Answers are grounded in this paper and cite their sources.
Direct physics-informed training of spiking networks with genuinely binary internal spikes is widely reported to be infeasible, with the blame on surrogate gradients. On viscous Burgers we localize the failure: on a hard-threshold quadratic integrate-and-fire (QIF) backbone the classical u_xx strong form is unstable, its loss climbing seventeen orders of magnitude, while the identical u_xx computed by central differences trains. Rewriting the PDE as two first-order flux equations with a pointwise closure restores trainability under the same surrogate autodiff. Conservative central differences of the network's own forward passes match it on shock formation, train from scratch on one CPU, and are decisive on shock propagation, with six times lower error over three seeds. Autodiff loses the front by more than a shock width at comparable residuals. We show that the conservation residual is blind to the gauge q -> q + h(t), so the closure is mandatory. Removing it drives the loss 300x below the honest optimum on badly wrong solutions. The core hypotheses and pass/fail criteria were pre-registered and later diagnostic arms carry pre-run records. Falsifications are reported as findings.
No takes yet. Share an insight, caveat, or question.
Gullapalli et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: