mathfoundationsnumerics
Numerical methods
Floating-point arithmetic, conditioning, iterative solvers — why training loops go NaN and how to prevent it.
Уровни глубины
L0Intro~3ч
Knows that floats have limited precision; reads IEEE-754 overflow/underflow symptoms.
L1Basics~15ч
Understands condition number, catastrophic cancellation; uses log-sum-exp and stable softmax.
L2Working~25ч
Implements Newton, conjugate gradient, GMRES; debugs NaNs with proper mixed-precision practice.
L3Advanced~40ч
Analyses backward stability; designs numerically robust neural-network kernels.
L4Research~80ч
Contributes to low-precision training theory or numerically stable attention kernels.
Ресурсы
L1 — Basics
L2 — Working