As for your question about Marr’s levels of analysis, you are of course correct in saying that the first level (computational) will dictate the second (algorithmic). This is also true of the second level’s relationship to the third by the way. I think what Marr was getting at was that by distinguishing these three levels, we can be very clear in our minds which level we are trying to explore or explain the system at. This can avoid much mental confusion. Nobody is saying that each level is completely independent of the others, but you can imagine that at each level there is a certain degree of wiggle room, or alternative possiblities. In the article I used the example of motion detection as a computation to be performed by the visual system. The algorithm used to achieve this task could be selected from any number of candidates, like the Reichardt detector, or maybe some higher order statistical technique. The eventual neural implementation will of course be determined by the chosen algorithm, but again there may be more than one candidate neural circuit to carry out the task. A concise description of this idea which I like is at: http://www.nici.kun.nl/~peterh/doc/marr_levels.html

Regarding the 3 stages, wouldn’t the nature of the computations performed be determined by the algorithm used? I mean, you wouldn’t compute something if the idea you’re using doesn’t need it, right?