Fig. 4: The behaviour of the monkeys is consistent with the dynamics of a CAN model.
a, Model schematic. A LM driven by external landmarks (Iext) interacts bidirectionally with grid-cell (GC) modules (m1, m2, …). The GC to LM input (Iint) has plastic synapses (‘+’). GC modules integrate motion through asymmetric centre-surround connectivity and velocity inputs (top). b, During visual navigation, GC–LM connections change. The synaptic drive from GC cells whose periodicity and phase match that of Iext is gradually strengthened (black dot: Iext periodicity and phase). The network maintains its selectivity after Iext is removed. c, Network state trajectory across 50 simulations under noisy velocity input, with (red) and without (black) landmark inputs (dotted white lines: reset events due to endogenous landmarks). d, Standard deviation (s.d.) grows linearly with temporal distance in the absence of landmark inputs and sublinearly in their presence (model: s.d. = a × meanb + c; H0: b = 1; H1: b < 1; one-tailed t-test(999) = −11.29, P << 0.0001). e,f, Two models of behavioural variability. In the model without reset (e), the s.d. increases linearly (e, inset). In the model with reset (f), it grows sublinearly (f, inset). Consequently, the distribution of produced temporal distances is wider for the no-reset model (top black Gaussian) than the reset model (top red Gaussian). g, The model with reset (ordinate) provides a better fit to the monkey’s behaviour compared with that without reset (abscissa) (paired t-test, monkey A: t(77) = 7.93, P << 0.0001, monkey M: t(101) = 16.56, P << 0.0001). h, Distribution of Fano factor periodicity of neurons with a significant Fano factor PI compared with their corresponding (Poisson) null data. Dotted lines denote the window in which the significance of the Fano PI was tested. See Extended Data Fig. 8 for details.
