猫视觉皮层中的感受野、双眼交互作用与功能构筑

Building on Stephen Kuffler's finding that retinal cells have concentric centre–surround receptive fields, and on their own 1959 study of the cat's striate cortex, Hubel and Wiesel recorded from single neurons in the primary visual cortex of lightly anaesthetized cats while projecting spots, bars and edges of light onto a screen in front of the animal. They classified each cell by how its firing depended on a stimulus's shape, orientation, position, movement and which eye received it. (Paraphrased structure follows; quotations are not reproduced here — read them at the source.)

One class of cell had a receptive field divided into distinct excitatory and inhibitory regions laid side by side in parallel stripes. Light falling in an excitatory region raised the firing rate; light in an inhibitory region lowered it; the two opposed each other. The best stimulus was therefore a line — a slit, a bar or an edge — at the one orientation and position that filled the excitatory stripe while sparing the inhibitory ones. The whole response could be predicted from the map of the field. Hubel and Wiesel called these 'simple' cells.

A second class was also tuned to orientation but had no separable ON/OFF map. A correctly oriented edge drew a response anywhere within a larger field, and moving the edge across it was especially effective. These 'complex' cells kept the orientation preference of simple cells but threw away the dependence on exact position.

Most cortical cells could be driven through either eye, usually with one eye dominant by a graded amount — a single-cell substrate for combining the two views into one, a prerequisite for stereoscopic depth.

Driving a microelectrode straight down through the cortex, the cells encountered in one vertical penetration tended to share the same preferred orientation and the same eye preference — orientation columns and ocular-dominance columns. Moving across the surface, the preferred orientation shifted in orderly steps. The cortex is not a blurred copy of the retinal image but a systematic map of features.

In the discussion they proposed a wiring scheme: a simple cell could be built by summing the outputs of a row of aligned centre–surround cells, and a complex cell by summing many simple cells of the same orientation but different positions — a feed-forward hierarchy that converts the retina's points of light into edges, and edges into position-tolerant detectors.