Causal Abstraction

Causal abstraction aims to create simplified, higher-level representations of complex causal systems while preserving crucial causal relationships. Current research focuses on developing methods to learn these abstractions from data, often employing techniques like optimal transport, neural networks, and autoencoders to map between different levels of granularity in structural causal models. This work is significant for improving the interpretability and scalability of machine learning models, enabling more efficient reinforcement learning, and facilitating causal inference in high-dimensional settings across diverse applications like robotics and online advertising.