From data to candidate explanations
Scientific discovery research asks whether a pattern in data can be expressed as a model that people can examine. At MIT CSAIL, Dr. Costilla Reyes contributed to work on neurosymbolic programming for science, which explores how learning can be combined with programs, scientific context, and domain constraints.
The aim is to make candidate models explicit enough for researchers to question, compare, and revise. A generated expression is a hypothesis that researchers can then test against data, which is what makes explicit models useful for scientific work.
Reusable methods for model discovery
InceptionSR reuses partial equations as features during later rounds of symbolic regression, the search for mathematical equations that fit a dataset. A separate learned concept-library method studies how reusable abstractions can guide equation search. These approaches connect program-synthesis ideas, in which a computer writes programs that meet a goal, with the search for compact mathematical models.
Related projects explore reusable programs for mathematical reasoning and concise solution strategies. Together, the publications cover several ways to represent and reuse structure, from reusable equation fragments to learned mathematical abstractions.
Evaluation and scientific judgment
A compact, readable equation gains scientific meaning when it is compared with relevant baselines and reviewed by people who understand the domain. Recent work examines how language models can help clinicians compare symbolic-regression formulas for body-fat estimation: in a clinician-evaluated preprint, clinicians rated comparative rankings of the formulas more favorably than explanations of single equations.
This research develops methods for generating and inspecting candidate models, with scientific interpretation grounded in the data and the domain expertise of the people who review them.