Shift Invariance

[Edelman Shimon]

> Date: Fri, 8 Mar 1996 18:25:00 -0800
> From: Irving Biederman <ib at rana.usc.edu>
> 
>         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.
> ...
[ rest of Biederman's message omitted ]
> ...

Many thanks to Irv Biederman for posting the details of his findings,
along with a comparison with the results of Nazir & O'Regan. His
effort should reduce the chance of the readers of this list jumping to
premature conclusions.

Note that the purpose of my previous posting was to advocate caution,
certainly not to argue that all claims of invariance are wrong.
Fortunately, my job in this matter is easy: just one example of a
manifest lack of invariance suffices to invalidate the strong version
of invariance-based theory of vision, which seems to be espoused by
Goldfarb:

> If we 1) DO NOT FORGET that the biological systems have at their disposal
> quite adequate means to extract symbolic (structural) representation right
> from the very beginning and 2) FORGET about our inadequate numeric models,
> then the question would not have arisen in the first place. Symbolic 
> representations EMBODY shift invariance. 

So, here it goes... Whereas invariance does hold in many recognition
tasks (in particular, in Biederman's experiments, as well as in the
experiments reported in [1]), it does not in others (as, e.g., in [2],
where interaction between size invariance and orientation is
reported). A recent comprehensive survey of (the far from invariant)
human performance in recognizing rotated objects can be found in
[3]. Furthermore, not only recognition, but also perceptual learning,
seems to be non-invariant in some cases; see [4,5].

FORGETTING about experimental findings will not make them go away,
just as pointing out that symbolic representations EMBODY invariance
will not make biological vision embrace a symbolic approach if it has
not done so until now.

-Shimon

Dr. Shimon Edelman, Applied Math. & Computer Science
Weizmann Institute of Science, Rehovot 76100, Israel
The Web:  http://eris.wisdom.weizmann.ac.il/~edelman
fax: (+972) 8 344122   tel: 8 342856   sec: 8 343545

-----------------------------------------------------------------------------
References:

[1] 
@article{BricoloBulthoff92,
author="E. Bricolo and H. H. {B\"ulthoff}",
title="Translation-invariant features for object recognition",
journal="Perception",
volume="21 (supp.2)",
year = 1992,
pages = "59"
}

[2]
@article{BricoloBulthoff93a,
author="E. Bricolo and H. H. {B\"ulthoff}",
title="Further evidence for viewer-centered representations",
journal="Perception",
volume="22 (supp)",
year = 1993,
pages = "105"
}

[3]
@InCollection{JolicoeurHumphrey94,
  author = 	 "P. Jolicoeur and G. K. Humphrey",
  title = 	 "Perception of rotated two-dimensional and
		  three-dimensional objects and visual shapes",
  booktitle =	 "Perceptual constancies",
  publisher =	 "Cambridge University Press",
  year =	 1994,
  editor =	 "V. Walsh and J. Kulikowski",
  chapter =	 10,
  address =	 "Cambridge, UK",
  note =	 "in press"
}

[4]
@article{KarniSagi91,
author="A. Karni and D. Sagi",
title="Where practice makes perfect in texture discrimination",
journal=pnas,
volume="88",
pages="4966-4970",
year="1991"
}

[5]
@article{PoggioFahleEdelman92,
author="T. Poggio and M. Fahle and S. Edelman",
title="Fast perceptual learning in visual hyperacuity",
journal="Science",
year="1992",
volume="256",
pages="1018-1021",
}