Input-space geometry shapes adaptation dynamics underlying repetition suppression in a neural network model of relatedness priming
A key function of the human brain is its ability to dynamically adapt to novel contexts by integrating prior experience. While neural adaptation is widely observed across cortical systems, the microcircuit-level mechanisms governing its intensity and temporal dynamics remain unclear. To bridge this gap, we develop a recurrent network of adaptive exponential integrate-and-fire neurons governed by a triplet spike-timing-dependent plasticity rule, designed to reproduce neural dynamics recorded via intracranial electrophysiology and single-unit recordings from the medial temporal lobe of neurosurgical patients performing a priming task. Using input organizations inspired by the experimental paradigm, we systematically vary input geometry to investigate its impact on adaptation dynamics. We find that neural adaptation emerges from recurrent dynamics shaped by learned connectivity, with both its magnitude and temporal profile depending on the geometry of the input space. This dependence gives rise, at the simulated single-neuron level, to a continuum of response regimes ranging from sharpening-like to fatiguing-like dynamics. Sharpening-like responses dominate when inputs exhibit high within-meta-category similarity and strong between-meta-category separation, whereas fatiguing-like responses emerge under the opposite regime.