The retina is a neural network at the back of the eyeball and constitutes the first stage of visual processing. It has long been established as one of the most popular model systems for studying neural coding. One primary reason for this is that the input and output of the system are well-defined and can be exquisitely controlled and monitored in experiments. The input is given by the light signals that fall onto the retina and are transduced by photoreceptors into electrical signals. The retina’s output is defined by the patterns of spikes produced by retinal ganglion cells, whose axons form the fibers of the optic nerve. These spike patterns encode all visual information available to the rest of the brain. Between photoreceptors and ganglion cells, a complex network of various neuronal types processes the visual information (Fig. ) and thus produces the neural code of the retina, which sets the stage for all further visual processing. The neural responses can be recorded by fine electrodes that are brought near or in contact with individual cells. In particular, the use of planar multielectrode arrays, onto which isolated retinas can be placed, has allowed the recording of spikes from many individual ganglion cells simultaneously while stimulating the retina by projecting light patterns onto the photoreceptors (
Meister et al., 1994;
Segev et al., 2004;
Petrusca et al., 2007). The multielectrode arrays not only increase the yield for recording activity of individual neurons, but also provide the means to study concerted neuronal activity by pairs and groups of neurons (
Meister et al., 1995;
Schneidman et al., 2006;
Shlens et al., 2006;
Pillow et al., 2008).
The precise nature of the neural code in the retina is the subject of extensive ongoing research. Several hallmarks of retinal neural responses, however, have been well characterized. One of the most striking features of spike trains recorded from retinal ganglion cells is the high degree of reliability and temporal precision. This is shown in Fig. for a ganglion cell that was recorded under repeated stimulation with the same sequence of spatiotemporal flicker. The responses are characterized by distinct spiking events, typically brief bursts of spikes at short intervals. Over stimulus repeats, the spike bursts are quite reliable in the number of spikes and display temporal precision in their onsets down to a few milliseconds (
Berry et al., 1997;
Uzzell and Chichilnisky, 2004). A second striking feature of retinal responses is that, when several ganglion cells are recorded simultaneously, one often finds synchronized spiking on various time scales and both in the presence and absence of visual stimulation (
Meister et al., 1995;
DeVries, 1999). These characteristics illustrate the wide range of possibilities by which retinal ganglion cell spiking could carry visual information: by rate, precise timing, relation to spiking of other cells, or any combination of these. With these general features in mind, let us return to the issues of short fixation periods during saccadic vision and high processing speed in the visual system. One requirement for the retinal neural code here should be that it transmits at least part of the visual information in a rapidly accessible way.
The responses of retinal ganglion cells to saccadelike stimulation have been the subject of a number of recent studies. First, there is the question of how ganglion cells respond to the global motion signals during the saccade itself. By playing natural movies that contain saccadelike sudden shifts of the full image, Roska and Werblin showed that specific types of ganglion cells in the rabbit retina are strongly inhibited during the saccade, but typically respond with bursts of spikes after the saccade is over (
Roska and Werblin, 2003). With these bursts, the cells may likely encode the newly encountered image. Other ganglion cell types, on the other hand, were found to respond with increased activity during the saccade, potentially signaling the features of these motion signals themselves. The diversity of responses by different types of ganglion cells was also emphasized by a study that characterized responses in the rabbit retina to sudden changes in mean luminance as can occur as a result of saccades (
Amthor et al., 2005). The induced changes in activity by the luminance steps ranged from strong suppression to strong activation, underscoring the complexity of visual information transmission in the presence of such dynamic stimuli.
The effect of saccades, however, goes beyond mere modulations of the level of activity; they can alter fundamental response features of single neurons. This was revealed by a study that showed how saccadelike stimulation can change the classical distinction of neurons into ON and OFF types. Typically, these types denote classes of neurons that primarily respond to onsets and offsets in light intensity, respectively. A study by Geffen et al., however, found certain retinal ganglion cells in the salamander retina that behaved like OFF-type cells without saccadelike stimulation and like ON-type cells right after a simulated saccade (
Geffen et al., 2007). Like most neurons in the early visual system, these cells respond primarily to light stimuli within a small spatial region, the neuron’s “receptive field.” The study monitored the response characteristics of these neurons to flickering light within their receptive fields while applying sudden shifts of a striped pattern in the spatial periphery. These stimulus shifts mimicked the global motion signals induced by saccades. In the absence of a saccadelike shift, the investigated neurons responded on average shortly after brief decreases in light intensity, as one would expect from OFF cells. But, for a short period immediately following the shift, these cells switched to responding when the light intensity in the center had shown a brief increase. Consequently, the message conveyed by a spike from such a neuron seems to be different at different times after a saccade. Elucidating the functional significance of this switch in the neural code is among the pending challenges for understanding the operation of the visual system in the presence of saccades.
Does the retina make use of its potential to produce temporally precise spikes for transmitting visual information after a saccade? For a particular subtype of ON-OFF ganglion cells in the turtle retina, this was investigated by stimulation with light intensity steps (
Greschner et al., 2006;
Thiel et al., 2006). The investigated cells reacted with two precise bursts of spikes whose timing depended on the applied stimulus. Whereas the first burst had a typical dependence on the light intensity step—it came with shorter latency for larger steps—the timing of the second burst, which followed some tens of milliseconds later, showed a surprising nonmonotonic dependence on the step size; minimal latency of the second spiking event was encountered for an intermediate light-intensity step. Moreover, the timing of both spiking events indeed contained substantial stimulus information. This followed from the fact that discrimination of the applied step sizes based on the ganglion cell spikes worked best when the spike counts as well as the latencies of both spiking events were taken into account (
Greschner et al., 2006).
Besides a change in light intensity, saccades typically also lead to substantial changes in the spatial pattern of the fixated image. It seems likely that visual systems have evolved to quickly transmit information about the details of the new spatial structure encountered after a saccade. That saccadic stimulation can play an essential role in spatial scene analysis has been demonstrated in a recent study of visual behavior and retinal coding in the archer fish (
Segev et al., 2007). These fish can accurately squirt water at insects above the water surface and thereby “shoot them down.” During the localization of the target, the fish’s direction of gaze is shifted by saccades. In recordings from ganglion cells, Segev et al. investigated how well targets of different sizes could be distinguished based on the cells’ spike trains. They found that the best distinction was obtained for the time period immediately following a saccade. The sudden changes in the image that the cells experience after the saccade activate them strongly and reliably, whereas responses during fixation were comparatively weak and unreliable and only slightly improved when small fixational eye movements were included. The authors conclude that saccadic vision provides the opportunity for the visual system to experience extensive global activation, thus acquiring “snapshots” of the environment.
This snapshot concept may be aided by the fact that several ganglion cell types are suppressed during the saccade as discussed above (
Roska and Werblin, 2003). The suppression could function as a reset of the message stream sent by the ganglion cells. The reduced activity may mark the beginning of a new signal transmission episode, distinct from previous episodes of the neural code. For reading out the neural code, this has the advantage of providing a specific context for each observed spike as being part of well-demarcated events. For example, this allows for robust signal transmission via the timing of the spike events at the beginning of fixation.
That the arrival times of the first spikes at the onset of a stimulus could form a powerful and rapid neural code for the visual system had been proposed and investigated in a series of theoretical studies (
Thorpe and Imbert, 1989;
Thorpe, 1990;
Gautrais and Thorpe, 1998;
VanRullen et al., 1998;
Delorme and Thorpe, 2001). These investigations were motivated by the observation that neurons in higher visual areas can show highly selective responses to visual stimuli already around 100 ms after the onset of the stimulus (
Perrett et al., 1982). Given that the visual information has to cross several neuronal stages in that time, it was argued that the required information transmission and computations must be based on few spikes and potentially a single spike per neuron at each stage (
Thorpe, 1990).
A recent experimental investigation showed that a population of ganglion cells measured in the salamander retina can indeed transmit considerable amounts of information by the latencies of their first spikes (
Gollisch and Meister, 2008). This study analyzed how the retina can encode a visual image when it is only briefly displayed. To this end, ganglion-cell spike trains were recorded in response to flashed images. Again, cells responded with bursts of spikes shortly after the onset of the image. The fastest responses came from specific types of ON-OFF cells that also displayed high precision in the timing of the first spike. Most importantly, however, these cells responded with spike bursts of nearly identical numbers of spikes to very different images, whereas the latency of the first spike could differ by about 40 milliseconds. Examples of these responses are shown in Fig. . The precise and systematic shifts in spike timing may constitute a powerful channel for information transmission. Indeed, it was shown that, for the vast majority of recorded neurons, the first-spike latency provided considerably more information than the number of elicited spikes. For these cells, a latency code is thus a reliable source of information. Moreover, it also represents a particularly fast code—the information is already available with the first spike, thus providing a potential substrate for rapid image processing.
The same study also considered the relative timing of the first spikes from pairs of ganglion cells and found that this provided an even better source of information. Information in relative spike timing could be directly read out by downstream brain regions, for example, through delay lines and coincidence detection. By contrast, information contained in the spike timing of a single cell alone requires an additional reference signal, which—in the case of saccades—could be supplied by a corollary discharge that accompanies the relevant motor commands (
Sommer and Wurtz, 2002). It is still not quite clear, however, to what extent the brain uses the corollary discharge of eye movements.
A code based on relative spike timing may have other advantages. In the case of salamander retinal ganglion cells, it was shown that this relative timing is to a large degree independent of visual contrast, thus providing direct information about the image structure independent of illumination conditions (
Gollisch and Meister, 2008). Furthermore, the relative timing was surprisingly robust; fluctuations in relative timing, measured over repeated stimulus presentations, were often smaller than for the spike times of the individual cells. In other words, the first-spike times of simultaneously recorded neurons were often strongly correlated; when one of them happened to fire a bit earlier in a given trial, the other tended to follow suit. Different mechanisms can be envisioned that may underlie these correlations. Neighboring ganglion cells can be coupled through electrical gap junctions, and these can have a synchronizing effect in the millisecond range (
Brivanlou et al., 1998;
Hu and Bloomfield, 2003). Alternatively, shared input into ganglion cells from bipolar or amacrine cells may result in correlated neuronal activity (
Mastronarde, 1983;
Levine, 1997;
Murphy and Rieke, 2008). Regardless of the mechanism, the intercellular correlations in spike timing underscore the potential for temporal coding in the early visual system.
This article has been 
LIKE STIMULATION
