Research map / Reinforcement learning / Visual and representation RL
Palm up: Playing in the Latent Manifold for Unsupervised Pretraining
Hao Liu, Tom Zahavy, Volodymyr Mnih, Satinder Singh
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 PALM exploratory pretraining instead of CSI | CIFAR-10 classifiers · CIFAR-10 → SVHN, CIFAR-100, and Interp.; CIFAR-10 → LSUN and CelebA | AUROC (%): 96.5\stdv{0.2} -> 97.9\stdv{0.2}; 90.5\stdv{0.1} -> \bf 91.5\stdv{0.2}; 78.5\stdv{0.2} -> \bf 79.9\stdv{0.2}; \bf 96.3\s | Mixed |
| Use PALM pretraining instead of DrQ's no-pretraining setup | Atari 100k agents · Atari 100k | Median HNS: 0.268 -> \bf 0.298 | Improved |
| Use PALM exploratory pretraining instead of DrQ's no-pretraining setup | Atari 100k agents · Atari 100k | Mean HNS: 0.357 -> \bf 0.411 | Improved |
| Use PALM instead of offline-data SGI pretraining | Atari 100k agents · Atari 100k | Median HNS: \textbf{0.753} -> \bf 0.298 | Hurt |
| Use PALM instead of offline-data SGI pretraining | Atari 100k agents · Atari 100k | Mean HNS: \textbf{1.598} -> \bf 0.411 | Hurt |
| Add the agent's action to the latent transition rather than ignore it (ablation) | PALM · CIFAR-100 classification | Accuracy | Improved |
| Use agent-selected actions rather than random actions (ablation) | PALM · CIFAR-100 classification | Accuracy | Improved |
| Add temporal state persistence to the latent transition (ablation) | PALM · CIFAR-100 classification | Accuracy | Improved |
| Mix some random latent into the action rather than use action alone (α=1.0) (ablation) | PALM · CIFAR-100 classification | Accuracy | Improved |
| Use temporally persistent dynamics rather than reduced persistence (α=0.5, β=0.5) (ablation) | PALM · CIFAR-100 classification | Accuracy | Improved |
| Use PALM's explored ordered experience rather than randomly sampled consecutive offline observations (ablation) | Passive PALM · CIFAR classification | Accuracy | Improved |
| Interactively explore during training rather than train passively on the same ordered experience (ablation) | PALM · CIFAR classification | Accuracy | No clear effect |
Closest papers by meaning
- Towards Principled Representation Learning from Videos for Reinforcement Learning (ICLR 2024 · similarity 0.88)
- Foundation Policies with Hilbert Representations (ICML 2024 · similarity 0.86)
- Zero-shot World Models via Search in Memory (NeurIPS 2025 · similarity 0.86)
- METRA: Scalable Unsupervised RL with Metric-Aware Abstraction (ICLR 2024 · similarity 0.86)
- Rich-Observation Reinforcement Learning with Continuous Latent Dynamics (ICML 2024 · similarity 0.86)
- Learning Latent Dynamic Robust Representations for World Models (ICML 2024 · similarity 0.85)
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.
