Automated Numerical Stability Analysis of Deep Learning Operators

The paper introduces a software tool that integrates CESTAC to analyze the numerical stability of deep learning operators. It is designed to validate numerical behavior in a single computation pass, identify sources of instability, and monitor stability during both training and inference. The authors report tests on polluted operators with injected numerical instabilities across various tasks. The stated goal is to support the development of more numerically stable and efficient deep learning kernels, where finite-precision arithmetic can cause approximation errors or amplify instability.