Receptive fields in visual cortex

Hubel and Wiesel were recording from cat visual cortex with tungsten microelectrodes and getting almost nothing, until the shadow cast by the edge of a glass slide, sweeping across the retina as they slid it into the ophthalmoscope, set a cell firing vigorously. The cells were not interested in spots. They were interested in oriented bars at a particular place in the visual field, preferably moving — and some of them moving in one direction only.

Starting from the retina

The plan was to carry a result from the eye into the brain. In 1953 Stephen Kuffler, shining small spots on the cat’s retina while recording from its ganglion cells, had found concentric receptive fields: an excitatory centre with an antagonistic surround, or the reverse, so that light over the whole field drives a cell less well than a spot just filling the centre — as Barlow, FitzHugh and Kuffler went on to show in 1957. Hubel joined Wiesel in Kuffler’s laboratory at Johns Hopkins in 1958, and the two meant simply to extend that work to the cortex, mapping fields with small spots. Their stimulator was Kuffler’s too: a multibeam ophthalmoscope designed by Samuel Talbot and Kuffler, which projected spots onto the retina from metal or glass slides pushed into a slot.

For about a month the cortical cells would not respond to spots or rings. The glass-slide cell came from an unusually stable recording that held it for nine hours. Several hours in, as a slide carrying a black dot went into the slot, the cell, in Hubel’s words, “went off like a machine gun” over the audio monitor. The dot was irrelevant; the cell was answering the line of shadow, and only within a narrow band of orientations. Hubel later judged it almost certainly a complex cell.

Their first full paper, in 1959, described fields divided into excitatory and inhibitory regions, as Kuffler’s were, but laid side by side rather than one inside the other; light covering a whole field drove most cells poorly, because the regions cancelled.

What they found

Simple cells respond to a line or edge of a specific orientation at a specific position, with adjacent excitatory and inhibitory subregions — essentially an oriented filter. Small spots can still map the subregions, which is how the fields were drawn; whether a slit of light, a dark bar or an edge then works best follows from how they are laid out.

Complex cells respond to the same orientation but tolerate position, as though pooling the outputs of several simple cells with the same preference at different places. The 1962 paper suggested exactly that wiring, as a hierarchy: simple cells possibly fed directly from the lateral geniculate nucleus, complex cells by simple cells within the same column.

And the arrangement is orderly: orientation preference varies smoothly across the cortical surface, cells in a column share a preference, and the whole structure is a map.

In the cat in 1962 the columns were about half a millimetre across and ran from the surface to the white matter. The smooth drift first showed itself clearly in a monkey: in a 1961 penetration angled at about 45° through the cortex, preferred orientation turned by about 10° for every 50 µm the electrode advanced.

In 1981 Hubel and Wiesel shared half of the Nobel Prize in Physiology or Medicine, for their work on how the visual system processes information; the other half went to Roger Sperry.

What it took to see it

This is neuroscience that existed only because the instrumentation existed. Resolving one cell’s response to a controlled stimulus needed microelectrodes with the impedance and noise performance to isolate a single unit, and amplifiers quiet enough to hear it — the same lineage as the patch clamp and microelectrode arrays.

The dependence is stronger than it first appears. Finding an orderly map required not just hearing one cell but hearing many, one at a time, holding the recording long enough to characterise each, and knowing where the electrode was to within a column. Every one of those is an instrumentation specification. Without them the cortex looks like noise, which is roughly what it looked like before.

The electrode was Hubel’s own, developed at Walter Reed for recording single cells in cats with chronic implants, awake and free to move, and described in Science in 1957: tungsten wire, sharpened electrolytically and insulated with vinyl lacquer. In those cats the tips were under half a micrometre across, with only the last 15–25 µm left bare for visual cortex, and sturdy enough to pass through the membranes over the brain weeks after implantation. Tungsten’s advantage was stiffness: hardened stainless steel, which he also tried, recorded just as well but was less stiff.

Holding a cell was a mechanical problem, and Hubel adapted his chronic rig to the acute experiments. A hydraulic drive on a peg fixed to the skull advanced the electrode, and the cortex was sealed from the air in a closed chamber, an idea of P. W. Davies, to damp the movement of breathing and pulse — steady enough, Hubel recalled, that only an earthquake would have shifted it. The 45 cells analysed in the 1959 paper were each held for 2 to 9 hours.

The input stage was a cathode follower, the valve version of a unity-gain buffer — in Hubel’s chronic rig, a design by A. F. Bak — and it fed a capacitor-coupled preamplifier. Two details in his methods are the ones a circuit designer would look for. The shield of the 20 cm cable from the electrode holder was connected to the cathode, so it moved with the signal; with almost no voltage across the cable’s capacitance, almost no signal leaked through it from a source of high impedance. And grid current, which flows through the electrode, was checked and kept small, because metal electrodes become noisy when it is excessive.

The audio monitor was part of the method. Between isolated cells, the 1962 paper describes an unresolved background of activity, heard on the monitor as a crackle, that an optimally oriented slit made swell and diffuse light left unchanged. It told them whether the cells near the tip shared an orientation, and because the retina was stimulated continuously as the electrode advanced, cells were found by their responses, not only when they fired on their own.

Position was settled afterwards, with the same electrode. Current passed through the tip left lesions 50–100 µm across, enough to place it to the nearest cortical layer — Hubel had worked the method out by watching coagulation at an electrode tip in egg white — and the other cells along a track were placed from depth readings corrected for shrinkage. When the electrode held two or three cells at once, told apart by the sizes of their spikes — spike sorting at its simplest — their fields invariably shared an orientation, one of the first signs that cells were grouped by orientation.

One specification pulls against the others. Regular orientation sequences were recognised as the rule only in the mid-1970s, probably, Hubel thought, because very oblique penetrations began only then; seeing them also takes electrodes coarse enough to record activity all along a track, not a cell every 100 µm or so. Isolating one cell and hearing the neighbourhood trade against each other at the tip.

What has aged

The measurements have held up better than the interpretation. Calling simple cells oriented edge detectors is now understood as too clean: responses depend heavily on stimulus context and on what the rest of the scene is doing, and natural images are not moving bars. The map is real; the filter story is a first approximation.

Hubel was already wary of the detector language in his Nobel lecture. The circuit that turns centre–surround input into orientation selectivity was still unknown, he said; simple and complex were not cleanly separated categories; and the cells should not be thought of as line detectors, since perceiving a line presumably depends on a whole set of them — a population, not a detector.

The map, too, turns out to belong to some brains rather than to visual cortex as such. Imaging thousands of cells individually with two-photon microscopy, Ohki and colleagues found in 2005 that maps in cat visual cortex are precise down to single cells, while in rat primary visual cortex strongly orientation-selective neurons sit with no discernible local order, neighbours often preferring different orientations.

The deeper reframing came from asking why these filters and not others — which is where efficient coding picks the question up, and answers it from the statistics of natural images rather than from anatomy.