In what sense do �feature detectors� detect features?

Our perceptual system provides us with an internal representation of the world, constructed after the sense data are received through our sense accessories, where they are transduced by receptors into electrical impulses in a data-reducing, neural coding process culminating in the firing of action potentials. Debate about the development of our perceptual mechanisms can be split into the nativist-empirist camps, although most psychologists would now acknowledge at least some interaction between biological and learnt processes; alternatively, top-down or bottom-up, i.e. knowledge vs data-driven views of how we perceive objects, although again some interplay between the two seems necessary and can be demonstrated by �priming subjects�.

That visual perception is not an easy task did not become apparent until similar feats of pattern recognition were attempted with computers on images. Our visual system starts to shape the raw data even as an image is formed on our retinas. Marr outlined four major stages in vision processing, the first being a �grey-level description�, representing the intensity of light at each point in the retinal image in order to discover regions in the image and their boundaries.

A simpler example of delineating an object�s boundaries might be the recognition of a square. The theoretical explanation behind this involves a hierarchical feature net, where detectors for the simpler elements trigger detectors for the more complex elements. In this case, our feature detectors first recognise four straight lines of perpendicular orientations, then four right-angles, all of which together finally trigger the square detector which recognises the composite figure.

Before discussing what detection mechanisms exist, I want to first consider the basis for neural coding of the photoreceptors� input. The evolution of the eye can be seen as the culmination so far of a string of tiny steps all the way from complete insensitivity to light intensity. Like a camera, light streams in towards the retina, focused by the lens, with the level controlled by a diaphragm (the iris) and into a black-lined (the choroid coat, which forms the pupil at the front) area where it stimulates photoreceptive cells (120 million achromatic higher-sensitivity rods or 6 million trichromatic cones mainly found in the fovea).

These receptors are linked to the ganglion cells by bi-polar cells (which by an anatomical curiosity are in between the retina and the lens), with between 10 and 1000 receptors per ganglial cell, depending on how central the receptor is. Each of these groupings forms a �receptive field� (which was defined by Kuffler & Barlow (1953) as �that area of the retina such that light (or a pattern of light) falling within that area affects the pattern of firing�), with excitatory and inhibitory inputs being spatially summed.

Ernst Mach and Edward Hering provided evidence in the 19th century of lateral inhibition with their Mach bands:

Even though the brightness intensity of the middle section varies uniformly, our eyes perceive stripes of a particularly dark area of black and a particularly bright area of white on either side. This is due to on-centre/off-surround receptive fields, whose inhibitory rings overlap and are laterally inhibited by their neighbours, causing the spike rate to be particularly low on the dark edge and particularly high on the bright edge. By coding for change (in the case of lateral inhibition, spatial, but also temporal) the amount of data required to describe a scene is drastically reduced. Selective coding of information about pattern or the changes of light and dark at the retina (Ratliff & Hartine, Barlow) is an example of redundancy reduction.

Huber & Wiesel (1959) identified three types of cells, which respond to only very low-level, basic features, such as lines of differing orientations, lengths and directions:

1. simple cells	orientation-tuned separate excitatory and inhibitory regions
2. complex cells	orientation-tuned often directional-selective no separate excitatory or inhibitory regions
3. hyper-complex cells	orientation-tuned often directional-selective no separate excitatory or inhibitory regions selective for length (either one or both ends)

However, these on their own are not feature-detectors, since although they are selective to certain lines/bars, the responses of single cells are necessarily ambiguous, and modulated by a variety of different stimulus dimensions (e.g. length, position, luminance, contrast, wavelength, motion). Moreover, information about orientation is distributed coded, i.e. preserved in the pattern of firing across many cortical cells, rather than any single one.

Lettvin et al (1959) first provided evidence of a frog�s �bug detector�, certain ganglion cells in a frog�s optic nerve which respond intensely to a small, dark object which moves around within a certain retinal region. The obvious evolutionary explanation is of the frog�s need to rapidly identify insects flying about within range. However, the level of visual analysis varies, with higher animals such cats and primates having relatively simple retinal detectors and far more work being carried out in the visual cortex.

Perception is not simply a matter of reproducing a 2-dimensional image like a photograph, but rather extracting from it to form an internal representation of its 3-dimensional content. In humans, this appears to come about through a multi-layered system of redundancy-reduction at the retinal level, then simple and complex cells in the visual cortex, whose rate of firing in conjunction with each other builds up an increasingly sophisticated coding of the inputs contours, shapes and motions; this information is then processed with reference to memory to develop recognised patterns into objects and then scenes.