The concept of a feature detector developed under the influence of the discovery of ``bug detectors'' in the frog retina [Lettvin et al., 1959], and was linked to the notion of behavior-releasing mechanism, borrowed from ethology. It was quickly generalized to encompass higher perceptual functions such as shape recognition. A well-known proposal for an object recognition scheme based on feature detectors --- the Pandemonium [Selfridge, 1959,Lindsay and Norman, 1977] --- consisted of a three-level hierarchy: feature demons (responsible for the detection of lines, corners, etc.), cognitive demons (responsible for entire objects) and a master demon (responsible for the recognition decision; cf. Figure 8).
The proponents of the so-called computational view of vision (which should be more properly termed reconstructionist) did not seem to believe in the possibility of implementing any visual function of interest using feature detectors as building blocks (see [Marr, 1982], ch.7). The main objection leveled by Marr ( op. cit., p.341) was that ``the world is just too complex to yield to the types of analysis suggested by the feature detector idea.'' This objection, however, is being proved unfounded in more and more areas of visual perception.
As an example, consider the perception of coherent motion, a visual task intensively studied on all levels, including that of neurobiological implementation. In the middle temporal (MT) cortical area in the monkey, one can find cells with receptive fields tuned to coherent motion in a particular direction [Newsome and Paré, 1988]. The ensemble of activities of these cells may be regarded as representing the motion of the visual field seen by the animal, in a distributed non-reconstructionist sense. The ability of experimenters to bias the perceptual decision of the monkey by electrical stimulation of single MT cells [Salzman et al., 1990] indicates that the activity of an MT cell indeed constitutes a representation of the stimulus motion (for a philosophical perspective on this issue, see [Albright, 1991]). Specifically, because the activity of a given MT cell co-occurs with a certain well-defined motion in the visual field, and because artificial stimulation of the cell causes a behavioral response similar to the one precipitated by a real moving stimulus, the cell's firing for all practical purposes represents the motion event as far as the monkey is concerned. The visual motion, however, can hardly be considered as reconstructed in the activity of an MT cell. Moreover, it is not clear whether the pattern of activity of an ensemble of MT cells (or of cells in areas that precede MT in the M pathway) can or need to carry out such reconstruction.
How many of the motion detector cells are necessary to support perceptual performance similar to that of the entire organism? Newsome and his collaborators estimate this number to be about 30. Interestingly, combined responses of a similar number of shape-selective cells in the inferotemporal (IT) cortex in monkeys can be used to replicate the psychophysically defined threshold in a shape matching task [Miller et al., 1993]. These results vindicate Barlow's doctrine which assigned to single cells a central role in supporting perception [Barlow, 1972,Barlow, 1979a]. They also demonstrate the viability of feature-detection theories of sensory coding and of visual recognition, and account for the renewed interest in Selfridge's Pandemonium (discussed below, and again in section 5.1).