Tag
A NeurIPS 2026 paper introduces topological out-of-domain generalization for dynamical systems reconstruction, identifying failure modes in hierarchical DSR models and fixing them with feature-splitting and physical sparsity priors, enabling correct prediction of bifurcations and regime changes without explicit control-parameter knowledge.
GradRepair-ODE introduces a reliability framework for certifying and repairing gradients in Neural ODE training to address numerical stability issues in scientific machine learning and generative models.
This paper introduces Response Renormalization, a backward-pass framework to improve training stability in Deep Equilibrium Models by addressing near-singular Jacobian issues, demonstrated across various multiphysics applications.
This paper evaluates five model families for macroeconomic forecasting, finding that less-constrained models like ARIMA and Neural ODE outperform structurally-prioritized models like PINNs, highlighting that structural priors can act as misregularizers when mismatched.
This paper presents a memory–stability–expressivity trilemma for trainable dissipative oscillator networks, showing that damping governs all three and limits trainability, with experimental validation on a 20-oscillator network confirming the theoretical bounds.
The author compares Structural Equation Modeling, Neural ODEs, and AI Programs like DSPy as declarative frameworks for defining and optimizing computational graphs, arguing that structured flows are essential for trustworthy AI agents.
This paper introduces Dynamical Physics-Modeled Neural Networks (DynPMNNs), a continuous-time deep learning architecture where hidden layers are defined by ordinary differential equations. It presents a biologically inspired approach grounded in Reproducing Kernel Banach Spaces, demonstrating competitive performance on the California Housing dataset with fewer parameters than standard Neural ODEs.