Research map / Architectures and efficiency / RNNs and normalization
Theseus: A Library for Differentiable Nonlinear Optimization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Replace the dense linear solver with sparse solvers, especially BaSpaCho (ablation) | Theseus on PGO · PGO Cube, batch size 128, 256 poses | forward+backward time: 20.81s (dense) vs. 2.25s (BaSpaCho) | Helped |
| Use BaSpaCho with batching and vectorization instead of Ceres | PGO optimizer · PGO, 2048 poses, batch size 256 | speedup over Ceres: ∼23x faster | Helped |
| Use Theseus rather than Ceres on small problem scales | PGO optimizer · PGO, 256 poses, batch size 16 | runtime: 25x faster | Hurt |
| Use implicit differentiation instead of unrolling (ablation) | Theseus backward pass · Tactile state estimation | validation loss | Helped |
| Use implicit differentiation rather than unrolling (ablation) | Theseus backward pass · Tactile state estimation | backward memory: ∼34MBs to ∼262MBs; ∼28MBs | Helped |
| Increase the number of inner-loop iterations for truncated differentiation (ablation) | Theseus truncated backward pass · Tactile state estimation | validation loss | Helped |
| Increase inner-loop iterations for direct loss minimization (ablation) | Theseus DLM backward pass · Tactile state estimation | validation loss | No clear effect |
| Use vmap instead of dense Jacobian computation (ablation) | AutoDiffCostFunction · Homography estimation, batch size 128 | Jacobian computation time: ∼337ms to ∼15ms; 22x speedup | Helped |
| Use vmap instead of dense Jacobian computation (ablation) | AutoDiffCostFunction · Homography estimation, batch size 128 | Jacobian memory: ∼9.6GBs to ∼1.2GBs | Helped |
| Use dense Jacobian computation instead of vmap (ablation) | AutoDiffCostFunction · Homography estimation, batch size 256 | memory feasibility: batch size 256 | Hurt |
Closest papers by meaning
- Differentiable Convex Optimization Layers (NeurIPS 2019 · similarity 0.85)
- Moreau Envelope for Nonconvex Bi-Level Optimization: A Single-Loop and Hessian-Free Solution Strategy (ICML 2024 · similarity 0.84)
- Prof. Robot: Differentiable Robot Rendering Without Static and Self-Collisions (CVPR 2025 · similarity 0.83)
- BPQP: A Differentiable Convex Optimization Framework for Efficient End-to-End Learning (NeurIPS 2024 · similarity 0.83)
- A Study of First-Order Methods with a Deterministic Relative-Error Gradient Oracle (ICML 2024 · similarity 0.83)
- FOSI: Hybrid First and Second Order Optimization (ICLR 2024 · similarity 0.83)
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.