Deaglan Bartlett: Implicit Likelihood Inference: From Emulators...(September 17, 2026)

Deaglan Bartlett: Implicit Likelihood Inference: From Emulators...(September 17, 2026)

Title history
  • Deaglan Bartlett: Implicit Likelihood Inference: From Emulators...(September 17, 2026) — observed on 20261113-2201 UTC (YouTube)
🎙 Deaglan Bartlett 👥 62K 📅 November 13, 2026 ⏱ 44 min 👁 4 📄 original study 🧭 2026-11-13
Available in: English (current) Français

Keywords

Implicit Likelihood Inferencegalaxy clusteringemulatorsposterior learningcosmological parameters

Summary

This talk, given by Deaglan Bartlett at the Simons Collaboration on Learning the Universe annual meeting, presents the work of the Implicit Likelihood Inference (ILI) group. The speaker introduces the concept of simulation-based inference, where the likelihood is not written explicitly but a neural network learns the posterior from forward simulations. He discusses strategies to improve efficiency, including tuning the sampling prior and using agent-based optimization. The main application is to galaxy clustering using SDSS data, where a machine-learning-based bias model and cheap simulations enable inference on cosmological parameters like Omega_m and sigma_8. After initial issues where constraints degraded upon adding information, the group refined the forward model and summary statistics, now yielding good results. Extensive validation on external simulations (Cahoots, Millennium-TNG) demonstrates robustness. The talk also covers hybrid summary statistics for weak lensing (DES) and amortized inference for initial conditions using diffusion models. Finally, he emphasizes the importance of model-misspecification checks and discusses learning physics from degeneracies.

158 words

Critical Evaluation

Value of the Information & Strength of the Argument

The talk provides substantial value by showcasing a complete workflow for large-scale cosmological inference with implicit likelihood, backed by detailed methodological improvements and validation tests. The argumentation is solid, as each step is justified with examples and pitfalls are acknowledged. The inclusion of self-consistent tests and tests on independent simulations strengthens credibility. The speaker carefully avoids overclaiming, emphasizing the progress and remaining challenges.

Scientific Rigor, Source Quality, Title Accuracy

The scientific rigor is high: the method is grounded in established concepts of simulation-based inference, and the talk includes thorough posterior predictive checks and calibration tests. Sources are primarily internal to the collaboration, with only the event page referenced in the description. The title/content match is adequate, though the ellipsis could obscure the breadth. The presentation is technical and assumes prior knowledge, but the reasoning is transparent.

145 words

Title / Content Match

The title accurately reflects the core topic—implicit likelihood inference and emulators—though the ellipsis truncates the full scope. The content aligns well.

Quality & Reliability

8/10

Presentation at a Simons Foundation collaboration meeting, by a researcher from the collaboration. The talk details methods, includes validation tests, and openly discusses issues and limitations, showing scientific rigor.

Key Moments

Cited Sources

Contribution & Novelties

The talk contributes progress in applying simulation-based inference to large-scale cosmological surveys, addressing computational challenges via emulators and machine learning. It highlights the importance of careful prior design and validation.

Pour aller plus loin :

64 words

Radar Profile

The radar profile shows high scores across quantity, quality, and technical level, with slightly lower but still strong reliability. This indicates a dense, well-structured technical presentation with robust methodology, though perhaps not fully accessible to a broader audience.

Reliability 8/10