Samantha Kleinberg, Stevens Institute of Technology
Finding why things happen is a core goal for nearly all sciences, from determining why some people recover after stroke and not others to investigating whether phenomena such as divorce are contagious. Causality is often implied, rather than explicitly expressed and our knowledge of what it is and how to find it is spread across multiple fields. Even after centuries of work by philosophers and scientists, there is still not a single agreed-upon definition of causality. Yet, it is something that even non-scientists must reason with on a daily basis. When we hear that cancer rates are higher among heavy drinkers, but red wine is linked to lower rates of heart disease, for example, we need to understand whether these relationships are causal to be able to make personal health decisions and effective public policies. This session aims to explore methods for causal inference, challenges in finding causes from observational data, and mechanisms underlying our understanding of causal relationships.
This session is particularly timely given the recent proliferation of “big data” and machine learning tools for knowledge discovery within large datasets. Many of these methods, such as deep learning, focus on prediction, rather than the interpretable, causal, models needed for decision-making. Predicting a medication side effect with high accuracy does not tell us whether to discontinue the medication; we must know whether the relationship is causal to make treatment decisions. At the same time, with growing concern over unreproducible results in science, reliance on observational data brings new challenges in both reproducing studies and interpreting results, requiring more domain expertise. Humans are able to reliably infer causes from far fewer examples than computational methods require, and may provide inspiration to computational programs faced with the challenge of a large search space. However, human judgment is also sensitive to small changes in the presentation of a problem, and memory of events may be unreliable. Ultimately, integrating perspectives from computation, statistics, psychology, and linguistics may lead to more robust inferences as well as better understanding of the human mind and its relationship to understanding the physical laws that govern it.