CAT LGN SPIKE VARIABILTY

Robert C. Liu

Sloan Center for Theoretical Neurobiology

University of California at San Francisco

Progress in understanding neural coding has generally relied on studying averages. The most basic approach is to simply look at the average spike response to a stimulus in a peristimulus time histogram (PSTH). Changes in the average spike rate that are associated with stimulus changes are then assumed to convey information. However, quantifying the information transmitted requires looking beyond simple averages and considering the variability of the neural responses. For a neuron to convey information about a stimulus, it should do so with a level of reliability and precision; both elements contribute to a higher capacity for transmitting information by reducing the amount of noise in the response. Using full-field, m-sequence (psuedo-random binary) stimuli, we have investigated the variability of neural responses in the lateral geniculate nucleus (LGN) of barbiturate anesthetized cats. This stimulus has the convenient property that sub-sequences (up to some limit) of all possible combinations of binary frames occur in equal number of times, making it useful for statistical characterizations of variability. For a specified number of frames, N, we can generate a PSTH matrix of responses to all possible N-frame sub-sequences, allowing us to visualize structure in the responses to different sequences. We can then quantify the spiking reliability and spike timing and count precision of these different responses. We have also employed the direct entropy method to calculate noise entropies and information rates, a more generalized measure of variability. The main conclusions are that the responses can be highly reproducible in their spike timing (~1.3 ms standard deviation) and spike count ~0.3 variance normalized by the mean) and the coding efficiency from these responses can be quite high (~67%). Moreover, in many cases, this level of precision can be achieved even when the reliability in spiking is low. Finally, simple models of spiking based on the temporal kernel predict much higher variability than what is observed.