When invisible noise obscures the signal: the consequences of nonlinearity in motion detection

摘要

Many models of neural processing consist of series of linear-nonlinear cascades, where at each stage inputs are pooled linearly and then undergo a nonlinearity such as squaring. One example is the motion energy model, the standard model of motion detection in animals from beetles to humans. Despite the energy model's nonlinearity, linear system analysis continues to be successfully applied in motion perception, as well as in other domains of visual neuroscience such as contrast and disparity. A critical assumption of many linear systems approaches is that noise injected at a frequency to which a sensory system does not respond has no effect on the system's ability to detect a signal. Even simple nonlinearities, as used in the energy model, mean that this assumption does not necessarily hold. We show that when early spatial filtering is lowpass, as in insect vision, the nonlinear nature of the energy model predicts that motion detection will be impaired by "invisible"' noise, i.e. noise at a frequency that elicits no response from the animal when presented on its own as a signal. We confirm this surprising prediction using the optomotor response of the praying mantis Sphodromantis lineola. Conversely when early filtering is spatially bandpass, as in mammalian vision, the effect does not occur and invisible noise has no effect. This means that masking techniques, which examine what frequencies of noise impair the detection of a signal, are able to reveal the sensitivity of motion channels in mammals but not insects. Thus, although the computations extracting motion have the same structure in both insects and mammals, differences in the early stages of visual processing produces radically different responses to noise. Counter-intuitive effects such as masking by invisible noise may occur in neural circuits wherever a nonlinearity is followed by a difference operation.