Simulated Bayesian Inference for Social Science and Network Models

[CameronFen]

• Abstract (2-3 lines)

Simulation-based inference is a technique that uses normalizing flows, GANs, and variational inference to perform likelihood-free Bayesian machine learning. SBI has applications in fields as diverse as physics, biostatistics, machine learning, and economics. Following novel papers written by the presenter, this presentation will discuss the benefits of SBI with application to social science and network analysis.

• Brief Description and Contents to be covered

  Bayesian Inference Basics Traditional Methods (Metropolis-Hastings) Normalizing Flows The simulation-based inference approach (Sequential Neural Posterior Estimation) Maybe briefly discuss Graph Convolutional Neural Networks Comparison of SNPE with MCMC for social science and network problems

• Pre-requisites for the talk

  Some understanding of Bayesian methods would be helpful Understanding what a feed-forward neural network is would be important

• Time required for the talk I would like to do an hour or more, but I can present for as few as 20 mins.

• Link to slides These are the 20-minute version of the slides. Also this slide show is directed to an academic audience. For the meetup, I would have more pictures, intuitions, and less writing and equations.

https://cameronfen.github.io/files/sbi_pres.pdf

• Will you be doing hands-on demo as well? If you want me to I can, but never done one before so if you can provide a video of what you want in a demo (jupyter notebook, people coding along etc.) I would be happy to learn.

• Link to ipython notebook (if any) None

• About yourself I’m a macroeconomist and my work lies at the intersection of deep learning and macroeconometric modeling. I think a lot about the limitations and improvements to dynamic macro models including how machine learning can help. I've presented at major conferences like ASSA, EcoMod, and Econometric society conferences. I've also presented at meetups like Ann Arbor Tech. I've been interviewed for quite a few machine learning articles for my research and my work as a data scientist:

1. https://beta.informationweek.com/ai-or-machine-learning/machine-learning-basics-everyone-shouldknow
2. https://searchenterpriseai.techtarget.com/feature/10-AI-tech-trends-data-scientists-should-know
3. https://medium.com/authority-magazine/cameron-fen-of-ai-capital-management-on-the-future-ofrobotics-over-the-next-few-years-b03a11be48c6

• Are you comfortable if the talk is recorded and uploaded to PyData Delhi's YouTube channel ? Yes that's fine. Please give me the link so I can post it on my website.
• Any query ? I'm happy to do a virtual meetup since I'm not in Delhi.

[MSanKeys963]

@shagunsodhani, please have a look.

[shagunsodhani]

Hey @CameronFen Thank you for proposing the talk! I think it is a pretty interesting topic. However, in the current form, it might be too difficult for the audience to appreciate the intricacies of the topic. Our audience is generally not from academia and will likely need some background material to understand concepts like SNPE. One way could be to make it into a multi-part workshop where the first part sets the background and the other parts build on that.