[Martin Taylor 920309 16:30]
(Bruce Nevin 92019 13:20:24) I think 920309, but that's what the heading says.
It seems to me more plausible that the "adapting stimulus" is used by
the control system to establish a reference signal appropriate for a
particular speaker. After several variations with little or no
variation, the hearer "expects" that anything phonetically different is
categorically different (a [d] instead of a [b]). I wonder what happens
if the listener is presented with a series of [ba]s and [da]s
interspersed, setting consistent reference values for both, and then is
presented with the continuum, with the option of saying "neither."
I quite agree with Bruce on this. I have a typescript that I intended to turn
into a paper some time in the 1970s on the topic, from which I will quote
some passages at the end of this posting. I simply do not believe that any
of the perceptual effects attributed to "fatigue" are properly attributed.
This applies to figural aftereffects, of which the shift of the phonetic
category boundary is probably an instance. It applies specifically to
reversing figures, which are often attributed to the fatigue of one percept
with the consequent appearance of the other when the fatigue progresses far
enough. The timings of the changes of percept are quite inconsistent with
a fatigue interpretation and are consistent in detail with a random walk of
a small stable number of "detectors." In the case of phoneme detectors it
used to be fashionable to say that they could be detected by fatiguing them
selectively. I disputed this interpretation of the data, and asserted that
any "algorithmic" detector based on subjective probability would reach the
same results.
In the quote that follows, the phoneme boundary is between /j/ and /d/, which
differ in the artificial stimuli by the length of the noise burst that
follows the stop release. After an introductory description of the effect--
repeated presentation of /j/ moves the boundary j-ward--the text continues:
"This effect almost always occurs when such an experiment is run, and its
occurrence is taken as evidence for the existence of feature detectors for
the phoneme of characteristic in question. It is not. Any sensible
algoritmic classifier that takes some notice of recent history will do the same.
"In order to see why an algorithmic classifier would shift the transition
region toward the 'fatigued' phoneme in a selective adaptation experiment,
consider the distribution of occurrences of the characteristic (e.g. burst
length) in examples of phonemes from natural speech. Some examples of
/j/ will have short noise bursts, some examples of /d/ longish ones. In
the experiment, the subject must choose on the basis of the length of the
noise burst whether the sound is /j/ or /d/. If he has a feature detector,
he will select the one whose feature detector gives the higher output. If
he detects on an algorithmic basis, his response will be determined by which
phoneme is more likely, or to which the given sound is more similar if it
falls outside the range of either real phoneme. The situation is as shown
in Figure 1 [Picture of sharply peaked /d/ distribution of burst lengths
on left, overlapping wide, low /j/ distribution of burst lengths to its right].
If there exists a feature detector, it must be more broadly tuned for /j/
than for /d/, since the range of natural burst durations is greater for /j/
than for /d/. Either feature detector probably has some output for burst
durations beyond the natural range of "its" phoneme. If the detector is
algorithmic, it must operate on the basis of some kind of "similarity" index
related to the probability that a particular phoneme could be represented by
the sound in question. This simiarity index probably looks like, but might be
broader than the occurrence probability distribution. If the index is
likelihood, as it would be for a Bayesian classifier, then it is identical
to the occurrence distribution.
The distributions of Figure 1 apply to normal speech. In the experiment,
the distributions are quite different. One specific example of the phoneme
is presented over and over again. The variance of the occurrence
distribution of this phoneme is drastically reduced, by an amount that depends
on how sensitive the system is to long-term trends and how sensitive to
local context. Considering both local context and history, the distributions
(assuming /j/ is being "fatigued") become something like those for Figure 2a
[like Figure 1, but now the /j/ distribution is sharply peaked well to the
right of the /d/ distribution]. The /d/ distribution is unaffected, but the
/j/ distribution is reduced except for the central peak that corresponds to
the distribution of experimental presentations.
The responses of the models change, as well. [discussion of feature
detector fatigue response omitted]. The algorithmic classifier's index
changes to match or to track the changes in the occurrence distributions.
It, in effect, says "Now /j/ never gets as short as it used to do, so
I know that middle-length bursts that used to be /j/ are now /d/, since
the /d/ distribution still hasn't changed." [...] The effect on the
experiment is that both the feature detector and the algorithmic detector
give the same result, and the standard selective adaptation procedure
cannot distinguish them.
[The text continues by proposing that the experiment be modified to
retain the natural distribution of burst length for the "fatigued"
phoneme during the "fatigue" process, or to increase the natural range
of burst duration. In the former case, there should be no change in
the perceptual boundary if an algorithmic classifier is used. In the
latter, the result depends on the degree to which the natural ranges
of the two phonemes overlap. If the overlap is small, the wider range
will not much affect which phoneme is perceived during the "fatiguing"
trials, and the boundary should move in the direction opposite to the
normal finding.]
I think that the same analysis could easily be applied to Bruce's proposal.
If we use the term "reference level" as Bruce does, the analysis means
that the continuum of perception levels is being re-scaled, by being
expanded in the neighbourhood of each presented phoneme. I find it
interesting that I used exactly this principle to explain the consistent
errors people make in reproducing figures, in my 1960 thesis (Effect of
anchoring and distance perception on the reproduction of forms. Perceptual
and Motor Skills, 1961, 12, 203-230), and later to account for more
orthodox figural after-effects (Figural after-effects: a psychophysical
theory of the displacement effect, Canadian J. Psychol., 1962, 16, 247-277).
You might like to look at this latter one, because it develops the idea
of subjective probabilities in perception better than I was able to do
in our discussions before Xmas.
Incidentally, though I never measured correlations, I think from looking
at the figures that the figural after-effect paper might satisfy Tom
Bourbon's request for an instance in which subjective probability
considerations produce precise predictions of experimental data. Indeed,
in one later case, the theory provided a good fit to the data using no
(zero) degrees of freedom (Psychol Review, 1963, 70, 357-360) even
though the variable affecting the size of the effect was not one that
had been used in previous one- or two-degree-of-freedom data fits.
That's pretty much a side issue, I suppose, but it has little hooks into
several previous discussions.
Martin