asdex: Automatic Sparse Differentiation in JAX
asdex cuts sparse Jacobian and Hessian work in JAX by grouping independent derivative columns or rows into shared AD passes.
The paper describes automatic sparse differentiation as a four-step pipeline: detect the sparsity pattern, color the graph, compute compressed derivatives, then decompress them. Its core claim is that the number of AD passes can depend on the sparsity structure rather than the full matrix size. The authors present asdex as the first standalone ASD toolkit for JAX, with sparse drop-in replacements for JAX Jacobian and Hessian routines. ArXiv · AI/CL/LG's note
The paper describes automatic sparse differentiation as a four-step pipeline: detect the sparsity pattern, color the graph, compute compressed derivatives, then decompress them. Its core claim is that the number of AD passes can depend on the sparsity structure rather than the full matrix size. The authors present asdex as the first standalone ASD toolkit for JAX, with sparse drop-in replacements for JAX Jacobian and Hessian routines. ArXiv · AI/CL/LG's note
score 5