Research map / Architectures and efficiency / RNNs and normalization
L-SR1: Learned Symmetric-Rank-One Preconditioning
Gal Lifshitz, Shahar Zuler, Ori Fouks, Dan Raviv
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 |
|---|---|---|---|
| add secant penalty (PGSM) during meta-training (ablation) | L-SR1 · 3DPW test | PA-MPJPE (mm): 52.80 -> 51.74 | Improved |
| replace LGD with L-SR1 at 4 steps | learned optimization HMR · 3DPW test | PA-MPJPE (mm): 55.90 -> 51.74 | Improved |
| use optimal-horizon steps instead of 4 steps | L-SR1 · 3DPW test | PA-MPJPE (mm): 51.74 -> 51.58 (364 ms -> 1183 ms) | Improved |
| use element-wise instead of scalar learning rate (ablation) | L-SR1 · 3DPW | PA-MPJPE (mm): 51.80 -> 51.58 | Improved |
| use element-wise instead of scalar learning rate (ablation) | L-SR1 · Quadratic N=10 | mean inner loss: -0.86 -> 7.33 | Hurt |
| add secant penalty during meta-training (ablation) | L-SR1 · Quadratic test, buffer sizes 2/4/8 | test loss over first 20 iterations | Improved |
| increase hidden dimension from 64 to 128 (ablation) | L-SR1 · Quadratic validation | validation loss: -1.90 -> -2.56 | Improved |
| add p and q encoder inputs and d (all five inputs) (ablation) | L-SR1 · Quadratic validation | validation loss: -2.09 -> -2.56 | Improved |
| add p and q inputs to x,d,g (ablation) | L-SR1 · Quadratic validation | validation loss: -2.16 -> -2.56 | Improved |
| add d input to x,p,g,q (ablation) | L-SR1 · Quadratic validation | validation loss: -2.51 -> -2.56 | Improved |
| train on 10% instead of 80% of AMASS | L-SR1 · 3DPW | PA-MPJPE (mm): 51.67 -> 53.03 | Hurt |
| train on 1% instead of 10% of AMASS | L-SR1 · 3DPW | PA-MPJPE (mm): 53.03 -> 56.14 | Hurt |
Closest papers by meaning
- Layerwise LQR for Geometry-Aware Optimization of Deep Networks (NeurIPS 2026 · similarity 0.86)
- On The Concurrence of Layer-wise Preconditioning Methods and Provable Feature Learning (ICML 2025 · similarity 0.86)
- Solving hidden monotone variational inequalities with surrogate losses (ICLR 2025 · similarity 0.85)
- Riemannian Preconditioned LoRA for Fine-Tuning Foundation Models (ICML 2024 · similarity 0.85)
- Error Feedback Can Accurately Compress Preconditioners (ICML 2024 · similarity 0.85)
- AdaPreLoRA: Adafactor Preconditioned Low-Rank Adaptation (NeurIPS 2026 · 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.
