PhD Candidate: Causality-Inspired ML/RL

Are you interested in applying ideas from causal inference in machine learning research? The Intelligent Data Engineering Lab  (INDELab) at the University of Amsterdam is seeking a PhD  candidate in causality-inspired machine learning, i.e. the application of ideas from causality to different areas of ML and RL, under the supervision of Dr. Sara Magliacane.

As powerful as today’s AI systems are, nearly all of them are only able to see correlations – they can find patterns and apparent relationships in data, and use these patterns to make predictions and decisions. However, correlation is not causation, e.g., an expensive drug may appear to cure a disease until it is discovered that it is only prescribed to patients from more affluent backgrounds that have access to better healthcare.

In a world that is increasingly reliant on AI algorithms for mission-critical decisions, an AI that cannot distinguish between correlation and causation can lead to poor decision making, inefficiency and unfairness. Questions like “Would I have been hired if I was a different gender?” or “Why was my creditapplication denied?” require a fundamental understanding of causality, as does the application of an AI system developed for one context (e.g. a movie recommendation algorithm trained to target a student population) to a different context (e.g. targeting the general public).

What are you going to do

The focus of this PhD position is: how can we use insights from causal inference (a field with an extensive history in statistics, epidemiology and computer science) to improve machine learning and reinforcement learning algorithms? In particular, we are looking for PhD students that are interested in exploring the connections between causal inference, transfer learning and  reinforcement learning.

While recent works (e.g. on Predicting Invariant Conditional Distributions ) have shown that causal insights may help in identifying features that transfer across different context, even in some apparently hopeless cases – for example, when the new context is substantially different in many aspects from the original context and where there are no examples (labels) in the new context – many of these methods still consider toy examples and leave many open questions on how to implement these insights in a real-world system.

While most ML starts with an abundant set of fairly clean data, one of the aims of INDELab is to tackle machine learning problems in which the data is heterogeneous, noisy, missing or with few labels. The lab is situated within the larger Amsterdam data science and artificial intelligence ecosystem (e.g. Amsterdam Data Science  and ICAI ) and values practice-informed and interdisciplinary research and outreach.