Machine learning discovers numerous new computational principles supporting elementary motion detection.

Motion direction detection is a fundamental visual computation that transforms spatial luminance patterns into directionally tuned outputs. Classical models of direction selectivity rely on temporal asymmetry, where motion detection arises through either delayed excitation or inhibition. Here, I used biologically inspired machine learning applied to retinal and cortical circuits to uncover receptive field architectures capable of direction selectivity. These include mechanisms based on asymmetric synaptic properties, spatial receptive field variations, new roles for pre- and postsynaptic inhibition, and previously unrecognized kinetic implementations. Conceptually, these circuit architectures cluster into eight computational primitives underlying motion detection, four of which are previously undescribed. Many of the solutions rival or outperform classical models in both robustness and precision, and several exhibit enhanced noise tolerance. All mechanisms are biologically plausible and correspond to known physiological and anatomical motifs, offering fresh insights into motion processing and illustrating how machine learning can uncover general principles of neural computation.