What it means to “understand” the retina

The retina is a thin sheet of brain tissue (100 to 250 µm thick) that grows out into the eye to provide neural processing for photoreceptor signals (Fig. 17.1). In cats and macaque monkeys, it weighs about 0.1 g and covers 800 mm², about twice the area of a U.S. quarter (Packer et al., 1989). The retina includes both photoreceptors and the first two to four stages of neural processing. Its output projects centrally over many axons (1.6 × 10⁵ in cats [Williams et al., 1993]; 1.3 × 10⁶ in humans; and 1.8 × 10⁶ in macaques [Potts et al., 1972]), and analysis of these information channels occupies about half of the cerebral cortex (van Essen et al., 1992; Baseler et al., 2002). Because the retina constitutes a signifi-cant fraction of the brain (roughly 0.3%), to “solve” it completely would be a significant achievement for neuroscience. This overview considers what a “solution” would entail and summarizes progress toward that goal.

Figure 17.1..  

Radial section through monkey retina about 5 mm (∼25°) from the fovea. The synaptic layers span only 60 µm. Cone and rod inner segments are easily distinguished from each other, as are their terminals in the outer plexiform layer. Pigmented cells of the choroid layer (Ch) convert vitamin A to its photoactive form and return it to the outer segments. Pigmented cells also phagocytose membrane discs that are shed daily from the outer segment tips. OS, outer segment; IS, inner segment; ONL, outer nuclear layer; CT, cone terminal; RT, rod terminal; OPL, outer plexiform layer; INL, inner nuclear layer; IPL, inner plexiform layer; GCL, ganglion cell layer; B, bipolar cell; M, Müller cell; H, horizontal cell; A, amacrine cell; ME, Müller end feet; G_ON and G_OFF, ganglion cells. (Light micrograph by N. Vardi; modified from Sterling, 1998.)

First, we need the basic patterns of connection. The retina's two synaptic layers span only 60 µm, and most lateral processes span only tens to hundreds of micrometers (versus millimeters to centimeters in cortex). Therefore it has proved technically straightforward to identify, trace, and quantify the neurons and many of their synaptic connections. This reveals the retina to comprise about 75 discrete neuron types connected in specific, highly stereotyped patterns. Second, we need the “neurochemical architecture”. Although information is far from complete, the main neurotransmitters and their receptor types have been identified with the key cell types and synapses. Third, we need the basic response of each cell type. Because the intact retina can be maintained in vitro, a cell's light response can be recorded and the cell filled with dye to determine its morphology. This has permitted the neuroanatomical/neurochemical connections to be interpreted as functional “circuits” (Fig. 17.2).

Figure 17.2..  

The basic circuits that relay rod and cone signals through the retina to ganglion cells are known. Cone signals modulate ON and OFF cone bipolar cells (CB) that excite ON and OFF ganglion cells (GC). Rod signals modulate cone terminals via electrical synapse and relay single-photon signals via a private rod bipolar cell (RB) that excites the AII amacrine cell. The AII is bifunctional, inhibiting the OFF ganglion cell with glycine and exciting the ON ganglion cell via electrical synapse to the ON bipolar terminal. IPL, inner plexiform layer. (Modified from Sterling, 1998.)

Such circuits explain both intrinsic retinal mechanisms and also visual performance. For example, known circuits can explain reasonably well how a ganglion cell achieves its “center-surround” receptive field, and how one type of ganglion cell produces a linear, sustained response while a different type yields a nonlinear, transient response. Still other circuits explain how at night a ganglion cell manages to fire 2 to 3 spikes to a single photon, while by day it fires a similar number of spikes to an increment of 10⁵ photons. Finally, we know how different retinal circuits specialize for spatial acuity, motion, and opponent perception of hue (Calkins and Sterling, 1999; Wässle and Boycott, 1991).

Crossing from “How” to “Why”

Deep understanding of any brain region requires that we go beyond mechanism (how) to consider the computational purpose (why). For instance, why do we need a neural processor within the eye—since all other sense organs transmit spikes directly to the brain (Fig. 17.3)? And what explains the particularities of retinal design? For example, why are mammalian photoreceptors small, whereas in many cold-blooded species they are large? And why do photoreceptor and bipolar cells use presynaptic “ribbons,” since these are absent from the brain itself?

Figure 17.3..  

Only the visual sense requires neural processing at the site of transduction. The mammalian cone (upper left) requires lateral integration at its output (horizontal cells [H]), followed by 8 to 10 parallel circuits for a second stage (cone bipolar cells [CB]). Then, it requires more lateral integration (amacrine cells [A]) and finally, 10 to 20 parallel lines (four are shown; ganglion cells [G]) to carry action potentials to the brain. This chapter considers why photoreceptors require such extensive integration and so many parallel circuits before projecting centrally.

Many biologists dislike asking “why?” because it can be hard to prove that a given feature is truly adaptive, rather than merely decorative or a historical leftover of some evolutionary/developmental program (Gould and Lewontin, 1979). For an engineer, however, “why” is no problem. Every aspect of his design implies some specification of performance, constraints in cost (energy, materials, labor), and many compromises. If an engineer understands his design, he can explain exactly why his bridge was built a particular way (Glegg, 1969; Petroski, 1996). If he cannot explain this, stay off! So, the ability to answer ”why” measures the depth of our comprehension.

Another test of comprehension is to ask whether circuit components match optimally, a condition termed “symmorphosis“ (Diamond, 1992, 1993; Diamond and Hammond, 1992; Taylor and Weibel, 1981; Weibel, 2000). Where a neural system can be shown to satisfy the principle of symmorphosis across many levels (behavior ↔ circuit ↔ cells ↔ molecules), it can be chalked up as virtually understood. Many biologists do not care for this question either, reasoning that because evolution is ceaseless, how can we know whether a given feature has reached optimality?

But when a mechanism is shown to meet some physical limit, such as diffraction or diffusion, then natural selection has clearly hit the wall. And where several physical constraints conflict, neural design must reflect their compromise. In short, where actual performance approaches “ideal” performance calculated from physical limits, there is a genuine opportunity to address the “why” of a design. Although for most brain regions this is a distant goal, for mammalian retina such questions can now be addressed, and they provide the framework for this overview.

Consider that in nature the visual system operates near threshold. This is easily forgotten living under artificially bright light and viewing mostly high-contrast images, such as newsprint or the computer screen. But go bird watching or hunting (heaven forbid!), and you are quickly reminded that our ancestors strained to see the finest detail at the lowest contrast in the poorest light. To maximize sensitivity their eyes were selected to make each stage—from the optical image to the ganglion cell spike train—as efficient as possible. Thus each stage should approach the limits set by physical laws and by compromises required by the organism's “niche.” Thus every stage is a potential “bottleneck,” and the purpose at each stage must be to staunch the loss of information up to the physical limit. This hypothesis sets a framework for interpreting the functional architecture.

The central idea of this chapter is that the retina evolved to maximally extract information from the natural spatiotemporal distribution of photons and to convey this information centrally, with minimal loss. Upon this broad goal there are functional constraints: cover a wide range of intensities (10¹⁰); respond to very low contrasts (∼1%); integrate for short times (∼0.1 second); keep tissue thin (∼0.2 mm); and maintain the metabolic rate no higher. There are also basic constraints on biological computation: signal amplitude and velocities are set by properties of biological membranes and the speed of aqueous diffusion; accuracy and reliability of synaptic transmission are constrained by its quantal and Poisson nature. The retina's functional architecture reflects numerous compromises shaped by the interplay of these major factors as they contribute to the organism's overall success in its environment.

Why Natural Images Need Lots of Light

Both prey and predators try to merge with the background, so in nature contrast for significant objects tends to be low. Consider the bighorn sheep among the cottonwoods (Fig. 17.4A). The retinal image is represented as peaks and troughs of intensity that differ from the local mean by only ∼20%, and much fine structure exhibits far lower contrast, only a few percent (Fig. 17.4B). This range is common in nature (Laughlin, 1994; Srinivasan et al., 1982), and thus our visual threshold for a small stimulus, such as one spanning a single ganglion cell dendritic field, is ∼3% contrast (Watson et al., 1983; Dhingra et al., 2003).

Figure 17.4..  

How narrow-field and wide-field ganglion cell arrays “filter” the transduced image of a natural scene. A, Photograph of a bighorn sheep among the cottonwoods. Spatial detail is represented as peaks and troughs of intensity around some mean level (arrowheads mark the area scanned in B). B, Photometer scan across the middle of the image. Much discernible structure, for example, fine branches, differs from the mean by only a few percent. Were this scene viewed by a cat at 10 meters, 1 pixel would correspond roughly to one cone, and the intensity axis would correspond roughly to the signal amplitude across the cone array. Dimensions and spacings of the narrow and wide receptive fields are also indicated. C, Signal amplitude after filtering by narrow-field array. Subtraction by the surrounds of the shared signal component has reset the mean to zero: pooling by the centers has removed the noisy fluctuations. D, Signal amplitude after filtering by the wide-field cell array. Again, a zero mean, but the broad pooling and sparse sampling has removed all but the coarsest spatial detail—thereby clearing the wide-field cell dynamic range to efficiently encode motion. (Photograph by A. Pearlman; computations for B to D by R. Rao-Mirotznik and M. Eckert; modified from Sterling, 1998.)

To create an optical image at low contrast requires many photons (Rose, 1973; Snyder et al., 1977). Because light is quantized, a small difference from the mean, say 1%, implies that the mean itself must contain at least 100 photons. But photons at each image point arrive randomly in time (Poisson distribution). So even when an image is perfectly still on the retina, the intensity at every point varies temporally, with a standard deviation equal to the square root of the mean. Because the minimum detectable contrast (Δn) must differ from the mean by at least one standard deviation, the ability to detect a contrast of 1% implies a mean of at least 10,000 photons:

This root-mean-square fluctuation (√n) is termed “photon noise.”

One might think that daylight would provide plenty of photons to represent any scene. But this depends on the extent of photon integration: fine spatial detail implies limited spatial pooling and thus relatively large fluctuations from photon noise (Fig. 17.5A). This might be avoided by increasing temporal integration, but because mammals move swiftly, prolonged integration would blur the spatial image. Thus temporal integration is constrained to ∼100 msec (Schneeweis and Schnapf, 1995). Although daylight contains enough photons/100 msec to cast a fine image on the cornea, the excess is not large, nor does it extend to even slightly dimmer situations, for example, when a cloud obscures the sun. The need for intense light to register fine detail at low contrast partly explains why athletes, bird watchers, and the like do not wear sunglasses (Sterling et al., 1992).

Figure 17.5..  

The foveal receptor mosaic is optimized for spatial resolution and contrast sensitivity; the peripheral mosaic is optimized for temporal resolution by day and absolute sensitivity by night. A, Human fovea, radial view. Cone inner segments are narrow and gently tapered; outer segments are long and fine. B, Human fovea, tangential view through the base of the inner segments. Hexagonal cone packing provides the finest possible spatial sampling in daylight, but the absence of rods renders the fovea blind from dusk to dawn. C, Human, near periphery, radial view. Cone inner segments taper sharply and are surrounded by rod inner segments, which are much finer and untapered. D, Human, 20° nasal, tangential view. Large cone inner segments enhance sensitivity to high temporal frequencies, yet “spill” photons at night to surrounding rods. E to H, See text for explanation. IS, inner segment; OS, outer segment; ECS, extracellular space. (A–D, Video-DIC images from Curcio et al., 1990; E and F, Replotted from Packer et al., 1989. Used with permission.)