Causal Representation Learning

Overview

Machine learning (ML) has shown great success in learning low-dimensional and semantically interpretable representations of high-dimensional data. Recent advancements in transformer design have further expanded the scope of representation learning. Despite such success, strong generalization — the transfer of learned representations to new problems — remains an unsolved problem. Addressing strong representation requires moving away from learning good enough representations to learning ground truth representations. As a key step toward strong generalization, causal representation learning (CRL) has emerged as a cutting-edge field that merges the strengths of statistical inference, machine learning, and causal inference. Its objective is to estimate the ground truth latent representation of the data and the rich structures that model the interactions among the variables in the latent space.

In this talk, we will explore the latest advancements in the emerging field of CRL. We will introduce the foundational concepts and motivations behind combining representation learning with causal inference. Following a brief history of CRL, we will outline its primary objectives and the theoretical challenges it faces. We will then review the key approaches to address these challenges, including CRL with multi-view observations, CRL with interventions on latent variables, and CRL applied to temporal data. We will also highlight real-world application opportunities, discuss the challenges of scaling CRL to practical use cases, and explore open questions related to CRL from both theoretical and empirical viewpoints.