Alexander Lyon¹ (), Oliver Layton¹; ¹Colby College

Optic flow provides rich information about the self-motion of an observer and the layout of the surrounding environment. Over the past several decades psychophysical studies have characterized the accuracy with which humans perceive their self-motion from optic flow and neurophysiological studies have linked self-motion perception to primate brain area MSTd. Despite this progress, most studies rely on optic flow generated from minimal dot environments that differs substantially from the rich and complex patterns encountered during real-world self-motion (Matthis et al., 2022). Moreover, naturalistic self-motion is not uniformly distributed; some types of motion are more common than others. For instance, studies on optic flow experienced by infants and their mothers reveal distinct asymmetries, such as higher rates of expansion compared to contraction (Raudies et al., 2012; Gilmore et al., 2015). In the present work, we examined the extent to which ecological optic flow and biases in self-motion statistics may shape the neural mechanisms underlying self-motion perception. Specifically, we explored how well deep artifical neural networks (DNNs) capture optic flow tuning in MSTd when trained on optic flow generated from minimal dot or realistic virtual environments. Using the 3D game engine Unity, we created large-scale video datasets of simulated self-motion through realistic cluttered warehouse and outdoor scenes with ground-truth labels. We found that training DNNs on optic flow from realistic environments improves their alignment with MSTd optic flow tuning properties. Similarly, incorporating the nonuniform self-motion statistics from Raudies et al. (2012) enhanced the consistency between the DNNs and MSTd, for example with respect to translation tuning preferences. Our work takes a step toward characterizing how ecological self-motion statistics and optic flow generated from realistic environments may shape the neural mechanisms that underlie self-motion perception.

Acknowledgements: The Colby College Provost's Office and Department of Computer Science