The Statistical Structure of Identifiable Generative Models | Johnny Xi
Valence Portal is the home of the TechBio community. Join for more details on this talk and to connect with the speakers: https://portal.valencelabs.com/care
Abstract: A major goal of causal representation learning is to process unstructured low-level data into high-level causal units, on which we can apply our standard toolboxes to make inferences. The assumption that such high-level factors are probabilistically realized, but merely unobserved, underlies the motivation to fit and perform inference over generative models for causal representation learning. This talk will discuss the identifiability of such models, a foundational property that describes whether inferences can be made uniquely, or at least up to tolerable ambiguities, based on observations and suitable assumptions. Significant progress has been made recently in exploring specific assumptions and techniques for identifiability---I will take a step back and attempt to examine the statistical structure of the identifiability problem itself. Specifically, I will introduce statistical modelling and identifiability from the ground up, and apply this framework to analyze generative models as a class of statistical models, obtaining generic identification results that describe the properties of the identifiability problem without assuming any specific model. Along the way, I will also discuss some historical notes from factor analysis and ICA, why these approaches differ in their definition of "strong" identifiability, and how non-linear generators can in fact admit perfectly unique solutions simply by fixing multiple latent distributions, in a modern take on factor models.
Speaker: Johnny Xi - https://portal.valencelabs.com/member...
Twitter Chandler: / chandlersquires
Twitter Dhanya: / dhanya_sridhar
Twitter Jason: / jasonhartford
~
Chapters
00:00 - Intro + Background
07:35 - Generative Models as Statistical Models
14:03 - Indeterminacy in Generative Models
30:04 - Strong Identifiability
37:36 - Conclusions
41:52 - Q&A