This project aims to develop machine learning models that can learn data structures to induce programs that more accurately represent how humans think in the mathematical domain. An example of such a data structure is addition, which is binary in most representations but nary for most human beings. After designing learning goals for the model as well as creating an appropriate reward system, ideas from Bayesian inference, program synthesis, and language can be implemented to build such a model. The model can then be improved by trying alternative program abstraction techniques and neural network architectures, using other language models such as transformers, and using inductive, as well as deductive approaches.

I am participating in SuperUROP to study the field of program synthesis and better understand how we can teach machine learning models to reason. In the past, I have published papers in applied machine learning applying existing models to problems such as flood forecasting and phishing website detection, and I am excited to work on designing a new machine learning model.