Dissertation
Analogical inference is a type of inductive inference that identifies similarities between a source and target
domain, and uses those similarities to infer that some further feature of
the source domain can be generalized to the target domain. My dissertation identifies several challenges for current normative accounts of analogical inference and provides a positive account that resolves these challenges.
Current
normative accounts of analogical inference in philosophy of science and cognitive science fall into two main categories: two-dimensional accounts and empirical accounts.
The two-dimensional accounts assume that an analogical inference’s plausibility is
determined by it having two essential characteristics, which I call Source Domain
Salience and Overlap. I use examples from archaeology, physics, and astrobiology to
show that these two characteristics are neither sufficient nor necessary for plausibility.
I provide an analysis that demonstrates that there is a further essential characteristic
that affects plausibility, which I call Target Domain Salience.
I show that empirical accounts of analogical inference also encounter significant challenges. These empirical accounts are domain-specific and fragmentary, offering narrow critiques of specific analogies in specific domains. The most influential empirical account, Norton’s [2003] material account, does not offer normative criteria, and the
method of evaluation that it recommends fails to capture essential criteria that scientists use when evaluating analogical inferences.
In other work, I identify a type of analogical inference, which I call reverse analogical
inference, that is not recognized in the analogical inference literature. Under current
influential accounts, an analogical inference’s plausibility is determined by it having
(1) a meaningful representation in the source domain and (2) relevant similarities
between the source and target domain [Hesse 1966, Bartha 2009, Gentner 1997].
However, this results in a puzzle. The degree of similarity between source and target domain is determined by how the source and target domain are represented, and
in cases of underdetermination there are multiple possible ways for the source and
target domain to be represented. Thus, there are two possible ways of interpreting
dissimilarity when evaluating an analogical inference. First, dissimilarity between
source and target domain could indicate that the analogical inference is implausible
(i.e. that the further feature of the source domain cannot be generalized to the target
domain). This first type of interpretation is well discussed in the analogical inference
literature. Alternatively, dissimilarity could be taken as justification for altering the
representation in the source or target domain in order to increase the degree of similarity between source and target domain. This latter type of interpretation has not
been examined in the analogical inference literature.
Using examples from biology and physics, I demonstrate instances where scientists respond to a lack of similarity between source and target domain in this latter way—not
by questioning the plausibility of the analogical inference, but by altering the representation in the source domain. This gerrymandering of the source domain—which I
call reverse analogical inference—can entail reevaluating what level of description is
meaningful, what causal relations are important, or what the boundaries of different
entities are. Acknowledging the existence of this unrecognized type of analogical inference offers new insight into the characteristics a normative account of analogical
inference would need to have in order to be successful.
My research provides a positive account that bridges the gap between the two-dimensional and empirical accounts, and includes criteria for reverse analogical inferences. My positive account includes criteria for how background assumptions are
used, identification of innate cognitive biases that can influence scientists’ similarity
and salience judgments (like those discussed in the stimulus generalization literature),
and guidelines for how to mitigate problematic biases when non-epistemic values inevitably influence our Target Domain Salience judgments. I am co-supervised by Prof. Jeffrey Barrett and Prof. Lauren Ross and will defend in Spring 2025.
Other Research
My other research explores the constructive nature of the cognitive faculties involved in analogical reasoning. The philosophical literature already contains abundant discussion of how cognitive biases can affect scientific representation, theory choice, and observation. My research focuses on a new problem: how cognitive biases affect our similarity judgments when generalizing templates or models to new domains. Studies in cognitive science, particularly the stimulus generalization and predictive processing literature, show that cognitive biases shape similarity judgments both at the conscious and preconscious level [Ghirlanda 2006, Enquist 1997, Bruner 1957]. I look at known cognitive biases identified in cognitive science and see if they present among scientists in scientific reasoning. I also examine evolutionary narratives that explain and explicate the constructive nature of human reasoning by describing the evolutionary constraints that led to human reasoning having these characteristics. My goal in this research is to integrate findings from anthropology and cognitive science with philosophical accounts of analogical inference. I give an account of the ways that similarity relations are constructed by our cognition, and what this implies for normative accounts of analogical inference.
This project has practical implications both in the scientific and sociopolitical realm. How does the constructive nature of our similarity judgments make us susceptible to misinformation and conspiracy theories? My research also has utility in jurisprudence: how does the constructive nature of our similarity judgments affect applications of legal precedent to generalize from past court rulings to present cases? Finally, a goal in machine learning has been to improve the analogical reasoning ability of AI systems.
Cognitive science findings have often had applications for refining and improving machine learning. Integrating cognitive science with normative accounts of analogical inference promises to be fruitful for the goal of increasing analogical reasoning capacity in machine learning.
