Alicia Curth, a machine learning researcher specializing in causal machine learning, discusses topics such as the double descent phenomenon, conditional average treatment effect estimators, challenges in working with Kate models, curiosity-driven studies, sensitivity analysis in causal research, and contrasting approaches in machine learning and statistics/econometrics.
Understanding the underlying structure and assumptions in machine learning models is crucial for comprehensive problem analysis.
Uncovering the root causes and mechanisms behind complex phenomena in causality leads to a deeper understanding of the problem.
Improving benchmarking practices and developing better ways to evaluate models are essential for advancing causal AI in real-world applications.
Deep dives
The importance of understanding the problem in causal analysis
In this podcast episode, the speaker emphasizes the significance of stepping back and critically analyzing the problem at hand. Often, researchers focus on developing new methods without fully understanding the existing ones and the limitations they possess. By asking the fundamental question of 'what are we actually doing?', researchers can gain a more comprehensive understanding of the problem, identify gaps in knowledge, and explore different perspectives and approaches.
Examining the why in causal analysis
The podcast episode explores the speaker's approach to asking why in their research. In particular, they highlight the importance of understanding why certain methods or approaches work, rather than solely focusing on their effectiveness. The speaker mentions their fascination with statistics and building intuition for why things work, and highlights the need for the machine learning community to prioritize understanding the underlying structure of problems and the assumptions made in models.
Research on understanding complex phenomena
The podcast episode discusses the speaker's research on complex phenomena in causality, such as the double descent phenomenon and survival analysis. They explain their aim to uncover the root causes and mechanisms behind these phenomena, rather than simply observing their occurrence. Through their work, the speaker explores the nuances and complexities of problems like missing data, censoring, informative sampling, and more. They emphasize the need for a unified perspective that treats these challenges as varying forms of missingness, highlighting the potential for sensitivity analysis and robust modeling approaches.
Challenges in model evaluation and benchmarking
The podcast episode acknowledges the challenges in evaluating and benchmarking causal models. The speaker discusses the limitations of current benchmark data sets and how they can favor certain types of estimators over others. They argue for the importance of developing authoritative statements about likely data-generating processes and realistic problem structures to improve benchmarking practices. Additionally, the podcast emphasizes the need for better ways to evaluate models in practice, given the untestable assumptions and the complexity of causality in real-world applications.
The future of causality and machine learning
The podcast episode highlights the increasing interest and integration of causality in machine learning research. The speaker envisions a future where causality plays a more prominent role in mainstream machine learning literature, particularly in areas such as generalization and building reliable automated decision systems. They emphasize the potential of causal thinking for advancing the field and improving the robustness and interpretability of models.
Video version available on YouTube Recorded on Nov 29, 2023 in Cambridge, UK
Should we continue to ask why?
Alicia's machine learning journey began with... causal machine learning.
Starting with econometrics, she discovered semi-parametric methods and the Pearlian framework at later stages of her career and incorporated both in her everyday toolkit.
She loves to understand why things work, which inspires her to ask "why" not only in the context of treatment effects, but also in the context of general machine learning. Her papers on heterogeneous treatment effect estimators and model evaluation bring unique perspectives to the community.
Her recent NeurIPS paper on double descent aims at bridging the gap between statistical learning theory and a counter-intuitive phenomenon of double descent observed in complex machine learning architectures.
Ready to dive in? ------------------------------------------------------------------------------------------------------ About The Guest Alicia Curth is a Machine Learning Researcher and a final year PhD student at The van der Schaar Lab at Cambridge University. Her research is focused on causality, understanding machine learning methods from ground up and personalized medicine. Her works are frequently accepted at best machine learning conferences (she's a true serial NeurIPS author).