https://github.com/smsharma/sbi-lecture-mit

Notebook to go along with a lecture for the MIT course 8.16: Data Science in Physics on neural simulation-based inference.
https://github.com/smsharma/sbi-lecture-mit

Last synced: 3 months ago
JSON representation

Notebook to go along with a lecture for the MIT course 8.16: Data Science in Physics on neural simulation-based inference.

• Host: GitHub
• URL: https://github.com/smsharma/sbi-lecture-mit
• Owner: smsharma
• License: mit
• Created: 2023-04-24T01:54:00.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2023-10-31T03:27:49.000Z (about 1 year ago)
• Last Synced: 2024-05-22T14:31:31.973Z (6 months ago)
• Language: Jupyter Notebook
• Homepage:
• Size: 6.18 MB
• Stars: 42
• Watchers: 1
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE.md

Awesome Lists containing this project

• awesome-neural-sbi - SBI Tutorial - on tutorial introducing basic SBI concepts and methods. (Tutorials)

README

        
# SBI Lecture/Tutorial

[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/smsharma/sbi-lecture-mit/blob/main/tutorial.ipynb)

[![License: MIT](https://img.shields.io/badge/License-MIT-red.svg)](https://opensource.org/licenses/MIT)

A gentle introduction to some neural simulation-based inference methods. Jupyter notebook to go along with a guest lecture for the MIT course 8.16: Data Science in Physics.

## Simulation-based inference

Simulation-based inference (SBI) is a powerful class of methods for performing inference in settings where the likelihood is computationally intractable, but simulations can be realized via forward modeling. 

In this lecture we will

- Introduce the notion of an implicit likelihood, and how to leverage it to perform inference;

- Look at a "traditional" method for likelihood-free inference, Approximate Bayesian Computation (ABC);

- Build up two common modern _neural_ SBI techniques: neural likelihood-ratio estimation (NRE) and neural posterior estimation (NPE);

- Introduce the concept of statistical coverage testing and calibration.

As examples, we will look at a simple Gaussian-signal-on-power-law-background ("bump hunt"), where the likelihood is tractable, and a more complicated example of inferring a distribution of point sources, where the likelihood is computationally intractable.