Research map / Learning theory and algorithms / Combinatorial and search optimization
SoftJAX & SoftTorch: Empowering Automatic Differentiation Libraries with Informative Gradients
Anselm Paulus, Andreas René Geist, Vít Musil, Sebastian Hoffmann, Georg Martius
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply straight-through estimation to each function separately before multiplying (ablation) | STE-wrapped relu product · relu(x)*relu(y) gradient | gradient informativeness | Hurt |
| apply straight-through estimation to the composite function instead of each factor (ablation) | STE-wrapped relu product · relu(x)*relu(y) gradient | gradient informativeness | Improved |
| use sorting network instead of other soft sort methods | sj.sort smooth mode · n=4096 runtime | time (ms): 1.0 ms vs SoftSort 16 ms, NeuralSort 37 ms | Improved |
| use FastSoftSort instead of matrix-based soft sorts | sj.sort smooth mode · n=4096 memory | peak memory: 420 kB at n = 4096 | Improved |
| use SmoothSort or OT-based sort | sj.sort smooth mode · n=4096 runtime | time | Hurt |
| use soft abs inside NeuralSort instead of hard abs | NeuralSort · smoothness | differentiability at ties | Improved |
Closest papers by meaning
- Efficient and Modular Implicit Differentiation (NeurIPS 2022 · similarity 0.86)
- Automatic Differentiation of Programs with Discrete Randomness (NeurIPS 2022 · similarity 0.86)
- Automatic Differentiation in Machine Learning: a Survey (JMLR 2018 · similarity 0.85)
- SoftSort: A Continuous Relaxation for the argsort Operator (ICML 2020 · similarity 0.84)
- Betty: An Automatic Differentiation Library for Multilevel Optimization (ICLR 2023 · similarity 0.84)
- Tangent: Automatic differentiation using source-code transformation for dynamically typed array programming (NeurIPS 2018 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
