Towards Causal Representation Learning

Existing machine learning algorithms mainly rely on Correlation in data, which is based on the i.i.d. (independent and identically distributed) assumption. However, in real-world environments, Domain Shift or changes in temporal structure—where the data distribution shifts—occur frequently, causing a degradation in the model's Robustness.

Furthermore, existing models lack the ability to reuse prior knowledge when faced with new environments or tasks. In contrast, an approach that learns physical causal mechanisms in a modular way can achieve efficient Generalization by reusing existing modules or training only part of them when performing a new task.

This article addresses three major challenges in machine learning from the perspective of causal inference.

• Robustness: Robust prediction that responds to changes in data distribution
• Learning Reusable Mechanisms: Knowledge reuse through modularized causal mechanisms
• Causality Perspective: Understanding causal relationships beyond correlation

To this end, it explores methodologies for modeling physical systems, key principles for estimating causal relationships, and how existing techniques such as data augmentation and self-supervised learning can be reinterpreted through Causal Representation Learning.