Research map / Reinforcement learning / Policy gradient and actor-critic
Spotlight: Optimizing Device Placement for Training Deep Neural Networks
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 policy gradient with PPO-based Spotlight algorithm for device placement | device placement RL · CIFAR-10 (2 GPUs + 1 CPU) | training time to 87% accuracy: 70 min vs 60.9% of GRL time | Helped |
| Replace policy gradient with PPO-based Spotlight algorithm | device placement RL · NMT and RNNLM (4 GPUs) | per-step training time: RNNLM 3.15 to 2.27 s; NMT 4.74 to 3.62 s | Helped |
| Replace policy gradient with PPO-based Spotlight algorithm | device placement RL · NMT end-to-end (12000 iterations) | end-to-end training time: 12.5 h vs 16 h (21.9% reduction) | Helped |
| Spotlight placement vs synchronous towers on 4 GPUs | device placement RL · CIFAR-10 (4 GPUs + 1 CPU) | training time: 58 min vs 95 min (~40% better) | Helped |
| Spotlight vs GRL in sample efficiency on CIFAR-10 placement search | device placement RL · CIFAR-10 (4 GPUs + 1 CPU) | placement search progress: Spotlight beats single-GPU placement after 100 minutes; GRL never does | Helped |
| Group operations by two-level name prefix into 86 super operations (co-location) | Spotlight RNN policy on CIFAR-10 · CIFAR-10 | qualitative (RNN reading operations) | No clear effect |
Closest papers by meaning
- Post: Device Placement with Cross-Entropy Minimization and Proximal Policy Optimization (NeurIPS 2018 · similarity 0.91)
- Device Placement Optimization with Reinforcement Learning (ICML 2017 · similarity 0.90)
- Integrated Hardware Architecture and Device Placement Search (ICML 2024 · similarity 0.86)
- A Structure-Aware Framework for Learning Device Placements on Computation Graphs (NeurIPS 2024 · similarity 0.85)
- Decision SpikeFormer: Spike-Driven Transformer for Decision Making (CVPR 2025 · similarity 0.83)
- Accelerating neural network training: An analysis of the AlgoPerf competition (ICLR 2025 · 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.