Review for Cognitive Psychology MS 99-068 Steven Lehar: Computational Implications of Gestalt Theory
The goal of this paper is to model the subjective experience of illusory contours, and also some brightness phenomena. The approach, we are told, is unique because it is unfettered by biological plausibility and it has feedback mechanisms as well as feedforward mechanisms. There are also claims to model brightness constancy and simultaneous brightness contrast. I think that the potential of using a formal/simulation model to understand this important perceptual phenomenon is great. However, in its present form, I can not recommend it for publication in Cognitive Psychology. There are three areas where the paper could be enhanced and therefore increase its appeal. First, given the goals of the paper, I was hoping for a theory that accounted for many aspects of subjective contours. However, I did not see a serious attempt to account for the literature on this phenomena.
Let me give two examples of the type of phenomena that I would like seen applied to the theory. Park and Pendergrass (Perception & Psychophysics, 1982) have created some interesting examples of subjective contours where there is no spatial frequency component in the orientation of the contours. I wonder if the model would detect the subjective contours in these stimuli. These stimuli might be invisible to the elongated spatial kernels in the model. A second phenomena that I wonder about is that if the subjective contours in a Kanizsa figure are replaced with thin luminance contours, the subjective brightening disappears. Would the present model predict this effect? There are a plethora of subjective contours phenomena, and I would want to know which the theory can account for, and which is can not account for. In other words, present many figures to the model and compare the output of the model with what observers perceive. Sometimes a models failures are as informative as its successes.
Second, the paper would be improved if it was clear in what ways this theory is unique. The theory seems to put together bits of older theories, like old wine, from various vintages, in a new bottle. For example, Marr tried to account for some of the same phenomenon, with similar mechanisms. The author claims to use "computational algorithms which might seem neurophysiologically implausible" (p. 4). I did not know which aspect of the model were implausible, nor why this is an advantage. One aspect of the present model is that it has feedback mechanisms. However, feedback mechanisms are not neurophysiologically implausible. My physiology friends tell me that backward projections in the cortex are the rule, not the exception. Hence, a theory without reciprocal projections might be implausible. Having said that, I was not sure how important reciprocal feedback was in the present model. Reciprocal feedback is not introduced until page 19 - after some pretty nice subjective contours are created. Hence, I was not clear as to how these particular feedback mechanisms affect perception.
Finally, I would like to know what aspects of the model are important to explain what aspects of subjective contours. For example, the first 12 pages of the article presents some fairly typical front end processing (DOG convolutions). Is this particular front end processing critical? As I mentioned above, it was not clear what aspect of the model was played by the feedback mechanisms. What would be the consequences if no such mechanisms were postulated.
In summary, there is potential here in increasing our understanding of subjective contours. However, the paper could be enhanced in several areas: (1) take the psychological data more seriously; (2) explain in what ways the model makes a unique contribution; (3) explain what aspects of the model contribute to what aspects of phenomenal experience.