Research map / Probabilistic and scientific modeling / Molecules and proteins
Active Target Discovery under Uninformative Priors: The Power of Permanent and Transient Memory
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes EM-PTDM, an active target discovery method that works without domain prior samples by pairing a frozen pretrained diffusion model (permanent memory) with a lightweight Doob's h-transform module (transient memory) trained via an EM-style update.
Problem
Active target discovery under a strict query budget in partially observable environments where no domain-specific prior samples exist, so learned diffusion priors are uninformative.
Approach
Decomposes the posterior score into a frozen pretrained diffusion score (permanent memory) plus a learned Doob's h-transform correction (transient memory) trained with a denoising score-matching loss, updated with an EM-style procedure that claims monotonic improvement of expected log-evidence; sampling ranks cells by a combined exploration (posterior sample disagreement) and reward-weighted explo
Key results
- On CS species discovery with GG prior at budget 150, EM-PTDM reaches SR 0.4983 vs DiffATD 0.3420 and GA 0.4061.
- On DOTA overhead objects with ImageNet prior at budget 350, EM-PTDM achieves SR 0.8256 vs DiffATD 0.7348 and RS 0.3207.
- Updating permanent memory with posterior samples across sequential tasks raises SR at budget 350 from 0.8256 to 0.8461.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use EM-PTDM instead of the strongest baseline, Greedy-Adaptive | Active discovery of CS species with GG prior · CS species discovery, budgets 150, 200, and 250 | SR: 0.4061 -> 0.4983; 0.5067 -> 0.6495; 0.5567 -> 0.6989 | Helped |
| Use EM-PTDM instead of the strongest baseline, DiffATD | Overhead object discovery with ImageNet prior · Overhead object discovery, budgets 250, 300, and 350 | SR: 0.5143 -> 0.5620; 0.6391 -> 0.7013; 0.7348 -> 0.8256 | Helped |
| Update permanent memory between consecutive tasks (ablation) | EM-PTDM · Overhead object discovery with ImageNet prior, budgets 250, 300, and 350 | SR: 0.5620 -> 0.5859; 0.7013 -> 0.7194; 0.8256 -> 0.8461 | Helped |
| Add transient memory via the h-model (ablation) | EM-PTDM posterior estimation · Novel ATD tasks from sparse early observations | L2 semantic dissimilarity to ground truth | Helped |
| Update the h-model after every observation rather than after accumulating several observations (ablation) | EM-PTDM h-model update scheduler · Toy experiment, early discovery phase | Posterior sample stability | Hurt |
Surprising findings
- The authors report that EM-PTDM posterior samples have lower overall reconstruction quality than DiffATD yet discover more targets, suggesting discovery depends on modeling target-rich regions rather than full reconstruction.
Limitations
- Results reported only on two experimental settings (species distribution and overhead object discovery) with a newly proposed task setup.
- Reconstruction quality of posterior samples is lower than DiffATD, a trade-off the authors note.
- Permanent memory update and budget-dependent alpha are heuristic choices.
Experimental setup
Datasets: iNaturalist DOTA ImageNet
Benchmarks: iNaturalist species distribution (CS vs GG prior) DOTA overhead objects with ImageNet prior
Metrics: success rate (SR) L2 semantic dissimilarity
Baselines: Random Search DiffATD Greedy-Adaptive
Builds on: DiffATD Doob's h-transform denoising score matching
Closest papers by meaning
- Online Feedback Efficient Active Target Discovery in Partially Observable Environments (NeurIPS 2025 · similarity 0.92)
- Adapting Actively on the Fly: Relevance-Guided Online Meta-Learning with Latent Concepts for Geospatial Discovery (NeurIPS 2026 · similarity 0.85)
- Provable Maximum Entropy Manifold Exploration via Diffusion Models (ICML 2025 · similarity 0.85)
- ALINE: Joint Amortization for Bayesian Inference and Active Data Acquisition (NeurIPS 2025 · similarity 0.84)
- Can In-Context Learning Support Intrinsic Curiosity? (NeurIPS 2026 · similarity 0.84)
- Causal Deciphering and Inpainting in Spatio-Temporal Dynamics via Diffusion Model (NeurIPS 2024 · 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.