Published in: Michael A. Arbib (Ed.): The Handbook of Brain Theory and Neural Networks. Bradford Books/MIT Press, 1995.
email: rob@retina.anatomy.upenn.edu
At the most basic level, the retina transduces light intensity signals over space and time and transmits them to the brain. However the signal transmitted by the retina does not code intensity directly. Instead, a variety of neuron types transform visual signals in a multitude of ways to code properties of the visual world such as contrast and motion. The purpose of this chapter is to develop a conceptual theory, based on the actual problems faced by neurons, to explain why the retina codes visual signals and how the structure of the retina is related to its coding function.
The vertebrate retina reliably responds to light contrast as low as 1% (Shapley and Enroth-Cugell, 1984). Yet as the delicate visual signal is amplified in its passage through the retina from photoreceptors to the ganglion cell, the biological limitations of neural processing add distortion and noise with every neuron. The ease with which we see fine gradations in the presence of such biological limitations suggests the hypothesis that a major function of intricate retinal circuits is to maintain the signal's quality in spite of the limitations. This hypothesis would predict that the much of the retina's signal coding and structural detail is derived from the need (intrinsic to the retina) to optimally amplify the signal and eliminate noise.
The retina is a thin (100-200 um) tissue at the rear surface of the eye
consisting of 3 layers of neurons and glial cells (Figure 1; see Dowling,
1987; Rodieck, 1988; Sterling, 1990). Neurons in the "outer nuclear layer"
(ONL) are exclusively photoreceptors. The "inner nuclear layer" (INL) (i.e.
the middle layer) contains the cell bodies of horizontal cells (H), bipolar
cells (B), and amacrine cells (A). Between these two layers lies the "outer
plexiform layer" (OPL) in which bipolar and horizontal cells extend
dendritic processes laterally to receive synaptic contacts from
photoreceptors. The innermost cell layer, called the "ganglion cell layer"
(GCL), contains cell bodies of ganglion cells and amacrine cells. Between
the INL and GCL lies the "inner plexiform layer" (IPL) where bipolar,
amacrine and ganglion cells are synaptically connected. Ganglion cells
send their output to the brain through axons that lie on the inner surface
of the retina.
Each class of neuron described above comprises several cell types, and overall the retina comprises several dozen (Sterling, 1990). A cell type is defined by a distinctive morphology, distribution, and synaptic connection pattern (Rodieck, 1988), or distinctive physiology or immunocytochemical staining pattern. Although the retina of one species may contain cell types not present in another, the same 5 retinal cell classes exist in all vertebrate species (Kuffler et al., 1984; Dowling, 1987; Sterling, 1990). Therefore all vertebrates likely share similar neural circuit organization.
To understand the physiological function of a retinal neuron, investigators often measure its "receptive field" (the region in space and time over which it responds to light with a change in voltage). Receptive fields of retinal neurons consist of a circular region in visual space to which the neuron is most sensitive, called the "center", and a larger but weaker antagonistic region concentric with the center, called the "surround" (Kuffler et al., 1984). The receptive field of a neuron is determined by the receptive fields of neurons presynaptic to it and the signal-filtering properties of its input synapses, as well as signal- processing performed inside the neuron. For example, a ganglion cell's receptive field must reflect not only its own morphology and membrane channels but receptive field properties of the bipolar cells that contact it, and these in turn must originate to some extent in the receptive field properties of all the photoreceptors, horizontal cells and amacrine cells presynaptic to bipolar cells.
While receptive field analysis is a powerful method for studying the function of a neural circuit (see Kuffler et al, 1984), the origin of a receptive field's components is difficult to grasp in a circuit that includes several layers of neurons, even when studied by modern engineering analysis methods (Shapley and Enroth-Cugell et al, 1984). The problem seems to be the difficulty of relating components of a physiologically-based model to the details of cell morphology, synaptic connectivity, and membrane channels (e.g. see RETINAL DIRECTION SELECTIVITY). However, by computationally simulating these biophysical details, it is possible to test specific hypotheses about neural circuit connectivity with only partial knowledge of the circuit's detail (Teeters and Arbib, 1991; Freed et al., 1992). The analysis of retinal circuitry has been also extended to an information-theoretic approach (Atick and Redlich, 1992; Laughlin et al, 1987).
Since neurons of each cell type are spaced in a regular array across the retina (see Figure 1), one might hope to decompose the arrays of different types into a repeating structure that would correspond to a "functional module", i.e. a small neural circuit duplicated over the retina that contains several neuron types and performs a specific signal-processing function such as generating a ganglion cell's receptive field. The problem is that each cell type is distributed with a different density, so individual neurons in one array cannot easily be grouped with those of other arrays on the basis of their proximity (Rodieck, 1988; Sterling, 1990). The key to identifying functional modules is, of course, their synaptic connections. By tracing each neuron's presynaptic circuit in a series of electron micrographs, an essentially complete circuit of the cat beta ganglion cell was reconstructed (Sterling, 1990).
The outer segment (OS) of a vertebrate photoreceptor transduces light via a multi-step biochemical cascade (Liebman, Parker, and Dratz, 1987) into an electrical signal that is conducted through the photoreceptor's axon to its terminal in the OPL. In response to a flash of light, the photoreceptor closes channels in its outer segment, i.e. it hyperpolarizes. Over a limited range of stimulus intensity the change in the light- modulated signal is proportional to the change in the light stimulus. The advantage of this coding function is that a photoreceptor responds well to small intensity differences, i.e. low contrast signals, that are common in the visual world. The disadvantage is that at intensities outside this limited range the photoreceptor responds poorly. At low intensities, the photoreceptor's transduction gain (i.e. proportion of change in its output signal to a change in input) is insufficient, and at high intensities, the photoreceptor's response saturates. To solve such saturation problems a photoreceptor continuously adjusts the intensity range to which it responds, in a process called "adaptation". In some species, adaptation in a photoreceptor can modulate transduction gain by up to 4 log units.
Photoreceptors comprise two classes, rods and cones (see Rodieck, 1988). They differ in that rods are sensitive to single photons and are bleached by daylight (Rodieck, 1988), but cones are about 2 log units less sensitive and can regenerate their pigment in daylight (photopic intensity range). At twilight (the mesopic intensity range), cones do not respond well, so rods are coupled via gap junctions to their neighboring cones, causing the rod signal in twilight to pass directly into cones where it is carried by the lower-gain cone pathway (Daw et al, 1990). For the low intensity range encountered at night (scotopic intensity range), a special "rod bipolar" pathway (RB in Figure 1) carries quantal "single photon" signals, removes dark noise, and adapts over an extra 3 log units of intensity (Sterling, 1990).
The axon terminal of a cone transmits its signal to bipolar cells with a chemical synapse which increases signal gain at the cost of adding extra noise and reducing the intensity range over which the bipolar cell can respond (Laughlin et al., 1987). Here we consider how the function for the OPL circuit might be related to the synapse's limitations.
To reduce the tendency of the cone signal to saturate its synapse, the OPL filters the signal both spatially and temporally (Laughlin et al, 1987). The filter consists of 2 components, a spatial low-pass filter constructed from lateral electrical connections, and a spatio-temporal high-pass filter constructed by horizonal cells. The low-pass filter is constructed by electrically coupling a cone directly to its neighbors with gap junctions which tends to remove uncorrelated noise from the cone's response. Although the coupling also causes "neural blur", this is useful to provide an "anti-aliasing" filter for the next stage of processing in the IPL.
The high-pass filter is constructed by subtracting a "local average" from the cone by negative feedback. Horizontal cells sum inputs from many cone terminals and provide negative feedback to each via a GABA-ergic synapse (Sterling, 1990). The synaptic structure that performs this function, called a "triad", is complex because it is both "feedforward" and "feedback", and thus is termed a "reciprocal" feedback connection (Dowling, 1987). This type of coding has been termed "predictive" (Srinivasan et al, 1982) because the ideal signal to subtract from each cone would be a "local mean" signal produced by averaging signals from the immediately surrounding cones over a short time interval.
The glutamatergic synapse that transmits a cone's signal to bipolar and horizontal cells adds noise originating in the random fluctuation of synaptic vesicle release rate (Kuffler et al., 1984; Sterling, 1990). To reduce the amount of noise relative to the signal, the vesicle release rate needs to be high. Therefore the chemical synapse of the photoreceptor contains a special structure, called a "ribbon", that functions as a docking site and reservoir for vesicles (Kuffler et al, 1984; Rodieck, 1988; Sterling, 1990) allowing a high vesicle release rate.
Mammalian retina contains about 10 types of bipolar and more than 20 types of amacrine cell (Sterling, 1990; Kolb et al, 1981). Dendritic trees of bipolar cells arborize in the OPL to receive contacts from multiple photoreceptors, and their axons terminate in the IPL. Amacrine cells, in contrast, make no connections in the OPL and extend their dendrites in the IPL to contact bipolar cells, other amacrine cells, and ganglion cells.
Bipolar cells respond to light as photoreceptors do with a voltage proportional to the intensity change, but their response range is narrower and they adapt over a wider range of stimuli. Bipolar cells contact ganglion cells with glutamatergic synapses, so they employ the same synaptic ribbon structure found in the photoreceptor to allow high release rates and reduce noise. A bipolar cell may contact several ganglion cell types, each with different numbers of synapses (from 2 to several dozen), and this implies a specific coding of the bipolar signal (Sterling, 1990; Teeters and Arbib, 1991).
The need for several bipolar cell types may be related to a need for different coding requirements by different types of ganglion cells. For example, some bipolars respond transiently with a 20-50 msec. transient at the beginning of a light flash, while others are slower responding. The circuitry unique to a bipolar cell type therefore might define a special coding property such as chromatic, velocity or direction sensitivity. This specialization increases signal/noise ratio and reduces distortion which is an advantage for a visual signal that is destined to pass through another synapse.
Amacrine cells are a diverse group in both morphology (Kolb et al, 1981) and neurochemistry (Masland, 1988; Rodieck, 1988). Many amacrines have a large (1-2 mm) but sparse dendritic field with very fine dendritic processes (0.2 um dia) that look much like axons (Kolb, et al, 1981; Dowling, 1987). In contrast to the generally passive membrane properties of bipolar cells, these amacrine cells fire action potentials which allows them to transmit signals laterally over the extent of their large dendritic field (Masland, 1988).
Amacrine cells are generally either GABAergic (Kuffler et al, 1984; 1992; Rodieck, 1988) or glycinergic, which implies that they perform subtractive or shunting control functions. Some, e.g. the cholinergic "starburst" amacrine, are involved in temporal processing, and respond transiently to light (Masland, 1988). Amacrine circuitry is thought to be responsible for directional selectivity in ganglion cells (see RETINAL DIRECTION SELECTIVITY), excitatory transient and peripheral effects, and several types of gain control (Shapley and Enroth-Cugell, 1984; Dowling, 1987).
Amacrine cells receive synaptic contacts from bipolar cells at a structure called a "dyad" where a bipolar ribbon synapse contacts two postsynaptic cells, either a ganglion cell and an amacrine cell, or 2 amacrine or ganglion cells (Rodieck, 1988). The similarity between the synaptic "dyad" in the IPL and the "triad" in the OPL is striking. Both contain synaptic "ribbons" and both include reciprocal feedback from a lateral neuron. The reason may be the identical problem of noise. The reciprocal feedback from an amacrine varicosity to the bipolar cell that feeds it can spatio-temporally filter the bipolar signal, reducing the signal's range before transmission to ganglion cells (Masland, 1988; Dowling, 1988).
The salient feature of a ganglion cell's response to light is its exquisite sensitivity to low contrast stimuli over a wide range of light intensity (Kuffler et al, 1984). Ganglion cells are specialized into a diverse set of cell types that code different properties of the visual world (see RETINAL DIRECTION SELECTIVITY, Kolb et al., 1981; Maturana et al, 1960; Rodieck, 1988). Some give a "tonic" response to stationary stimuli (e.g., the X cell of cat retina), and others give a more "phasic" response to signal the presence of flashing or moving stimuli (e.g., the Y cell of cat retina). Many species (lower vertebrates but also mammals) possess ganglion cells with more complex receptive fields, e.g. they respond only to small moving objects that could be insects (Maturana et al, 1960; Teeters and Arbib, 1991). In some species (e.g. primates, turtles) "color-opponent" ganglion cells provide excellent color coding (Rodieck, 1988; Dowling, 1987). In most cases, summation of inputs by a ganglion cell is linear, so a "nonlinear" ganglion cell is dependent on its presynaptic amacrine and bipolar circuitry to produce complex properties (Freed et al., 1992).
To transmit a signal to the brain, the ganglion cell codes its intracellular voltage as a firing rate of action potentials along its axon (Kuffler et al., 1984). This process is limited by noise and dynamic range in a manner similar to synaptic coding. Noise in the "spike generator" of ganglion cells causes the action potential frequency to vary, which can obscure the ganglion cell's signal at low firing rates. To cope with the problem of noise, the retina contains two subclasses of ganglion cell, called "on" and "off". These cells respond with opposite polarity to a light stimulus, the "on" cell increasing its firing rate and the "off" cell reducing its rate. Responses of on- and off-ganglion cells in many species are symmetric which allows the retina to code bright and dark objects without much distortion.
To supply on- and off-ganglion cells with appropriate signals, the IPL is organized into on- and off- sublaminae. To split the light signal into two symmetrically opposite responses, two bipolar cell subclasses respond oppositely to the neurotransmitter (glutamate) released by cones. The number of bipolar and amacrine cell types are divided roughly equally between on- and off- sublaminae, although some arborize (e.g., an "on-off amacrine") in both.
An on-bipolar contains in its dendritic membrane "metabotropic" receptors, which, when bound by glutamate released by a photoreceptor, signal a cytoplasmic second-messenger system after a short time delay to turn off the synapse's ionic channels (Sterling, 1990; Dowling, 1987). Thus the on-bipolar depolarizes when the photoreceptor decreases its glutamate release (i.e. in response to light). An off-bipolar contains ionic channels that open when glutamate binds to the channel's receptor site. The on- and off-bipolar cells therefore have the proper responses to directly excite their respective ganglion cells, i.e. the on-bipolar depolarizes to a bright spot and the off-bipolar depolarizes to a dark spot.
There are several reasons for the diversity of retinal circuitry. Although all ganglion cell circuits receive the same information from the OPL, by discarding part of the information a ganglion cell can specialize in coding one or more special properties of the signal, i.e. contrast, motion, bright, dark, or colored light flashes, etc. The exact details of coding are probably related to the ecological niche occupied by the organism. Rod signals because of their quantal nature are qualitatively different from cone signals over a range of more than 3 log units so there is an advantage in a separate rod pathway. Such specialization in coding increases the signal/noise ratio and makes better use of the limited dynamic range of neurons, synapses, and the spike train in the ganglion cell axon (Srinivasan et al., 1982). Specialization in coding also simplifies the task of brain circuitry in visual segmentation (Atick and Redlich, 1992; Kuffler et al, 1984), so it appears that the brain's need for special coding contributes to the retina's circuit structure.
Yet receptive fields of many retinal neurons, and ganglion cells in particular, share important properties: their center-surround organization, high contrast sensitivity, and wide-ranging adaptation. The high sensitivity of the retina is achieved at the cost of complexity: to the extent that each retinal circuit amplifies the signal, it must adapt to reduce the signal's dynamic range, and this implies circuit complexity. For example, the net effect of the OPL circuit is to create for the photoreceptor a receptive field with a broad center region and a wide antagonistic surround (Sterling, 1990, 1992) that adapts temporally and spatially. By removing information about absolute light intensity, the OPL circuit transmits what is left, i.e. information about contrast. In turn, the IPL circuit removes more information about absolute light intensity, accentuating the center-surround receptive field in bipolar and amacrine cells, and in the process regulates its gain to prevent saturation at high contrast (Shapley and Enroth-Cugell, 1984). Thus it appears that circuits along the retinal pathway all contribute to the ganglion cell's receptive field properties for a similar reason: to prevent noise or saturation from degrading the signal.
The well-known antagonistic center-surround and adaptation properties of the ganglion cell receptive field, therefore, seem less designed for the brain's ulterior purposes than to preserve the quality of retinal signals. To accomplish this, the circuitry of both OPL and IPL increase the lateral extent of center and surround, but the need for high visual acuity mandates that OPL and IPL circuits not extend laterally too far. Thus the retina is shaped by a compromise between spatial acuity and accuracy of coding in the need to compensate for its limitations.
Although knowledge of the biophysical components of retinal circuitry and its receptive fields is progressing at an accelerating pace, such knowledge does not guarantee a useful theory. For example, the biophysical properties and visual responses of bipolar and amacrine cells presynaptic to a ganglion cell are currently being measured with whole- cell patch recordings, and these presynaptic responses presumably correspond to components of the ganglion cell's receptive field. Yet such knowledge alone cannot answer the question of function in design: why the individual components exist. The answer can only be derived from synthetic models that integrate details of the retina's neural circuitry with the noise and dynamic range limitations inherent to neural biology.
Computational modeling promises to help find the answers (Teeters and Arbib, 1991; Freed et al, 1992). Once the basic signal flow and function in a retinal circuit have been established, a series of simulations can determine what biological limitations are most serious to the circuit under different conditions. The effect of synaptic noise on the retina's performance can be tested by simulating noise of various types to the signal pathway and comparing the resulting signal/noise ratios as a measure of signal quality.
ACKNOWLEDGEMENTS Thanks to V. Ciaramitaro and Drs. M. Freed, K. Morigiwa, N. Vardi and L. Palmer for their comments. Supported by MH48168.
Atick, J.J. and Redlich, A.N., 1992, What does the retina know about natural scenes?, Neural Comp., 4:196-210.
*Daw, N., Jensen, R.J., and Brunken, W.J., 1990, Rod pathways in mammalian retinae, Trends Neurosci. 13:110-115.
*Dowling, J.E., 1987, The Retina: An Approachable Part of the Brain, Harvard University Press, Cambridge, MA.
Freed, M.A., Smith, R.G., and Sterling, P. (1992) Computational model of theon-alpha ganglion cell receptive field based on bipolar cell circuitry. Proc. Nat. Acad. Sci. 89: 236-240.
Kolb, H., Nelson, R. and Mariani, A., 1981 Amacrine cells, bipolar cells, and ganglion cells of the cat retina: a Golgi study. Vision Res. 21:1081-1114.
*Kuffler, S.W., Nicholls, J.G., and Martin, R.A., 1984, From Neuron to Brain: A Cellular Approach to the Function of the Nervous System, Sunderland, MA: Sinauer Associates Inc.
Laughlin, S.B., Howard, J., and Blakeslee, B., 1987, Synaptic limitations to contrast coding in the retina of the blowfly Calliphora., Proc. R. Soc. Lond., B 231:437-467.
*Liebman, P.A., Parker, K.R., and Dratz, E.A., 1987, The molecular mechanism of visual excitation and its relation to the structure and composition of the rod outer segment, A. Rev. Physiol., 49:765-791.
*Masland R H., 1988, Amacrine cells, Trends Neurosci, 11:405-410.
Maturana, H.R., Lettvin, J.Y., and McCulloch, W.S., 1960, Anatomy and physiology of vision in the frog (rana pipiens). J. Gen. Physiol, 43:129- 175.
*Rodieck, R.W., 1988, The Primate Retina, Comp. Primate Biol., 4:203- 278.
*Shapley, R.M., and Enroth-Cugell, C., 1984, Visual adaptation and retinal gain controls, Progr. in Retinal Res., 3:263-346.
Srinivasan, M.V., Laughlin, S.B., and Dubs, A., 1982, Predictive coding: a fresh view of inhibition in the retina, Proc. R. Soc. Lond., B 216:427-459.
*Sterling, P., 1990, Retina, in The Synaptic Organization of the Brain, Third ed., (Gordon M. Shepard, Ed.), New York: Oxford Univ. Press.
Teeters J. L. and Arbib, M. A., 1991, A model of anuran retina relating interneurons to ganglion cell responses, Biol Cybern., 64 3:197-207.