M. J. Berry, D. K. WarlandPresent address: 1301 Orchard Park Circle, Apt. Y7, Davis, CA 95616 and M. Meister
MCB Department, Harvard University
Previous work has demonstrated that retinal ganglion cells respond to random flicker stimulation with highly precise periods of firing separated by intervals of complete silence (1, 2). Here, we explore what aspects of this signal dominate the transmission of visual information to the brain. Ganglion cell action potentials were measured extracellularly from an isolated tiger salamander retina using a multi-electrode array. The retina was driven by spatially uniform random flicker with a Gaussian intensity distribution. The entropy of a ganglion cell spike train and its mutual information with the visual stimulus were calculated by the method of S. Strong et al. (3). In brief, the local sequence of spikes was represented by a binary word, and the variability of such words across time was compared to the variability across multiple presentations of the same stimulus.
On average over 17 ganglion cells, the full spike train conveyed information at a rate of = 3.4 0.7 bits/spike (average 1 std. dev.). The coding efficiency, namely the fraction of the entropy used for information transfer, was = 0.54 0.07. To assess the significance of firing events, we considered a reduced description of the spike train, in which only the number of spikes in each firing event N and the time of its first spike T were recorded. The information rate for this reduced neural code was found to be nearly equal to that for the full spike train, = 0.95 0.03. Thus, the detailed spike times within an event contribute very little to transmission of visual information.
The identification of firing events makes it possible to distinguish information about what kind of visual stimulus pattern occurred from when it occurred. The ``when'' information is carried by the time of a firing event and the ``what'' information is conveyed by all other features of an event. To compare the two contributions, the event train was reduced even further, by retaining only the time T of each firing event without its spike count N. For this code, the information rate was still a large fraction of that for the full spike train, = 0.84 0.07. Thus ``when'' information dominates ``what'' by a factor of 5 under these stimulus conditions. In addition, the coding efficiency was higher for the event time = 0.65 0.09 than for event spike number = 0.44 0.13. Thus, the timing of firing events is, bit for bit, the most valuable response variable of these retinal spike trains.
Note: see related abstract at NIC 97 by Smirnakis, et al. Supported by a NRSA from the National Eye Institute to MJB, and a Presidential Faculty Fellow award to MM.
1. Berry, M. J., D. K. Warland, and M. Meister. 1996. The
Precision of Retinal Spike Trains. Soc. Neurosci. Abstr. 22:493.
2. Smirnakis, S. M., D. K. Warland, M. J. Berry, and M. Meister. 1996. Spike Bursts in Visual Responses of Retinal Ganglion Cells, Soc. Neurosci. Abstr. 22:494.
3. S. Strong, R. Koberle, R. R. de Ruyter van Steveninck, and W. Bialek, REntropy and Information in Neural Spike TrainsS, submitted to Phys. Rev. Lett.