Computational models reveal non-linearity in integration of local motion signals by insect motion detectors viewing natural scenes

Motion detection in animals and humans employs non-linear correlation of local spatiotemporal contrast induced by movement through the environment to estimate local motion. An undesirable consequence of this mechanism is that variability in pattern structure and contrast inherent in natural scenes profoundly influences local motion responses. In fly motion detection, this `pattern noise' is mitigated in part by spatial integration across wide regions of space to form matched filters for expected higher order patterns of optical flow. While this spatial averaging provides a partial solution to the pattern noise problem, recent work using physiological techniques highlights contributions to velocity coding from static non-linear spatial integration mechanisms (spatial gain control) and dynamic temporal gain control mechanisms. Little is known, however, about how such non-linearities co-ordinate to assist neural coding in the context of the motion of natural scenes. In this paper we used a simple computational model for an array of elaborated elementary motion detector (EMDs) based on the classical Hassenstein-Reichardt correlation model, as a predictor for the local pattern dependence of responses to a set of natural scenes as used in our recent work on velocity coding. Our results reveal that receptive field alone is a poor predictor of the spatial integration properties of these neurons. If anything, additional non-linearity appears to enhance the pattern dependence of the response.