Receptive fields in visual cortex

Cortical cells respond not to spots of light but to oriented edges at particular positions, and to progressively more abstract features further along the pathway.

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 screen as they changed it, set a cell firing vigorously. The cells were not interested in spots. They were interested in oriented bars, moving in a particular direction, at a particular place in the visual field.

What they found

Simple cells respond to an edge of a specific orientation at a specific position, with adjacent excitatory and inhibitory subregions — essentially an oriented filter.

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.

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.

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.

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.

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.