Causal Inference: Introduction to Partial Identification and Recent Advances by Jakob Zeitler, PhD Candidate in Foundational Artificial Intelligence at the Centre for Doctoral Training in Foundational AI at UCL, London
Abstract: Causal inference provides the fundamental causal reasoning that machine learning is missing to effectively tackle decision making problems. So far, full identification of causal effects has been the focus of the majority of research: Strong and mostly untestable assumptions, such as no unmeasured confounding, yield point estimates of how a sprint will increase my endurance by 2% or how $10k more in savings will get my loan application accepted. Ideally, we would want to make fewer strong assumptions, but still provide informative suggestions.
|An Institute for Data Science and Artificial Intelligence seminar
|19 October 2022
|14:00 to 15:00
|Streatham Court Old C
Hybrid delivery by Zoom.
Partial identification enables this by calculating lower and upper bounds on the true causal effect which are more trustworthy due to more realistic assumptions. Unfortunately, current partial identification methods practically do not scale due to super-exponential parameter growth in the number of variables. Hence, I am developing scalable methods that trade-off computational cost with tightness of bounds. Exact bounding approaches will be crucial to high-stake decision making problems such as AI fairness, which require provable guarantees. Approximate methods will find use in environments valuing execution cost over guarantees, such as personal exercise recommendations or prioritisation of user experience experiments. Both approaches will become fundamental building blocks for trustworthy, causal machine learning.
The talk will discuss recent work on:
The Causal Marginal Polytope for Bounding Treatment Effects
Jakob Zeitler, Ricardo Silva
Stochastic Causal Programming for Bounding Treatment Effects
Kirtan Padh, Jakob Zeitler, David Watson, Matt Kusner, Ricardo Silva, Niki Kilbertus
Jakob graduated from Exeter in 2017 with a BA in PPE, having worked with Travis Coan on text-mining problems in climate change discussion. Since 2019, he is researching causal inference at the AI Centre of University College London as part of their CDT on Foundational Artificial Intelligence. He is working on fundamental problems of causal inference and machine learning —such as partial identification — with Ricardo Silva (continuing from a PhD he started 2017 in the US at Syracuse University): read more on partial-identification.com.
Delivery and Registration:
To be delivered hybrid. To register, please click here. Registration closes: Wednesday, 19 October 2022 at 09:00 (BST).
Whilst we appreciate the flexibility that hybrid delivery brings, we would encourage you to come along in person where there will be tea and coffee afterwards.
If you have any queries, please contact email@example.com.
Streatham Court Old C