distance dependent synaptic connectivity

The first time somebody bothered to add space

Since then, countless variants (Ermentrout, Bressloff, Kilpatrick, etc.) have used distance-tuned kernels for waves, bumps, grid-cell rings, you name it.


Do people usually include it today?

Model family Typical connectivity assumption Reality check
Point-neuron “vanilla” RNNs (e.g. Sompolinsky chaotic nets, most machine-learning RNNs) IID random weights, no spatial term Ignores distance completely because math is easier and GPUs don’t care about microns.
Cognitive-task recurrent rate/spiking models Often still random or low-rank + random; some add functional structure (low-rank modes) but not real geometry Distance left out unless the task itself is spatial.
Large-scale cortical simulations (Blue Brain, Potjans-Diesmann, NEST “hex-grid” demos) Frequently use an exponential or Gaussian fall-off for local axons; long-range patchy projections modeled separately Distance kept, because they’re trying to match anatomy.
Neural-field / bump / wave models Distance-dependent kernel is the whole point Always included; kernel shape is a free parameter.

A 2022 review of connectivity modelling notes that most population or abstract network studies still default to uniform random connectivity unless spatial layout is central to the question.    Even in modern papers that do add geometry, inhibition is usually made broader than excitation, but exact decay constants are rarely tuned from data.


Why it gets dropped so often