Adaptation leads toward optimal encoding of wide-field motion in a fly motion sensitive neuron

Visual motion estimation is one of the best studied examples of neural computation. In flies, these motion estimates are encoded by activity in a population of easily accessible wide-field motion sensitive neurons, the lobular plate tangential cells (LPTC). It is well known that responses of these neurons adapt to changes in stimulus statistics, that these responses can be decoded to give extremely precise estimates of velocity versus time, and that there is a regime in which mean responses exhibit biases that mirror perceptual and behavioral biases at low signal-to-noise ratio. These biases in turn have the form predicted for the optimal estimator given the measured statistical structure of the visual inputs. We try to bring these different observations together with new measurements on H1, a yaw sensitive LPTC. We use naturalistic inputs to map the neural responses and compare these with the output of the optimal motion estimator. We find that the adaptive behavior of H1 makes its responses more dynamic than that of a fixed-parameter correlator or gradient estimator and that it approximates the predictions of the optimal estimator across a wide range of conditions.