Recognition: unknown
Forward-Mode Automatic Differentiation in Julia
read the original abstract
We present ForwardDiff, a Julia package for forward-mode automatic differentiation (AD) featuring performance competitive with low-level languages like C++. Unlike recently developed AD tools in other popular high-level languages such as Python and MATLAB, ForwardDiff takes advantage of just-in-time (JIT) compilation to transparently recompile AD-unaware user code, enabling efficient support for higher-order differentiation and differentiation using custom number types (including complex numbers). For gradient and Jacobian calculations, ForwardDiff provides a variant of vector-forward mode that avoids expensive heap allocation and makes better use of memory bandwidth than traditional vector mode. In our numerical experiments, we demonstrate that for nontrivially large dimensions, ForwardDiff's gradient computations can be faster than a reverse-mode implementation from the Python-based autograd package. We also illustrate how ForwardDiff is used effectively within JuMP, a modeling language for optimization. According to our usage statistics, 41 unique repositories on GitHub depend on ForwardDiff, with users from diverse fields such as astronomy, optimization, finite element analysis, and statistics. This document is an extended abstract that has been accepted for presentation at the AD2016 7th International Conference on Algorithmic Differentiation.
This paper has not been read by Pith yet.
Forward citations
Cited by 4 Pith papers
-
Stochastic Differential Dynamic Programming for Trajectory Optimization under Partial Observability
A new stochastic differential dynamic programming method optimizes coupled trajectory design and orbit determination under partial observability, producing navigation-aware solutions with lower fuel consumption than d...
-
Admissible Lax-Wendroff Flux Reconstruction Method with Automatic Differentiation on Adaptive Curved Meshes for Relativistic Hydrodynamics
An admissible Lax-Wendroff flux reconstruction method with automatic differentiation and subcell blending enables robust high-order simulations of relativistic hydrodynamics on adaptive curved meshes.
-
Propagating data-driven galaxy redshift distribution uncertainties in 3$\times$2-pt analyses
A five-parameter PCA model for n(z) uncertainties in Stage-IV 3x2-pt analyses degrades the S8 constraint by only 5% relative to shift-stretch models while halving biases on Omega_m and sigma_8, and all tested models a...
-
Sensitivities of Black Hole Images from GRMHD Simulations
Differentiable GRMHD image sensitivities create a structured error landscape that supports gradient-based parameter recovery for black hole imaging under idealized and noisy conditions.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.