Abstract
Linear receptive fields (RF) are the basis of most vision models. However, they present important flaws: they change with the input, depend on the specific basis functions used to estimate them and contradict studies on dendritic computations. A novel effort has modelled spatial RFs non-linearly [Bertalmío et al., 2020, Scientific reports], giving rise to the intrinsically non-linear receptive field (INRF). Unlike linear RFs, which must vary with the input to predict responses, the INRF can remain constant with varying stimuli and is in better agreement with physiology. Here, we extend the INRF formulation to the spatio-temporal domain so that it can also predict responses to moving stimuli. We perform simulations using bars and drifting gratings, which show that the model successfully explains both basic and complex physiological and psychophysical motion perception phenomena. Specifically, the INRF model signals with a phase-independent response the direction of motion of drifting stimuli; it replicates results about saturation at high contrasts in striate cell responses [Heeger, 1992, Visual neuroscience]; it predicts motion perception reversals in the missing-fundamental and reverse-phi illusions [Adelson & Bergen, 1985, JOSA A]; it is able to explain the direction of motion of second-order stimuli; and it predicts the puzzling illusion of motion reversals that happen at short stimulus durations when a static low spatial frequency pattern is added to a moving high spatial frequency pattern [Serrano-Pedraza et al., 2007, JOV]. To develop predictions, classical motion energy models commonly use cascades of linear-nonlinear stages, while, in a simple way that is consistent with physiology, the INRF model uses nonlinear computations that cannot be decoupled from linear ones. This way, our model reproduces not only key phenomena in motion perception, but also findings that classical models can only explain by incorporating additional stages that increase their complexity and might not be biologically plausible.