Re: Autocorrelation

[Peter Cariani <peter@xxxxxxxxxxxxxxxxxxxx>]

Christian Kaernbach wrote:
> > 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.
>
> Once we knew it's ellipses we would no longer state it's circles.

If the proponents of a theory believe in ellipses, one does not make a
model with circles, falsify it, and expect that they will agree that
their model has been falsified.

Nobody that I know of believes literally that a simple autocorrelation
of the stimulus is being computed, and this for one  frequency band. Roy
Patterson and others have repeatedly emphasized that one has to take
into account cochlear filtering and the weighting of frequencies that
ensues in order to get more realistic behavior. Who is talking
"circles"? Your paper never referenced whose model you were testing, I
think because nobody proposes this model, but the title of the paper is
"Psychophysical evidence against THE autocorrelation THEORY of auditory
temporal processing." as if there were one model and this is the one
that was tested and that the evidence falsifies the whole theory. There
is absolutely nothing wrong with testing and falsifying the simple
autocorrelation, and pointing out its limitations, but one does not want
to overstate the extent to which a whole theory is falsified.

Even if the theory were shown to be incorrect for unresolved harmonics
("mechanism B"), the population-interval model would still explain very
well the (more important) pitches produced by resolved harmonics
("mechanism A"), (which curiously, is analogous to what spectral-pattern
proponents argued in the wake of the discovery of the  dominance region,
but with a temporal mechanism explaining pitches of resolved harmonics.)

What do you think is the nature of the neural representation that
subserves mechanism A?
Is it (un)fair for me to assume that when people speak of dual
mechanisms for pitch, they have in mind a (strong) spectral pattern
mechanism (A) for the resolved harmonics and a (weak) interval-based
mechanism (B) for the unresolved ones?

> Rejecting autocorrelation models would not mean "to go back to spectral
> pattern mechanisms". I think we are in want of a temporal model that
> shows an asymmetry between first- and second-order regularities. I guess
> that such a model will not be realizable with any sort of (neural)
> autocorrelation. Sure, such a model should ...
> > involve the neural substrates of auditory percepts that encompass
> > phenomena other than the masking of these click trains
> ...but it should also encompass those little click train observations. I
> really don't want to overemphasize the importance of the findings of
> K&D, but when I am called to the defence of our paper, I defend it.

Defence is fine; one just shouldn't overstate the conclusions. As I said
before, if the click trains consisting of unresolved harmonics that
contain interspersed clicks don't produce 1 spike per click, contrary to
what you assumed but consistent with what I saw in high CF fibers, then
there are no interval peaks at the pitch period (10 ms) in either
all-order or first-order distributions, and the masking is consistent
with a central autocorrelation analysis on the output of the auditory
nerve array (however far up the pathway one wants to put it). The
population-interval representation is a central auditory representation
whose form reflects cochlear and neural processes. The absence of
interval peaks in the first-order distribution in the masked case also
means that this is also consistent with a first-order interval analysis
-- it appears that first- and all-order distributions yield similar
results, so one cannot distinguish the behavior of the two
representations using these stimuli. It appears to me that the
hypothesis of a central  analysis on population-wide all-order interval
distributions thus was not falsified, even for  stimuli with unresolved
harmonics. While autocorrelation alone does not account for the masking
(you are right, how could it?), I think cochlear filtering + neural
processing + central all-order interval analysis does.

The specific adjustments that we need to make in our assumptions involve
taking into account the kinds of temporal precedence effects that seem
to be operant in high-CF fibers when one has unresolved harmonics
(higher frequency components & higher harmonic numbers).

> > 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
>
> The use of low-pass noise is indispensable. A major argument to assume
> that there is a second mechanism involved in pitch perception working on
> unresolved harmonics (let's call it "mechanism B") is the increase of
> the JND once resolvable harmonics are excluded. Exclusion means:
> down-filtering _and_ masking. As long as harmonics well below 15 Fo (say
> at 10 Fo) are either present or not masked, the improvement in JND by a
> factor of at leat five indicates that we have not yet isolated
> "mechanism B". This is nicely illustrated in our paper under revision
> (Bering & Kaernbach). For the same reason, Houtsma & Smurzynski used
> pink noise that was not even low-pass filtered.
>
> Anyway, as we employed high-pass filtered click trains there should be
> regions where the signal is relatively undisturbed by the low-pass
> filtered noise. And the use of low-pass noise does not offer a plausible
> explanation of why first- and second-order regularities should be
> treated differently.

I agree that the low-pass noise is largely a side-issue, since the
pitch-masking effect is profound without the LP noise. (The robustness
of the pitch-masking effect in quiet rules out the situation where one
has reduced the salience of the pitch to near-threshold and then
introduced a small amount of masking that pushes its salience just below
threshold. The pitch-masking effect is very real and not due to
threshold manipulations.)

Whatever one chooses, if one uses LP noise in one's stimuli, then this
should also be explicitly incorporated in model predictions because low
frequency components can have significant effects on the temporal
discharge patterns of high-CF fibers.

-- Peter Cariani