Shift Invariance

[Irving Biederman]

        The communication by Shimon Edelman is, in my opinion, a bit
misleading.  In response to a posting by Eric Postma that listed papers by
Biederman & Cooper (1991) and Nazir & O'Regan (1990) as evidence for shift
invariance, Edelman writes:

"Putting Nazir & O'Regan on the same list with Biederman like that may
be misleading to someone who will not bother to read the paper. Nazir
& O'Regan actually found evidence AGAINST translation invariance in
human vision."

        One may distinguish a strong form of shift invariance, in which
there is no cost in performance from changing the position of a stimulus
with a weak form in which there is facilitation but not as much as when the
stimulus is presented at its originally experienced position.

        Eric E. Cooper and I (Perception '91) found virtually complete
(i.e., strong) shift invariance, as measured by the priming of briefly
presented (100 msec) object pictures.  [100 msec is too brief to make a
fixation onto the stimulus.]  Picture-naming RTs and error rates were
unaffected by a shift.  We did this by presenting the pictures either 2 deg
to the left and or 2 deg to the right of fixation.  The order of left-right
positions was random appearing.  In two experiments, when the pictures were
shown a second time, there was virtually no difference in performance if a
given picture was shifted or not.  A third experiment produced the same
result with 2 deg shifts above and below the fixation point.   That there
was perceptual and not just concept or name priming was evidenced by a
reduction in priming of pictures with the same name and basic-level concept
but a different shape (i.e., two different kind of chairs).  So we obtained
a strong form of shift invariance.

        The finding of strong left-right shift invariance on RTs was
replicated using contour-deleted pictures by Cooper, Biederman, & Hummel
(1992) that, for half the subjects, were also mirror-reversed when they
were shifted.  A slight, but reliable, increase in error rates was noted
only for pictures that were shifted but not reversed.  Shifted pictures
that were also reversed (so the fish, for example, is always facing toward
the fixation) showed no increase in error rates.  We proposed that when a
picture is shifted across the vertical midline, different features (e.g.,
parts) will be present at different eccentricities and therefore, receive
differentresolution.  Mirror reversing the stimulus preserves the
original relation between resolution and features.  If the features are
difficult to discriminate, then the modest variation in resolution could
produce an apparent shift cost.

        A subject in the Nazir & O'Regan ('90, Spatial Vision) experiment
was extensively trained to discriminate a symmetrical nonsense pattern from
two highly similar non-target patterns at 2.4 deg to the left of fixation.
(Other subjects would be trained with that pattern to the right of
fixation.)  The subject could be then tested at the learned position (which
was always peripheral), at central fixation, or on the opposite side.  As
Nazir & O'Regan noted there was an enormous amount of facilitation in all
conditions in all experiments.  So there was at least weak invariance.  Was
there strong invariance as well?  When they controlled potentially
confounding and contaminating factors, there was strong shift invariance,
Edelman's claim to the contrary.

        Under controlled conditions, when a stimulus was not presented
unless the eye was on the fixation point, there was no effect of a shift
from learned to opposite positions.  There was a cost, however, of shifting
from a learned (peripheral) to central position.  But this comparison
confounds resolution (from peripheral to central) with shift.  Although it
may be surprising that one would do worse with central as compared to more
peripheral positions, it is not implausible that a different set of
features were employed at different resolutions.  In three latter
experiments, where eye position was not controlling (it was
difficult to train subjects to do it), there were much larger costs but
there well could have been a bias to look at the learned location.   In
these latter three experiments, as Nazir & O'Regan noted, there was
considerable subject and
stimulus variability, perhaps reflecting various task strategies.
Certainly, a bias to monitor the trained location is not out of the
question.

        So we have four name-priming experiments documenting strong shift
invariance when resolution is controlled, three left-right and one up-down.
The Nazir & O'Regan research shows strong invariance under controlled
conditions.  So five well-controlled experiments document strong shift
invariance.  Weak invariance is obtained under less controlled conditions.

        Finally, let me note that, of course, it is not the case that every
representation is shift-invariant.  Under the identical conditions that
yielded invariance in object priming, subjects showed well-above-chance
explicit memory of where the picture was presented.  Cooper and I
hypothesized that position information may be specified by the dorsal
system.  Those who presume to test invariance of shift (or size or
orientation or reflection) should bear in mind the possibility that a
particular task, especially if it is extremely difficult so that subjects
are induced to undertake various strategies such as search, may tap both
shift-invariant and shift-specific representations.  For example, if I've
been extensively trained to search for a small distinguishing feature on
the left side of the display, I could readily show a shift cost if the
feature is no longer there.

References:

Cooper, E. E., Biederman, I., & Hummel, J. E.  (1992).  Metric invariance
in object recognition:  A review and further evidence.  Canadian Journal of
Psychology, 46, 191-214.

> Biederman, I. & Cooper, E. E.  (1991).
> Evidence for complete translational and reflectional invariance in visual
> object priming.
> Perception, 20, 585-593.
>
> Nazir, T.A. & O'Regan, J.K. (1990).
> Some results on translation invariance in the human visual system.
> Spatial Vision, 5, 81-100.


**************************
Irving Biederman, Ph. D.
William M. Keck Professor of Cognitive Neuroscience
Department of Psychology
University of Southern California
Hedco Neurosciences Building, MC 2520
Los Angeles, CA 90089-2520

ib at rana.usc.edu
(213) 740-6094 (Office); (213) 740-5687 (Fax);
(310) 823-8980 (Home); (213) 740-6102 (Lab)
Visit our web at:  http://rana.usc.edu:8376/~ib/iul.html