(11–14), and salamander (15,16). Each of these studies concluded that ganglion cells were not independent in their behavior. In particular, the spike trains from different ganglion cells were strongly correlated in absence of any visual stimulus, for example, in complete darkness or constant uniform illumination. Nearby cells tended to fire synchronously much more frequently than expected by chance, as evidenced by a central peak in the correlation function of their two spike trains (Fig. 1A). Such positive correlation was found primarily if both cells were ON type or both were OFF type. Two cells of opposite response type were often anticorrelated and tended to avoid generating simultaneous spikes.
In a remarkable study of pairwise recordings from cat retina, Mastronarde (12–14) identified three separate sources of correlations among the spike trains of two ganglion cells: Quantal fluctuations in a shared photoreceptor led to a broad peak or valley in the correlation function, with widths around 50 ms. These were observed only in darkness or very dim illumination, presumably requiring the high gain of phototransduction achieved by the dark-adapted retina. Shared synaptic input from a spiking neuron, possibly an amacrine cell, produced fast correlations on the time scale of 5–10 ms. These occurred at all light levels, provided that the receptive field centers of the two ganglion cells overlapped. Finally, ganglion cells appeared to be coupled to each other by gap junctions, and occasionally triggered each other with spike delays of <1 ms. Taken together, these correlations affected a large fraction of retinal activity—e.g., the rapid correlations from shared spiking input alone accounted for 80% of the maintained activity of Y cells. However, the experiments were all performed under constant illumination, and, thus, it remained unresolved whether such synchronous firing occurs during visual stimulation and how it might affect the transmission of information from the retina to the brain.
It is now feasible to record simultaneously from many retinal ganglion cells by placing the isolated retina on a flat array of metal microelectrodes (17). In this way, we have recently analyzed the concerted activity among ganglion cells in the tiger salamander retina (15,16). The efficiency of monitoring a large number of cells and the long lifetime of the amphibian preparation have allowed a more thorough statistical analysis of concerted firing patterns, as well as an assessment of their role in visual signaling. During spontaneous activity in darkness, nearby ganglion cells had a pronounced tendency to fire synchronously, within 10–20 ms of each other (Fig. 1A). This degree of synchrony is highly significant, and one can quantify its strength by the “correlation index”: the observed number of synchronous spike pairs (with delays <20 ms) divided by the number expected if the ganglion cells fired independently. The correlation index was found to depend strongly on the distance between the receptive field centers of the two neurons (Fig. 1B), decreasing from a maximum of ≈20 for neighboring cells to 1 at a separation of 0.4 mm. At greater distances, up to 1 mm, the correlation index dipped significantly below 1, indicating that distant neurons tended to avoid firing synchronously. The phenomenology of these correlations is remarkably similar to the fast pairwise correlations observed in cat retina (12).
A higher-order analysis revealed that the effects extended beyond pairwise synchrony (M. J. Schnitzer and M.M., unpublished results): larger groups of ganglion cells were found to discharge simultaneously. Such firing patterns involved up to seven neurons in an experiment that monitored about 10% of the ganglion cells overlying the electrodes. These groups of cells were usually not nearest neighbors, but dispersed over the retina within the 0.4-mm distance determined from the pairwise analysis. Each ganglion cell could participate in several such stereotyped firing patterns. Altogether, synchronized firing among two or more cells accounted for more than 50% of all recorded action potentials. This probably underestimates the overall importance of concerted firing since only a fraction of all ganglion cells was observed.
The narrow width of the correlation peak centered near zero delay (Fig. 1A) suggests that groups of synchronized ganglion cells share excitatory input from another spiking neuron: its action potentials could trigger the follower ganglion cells simultaneously with high reliability. This shared input may originate from a spiking amacrine cell (12), and each stereotyped firing pattern among ganglion cells may represent the postsynaptic field of such an amacrine cell. Alternatively, synchronized ganglion cells could be coupled to each other via gap junctions, a possibility that is less plausible (16), though not entirely ruled out by the available evidence. For the remainder of this article, an origin in spiking amacrine cells will be taken as a working hypothesis, although the basic conclusions do not depend on that assumption.
Concerted firing persisted when the retina was driven by various visual stimuli, such as periodic flashes, traveling gratings, and a randomly flickering checkerboard (16). The shape of the pairwise correlation functions and the distance dependence of the correlation index remained essentially unaltered under visual stimulation. The same stereotyped multineuronal firing patterns were found as in darkness, but their rate of occurrence was strongly modulated by the stimulus. We de-