Tutorial's Motivation

​

Machine learning (ML) has proliferated the progress in learning informative representations for high-dimensional data. While effective for learning the statistical associations (e.g., correlation) in data, ML cannot learn the intricate causal interactions across data coordinates. Causal inference transcends learning statistical associations and enables cause-effect reasoning. Such reasoning is often facilitated via external interventions (experiments).  ​In several fast-growing fields (e.g., robotics) the causal variables are often inaccessible and lie in an unobservable latent space. The emerging field of causal representation learning (CRL) aims to learn the latent causal structures by integrating representation. CRL, subsequently, facilitates causal reasoning, intervention, and planning. This tutorial will provide a thorough overview of the recent advances in CRL.

​

Tutorial's Outline

​

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

​

Content

​

The half-day tutorial will be structured as follows with presenters denoted by their initials.

​

Part 1:  Introduction and Background

• Representation learning and causality

• Background on causal inference and discovery

• Toward CRL: Problem formulation and objectives

• Identifiability question

• Independent component analysis (ICA) and disentangled representation learning

• Overview of data modalities

• Taxonomy of CRL approaches

​

Part 2: Multi-view CRL

• Overview of multi-view data settings

• CRL from paired data

• CRL under partial observability

• Identifiability results and algorithms

​

Part 3: Interventional CRL

• Interventional environments

• CRL under parametric assumptions

• Nonparametric CRL and score-based approach

• Multi-node interventional CRL

• Multi-domain CRL

• Finite-sample analysis

​

Part 4: Temporal CRL

• Overview of data generation settings

• Temporal CRL with temporal causal effects

• Temporal CRL with instantaneous causal effects

​

Part 5: Applications and Future Directions

• Genomics and biomedical applications

• Environment extrapolation

• Visual representation learning

• ​​Challenges for scalability and generalization

​​

​​​​​​