Connectionists: Best practices in model publication

[Shimon Edelman]

On Jan 30, 2014, at 12:07 AM, Brad Wyble <bwyble at gmail.com> wrote:

> Thanks for your input John, I appreciate it extremely.  One of the points that I think is perhaps most important to improve on in our field is the scope of the predictions.  In physics, predictions seem to occur on a much grander scale, and very inspiring, and this itself engenders respect for the theoretical work.  
> 
> In neuroscience, the predictions are expected to be easily testable, which leads necessarily to predictions that are smaller in scope, and testable with technology that is already at hand.  
> 
> -Brad


Speaking of predictions, here’s one, regarding the veridicality of the representation of some metric properties of the world in brains. In several papers and one book, published between 1994 and 1999, I predicted (from certain mathematical properties of smooth functions) that the ensemble activity of graded responses of neurons in the primate inferotemporal cortex will be found to capture similarity patterns over families of 3D shapes. Here’s how the prediction was formulated in my 1998 BBS paper:


"1. The cell will respond equally to different views of its
preferred object, but its response will decrease with parameter-
space distance from the point corresponding the
shape of the preferred object (three such cells have been
reported by Logothetis et al. 1995).
2. The responses of a number of cells, each tuned to a
different reference object, will carry enough information to
classify novel stimuli of the same general category as the
reference objects.
3. If the pattern of stimuli has a simple low-dimensional
characterization in some underlying parameter space (as in
Fig. 6, left), it will be recoverable from the ensemble response
of a number of cells, using multidimensional scaling.”

Prediction #1 was really an explanation of some findings that just began to emerge when it was formulated. Predictions #2 and #3, however, were really about the future :-)
I am particularly fond of #3. A couple of years later, it was tested and found correct ("Inferotemporal neurons represent low-dimensional configurations of parameterized shapes”, Hans Op de Beeck, Johan Wagemans and Rufin Vogels, Nature Neuroscience 4:1244, 2001). Funny enough, that paper opened with this sentence: "Behavioral studies with parameterized shapes have shown that the similarities among these complex stimuli can be represented using a low number of dimensions.” — apparently, a theoretical prediction wasn’t good enough; they (or, more likely, the Nat Neuro reviewers) preferred to kick off an empirical finding, as if that had existed in a theoretical vacuum…

Unless I am mistaken in my interpretation of the stances expressed by different contributors to this thread, this kind of prediction is supposed, on at least one account, to be impossible/imprudent/whatever in trying to understand the brain. Yet, here it is: veridicality as a mathematically guaranteed generic property of neural representations. I guess sticking exclusively to Genesis at the expense of an occasional glimpse at what Prophets are up to may not be the most productive way ahead, after all ;-)

Cheers,

—Shimon

p.s. My papers are all available here:
http://kybele.psych.cornell.edu/~edelman/archive.html


Shimon Edelman
Professor, Department of Psychology, 232 Uris Hall
Cornell University, Ithaca, NY 14853-7601
http://kybele.psych.cornell.edu/~edelman
@shimonedelman