Interdyscyplinarne Centrum Etyki UJ (INCET) zaprasza na seminarium badawcze w ramach projektu BIOUNCERTAINTY. Spotkanie odbędzie się w czwartek 28 stycznia o godzinie 17:30 na platformie MS Teams (link poniżej).

Abstrakt: Machine learning promises near-magical abilities to derive accurate predictions from large, messy data sets. But it does so without essential reference to underlying causal structures. Epidemiology, on the other hand, is traditionally obsessed with, and even defined in terms of, discovering causes (“determinants”) of disease. In this paper, we contrast two ways that ML could be applied to public health problems: a purely computational role supporting epidemiological investigations, or a new investigative approach that goes beyond providing merely computational support. Epidemiologists have tended to see it in the former role, but ML teams take a more radical approach, bringing an entire investigative methodology that is quite different from what epidemiologists traditionally do, with dramatic and high-profile results. In this paper, we focus on a fundamental difference concerning the role of causation. We ask whether ML should be “causally constrained”: whether they should be required to make (justified) causal inferences in support of recommended interventions. The idea that causal inference underpins effective intervention is engrained in epidemiology, and is a recurrent criticism of ML efforts in various contexts, including this one. Nonetheless, we argue that there should not be a causal constraint on ML. Associations that seem to us to be unprojectable and “grue-like” may not be so, while on the other hand the insistence on causal inference hardly protects us from error. When one analyses the motivations for asserting a causal constraint, whether realist, pragmatic or definitional, they all support a “try it and see” attitude. Moreover, the tenability of a sharp distinction between prediction and causal inference problems is doubtful, especially as the field increasingly adopts counterfactual approaches to causal inference, which make causal inference itself a species of prediction problem.

Alex Broadbent jest profesorem filozofii na Uniwersytecie w Johannesburgu. Jego zainteresowania naukowe dotyczą przede wszystkim filozofii epidemiologii, przyczynowości i wnioskowań przyczynowych oraz roli dowodów naukowych i statystycznych w prawie i procesie podejmowania decyzji politycznych.

Link do spotkania