or whole class of models. It is exceedingly difficult to rule any model out of hand, including rate-place-based spectral pattern models for pitch (this is why definitive experiments in the neurophysiology of perception are few and far between). I can imagine ways in which spectral pattern mechanisms might be operant in covert form, even though we have yet to observe any neural response patterns that would support such analytical mechanisms. All models need to be kept on the table, in their strongest, most plausible forms, and all possibilities need to be considered. In any case, we don't throw out the Copernican model because we observe that the orbits are ellipses not circles, and that the sun is not exactly at the center of the solar system. One doesn't go back to epicycles (or to spectral pattern mechanisms) because of such observations; one refines the models. -- Peter Cariani