Abstract
Purpose:
Theoretical work indicates that sensory encoding can be improved by appropriately adjusting noise correlations between pairs of neurons depending on the relative stimulus tuning of the pair, in effect introducing a system that minimizes discrimination errors for a given amount of noise. In particular, positive noise correlations improve discriminability for pairs with negative signal correlations and vice versa. We studied whether such a result is borne out in real neural systems. We measured spike correlations in pairs of On-Off Direction-Selective (DS) ganglion cells with various tuning relations and used statistical models of DS cell populations to assess the consequences of these spike correlations for direction encoding.
Methods:
Simultaneous spike responses to moving bar stimuli were recorded in pairs of DS cells from a flat mount mouse retina preparation. We investigated whether the relative tuning of a cell pair was predictive of its noise correlation. We then used a statistical model of eight DS cells (two cells for each of four preferred directions) that incorporated pairwise spiking correlations to compute various information theoretic measures of encoding fidelity. This model was used to assess the population encoding of directional information across a range of correlation strengths, including those observed experimentally.
Results:
We found that noise correlations in DS cell pairs did not depend strongly on tuning relation. In fact, we observed small positive correlations across all pairs. Surprisingly, statistical models of DS cell populations revealed that these observed correlations were near-optimal for encoding direction. We also observed a dependence of spiking correlations on firing rates, introducing a stimulus-dependence of correlations.
Conclusions:
The spike correlations measured in DS cell pairs are consistent with near-optimal encoding of direction, despite conflicting with the intuition based on cell pairs. The inability to extend optimal coding strategies for cell pairs to full populations stems from the interdependence of spiking correlations within a population and reflects a tradeoff among multiple beneficial spiking correlation strategies. Our modeling revealed that optimal encoding requires oppositely-tuned cells to have positive noise correlations. This prediction is confirmed by our data yet it remains unclear how DS circuitry achieves such an antagonism between signal and noise correlations.
Keywords: 531 ganglion cells •
473 computational modeling •
688 retina