Research map / Architectures and efficiency / RNNs and normalization
Transformer-Based Learned Optimization
Erik Gärtner, Luke Metz, Mykhaylo Andriluka, C. Daniel Freeman, Cristian Sminchisescu
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 |
|---|---|---|---|
| Use Optimus instead of DiffPhy's BFGS optimizer | DiffPhy human-motion reconstruction · Human3.6M validation | MPJPE-G: 139.1 -> 138.6 | Improved |
| Use Optimus instead of DiffPhy's BFGS optimizer | DiffPhy human-motion reconstruction · Human3.6M validation | MPJPE: 82.1 -> 82.8 | Hurt |
| Use Optimus instead of DiffPhy's BFGS optimizer | DiffPhy human-motion reconstruction · Human3.6M validation | MPJPE-PA: 55.9 -> 57.0 | Hurt |
| Use Optimus instead of DiffPhy's BFGS optimizer | DiffPhy human-motion reconstruction · Human3.6M validation | Foot skate: 7.2 -> 6.5 | Improved |
| Use Optimus instead of DiffPhy's BFGS optimizer | DiffPhy human-motion reconstruction · AIST dance subset | MPJPE-G: 150.2 -> 149.8 | Improved |
| Incrementally update the preconditioning matrix across iterations (ablation) | Optimus · Human3.6M validation | MPJPE-G: 36.5 -> 33.1 | Improved |
| Use the structured learned preconditioner rather than directly predicting the parameter update (ablation) | Optimus · Human3.6M validation | MPJPE-G: 37.3 -> 33.1 | Improved |
| Use the structured learned preconditioner rather than directly predicting the parameter update (ablation) | Optimus · Human3.6M validation | Runtime (ms): 10.7 -> 78.9 | Hurt |
| Increase optimization to a fixed 50 function evaluations without stopping early (ablation) | Optimus · Human3.6M validation | MPJPE-G: 33.1 -> 28.5 | Improved |
| Add BasinHopping perturbations (ablation) | Optimus · Human3.6M validation | MPJPE-G: 33.1 -> 25.6 | Improved |
| Increase the Adafactor MLP model size from 128x4 to 512x4 | Adafactor MLP · Human3.6M validation | MPJPE-G: 40.6 -> 40.0 | Improved |
| Use Optimus instead of Adafactor MLP | Human-motion reconstruction optimizer · Human3.6M validation | MPJPE-G: 40.0 -> 33.1 | Improved |
Closest papers by meaning
- Linear Transformers are Versatile In-Context Learners (NeurIPS 2024 · similarity 0.85)
- Unraveling the Gradient Descent Dynamics of Transformers (NeurIPS 2024 · similarity 0.84)
- Investigation into Training Dynamics of Learned Optimizers (Student Abstract) (AAAI 2024 · similarity 0.84)
- Preconditioners for the Stochastic Training of Neural Fields (CVPR 2025 · similarity 0.84)
- Linear Transformers Implicitly Discover Unified Numerical Algorithms (NeurIPS 2025 · similarity 0.84)
- Transformers Learn to Achieve Second-Order Convergence Rates for In-Context Linear Regression (NeurIPS 2024 · 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.
