UC San Diego

The direction of visual motion is encoded in a population of direction selective retinal ganglion cells (DSRGCs) that provide inputs to the lateral geniculate nucleus (LGN). In this work, a novel method of in vivo two-photon calcium imaging is developed and used to determine how direction selective information is organized in the superficial LGN. Neurons preferring a single direction of motion within the superficial LGN predominantly represent the horizontal directions of motion, and a population of cells exists which responds to both opposing directions along the horizontal axis. A simple statistical model of retina-LGN connections demonstrates that random wiring within the superficial layer is sufficient to produce the observed fractions of direction and axis of motion preferring LGN neurons. The random wiring model is consistent with previous experimental results on the fraction of LGN neurons receiving a single driving input, and makes quantitative predictions about retina-LGN connectivity. The efficient coding hypothesis suggests that sensory neurons should be adapted to carry as much information as possible about the statistics of the environment. In a theoretical study, we reconcile this hypothesis with the organization and function of the retinogeniculate motion selective circuit. Under this theory, the symmetries of the distribution of optic flows in natural scenes constrain optimal direction selective neurons to prefer only the cardinal directions of motion. The directional tuning curves of On-Off DSRGCs in mice are compared with the optimal tuning curves for jointly encoding orthogonal directions of motion, and are found to correspond closely. The optimal encoding scheme for two opposing direction selective neurons, as observed in our study of the superficial LGN, requires sharpening of tuning relative the retinal representation. The theory's predicted sharpening corresponds to our own data on direction selective responses in LGN, and predicts that tuning curves that in a vertical motion layer in the LGN should be more broadly tuned than those in a horizontal motion layer