This repository contains my structured research notes on causal machine learning, with a focus on causal representation learning and causal NLP.
The goal of this repository is not to provide tutorials, but to document conceptual clarifications, theoretical reflections, and open research questions encountered during my study of causal inference and machine learning.
The notes revolve around several core themes:
- Causality as a language for mechanism
- Distribution shift and mechanism change
- Representation learning vs causal representation
- Identifiability in text-based models
- Robustness vs causality
Many modern machine learning systems learn statistical correlations from observational data. However, such correlations may rely on spurious shortcuts that fail under distribution shifts.
Causal inference offers a framework for distinguishing:
- stable mechanisms
- spurious correlations
- selection effects
- intervention effects
These notes attempt to explore how such ideas apply to language modeling and NLP systems.
foundations/
Mathematical foundations such as probability, statistics, and optimization.
causality/
Structural causal models, graphical models, and causal identification.
representation/
Notes on representation learning, invariance, and causal representations.
nlp/
Reflections on shortcut learning, mechanism shift, and causal analysis of NLP models.
reading-notes/
Summaries and reflections on key papers and books.
research-questions/
Open problems and research directions.
My current research interest focuses on:
- Causal representation learning
- Mechanism shift under distribution change
- Identifiability in NLP models
- Causal probing of neural representations
The broader goal is to understand:
What kinds of causal mechanisms can be identified in language models, and which effects are fundamentally unidentifiable from text data?
M.Sc. student in Human and Artificial Intelligence University of Technology Nuremberg
Background in philosophy and logic, now focusing on causal machine learning and causal NLP.