Propagation of local parameter estimation errors in large-scale simulations

Ascertain how estimation errors in local parameters (such as synaptic conductances and membrane currents) produced by an ultrastructure-to-dynamics compiler compound in large-scale neural circuit simulations, and determine whether the non-catastrophic error accumulation observed in artificial neural networks under low-precision quantization extends to biophysical brain simulations.

Background

The compiler will inevitably produce some estimation error in local parameters. In artificial neural networks, similar errors often do not catastrophically degrade performance, but it is unknown whether this robustness applies to biophysically detailed brain models.

Understanding error propagation is critical for evaluating whether compiled parameters can support reliable circuit- or brain-scale simulations.

References

We don't know how such errors would compound. We know that in neural networks compounding is not catastrophic (and we can e.g., transfer a neural network to lower bit-depth65) but it is unknown if this would generalize to brains.

— Compiling molecular ultrastructure into neural dynamics  (2603.25713 - Kording et al., 26 Mar 2026) in Appendix 1: potential limitations — Modeling and scaling limits (Compounding errors)