Direction selectivity in rabbit retina

In the 1960s the neurophysiology of vision was starting to move forward from a static view of organization based on the compendium of structural information accumulated by the great pioneering morphologists (Cajal, 1892; Polyak, 1941). Neurophysiologically based pathway tracing and topographic projection mapping still had plenty of momentum, but the ground was being quietly prepared for change. Retinal ganglion cells could do ‘arithmetic’ on spatially distributed light stimuli (Hartline, 1940); or switch between cones and rods during dark adaptation (Granit, 1944); or subtract as well as add by using inhibition (Barlow, 1953; Kuffler, 1953). The idea that the retina might be doing much more than the earlier results had uncovered came from a paper with the enigmatic title: ‘What the frog's eye tells the frog's brain’ (Lettvin et al. 1959). An influential insight was that the classical stationary flashed spots of light for analysing ganglion cell action might be entirely ineffective with certain classes of ganglion cells.

Far into the future of that time were the tools of today's laboratories: isolated retina and slices, patch-recording, multicoloured labelling of interrelated subsets of retinal neurons, the extraordinary range of modern pharmacological agents and the unimagined scope of genetic engineering.

The classical paper introduced by this Classical Perspectives article appeared in the mid-1960s (Barlow & Levick, 1965). Serendipitously, the experimental work was based on rabbit rather than cat as the model mammalian retina. Until that time, radially symmetric centre-surround receptive fields (Kuffler, 1953) were regarded as the fundamental spatial filter through which signals reached higher centres. There was a surprise waiting! For some cells, hand-controlled scanning of a light spot across the receptive field (small sensitive patch in the external visual field) yielded a strong response in one direction, but silence for the reverse pass. This was a striking departure from the radially symmetric responses of a classical mammalian ganglion cell. Control experiments soon ruled out mundane optical explanations or peculiarities in the overall shape and arrangement of ON- or OFF-response regions. To the surprise of many, here was a mammalian retina extracting a complex visual property, namely the direction in which objects move in the visual scene. Previously, that capability was thought to require the elaborate circuitry of the visual cortex (Hubel, 1963).

Soon the basic visual properties of direction-selective (DS) ganglion cells (GCs) were worked out. These included: receptive field maps with mixed transient ON- and OFF-responses; polar directional response profiles of cardioid shape; robustness of directional property despite reversal of contrast and variations in speed, shape and size of stimuli. ON–OFF DSGCs constitute a complete system in that there are subclasses having their maximum response (‘preferred’) directions narrowly clustered around the principal visual directions: up, down, nasal, temporal. Every local region had representatives of the four subclasses, so the coordinated local responses would unambiguously encode direction of motion everywhere.

Why did this paper on mechanism become such a durable reference point for more than 40 years? Worth consideration is the conceptual representation of the DS process that was inspired by the experimental observations. The DSGC receptive field appeared to be built from an array of sequence-selective subunits, each responsive to the same sequence-order of stimulation. The deceptively simple block diagram below shows the inferred circuit fragment.

Input channels A, B, C, … provide both direct (straight down) and asymmetric lateral inhibitory connections to extensively replicated subunits (A.∼B′, B.∼C′.) that are combined by the DSGC to form the output. Separate paired mechanisms are needed for ON- and OFF-pathways. Primed symbols (B′, C′) refer to delay and stretching of signals. Subunit symbolism means: ‘Output is TRUE if at A it is TRUE that stimulus occurred AND at B it is NOT TRUE that stimulus occurred at earlier time’. Showing subunit action as a logical operator reflected the observation that targets moved in the null direction gate the output to zero. The idea was that visual perception and machine vision systems are analogous. Many algorithms in the latter utilize logical operations for efficient computation; why not perception also?

A notable innovation was the use of stationary masks so that the view of a moving edge was restricted to a slit of various widths or paired narrow slits at different separations. The effects on responses led to inferences on upper and lower bounds for the size of a subunit. Handcrafting the set-up allowed rapid reconfiguration of test situations. Direct manual control while listening to the impulses on a loudspeaker enabled efficient discovery of the most informative parameters. Thus results were distilled from a considerable variety of trials. Perhaps this conferred durability on the arguments constructed from them.

Provocatively, a generic term ‘Trigger Feature’ was introduced to refer to the class of complex stimulus patterns for which certain sensory neurons are selective. This later sparked friendly rivalry with colleagues espousing analysis of visual perception in terms of the spatio-temporal frequency content of scenes. A creative wit dubbed it ‘feature creatures versus frequency freaks’, but the issue is not really ‘either/or’.

Over the years, many colleagues have commented that reading the paper inspired the choice of a career in the neurophysiological analysis of visual mechanisms. It is a special privilege to have had the good fortune of refreshing that tradition.

Original classic paper

The original classic paper reviewed in this article and published in The Journal of Physiology can be accessed online at:

DOI: 10.1113/jphysiol.2006.120220

http://jp.physoc.org/cgi/content/full/jphysiol.2006.120220/DC1