From Causal Inference to Autoencoders, Memorization & Gene Regulation - Caroline Uhler, MIT

Explore a comprehensive lecture on the intersection of causal inference, autoencoders, memorization, and gene regulation presented by Caroline Uhler from MIT at the Alan Turing Institute. Delve into the development of a causal inference framework based on observational and interventional data in genomics, and learn about the first provably consistent algorithm for learning causal networks. Discover approaches for integrating different data modalities using autoencoders, and examine the theoretical analysis linking overparameterization to memorization in autoencoders. Gain insights into how overparameterized autoencoders implement associative memory and provide mechanisms for memorization and retrieval of real-valued data. This 42-minute talk covers topics ranging from structural equation models and interventional Markov equivalence classes to multidomain translation, lineage tracing, and the behavior of single-layer fully connected autoencoders.

Intro
Motivation: From single-cell measurements to mechanisms
Overview
Structural equation models
Markov equivalence classes on 3 nodes
Interventional Markov equivalence class
Causal inference and genomics
Multidomain translation & integration using autoencoders
Lineage tracing using autoencoders and optimal transport
Memorization in autoencoders
Single-layer fully connected autoencoders
Memorization of training images by iteration
Different interpolating solutions for autoencoders
Memorization: Storing and retrieving images
Memorization: Storing and retrieving sequences
Conclusions