Amber Campbell
Pronouns: she/her
Research Mentor(s): Patrick Grim
Research Mentor School/College/Department: Center for Study of Complex Systems / LSA
Program: UROP
Session: Session 5 (2:40pm – 3:30pm)
Authors: Amber Campbell, Patrick Grim
Abstract
In philosophy, scientific inference is often studied one proposition at a time–hypothesis by hypothesis. However, it can be argued that the ‘basic unit’ of science is not a single hypothesis, but rather a structured theory. Scientific theories can be viewed as ‘webs of belief,’ and subsequently modeled as Bayesian networks. Such networks entail a node arrow-structure, in which a node represents a proposition with certain probability (credence), and an arrow between nodes represents conditional support from one proposition to the other, with an accompanying conditional probability. Our model seeks to update a given theory, or representation, of the world by delivering a barrage of evidence from the world. This evidence is used to modify the theory, in terms of structure or credence at the nodes, or both. Our current work focuses on the visualization of different theories as causal networks, which can be compared in respect to structure. The foundation for this work is a program utilizing Python and portions of the Bayesian Network library pyAgrum to create, mutate, and compare networks in terms of structure, as well as encode and send evidence from world networks to representation networks. Further work will introduce the updating of credences and conditional probabilities within networks, resulting in a fully-fleshed Bayesian model resembling a scientific theory. Current findings suggest that a causal representation network that is structurally distinct from the world it attempts to model will invariably converge to the world structure, given an influx of evidence from the world. The overarching goal is to determine whether two theories attempting to model the same world–with radically different structures, credences, and conditional probability values–will eventually converge, given a sufficiently long stream of identical evidence. Such work will allow for a deeper understanding of scientific inference, and how ways of thinking are updated with new information.
