Research map / Label-efficient and robust learning / OOD and anomaly detection
AutoSciDACT: Automated Scientific Discovery through Contrastive Embedding and Hypothesis Testing
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 AutoSciDACT, a pipeline that learns low-dimensional contrastive (SupCon+CE) embeddings from simulated/labeled background data and runs NPLM kernel two-sample tests to statistically quantify novelty in scientific datasets.
Problem
Detecting novelties in large, noisy, high-dimensional scientific datasets while producing statistically rigorous (quantified significance) claims of discovery rather than only per-instance anomaly scores.
Approach
Combine supervised contrastive embedding (with domain-informed labels and augmentations) reducing data to 4 dimensions with the NPLM likelihood-ratio two-sample test, averaging p-values over six kernel widths, to detect distributional deviations with calibrated Z-scores.
Key results
- NPLM flags deviations with Z of roughly 3 or higher (p about 1e-3) at signal fractions as low as 1% across domains.
- For LIGO and JetClass, NPLM approaches the supervised limit at Z about 3 and outperforms the Mahalanobis baseline in all but the synthetic datasets.
- NPLM underperforms FID/MMD on histology, which the authors attribute to the small dataset sizes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use AutoSciDACT with NPLM rather than the Mahalanobis test | AutoSciDACT · Astronomy, Particle Physics, Histology, Images | Z-score | Helped |
| Use AutoSciDACT with NPLM rather than the Mahalanobis test | AutoSciDACT · Synthetic | Z-score | Mixed |
| Use NPLM rather than FID or MMD | AutoSciDACT · Particle Physics, Images | Z-score | Helped |
| Use NPLM rather than FID or MMD | AutoSciDACT · Astronomy | Z-score | No clear effect |
| Use NPLM rather than FID or MMD | AutoSciDACT · Histology | Z-score | Hurt |
| Add supervised cross-entropy loss to the contrastive objective (ablation) | AutoSciDACT embeddings · Scientific datasets | Embedding structure and class separation | Helped |
| Increase embedding dimension from 4 to 32 (ablation) | AutoSciDACT with NPLM · CIFAR-10, JetClass | Z-score | Hurt |
| Increase embedding dimension from 4 to 32 (ablation) | AutoSciDACT with NPLM · LIGO | Z-score | Helped |
| Increase embedding dimension from 4 to 32 (ablation) | Mahalanobis baseline · CIFAR-10, JetClass, LIGO | Z-score | Helped |
| Use multiple NPLM kernel widths instead of one width (ablation) | AutoSciDACT with NPLM · Synthetic, JetClass, LIGO, Histology, CIFAR-10 | Z-score | Mixed |
| Use the average over six kernel widths instead of the best-performing width (ablation) | AutoSciDACT with NPLM · Synthetic, JetClass, LIGO, Histology, CIFAR-10 | Z-score | Hurt |
| Add random, non-discriminating input dimensions without using a learned embedding (ablation) | Synthetic-data anomaly detection · Synthetic | Z-score | Hurt |
Surprising findings
- NPLM underperforms FID and MMD baselines on the histology dataset.
- Ideal supervised benchmark falls short of supervised on LIGO and CIFAR-10 due to class density in the 4D embedding.
Limitations
- Performance depends on label quality, and labeling large training sets can be laborious or impossible.
- Embedding dimension is limited to d=4 for tractable NPLM; sensitivity degrades at very high dimension.
- Assumes reference sample matches background; domain shift and epistemic uncertainties are not handled.
Experimental setup
Datasets: JetClass LIGO O3 gravitational wave data (simulated injections) Histology mouse/rat tissue dataset CIFAR-10 CIFAR-5m Butterfly Hybrid Detection challenge LHC Higgs four-lepton data (appendix) synthetic Gaussian mixture data
Benchmarks: Synthetic Gaussian mixture JetClass LIGO gravitational waves Histology CIFAR-10 Butterfly hybrid detection
Metrics: Z-score significance empirical p-value asymptotic p-value
Baselines: supervised MLP on embedding ideal supervised (retrained encoder with signal) Mahalanobis distance test Nystrom MMD Frechet distance (FID)
Builds on: SupCon SimCLR NPLM Falkon Mahalanobis distance test
Closest papers by meaning
- AutoSciLab: A Self-Driving Laboratory for Interpretable Scientific Discovery (AAAI 2025 · similarity 0.86)
- Universal Novelty Detection Through Adaptive Contrastive Learning (CVPR 2024 · similarity 0.85)
- Position: Data-driven Discovery with Large Generative Models (ICML 2024 · similarity 0.84)
- Exploiting LLMs for Automatic Hypothesis Assessment via a Logit-Based Calibrated Prior (NeurIPS 2025 · similarity 0.84)
- LLM-SR: Scientific Equation Discovery via Programming with Large Language Models (ICLR 2025 · similarity 0.83)
- MOOSE-Chem2: Exploring LLM Limits in Fine-Grained Scientific Hypothesis Discovery via Hierarchical Search (NeurIPS 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.