Spiking network simulations reveal the robustness of gradient boosted trees to detect transmission delays independent of feature dynamics.

A The layer of PoSub integrate-and-fire neurons (red dots) receives one-to-one input from a mirrored layer of neurons which determines their primary angular tuning curves (right T(PoSub) in blue dots) and inputs from a layer of ADn neurons (left T(ADn) in blue dots) with full connectivity. The synaptic weight from T(ADn) to PoSub is proportional to the angular difference between the respective tuning curves of ADn neurons and PoSub neurons. B Simulation of 15 s of data. Top row, real HD value of one animal. Middle, raster of spiking activity of T(ADn) (top) and T(PoSub) (bottom). Bottom, membrane potential of the PoSub neurons. C Cross-correlograms between the spiking activity of 10 T(ADn) neurons and 10 PoSub neurons sorted according to the angular peak of their tuning curves. The angular difference between the preferred firing directions is color-coded (0 in red, π in blue). D Centered standard deviation of the cross-correlograms at normal (full green line) and accelerated angular speed (dashed green line). Synaptic transmission is set at 0 in these simulations. Black lines show the best exponential fits. E Same as D, but using XGB peer prediction of PoSub spiking activity from T(ADn) activity. Note that the distribution peaks at 0 ms for both angular speeds. F Characteristic time decays of the cross-correlogram exponential fits as a function of angular speed. G Full width at half maximum (FWHM) of cross-correlograms and XGB learning gain as a function of angular velocity. H XGB gains as a function of synaptic delays of transmission between T(ADn) and PoSub.