Treating Models as Learned Optimizers: Multi-start Strategies in Training
YouJiacheng · x · 2026-08-04
The post explores a new perspective of treating models as "learned optimizers" and introduces restart/multi-start strategies common in non-convex optimization into the training process.
- Core idea: Iterating on multiple samples during training and backpropagating off the one with the lowest loss.
- Theoretical context: The quoted tweet notes that traversing down the energy landscape to the global minimum synthesizes techniques used during inference in Energy-Based Models (e.g., EBTs, V-JEPA2), but applies them during training.
Related event: Exploring Reinforcement Learning as a Learning-Based Optimizer(3 posts)→
More from Research
- SimpleAI releases HiFi-UMI: high-fidelity data engine enables direct real-robot deployment without teleoperation data — 机器之心 · 2026-08-04
- Qianxun AI study: legacy data shows 'emergent transfer' after robot hardware upgrades, with a critical threshold — 机器之心 · 2026-08-04
- QuerySplat: Decoupling Geometry and Appearance in 3DGS Prediction — zhenjun_zhao · 2026-08-04
- UniSim-SLAM: Feed-Forward SLAM with Unified Sim(3) Optimization — zhenjun_zhao · 2026-08-04
- Target-Free Self-Calibration for Rolling Shutter Cameras — zhenjun_zhao · 2026-08-04
- UniqueSplat: View-Conditioned 3D Gaussian Splatting for Generalizable Reconstruction — zhenjun_zhao · 2026-08-04