A multilevel stochastic-gradient neural solver for boundary integral equations
arXiv:2607.00560
2026
Training
1 ideas extracted · analyzed Aug 30, 2026
What the math gives to ML
The paper contains a transferable multilevel training principle: optimize an operator residual on a cheap coarse discretization, then warm-start the same network on progressively finer discretizations. Its mathematical explanation is that the empirical neural tangent kernel contracts residual modes at different rates, so fixed-resolution optimization can stall on slowly contracting modes while refinement exposes additional continuum modes. This suggests a practical continuation method for PINNs, neural operators, and boundary-integral networks in which discretization refinement replaces training from scratch at full resolution. The transfer is strongest when the loss is a discretized PDE, integral-equation, or collocation residual and the discretization can be refined without changing network parameters.
Ideas from this paper
Unverified
Re-invented
2026
Train an operator-residual network on a ladder of increasingly accurate discretizations instead of optimizing the finest residual from the beginning. At each level, reuse the previous parameters and introduce finer collocation or quadrature points only after the current residual has reached a plateau, allowing the network's spectral bias to act as a coarse-to-fine smoother.
Useful7/10
Difficulty4/10
Novelty6/10